Nucl - Phys.B v.670

Nuclear Physics B 670 (2003) 3–26
www.elsevier.com/locate/npe
Group-theoretical aspects of orbifold

and conifold GUTs
Arthur Hebecker, Michael Ratz
Deutsches Elektronen-Synchrotron, Notkestrasse 85, D-22603 Hamburg, Germany
Received 13 June 2003; accepted 28 July 2003
Abstract
Motivated by the simplicity and direct phenomenological applicability of field-theoretic orbifold
constructions in the context of grand unification, we set out to survey the immensely rich group-
theoretical possibilities open to this type of model building. In particular, we show how every
maximal-rank, regular subgroup of a simple Lie group can be obtained by orbifolding and determine
under which conditions rank reduction is possible. We investigate how standard model matter can
arise from the higher-dimensional SUSY gauge multiplet. New model building options arise if, giving
up the global orbifold construction, generic conical singularities and generic gauge twists associated
with these singularities are considered. Viewed from the purely field-theoretic perspective, such
models, which one might call conifold GUTs, require only a very mild relaxation of the constraints
of orbifold model building. Our most interesting concrete examples include the breaking of E7 to
SU(5) and of E8 to SU(4) × SU(2) × SU(2) (with extra factor groups), where three generations of
standard model matter come from the gauge sector and the families are interrelated either by SU(3)
R-symmetry or by an SU(3) flavour subgroup of the original gauge group.
 2003 Elsevier B.V. All rights reserved.
1. Introduction
Arguably, the way in which fermion quantum numbers are explained by SU(5)-related
grand unified theories (GUTs) represents one of the most profound hints at fundamental
physics beyond the standard model (SM) [1,2] (also [3]). In this context supersymmetry
(SUSY), usually invoked to solve the hierarchy problem and to achieve gauge coupling
unification, receives a further and maybe even more fundamental motivation: if the
underlying gauge group contains gauge bosons with the quantum numbers of SM matter,
E-mail address: mratz@mail.desy.de (M. Ratz).
0550-3213/$ – see front matter  2003 Elsevier B.V. All rights reserved.
doi:10.1016/j.nuclphysb.2003.07.021
4 A. Hebecker, M. Ratz / Nuclear Physics B 670 (2003) 3–26
SUSY enforces the existence of the corresponding fermions. This is very naturally realized
in a higher-dimensional setting, where the extra-dimensional gauge field components and
their fermionic partners can be light even though the gauge group is broken at a high scale
(see, e.g., [4–7]).
Thus, we adopt the point of view that, at very high energies, we are faced with a
super-Yang–Mills (SYM) theory in d > 4 dimensions which is compactified in such a
way that the resulting 4d effective theory has smaller gauge symmetry (ideally that of
the SM) and contains the light SM matter and Higgs fields. In the simplest models the
compactification space is flat except for a finite number of singularities. Although this
situation arises naturally in heterotic string orbifolds [8,9] and has thus been extensively
studied in string model building, it has only recently been widely recognized that many
interesting phenomenological implications do not depend on the underlying quantum
gravity model and can be studied directly in higher-dimensional field theory [10] (also [11–
15]).
In the purely field-theoretic context, one has an enormous freedom in choosing the
underlying gauge group, the number of extra dimensions and their geometry, the way
in which the compactification reduces the gauge symmetry (e.g., the type of orbifold
breaking), the possible extra field content and couplings in the bulk and at the singularities.
Although, using all this freedom, realistic models can easily be constructed, there is so
far no model which, by its simplicity and direct relation to the observed field content and
couplings, appears to be as convincing as, say, the generic SU(5) unification idea. However,
we feel that the search for such a model in the framework of higher-dimensional SYM
theory is promising and that a thorough understanding of the group-theoretical possibilities
of orbifold-breaking (without the restrictions of string theory) will be valuable in this
context. The present paper is aimed at the exploration of these possibilities and their
application to orbifold GUT model building. In particular, we are interested in methods
for breaking larger gauge groups to the SM, in possibilities for rank reduction, and in the
derivation of matter fields from the adjoint representation.
In Section 2, we collect some of the most relevant facts and methods of group theory,
which serves in particular to fix our notation and conventions for the rest of the paper.
In Section 3, we begin by recalling the generic features of field theoretic orbifold
models. It is then shown how orbifolding can break a simple Lie group to any of its
maximal regular subgroups. This implies, in particular, that any regular subgroup (possibly
times extra simple groups and U(1) factors) can be obtained by orbifold-breaking and opens
up an enormous variety of model building possibilities.
We continue in Section 4 by exploring rank reduction by non-Abelian orbifolding.
We show that simple group factors can always be broken completely. In cases where a
maximal subgroup contains an extra U(1) factor, this factor can only be broken under
certain conditions. We give a criterion specifying when the extra U(1) cannot be removed.
As an interesting observation, we note that under special circumstances rank reduction
based on inner automorphisms is also possible on Abelian orbifolds.
In Section 5, we discuss manifolds with conical singularities which cannot be obtained
by orbifolding. In particular, such ‘conifolds’ can have conical singularities with arbitrary
deficit angle. In addition, we consider the possibility of having Wilson lines with
unrestricted values wrapped around the singularities of orbifolds or conifolds. All this gives
A. Hebecker, M. Ratz / Nuclear Physics B 670 (2003) 3–26 5
rise to many new possibilities for gauge symmetry breaking and for the generation of three
families of chiral matter from the field content of the SYM theory.
Finally, Section 6 discusses three specific models, one with E7 broken to SU(5) and
two with E8 broken to SU(4) × SU(2) × SU(2) (with extra factor groups). In all cases,
three generations of SM matter come from the gauge sector. In one of the E8 models, the
families are interrelated by an SU(3) R-symmetry, while in the two other models an SU(3)
flavour subgroup of the original gauge group appears.
Section 7 contains our conclusions and outlines future perspectives and open questions.
2. Basics of group theory
This section is not meant as an introduction to group theory, but merely serves to remind
the reader of some crucial facts and to fix our notation. Relevant references include the
classic papers of Dynkin [16–18] (partially collected in [19]), various textbooks (e.g.,
[20–22]), and the review article [23].
For each finite-dimensional, complex Lie algebra g, the maximal Abelian subalgebra h,
which is unique up to automorphisms, is called Cartan subalgebra. Its dimension defines
the rank r of the Lie algebra and its generators will be denoted by {H i }ri=1 . They are
orthonormal with respect to the Killing metric, i.e., they fulfill the relation
tr(H i H j ) = λδij , (2.1)

where the trace is taken in the adjoint representation and λ is some constant.
The remaining generators can be chosen such that
[H i , E α ] = αi E α , (2.2)
and are called roots. They are normalized as in Eq. (2.1). Each root E α is determined
uniquely by the root vector α, which is an element of an r-dimensional Euclidean space,
called the root space. The set of all roots will be denoted by Σ. The E α obey the
commutation relations
[E α , E β ] = Nα,β E α+β , (2.3)

where the Nα,β are normalization constants, and Nα,β = 0 means that α + β ∈
/ Σ.
We introduce an order in the root space by
α−β >0 ⇔ first non-vanishing component of α − β > 0. (2.4)

Correspondingly, we will call a root ‘positive’ if the first non-vanishing component in the
root basis is positive. The smallest r positive roots are called simple and will be denoted
by {α(i) }ri=1 . They are linearly independent, and any root can be expressed by a linear
combination

r
α= k i α(i) (2.5)
i=1
Table 1
The classical Lie algebras and the corresponding extended Dynkin diagrams. The shorter
roots are hatched. If the simple roots α(i) and α(j ) enclose an angle of 90◦ , 120◦ or 135◦ ,
they are connected by 0, 1 or 2 lines, respectively
Name Real algebra Extended Dynkin diagram
An su(n + 1)
Bn so(2n + 1)
Cn sp(2n)
Dn so(2n)
with integer coefficients k i . Motivated by this, a basis

2
e(i) = α(i) , (2.6)
|α(i) |2
is introduced. The normalization factor will be justified later.
In this basis, the Euclidean metric of the root space is characterized by gij = e(i) · e(j ) .
It is useful to consider also the vector space dual to the root space which, given the
existence of a metric in the root space, can be identified with the root space by the canonical
isomorphism. It is spanned by the so-called fundamental weights µ(i) (1 i r) which
are defined by
µ(i) · e(j ) = δji . (2.7)
The components with respect to the µ(i) basis are called Dynkin labels. Correspondingly,
the µ(i) are frequently referred to as the Dynkin basis, in which case the e(i) are called the
dual basis. The constant λ in Eq. (2.1) is chosen such that |α(i) |2 = 2 for the longest of the
simple roots. Then the normalization factor in Eq. (2.6) ensures that the Dynkin label of
any weight (weights being the analogues of the vectors α in an arbitrary representation) is
integer valued.
Table 2
The five exceptional Lie algebras. In G2 , the two simple roots
enclose 150◦ , which is indicated by a triple line
Name Extended Dynkin diagram
G2
F4
E6
E7
E8
The Dynkin labels of each simple root are given by the corresponding row of the Cartan-
matrix
α(i) · α(j ) |α(i) |2
Aij = 2 = gij , (2.8)
|α(j ) |2 2
which encodes the metric of the root space.
It is well-known that there exist four infinite series of simple groups Ar , Br , Cr and Dr ,
corresponding to the classical groups, and the exceptional groups G2 , F4 , E6 , E7 and E8 .
The scalar products of the simple roots determine the Dynkin diagrams (cf. the captions of
Tables 1, 2).
For later convenience, we introduce the most negative root θ , which leads us to the
extended Dynkin diagrams as listed in Tables 1 and 2.
3. Obtaining all regular subgroups by orbifolding
Orbifold GUTs [10–15] are based on a gauge theory on R4 × M, where M is a manifold

with some discrete symmetry group K. In addition to the action of K on M, an action
in internal space can be chosen using a homomorphism from K to the automorphism
group of the Lie algebra of the gauge theory. If the classical field space is restricted by
the requirement of K invariance, a gauge theory on a manifold with singularities, M/K,

results in general. We assume that M/K, though not necessarily M, is compact. At the
singularities, which correspond to the fixed points of the space–time action of K, the gauge
symmetry may be restricted (orbifold breaking). An early review of the structure of such
models is contained in [24] (for more recent reviews see, e.g., [25,26]).
One of the main features of orbifold GUTs is the possibility of breaking a gauge group
without the use of Higgs fields. The orbifold field theory possesses the full (unified) gauge
symmetry everywhere except for certain fixed points. Although this fixed-point breaking
is ‘hard’, in the sense that the action does not possess the full gauge symmetry, gauge
coupling unification is not lost due to the numerical dominance of the bulk. Furthermore, it
is attractive for model building purposes that the symmetry—and hence the field content—
is characterized by different groups at different geometric locations, such as the various
fixed points and the bulk.
In this paper, we focus on inner-automorphism breaking, i.e., a homomorphism from
K to the gauge group G together with the adjoint action of G on itself is used to
define the transformation of gauge fields under K. Only gauge fields invariant under K
have zero modes. The corresponding generators define the symmetry of the low-energy
effective theory, which is a subgroup of G. We will assume that G is simple since it is
straightforward to extend our analysis to the product of simple groups and U(1) factors.
To discuss the breaking in more detail, consider a group element P which is the image
of some element of K. Any P ∈ G can be written as an exponential of some Lie algebra
element and is therefore contained in some U(1) subgroup of G. Constructing a maximal
torus starting from this U(1) and using the fact that the maximal torus in a compact Lie
group is unique up to isomorphism [27], it becomes clear that one can always write
P = exp(−2πiV · H ), (3.1)
with some real vector V . Hence, the action of the gauge twist on the Lie algebra is given
by
P E α P −1 = exp(−2πiα · V )E α , (3.2a)
P H i P −1 = H i . (3.2b)
We can also choose to write P = exp(−2πiξ T ), where T is a normalized Lie algebra

element, ξ ∈ R, and ξ T = V · H . For generic ξ , P commutes with precisely those Lie
algebra elements with which T commutes. Thus, the breaking is the same as would follow
from a Higgs VEV in the adjoint representation.
However, it is clear from Eq. (3.2a) that, for certain values of ξ , some of the E α may
pick up phases which are an integer multiple of 2π and are thus left invariant. In this case,
the surviving subgroup is larger than the one obtained from an adjoint VEV proportional
to T . This possibility is of particular interest since, in certain cases, such as the breaking
of SO(10) to SU(4) × SU(2) × SU(2), the relevant subgroup cannot be realized by using
Higgs VEVs in the adjoint or any smaller representation.
3.1. Orbifold-breaking to any maximal regular subgroup
We now show that, given a simple group G and a maximal regular1 subgroup H , there
exists a P ∈ G such that

H = g ∈ G; P gP −1 = g . (3.3)
In other words, every maximal regular subgroup can be generated by an orbifold twist.
In order to prove this statement, we first recall Dynkin’s prescription for generating
semi-simple subgroups. It starts with the Dynkin diagram, extends it by adding the most
negative root, and then removes one of the simple roots, the resulting Dynkin diagram
being that of a semi-simple subgroup. As demonstrated in [18] (cf. Theorem 5.3), any
maximal-rank, semi-simple subgroup of a given group can be obtained by successive
application of this prescription. Maximal subgroups can always be obtained in the first
application.
To implement Dynkin’s prescription and remove the simple root α(i) , one can use the
fundamental weight µ(i) and choose

2πi 2
P = exp µ (i)
· H . (3.4)
n |α(i) |2
Obviously, P commutes with all simple roots E α(j) where j = i. To discuss the roots α(i)
and θ , recall first that

r
θ =− ck α(k) (3.5)
k=1
with the ck being known as Coxeter labels. They can be read off from Table 3. Such group-
theoretical methods were used in [29] in the context of E8 breaking in string theory.
Thus, we can write the orbifold action on the two roots E α(i) and E θ as
P E α(i) P −1 = e2πi/n E α(i) ,

P E θ P −1 = e−2πici /n E θ ,
which shows that, for E θ to be invariant and E α(i) to be projected out, we need ci = 1.
Using Table 3 and the corresponding Dynkin diagrams, it is easy to convince oneself that
ci = 1 occurs only for those i where the Dynkin-prescription with removal of α(i) returns
the original diagram. Thus, all non-trivial subgroups accessible by the Dynkin-prescription
can be obtained by Zn orbifolding with n = ci .
An interesting and subtle observation can be made in those cases where ci is not prime
(only ci = 4 and ci = 6 occur). If ci = n = m · k, a Zm twist generated by P k is sufficient
to project out E α(i) while keeping E θ , yet the surviving subgroup is larger than for the
corresponding Zn twist P and its Dynkin diagram is not the one obtained by Dynkin’s
prescription. These are the famous five cases where Dynkin’s prescription produces a
1 In this paper, we concentrate on the breaking to regular subgroups. For a discussion of non-regular
embeddings (in the string theory context) see, e.g., [28].
Table 3
Highest weights of the adjoint representations, denoted by Λad , of the
simple groups in the Dynkin basis and their Coxeter labels
Group Dynkin labels of Λad Coxeter labels
An SU(n + 1) (1, 0, . . . , 0, 1) (1, 1, . . . , 1)
Bn SO(2n + 1) (0, 1, 0, . . .) (1, 2, 2, . . . , 2, 2)
Cn Sp(2n) (2, 0, 0, . . .) (2, 2, . . . , 2, 1)
Dn SO(2n) (0, 1, 0, . . .) (1, 2, 2, . . . , 2, 1, 1)
G2 (1, 0) (2, 3)
F4 (1, 0, 0, 0) (2, 3, 4, 2)
E6 (0, 0, 0, 0, 0, 1) (1, 2, 3, 2, 1, 2)
E7 (1, 0, 0, 0, 0, 0, 0) (2, 3, 4, 3, 2, 1, 2)
E8 (0, 0, 0, 0, 0, 0, 1, 0) (2, 4, 6, 5, 4, 3, 2, 3)
subgroup that is not maximal [30]. They occur when removing the 3rd root of F4 , the
3rd root of E7 , and the 2nd, 3rd or 5th root of E8 .
It is easy to see that for ci prime the produced subgroup is maximal. Indeed, the roots
of G which are not roots of the subgroup H can be classified according to their ‘level’
relative to α(i) , i.e., according to the coefficient of α(i) in their decomposition in terms
of simple roots. If a subgroup H with H ⊂ H ⊂ G exists, one of the levels below ci
(which is the highest level) and above 1 must be occupied (i.e., its roots belong to H ).
Let ) be the smallest of those levels. All multiples of ) are also occupied and, since ci is
not a multiple, the difference between ci and one of those multiples must be smaller than
). However, by the way in which the commutation relations are realized in root space, the
level corresponding to this difference must also be occupied. This is in contradiction to )
being the smallest occupied level in H .
Having dealt with all semi-simple maximal subgroups, we now come to maximal
subgroups containing U(1) factors. Given a maximal subgroup H with U(1) factor, i.e.,
G ⊃ H × U(1), we can always break to a subgroup H by an adjoint VEV along this
U(1) direction or a corresponding orbifold twist. It is obvious that H ⊂ H since, by the
definition of H , all its elements commute with the generator of the above U(1). Thus,
H = H and our analysis of orbifold breaking to all maximal-rank regular subgroups
is complete. The maximal regular subgroups and the corresponding twists are listed in
Table 4. We would also like to mention that the maximal subgroups with U(1) factors can
be obtained by removing one node of the original Dynkin diagram which carries Coxeter
label 1, and adding the U(1) factor.
Now that it is clear how a given maximal regular subgroup can be generated by
an orbifold twist, we can take the opposite point of view and ask to which subgroups
an arbitrary given gauge twist P = exp (−2πiξ T ) can lead. Since an adjoint VEV
proportional to T breaks to a maximal rank subgroup H × U(1), where the U(1) is
generated by T , we can classify all T ’s by such maximal subgroups. These are given
in various tables (see, in particular, [23]) together with the branching rules for the adjoint
representation
ad G → ad H ⊕ 1(0) ⊕ R 1 (q1 ) ⊕ R 2 (q2 ) ⊕ · · · . (3.7)

Table 4
Maximal subgroups of the simple groups and the corresponding Zn orbifold
twists. The five non-maximal subgroups which can be obtained by removing one
node of the extended Dynkin diagram are listed for the sake of completeness
Group Twist Symmetric subgroup Comment
SU(N + M) Z2 SU(N ) × SU(M) × U(1)
SO(N + M) Z2 SO(N ) × SO(M) N or M even
SO(2N ) Z2 SU(N ) × U(1)
Sp(2N + 2M) Z2 Sp(2N ) × Sp(2M)
Sp(2N ) Z2 SU(N ) × U(1)
G2 Z2 SU(2) × SU(2)
G2 Z3 SU(3)
F4 Z2 Sp(6) × SU(2)
F4 Z3 SU(3) × SU(3)
F4 Z4 SU(4) × SU(2) not maximal
F4 Z2 SO(9)
E6 Z2 SO(10) × U(1)
E6 Z2 SU(6) × SU(2)
E6 Z3 SU(3) × SU(3) × SU(3)
E7 Z2 SO(12) × SU(2)
E7 Z3 SU(6) × SU(3)
E7 Z4 SU(4) × SU(4) × SU(2) not maximal
E7 Z2 E6 × U(1)
E7 Z2 SU(8)
E8 Z2 SO(16)
E8 Z4 SU(8) × SU(2) not maximal
E8 Z6 SU(6) × SU(3) × SU(2) not maximal
E8 Z5 SU(5) × SU(5)
E8 Z4 SO(10) × SU(4) not maximal
E8 Z3 E6 × SU(3)
E8 Z2 E7 × SU(2)
E8 Z3 SU(9)
Here the R i are representations under H and qi the corresponding U(1) charges. Under
the gauge twist, the R i transform as R i → e2πiξ qi R i . This allows us to determine which
particular sets of generators R i survive for specific values of ξ , i.e., to identify those R i
for which ξ qi = 0 mod Z. Together with the generators of H × U(1), they form the Lie
algebra of the new surviving subgroup H ⊃ H × U(1). Thus, by analyzing all subgroups
H × U(1) and all values of ξ , our classification is complete.
Finally, we would like to comment on the minimal order of the twist required to achieve
the breaking G → H . A very useful approximate rule is that under a Zn twist
dim G − r
dim H r + . (3.8)
n
The reason is that the r Cartan generators survive the twist anyway, and the phases of the
roots are proportional to a level relative to a simple root α(i) , or linear combination of such
levels. Due to the symmetries of the root lattice, the phases are therefore almost evenly
Fig. 1. The breaking to the Pati–Salam and Georgi–Glashow subgroups of SO(10) can be illustrated by removing
the α(3) (or α(2) ) node of the extended Dynkin diagram (a) as shown in (b) or by removing α(5) (or α(4) ) as
shown in (c).
distributed among {0, 2π/n, 4π/n, . . ., (n − 1)2π/n} where an excess at 0 is possible if

the twist acts trivially on a certain part of the algebra. An inspection of Table 4 confirms
our rule which becomes the more accurate the larger the group is.
3.2. Some examples from the series SO(10) ⊂ E6 ⊂ E7 ⊂ E8
At this point, some examples are in order. Let us start with the SO(10) GUT which
contains the Georgi–Glashow group GGG = SU(5) ⊗ U(1) and the Pati–Salam group
GPS = SU(4) × SU(2) × SU(2) as subgroups. These properties are nicely illustrated by
using Dynkin’s prescription: starting from the extended Dynkin diagram (cf. Fig. 1), the
diagram of GPS is obtained by deleting the third (or second) node. Deleting the fourth (or
fifth) node of the original diagram, we arrive at GGG . According to Section 3.1, twists
which break to GGG and GPS , respectively, can be written as

PPS = exp πiµ(3) · H , (3.9a)

PGG = exp πiµ · H ,
(4)
(3.9b)
where we exploited the fact that |α(i) |2 = 2 in simply-laced groups.
In [14,15], it was shown that by identifying these two twists as generators of Z2 × Z2 ,
the gauge symmetry on a T2 /(Z2 × Z2 ) orbifold is reduced to GSM = SU(3) × SU(2) ×
U(1)Y × U(1)χ ⊂ SO(10). The resulting geometry can be visualized as a ‘pillow’ with the
corners corresponding to the fixed points.
The relevant group theory can be understood as follows: µ(3) · H and µ(4) · H are
the U(1) generators appearing in GSM . The corresponding decomposition of the adjoint
representation of SO(10) reads
45 → (1, 1)(6,4) ⊕ (3, 1)(−4,4) ⊕ (3, 2)(1,4) ⊕ (3, 2)(−5,0)

⊕ (1, 1)(−6,−4) ⊕ (3, 1)(4,−4) ⊕ (3, 2)(−1,−4) ⊕ (3, 2)(5,0)
⊕ (8, 1)(0,0) ⊕ (1, 3)(0,0) ⊕ (1, 1)(0,0) ⊕ (1, 1)(0,0) , (3.10)
where the SU(3) × SU(2) representations are given in boldface and the U(1) charges
(qY , qχ ) appear as index. The twist which is responsible for this breaking is generated
by a linear combination of the generators of the two U(1)’s, and rotates the charged
representations by a phase 2π(yqY +χqχ ). For some combinations of χ and y, the orbifold
breaking preserves a larger symmetry than adjoint breaking. For example, if (1, 1)(6,4) ,
(1, 1)(−6,−4) ,(3, 1)(−4,4) ) and (3, 1)(4,−4) survive (e.g., by taking χ = 0 and y = 1/2), the
resulting gauge group is GPS . If, on the other hand, (3, 2)(−5,0) and (3, 2)(5,0) survive (e.g.,
by taking χ = 1/8 and y = 0), the resulting gauge group is GGG . It is then clear that GSM
results as an intersection of gauge fields surviving PPS and PGG . The breaking to GSM can
also be realized on a single Z4 fixed point, e.g., by using χ = 1/16 and y = 1/4.
As a side-remark, let us restate the above discussion in terms of matrices: Consider the
adjoint VEV

0 1
v = diag(a, a, a, b, b) ⊗ , (3.11)
−1 0
which breaks SO(10) to GSM [22]. For the special case a = ±b, the remaining symmetry
is larger and equal to GGG . Alternatively, these breakings can be realized by a gauge twist
P = exp[2πiv] at an orbifold fixed point. In this case, taking a = 0 and b = 1/2 yields PPS
and a = b = 1/4 yields PGG .
Let us now turn to the task of extending the ‘pillow’ of Asaka, Buchmüller and Covi [31]
along the chain of exceptional groups SO(10) ⊂ E6 ⊂ E7 ⊂ E8 . A related discussion has
already appeared in [32]. However, as will become clear below, we disagree with some of
the results of that paper.
The obvious generalizations of (3.9) for the exceptional groups read
(r)
PGG := exp πiµ(r−1) · H , (3.12a)
(r)
PPS := exp πiµ (r−2)
·H . (3.12b)
In other words, the generalizations of PGG and PPS to higher groups along the above chain
remove the nodes α(r−1) and α(r−2) respectively. This is illustrated in Table 5.
The breaking patterns of SO(10), E6 and E7 are easily determined by the use of
Dynkin’s prescription because the Coxeter label corresponding to the nodes removed by
(r) (r) (8)
PGG and PPS are 1 and 2, respectively. In the case of E8 , it is also easy to see that PGG
(8)
breaks to E7 × SU(2). For PPS , the pattern is not so obvious: Since the 6th Coxeter label
(8)
is 3 (cf. Table 3), and we use a Z2 twist, the second level in terms of α(6) survives PPS ,
but θ is projected out. We see that the subgroup must contain E6 and SU(2), and it cannot
be E6 × SU(3) because this is not a symmetric subgroup of E8 . Hence, it must be also
E7 × SU(2).2 We checked this statement by using a computer algebra system. In [15], an
=P
interesting property of the SO(10) twists was pointed out: PGG GG · PPS breaks to a
different SU(5)×U(1) subgroup of SO(10), where the simple factor is often called ‘flipped
(r) (r) (r)
SU(5)’. This property is maintained for all three exceptional groups: PGG · PPS = PGG .
2 This is in contradiction to the breaking pattern given in [32]. Assigning negative parity to (27, 3) and (27, 3 )
is inconsistent since, as can be seen from the commutator [(27, 3), (27, 3)] ⊂ (27, 3 ), it does not correspond to
an algebra automorphism. This commutator does not vanish since (27, 3) ⊕ (27, 3 ) contains two positive levels
with respect to α(6) , linked by the raising operator E α(6) .
Table 5
Breaking patterns to the Georgi–Glashow and Pati–Salam like subgroups in SO(10) and the three exceptional
groups E6 , E7 and E8
PGG
SO(10) −−→ SU(5) × U(1)
PPS
SO(10) −−→ SU(4) × SU(2) × SU(2)
PGG ·PPS
SO(10) −−−−−→ SU(5) × U(1)
GGG ∩ GPS = SU(3) × SU(2) × U(1)2
(6)
PGG
E6 −−→ SO(10) × U(1)
(6)
PPS
E6 −−→ SU(6) × SU(2)
(6) (6)
PGG ·PPS
E6 −−−−−→ SO(10) × U(1)
(6) (6)
GGG ∩ GPS = SU(5) × U(1)2
(7)
PGG
E7 −−→ E6 × U(1)
(7)
PPS
E7 −−→ SO(12) × SU(2)
(7) (7)
PGG ·PPS
E7 −−−−−→ E6 × U(1)
(7) (7)
GGG ∩ GPS = SO(10) × U(1)2
(8)
PGG
E8 −−→ E7 × SU(2)
(8)
PPS
E8 −−→ E7 × SU(2)
(8) (8)
PGG ·PPS
E8 −−−−−→ E7 × SU(2)
(8) (8)
GGG ∩ GPS = E6 × U(1)2
(r)
Here PGG breaks to a subgroup linked by an inner automorphism to the subgroup left
(r) (r)
invariant by PGG . The reason is that PGG = exp[πi(µ(r−1) + µ(r−2) ) · H ] commutes
with E α(r−1) +α(r−2) which then becomes a simple root of the subgroup, and projects out
E θ . This root encloses an angle of 120◦ with α(r−3) so that the resulting Dynkin diagram
coincides with the one obtained by employing PGG . In the simple root system arising from
(r)
the substitution (α(r−2) , α(r−1) ) → (α(r−2) + α(r−1) , −α(r−1) ), PGG acts in the same way
(r)
as PGG in the original root system.
4. Rank reduction and non-Abelian twists
It is obvious from the discussion so far that using only one inner-automorphism orbifold
twist can never result in rank reduction. We therefore investigate the possibilities which
arise when two (or more) twists are applied. Rank reduction of the gauge group was
proposed in the context of string theory in [33]. Here, we will discuss this issue in the
context of field theory, where one has fewer group-theoretic and geometric constraints.
We assume that we have an additional orbifold symmetry,

P = exp −2πiξ(E β + E −β ) or P = exp 2πξ(E β − E −β ) , (4.1)
where i(E β + E −β ) and E β − E −β are real generators outside the Cartan subalgebra. For
simplicity, let us focus on the case that β is a simple root, i.e.,

Pj = exp −2πiξ(E α(j) + E −α(j) ) . (4.2)
Then the raising and lowering operators E ±α(j) form an SU(2) group together with
h = α(j ) ·H . Clearly, this linear combination of Cartan generators ‘rotates’ under the action
of P like
Pj −1 hPj = cos(4πξ )h − i sin(4πξ )(E α(j) − E −α(j) ), (4.3)

√
where we restricted ourselves to the case that α(j ) has length 2. Since a linear
combination of Cartan generators transforms non-trivially, it is obvious that rank reduction
is possible. Note also that these rotations yield an extension of the well-known Weyl
reflections, i.e., the reflections with respect to a plane perpendicular to a simple root.
It is straightforward, but somewhat tedious to derive the action on arbitrary roots E α .
In simply-laced gauge groups, the root chains have at most length two unless they contain
Cartan generators. Thus, Nα,±α(j) = 0 implies Nα,∓α(j) = 0. For the upper sign, we obtain
(for α = α(i) )
Pj −1 E α Pj = cos(2πNα,α(j) ξ )E α + i sin(2πNα,α(j) ξ )E α+α(j) , (4.4)

where we use the normalization constants Nα,β as defined in Eq. (2.3) with the convention
to choose them positive.
From the discussion so far, it is clear that we can break any simple group factor com-
pletely by non-Abelian twists: the roots can always be removed by suitable exponentials of
the Cartan generators, and the H i can be projected out by using Eq. (4.3). This observation
has an obvious application: let H ⊂ G be the subgroup that we want to obtain by orbifold-
ing. Let H ⊂ G be the maximal subgroup that commutes with H and the Cartan generators
of which are orthogonal to the Cartan generators of H . If H is semi-simple, an orbifold
breaking to H is always possible. In this context, it is interesting to observe that E8 is the
only simple group containing a maximal regular subgroup of the form SU(5) × H with H
semi-simple, namely E8 ⊃ SU(5)×SU(5). Thus, one could say that E8 is the smallest GUT
group larger than SU(5) which can be orbifolded to the SM without additional U(1) factors.
A further example is in order: It is clear that we can break an SU(2) factor completely
by the methods described above. Thus, since E6 ⊃ SU(6) × SU(2), we can achieve E6 →
SU(6) by taking P = exp[πiα(1) · H ] and P = exp[πi(E α(1) + E −α(1) )/2]. In addition,
we can modify P in a way so that the breaking is stronger, e.g., E6 → SU(5) × U(1).
However, if extra U(1) factors are contained in H , the story becomes more compli-
cated. One is tempted to conclude that such extra factors cannot be removed, given that
this is obviously not possible by adjoint VEV breaking. However, in the case of orb-
ifold breaking this is not true. Consider, for example, SO(5) which can be broken to
SU(2) × U(1) by using P = diag(−1, −1, 1, 1, 1). The extra U(1) can be destroyed by in-
voking P = diag(1, −1, −1, −1, −1). This example is particularly interesting since here
P and P commute although the rank is reduced (which is possible because the corre-
sponding generators do not commute). The above SO(5) example is special because P ,
which maps the U(1) generator to minus itself, acts on the other (real) representations in
a way consistent with SU(2) symmetry. If we deal with complex representations, i.e., the
adjoint of G branches as
ad G → ad H ⊕ 1(0) ⊕ R(q) ⊕ R(−q) ⊕ · · · , (4.5)

where R(q) and R(−q) are conjugate to each other, a flip of the U(1) charge carries R
and R into each other. The flip then acts non-trivially on H so that flipping the U(1) factor
without affecting H is impossible.
We emphasize that this excludes the possibility of orbifold breaking of the U(1) factor in
a large class of cases. Namely, let H × U(1) ⊂ G such that the U(1) is the maximal group
commuting with H . Clearly, any automorphism of G leaving H invariant has to map the
U(1) onto itself. Since the only non-trivial automorphism of U(1) is the above sign flip,
the presence of complex representations H in the adjoint of G (cf. Eq. (4.5)) excludes the
required H -preserving automorphism of G. The extension to H × H ⊂ G, where H is a
product group containing U(1) factors, is straightforward.
The above SO(5) scenario with P and P can, for example, be realized in 4 + 2
dimensions with compact space T2 /(Z2 × Z2 ). The Z2 generator acts on the torus as
a rotation by 180◦ , the Z2 generator acts as a shift by half of one of the original torus
translations (cf. Fig. 2(a)).
It turns out that the elements of Z2 × Z2 comply with the multiplication law of the
dihedral group D2 of order 4.3 While the dihedral group of order 4 is Abelian, higher order
dihedral groups are not. We illustrate a possible way of using the order 6 group D3 in an
orbifold construction in Fig. 2(b). It follows the T2 /(Z2 × Z2 ) construction up to the fact
that we now divide the cell into three parts instead of two. Embedding it into a gauge group
then allows for realizing non-Abelian twists.
These examples can be generalized in the following way: the orbifold can be interpreted
as O = Tn /R where R is a symmetry of the lattice, and the torus arises by modding out
flat space by discrete translations, Tn = Rn /Λ. By embedding the full symmetry group K,
containing the operations of R as well as the translations, into the gauge group, it is then
possible to achieve that the torus Tn , which arises as intermediate step in this picture,
carries Wilson lines [34]. Since the generators associated with the Wilson lines do not
necessarily commute with the twists corresponding to embedding the operations of R into
the gauge group, rank reduction is possible [33]. We believe that similar constructions will
be important for model building.
Let us briefly comment on non-regular embeddings. Consider the group SU(3)
which contains SO(3) (the subgroup of real matrices) as an S-subgroup (in Dynkin’s
3 Recall that the dihedral group of order 2n, called D , can be envisaged as the group generated by the rotation
n
of a regular n-polygon by 2π/n and the flip over one of its edges [35]. Clearly, the dihedral group always can be
embedded in an SO(3) SU(2). Anomalies of dihedral orbifolds are discussed in [36].
Fig. 2. Examples of (a) a T2 /D2 and (b) a T2 /D3 orbifold where rank reduction is possible. The action of the
dihedral group consists in a rotation by 180◦ around the origin, and in a translation by e1 in case (a), and a
translation by e1 or 2e1 in case (b).
terminology). Let us pick two generators of the embedded SO(3), for instance

0 1 0 0 0 −1
T1 = −1 0 0 and T2 = 0 0 0 . (4.6)
0 0 0 1 0 0
It is then straightforward to convince oneself that imposing the twists P1 = exp(2πiT1 /4)
and P2 = exp(2πiT2 /4) breaks SU(3) completely. Similar constructions can be used to
break larger groups with only a few twists. For instance, E8 has a maximal S-subalgebra
su(2) and can therefore be broken completely by only two twists, e.g., by embedding a
suitable dihedral group in the SU(2).
5. Conifold GUTs
We now want to continue the discussion of the generic structure of orbifold GUTs
given at the beginning of Section 3 and show that a mild generalization of the construction
principles leads to a much larger freedom in model building. Our main focus will be on 6d
models.
5.1. Geometry and gauge symmetry breaking
In 5 dimensions, the geometry is very constraining. Up to isomorphism, the only smooth

compact manifold is S1 , where one has the familiar problems of obtaining chiral matter
and of fixing the Wilson line, the value of which represents a modulus which, in the
SUSY setting, cannot be stabilized by perturbative effects. The only compact orbifold is
the interval, which can always be viewed as S1 /(Z2 × Z2 ) (with S1 /Z2 being a special
case). The gauge breaking at each boundary is determined by a Z2 automorphism and can
be interpreted as explicit breaking by boundary conditions. One may try to generalize the
setting by considering breaking by a boundary localized Higgs (in the limit where the VEV
becomes large) [37] or ascribing Dirichlet and Neumann boundary conditions to different
gauge fields (without the Z2 automorphism restriction) [24]. Furthermore, it is possible
to ascribe the breaking to a singular Wilson line crossing the boundary [24]. However, it
Fig. 3. Construction of a compact manifold with singularities from two triangles.
appears to be unavoidable that geometry is used only in a fairly trivial way and that the
breaking is confined entirely to the small-scale physics near the brane, outside the validity
range of effective field theory.
Group-theoretically, the 5d setting is also fairly constrained since the relative orien-
tation of the gauge twists at the two boundaries is a modulus. To be more specific, let
P1 = exp(T 1 ) and P2 = exp(T 2 ) be the two relevant twists. Even though this makes rank
reduction possible in principle, we are faced with the problem that, if the Wilson line con-
necting the boundaries develops an appropriate VEV, the situation becomes equivalent to
both T 1 and T 2 being in the Cartan subalgebra, in which case the symmetry is enhanced
to a maximal-rank subgroup. SUSY prevents the modulus from being fixed by loop cor-
rections.4
In 6 dimensions, the situation is much more complicated and interesting. Clearly, the
smooth torus has the same problems as the S1 discussed above. However, there is a large
number of compact manifolds with conical singularities. A simple way to envisage such
singular manifolds or, more precisely, conifolds is given in Fig. 3. The fundamental space
consists of two identical triangles. The geometry is determined by gluing together the edges
of the depicted triangles, thus leading to a triangle with a front and a back, a triangular
‘pillow’. It is flat everywhere except for the three conical singularities corresponding to
the three corners of the basic triangle. Each deficit angle is 2(π − ϕ), where ϕ is the
corresponding angle of the triangle. Obviously, in this construction the basic triangle can
be replaced by any polygon. If the polygon is non-convex, negative deficit angles appear.
Four specific polygons deserve a separate discussion. These are the rectangle, the
equilateral triangle, the isosceles triangle with a 90◦ angle, and the triangle with angles
30◦ , 60◦ , and 90◦ . The conifolds constructed in the above manner from these polygons
can alternatively be derived from the torus as a Z2 , Z3 , Z4 and Z6 orbifold, respectively.
Given that Zn cannot be a symmetry of a 2-dimensional lattice for n > 6, it is clear that
this last method of constructing conifolds is highly constrained when compared to the
generic conifold of Fig. 3 with an arbitrary polygon. However, from the perspective of
effective field theory model building, there appears to be no fundamental reason to discard
the multitude of possibilities arising in the more general framework.
Clearly, even more possibilities open up if, in addition to conical singularities one
allows for 1-dimensional boundaries. These arise in orbifolding if a Z2 reflection symmetry
(in contrast to the Zn rotation symmetries above) of the torus is modded out [15] (see
also [39]). However, in what follows we will concentrate on construction with conical
singularities only.
4 For more details and a discussion of the non-supersymmetric case see [38] and [25], respectively.
We now turn to the possibilities of geometric gauge symmetry breaking on conifolds.

Recall first that, if a given conifold can be constructed from a smooth manifold by modding
out a discrete symmetry group, i.e., as an orbifold, then an appropriate embedding of
this discrete group into the automorphism group of the gauge Lie algebra will lead to a
gauge symmetry reduction. Working directly on the fundamental space (as opposed to the
covering space) this gauge breaking can be ascribed to non-trivial values of Wilson lines
encircling each of the conical singularities.
It is now fairly obvious how to introduce this type of breaking in the generic construction
of Fig. 3 (possibly with the triangle replaced by an arbitrary polygon). First, we identify
one edge of the front polygon with the corresponding edge of the back polygon. Next,
when identifying along the two adjacent edges, one uses the freedom of introducing a
relative gauge twist P ∈ G. In more detail, if (x, y) and (x , y ) parametrize front and back
polygon near the relevant edge (such that the edge is at y = 0 or y = 0), one demands
A (x = x, y = 0) = P A(x, y = 0)P −1 for the gauge potentials A and A on the two
polygons. Continuing with the identifications, one finds that there is a freedom of choosing
n − 1 gauge twists Pi in the presence of n conical singularities. Technically, this is due
to the fact that the identification along one of the edges can always be made trivial using
global gauge rotations of one of the polygons. A geometric understanding follows from
the fact that the global topology is that of a sphere, in which case the Wilson lines around
n − 1 singularities fix the last Wilson line (we always assume the vacuum configuration,
i.e., A is locally pure gauge).
Clearly, we want to obtain a smooth manifold (except for the singularities) in the end so
that, to be more precise, the Pi have to be introduced in the appropriate transition functions
of the defining atlas. However, we believe that it is not necessary to spell out this familiar
construction in detail.
Instead of using only inner automorphisms described by Pi , we could have allowed for
outer automorphisms in the transition functions. In this case, which we will not pursue in
this paper, the corresponding vacua are clearly disconnected from those defined only by
inner automorphisms. The theory can then be thought of as defined on a generalization
of a principal bundle (in the commonly used definition of principal bundles the transition
functions involve only inner automorphisms).
We now want to analyze the gauge fields in a small open subset including one conical
singularity. A convenient parametrization is given by polar coordinates (r, ϕ) with 0 <
r < 5 and 0 ϕ < β, where the singularity is at r = 0 and the deficit angle is 2π − β. As
familiar from the Hosotani mechanism on smooth manifolds [40], we can trade the gauge
twist in the matching from ϕ = β to ϕ = 0 for a background gauge field which, for a twist
P = exp(T ), can be chosen as A = eϕ T /(βr). Here eϕ is the unit-vector in ∂ϕ direction
so that A is a Lie-algebra-valued vector. This simple exercise demonstrates explicitly that,
at least locally, the breaking can be attributed to a non-vanishing gauge field VEV in a flat
direction. However, in contrast to the Hosotani mechanism, the corresponding modulus can
be fixed without violating the locality assumption (which we consider as fairly fundamental
in effective field theory). Namely, the value of the Wilson line described by the above A can
be determined by some unspecified small-distance physics directly at the singularity. This
is similar to the boundary breaking in 5 dimensions. In contrast to the 5d case, however,
the breaking at the conical singularity is visible to the bulk observer, who can encircle the
singularity and measure the Wilson line without coming close to the singularity. Thus, one
might be tempted to conclude that this type of breaking has a better definition in terms of
low-energy effective field theory.
To conclude this subsection, we want to collect the generalizations of 6d field theoretic
orbifold models discussed above. First, one can work on conifolds, i.e., use deficit angles
that cannot result from modding out on the basis of a smooth manifold. The gauge twist at
each singularity may, however, be still required to be consistent with the geometric twist.
Second, one can insist on conventional orbifolds as far as the geometry is concerned but use
arbitrary gauge twists at each conical singularity, i.e., give up the connection between the
rotation angles in tangent and gauge space. Third, one may drop both constraints and work
on conifolds with arbitrary deficit angles and gauge twists. Obviously, such constructions
can also be carried out in more than 6 dimensions. The detailed discussion of those is
beyond the scope of the present paper.
5.2. Generating chiral matter
In general, compactification on a non-flat manifold can provide chiral matter if the

holonomy group of the compact manifold fulfills certain criteria. For example, it is well-
known that compactification of a 10d SYM theory on Calabi–Yau manifolds with SU(3)
holonomy [41] or on orbifolds [9] leads to N = 1 SUSY in 4d. Both constructions are not
unrelated as many orbifolds can be regarded as singular limits of manifolds in which the
curvature is concentrated at the fixed points. Since the reduction of SUSY is a matter of
geometry, compactification of a higher-dimensional field theory on a conifold can also lead
to N = 1 supersymmetric models in 4d.
Interesting models have been constructed using the fact that the vector multiplet of N =
(1, 1) SUSY in 6d corresponds to one vector and three chiral multiplets in 4d language,
A = (V , φ1 , φ2 , φ3 ). The fact that three copies of chiral multiplets appear automatically
may be an explanation of the observed number of generations [4]. The above 6d theory can
be interpreted as arising from a 10d SYM, in which case the scalars of the chiral multiplets
are the extra components of gauge fields [42], for example, φ1 A5 + iA6 , φ2 A7 + iA8
and φ3 A9 + iA10 (AM denote the components of the 10d vector). When defining our
6d models, we require the field transformations associated with going around a conical
singularity to be an element of an SU(3) subgroup of the full SO(6) SU(4) symmetry of
the underlying 10d SYM theory. Under this subgroup, which we call SU(3) R-symmetry,
the chiral superfields φi transform as a 3.5 The appealing feature of such a construction
is that matter multiplets are not put in ‘by hand’ but arise in a natural way from a higher-
dimensional SYM theory [4–7].
The action of the SU(3) R-symmetry transformation on the chiral superfields is not
completely arbitrary. For example, if φ1 A5 + iA6 , the transformation of φ1 is fixed by
geometry, e.g., when modding out a rotation symmetry, a corresponding rotation has to be
applied to the φ1 superfield. Thus, when going around a conical singularity, φ1 receives
a phase which is given by 2ϕ, where ϕ is the corresponding angle of the polygon. Since
5 For more details see, e.g., [42] as well as [4,43].

multiplying by the phase ei2ϕ corresponds to a rotation in the complex plane, we will call
2ϕ ‘rotation angle’ in what follows. Clearly, the rotation angle sums up with the deficit
angle to 2π .
6. Specific models
Let us now discuss three models in which some of the main features of the last sections
are exemplified. All these models are based on a SYM theory in 4 + 2 dimensions endowed
with (N1 , N2 ) = (1, 1) SUSY. In 4d we then deal with three chiral superfields φ1 , φ2 and
φ3 where we assume that φ1 = A5 + iA6 so that the action of the R-symmetry on φ1 is
fixed (cf. Section 5.2).
6.1. E7 → SU(5) × SU(3)F × U(1)
Consider a SYM theory based on an E7 gauge group. E7 contains SU(5) × SU(3)F ×

U(1), and the adjoint representation decomposes as
133 → (24, 1)0 ⊕ (1, 1)0 ⊕ (1, 8)0 ⊕ (5, 1)6 ⊕ (5, 1)−6
⊕ (10, 3 )−2 ⊕ (5, 3)−4 ⊕ (10, 3)2 ⊕ (5, 3 )4 , (6.1)
where we use a notation analogous to Eq. (3.10).
As explained in Section 3, the twist P which causes the desired breaking can be
understood as exponential of the U(1) generator. Under this twist, the multiplets appearing
in Eq. (6.1) acquire phases which are proportional to the U(1) charge. By taking the
proportionality constant to be −1/12, we arrive at the phases listed in Table 6 where here
and below phases are given in units of 2π .6
The smallest phase present is 1/6 so that P is a Z6 twist. Therefore, the R-symmetry
acts on φ1 as a −60◦ rotation in the 5–6 plane, and thus the (10, 3 )−2 possesses a zero-
mode. We choose the transformation of φ2 such that the (5, 3)−4 survives as well, and the
phase of φ3 is then fixed by the determinant condition. More explicitly, by taking

R = exp 2πi diag(−1/6, 2/3, −1/2) ∈ SU(3), (6.2)
we can achieve that 3 generations of 10 and 5 survive without any mirrors, indicated by
boldface phases in Table 6, and an N = 1 SUSY in 4d is preserved. It is also interesting to
observe that the only additional surviving superfields, namely (5, 1)6 and (5, 1)−6 which
acquire phases 1/2 and therefore have zero-modes due to the third diagonal entry of R,
carry the quantum numbers of the light Higgs fields in the supersymmetric SU(5) theory.
Thus, the SU(5) part of this model looks relevant for reality, and is in particular anomaly-
free.
The geometry of this model, which can be constructed as a standard orbifold T2 /Z6 , is
given by two triangles with angles 30◦ , 60◦ and 90◦ (cf. Fig. 4). The Z6 twist P (P6 = P in
6 The twist can be thought of as P = diag(ω, ω, ω, ω, ω, −ω), with ω defined as the 12th root of 1, acting
on the SU(6) embedded in E7 . Although the action of P on a fundamental representation of SU(6) would be
the one of a Z12 twist, its action on E7 is Z6 since the adjoint of E7 only contains antisymmetric and adjoint
representations of SU(6).
Table 6
Phases (in units of 2π ) for the different multiplets of SU(5) ×
SU(3)F × U(1) ⊂ E7 . Zeros correspond to the surviving gauge
bosons, other phases which are compensated by the R-symmetry
transformations are written in boldface
(24, 1)0 : 0 (5, 1)6 : 1/2 (5, 3)−4 : 1/3
(1, 1)0 : 0 (5, 1)−6 : 1/2 (10, 3)2 : 5/6
(1, 8)0 : 0 (10, 3 )−2 : 1/6 (5, 3 )4 : 2/3
Fig. 4. Example of a (4 + 2)-dimensional orbifold allowing for Z6 twists. The fundamental space consists of two
triangles. The geometry can be illustrated by gluing together the edges of the depicted triangles, thus leading to a
triangle with a front- and a backside.
Fig. 4) is associated with the first of these fixed points; the twists P 2 and P 3 are associated
with the remaining two fixed points. By construction, the order of rotation in the two extra
dimensions coincides with the order of the twist in the gauge group. It is then straight-
forward to determine the gauge groups which survive at these fixed points. In the actual
example, they turn out to be SU(6) × SU(3)F and SU(8), respectively. The content of non-
vanishing fields at these fixed points is also found to be anomaly-free under the relevant sur-
viving gauge group in both cases, which implies the absence of localized anomalies [44].
The complete model is, however, not free of anomalies. This is due to localized
anomalies at the P6 fixed points, where the gauge group is SU(5) × U(1) × SU(3)F .
However, the SU(5) part by itself is free of localized anomalies even at this fixed point.
Thus, if SU(3)F × U(1) is broken, as it has to be in order to describe reality, there are no
anomalies. The desired breaking of the unwanted symmetries may be due to fields which
live on the fixed points, however, discussing such possibilities is beyond the scope of this
study. Note also that if we were to break the additional symmetry by rank-reducing twists,
fewer matter fields would survive. That is also the reason why we do not use rank-reducing
twists in the next two models.
6.2. E8 → GPS × U(1)3
Under E8 → SO(10) × SU(4), the adjoint representation of E8 decomposes like

248 → (45, 1) ⊕ (1, 15) ⊕ (16, 4) ⊕ (16, 4 ) ⊕ (10, 6). (6.3)
SO(10) contains the Pati–Salam group [3] GPS = SU(4) × SU(2) × SU(2) whereby
45 → (15, 1, 1) ⊕ (1, 3, 1) ⊕ (1, 1, 3) ⊕ (6, 2, 2), (6.4a)
16 → (4, 2, 1) ⊕ (4, 1, 2), (6.4b)
10 → (6, 1, 1) ⊕ (1, 2, 2). (6.4c)
This breaking can be achieved by using the rotation −1 ∈ SU(2) for the first (or the second)
SU(2). In addition, we can now impose the twist

F = exp 2πi diag(1/3, −1/6, x, −1/6 − x) ∈ SU(4), (6.5)
where, e.g., x = 14 , in order to break SU(4) → [U(1)]3 . The charges of the 4 are then given
by qi ∈ {1/3, −1/6, x, −1/6 − x}. The charges of the 6 are qi + qj with i = j since the 6 is
the antisymmetric part of the 4 × 4 of SU(4), and finally the charges of the 15 are qi − qj .
Together with the R-symmetry transformation

R = exp 2πi diag(−1/3, −1/3, −1/3) , (6.6)
three chiral generations of matter and three Higgs, i.e., (1, 2, 2), survive. Only for certain
x, additional fields will possess zero-modes, and we will choose x to equal none of these
values. Here, the number of generations is due to dimensional reduction of (N1 , N2 ) =
(1, 1) SUSY in 6d to 4d. The surviving gauge group is GPS × [U(1)]3 . The geometry is
given by an equilateral triangle with the three corners corresponding to three identical fixed
points.
Obviously, for such a construction, the geometric twist, i.e., the rotation in the two
extra dimensions, is of a lower order than the group theoretical twist. This requires going
beyond the usual field-theoretic orbifold constructions (although the geometry is still an
orbifold). As proposed in Section 5, we define a field theory on a manifold with three
conical singularities, each of them possessing a deficit angle of 2π/3. This construction is
then an equilateral triangle. We then add Wilson lines such that the group-theoretical twist
P at two of the fixed points equals the one described above. The twist at the third fixed
point is then constrained to be P −2 by the global geometry.
At each singularity of the conifold, the Pati–Salam part of the gauge group is anomaly-
free. This is obvious for first two fixed points since the non-vanishing fields are those of
the standard model with three Higgs doublets. At the third fixed point, the gauge symmetry
is enhanced to the group SO(10), which has no 4d anomalies.7 Again, investigating mech-
anisms to break the extra U(1)s as well as GPS to GSM is beyond the scope of this paper.
Note finally that this particular model can be viewed as an extension of [4], where three
generations arise from the three chiral superfields present in the 4d description of a 10d
SYM theory, i.e., they follow from the presence of three complex extra dimensions.8 The
new points in our construction are the doublet–triplet splitting solution arising from the
breaking to the Pati–Salam group (see [48] and the recent related stringy models of [49])
and the realization of all rather than just part of the matter fields in terms of the SYM
multiplet.
7 Quite generally, the anomaly at a given conical singularity can be calculated from the zero-mode anomaly by
considering a conifold where this specific singularity appears several times (possibly together with other conical
singularities, the anomalies of which are already known) [45]. However, we do not investigate this further in the
present paper. For recent work on the explicit calculation of anomalies in 6d models see [36,44,46].
8 It has been claimed that this is related to the mechanism for obtaining three generations used in the string
theory models reviewed in [47].
Table 7
Table of the phase factors for the different multiplets of GPS ×SU(3)F ×U(1) ⊂
E8 . Zeros correspond to surviving gauge bosons; other phases which are
compensated by the R-symmetry transformation are written in boldface
(15, 1, 1; 10 ): 0 (1, 1, 1; 3−4/3 ): 2/3 (4, 1, 2; 3−1 ): 11/12

(1, 3, 1; 10 ): 0 (4, 2, 1; 31 ): 1/12 (4, 1, 2; 1−3 ): 1/4
(1, 1, 3; 10 ): 0 (4, 2, 1; 1−3 ): 3/4 (6, 1, 1; 31 ): 5/6
(1, 1, 1; 80 ): 0 (4, 1, 2; 31 ): 7/12 (6, 1, 1; 3−1 ): 1/6
(1, 1, 1; 10 ): 0 (4, 1, 2; 13 ): 1/4 (1, 2, 2; 31 ): 1/3
(6, 2, 2; 10 ): 1/2 (4, 2, 1; 3−1 ): 11/12 (1, 2, 2; 3−1 ): 2/3
(1, 1, 1; 34/3 ): 1/3 (4, 2, 1; 13 ): 3/4
6.3. E8 → GPS × SU(3)F × U(1)
Alternatively, we can obtain GPS from E8 and maintain an SU(3)F flavour symmetry by
breaking the extra SU(4) of the decomposition (6.3) to SU(3)F × U(1). In order to achieve
this breaking, we take a central element of SU(3),

P = exp 2πi(1/3, 1/3, 1/3, 0) . (6.7)
The phases which arise by combining this twist with exp(−2πi/41) ∈ SU(4) are listed in
Table 7. Now let us simultaneously impose an R-symmetry twist

R = exp 2πi diag(−1/12, −7/12, −1/3) . (6.8)
It is then easy to see from Table 7 that the zero modes which emerge in the matter sector
are three generations of SM matter, three Higgs and three additional neutrinos.
In order to realize such a model, we have again to relax the constraints of usual orbifold
models, and therefore consider a manifold with a conical singularity with deficit angle
2π · 5/12 instead (cf. Section 5). To be more specific, we envisage the geometry of the
model as an isosceles triangle with an angle of 2π · 5/12. Each corner corresponds to a
fixed point, and we are free to choose both π/12 fixed points identically. By construction,
the group-theoretical twist P at the π/12 fixed points generate a Z12 , i.e., P 12 = 1.
At the remaining 2π · 5/12 ‘corner’, we choose the twist P 10 = P −2 for consistency.
Interestingly, a quick inspection of Table 7 reveals that the there surviving gauge symmetry
is SO(10). Obviously, the SO(10) part of the gauge theory at this fixed point is anomaly-
free automatically.
Once more, discussing the breaking of the extra gauge symmetry is beyond the scope
of this study.
7. Conclusions
We have explored some of the group-theoretical possibilities in orbifold GUTs. In

particular, we showed that, given a simple gauge group G, the breaking to any maximal-
rank regular subgroup can be achieved by orbifolding.
We further studied rank reduction and found that simple group factors can always be
broken completely. This is possible when using non-Abelian twists, and also if twists
commute but the corresponding generators do not. Using such constructions in orbifolding
is made possible by embedding a non-Abelian (or even Abelian) space group into the gauge
group.
We then extended the familiar concept of orbifold GUTs by replacing the orbifolds
by manifolds with conical singularities. The possibilities we discussed include orbifold
geometries endowed with unrestricted Wilson lines wrapping the conical singularities,
manifolds with conical singularities with arbitrary deficit angles, and combinations thereof.
Finally, we presented three specific models where three generations of fields carrying
the SM quantum numbers come from a SYM theory in 6d. While the first one is a
conventional orbifold model illustrating the usefulness of our group theoretical methods,
the two others are based on the two new concepts mentioned above.
To summarize, we explored several new and interesting methods and possibilities which
can be used in orbifold GUTs and their generalizations.
As none of our models is yet completely realistic, more effort is required in order to
discuss phenomenological consequences. However, it is very appealing how easily three
generations can be obtained and the doublet-triplet splitting problem can be solved. Thus,
promoting our models to realistic ones in future studies appears to be worthwhile.
Note added
While this paper was being finalized, Ref. [50] appeared where Dynkin diagram
techniques were used as well. Aspects of our analysis not addressed by [50] include, in
particular, the breaking of any simple group to all maximal-rank regular subgroups, rank-
reduction, as well as several new field-theoretic concepts and models.
Acknowledgements
We would like to thank Fabian Bachmaier, Wilfried Buchmüller, John March-Russell,

Hans-Peter Nilles, Mathias de Riese and Marco Serone for useful discussions.
References
[1] H. Georgi, S. Glashow, Phys. Rev. Lett. 32 (1974) 438.

[2] H. Georgi, AIP Conf. Proc. 23 (1975) 575;
H. Fritzsch, P. Minkowski, Ann. Phys. 93 (1975) 193.
[3] J. Pati, A. Salam, Phys. Rev. D 8 (1973) 1240;
J. Pati, A. Salam, Phys. Rev. D 10 (1974) 275.
[4] T. Watari, T. Yanagida, Phys. Lett. B 532 (2002) 252, hep-ph/0201086.
[5] K.S. Babu, S.M. Barr, B.S. Kyae, Phys. Rev. D 65 (2002) 115008, hep-ph/0202178.
[6] G. Burdman, Y. Nomura, Nucl. Phys. B 656 (2003) 3, hep-ph/0210257.
[7] I. Gogoladze, Y. Mimura, S. Nandi, hep-ph/0304118.
[8] E. Witten, Nucl. Phys. B 258 (1985) 75.
[9] L.J. Dixon, J.A. Harvey, C. Vafa, E. Witten, Nucl. Phys. B 261 (1985) 678;
L.J. Dixon, J.A. Harvey, C. Vafa, E. Witten, Nucl. Phys. B 274 (1986) 285.
[10] Y. Kawamura, Prog. Theor. Phys. 105 (2001) 999, hep-ph/0012125.

[11] G. Altarelli, F. Feruglio, Phys. Lett. B 511 (2001) 257, hep-ph/0102301.
[12] L.J. Hall, Y. Nomura, Phys. Rev. D 64 (2001) 055003, hep-ph/0103125.
[13] A. Hebecker, J. March-Russell, Nucl. Phys. B 613 (2001) 3, hep-ph/0106166.
[14] T. Asaka, W. Buchmüller, L. Covi, Phys. Lett. B 523 (2001) 199, hep-ph/0108021.
[15] L.J. Hall, Y. Nomura, T. Okui, D.R. Smith, Phys. Rev. D 65 (2002) 035008, hep-ph/0108071.
[16] E.B. Dynkin, The structure of semi-simple algebras, Am. Math. Soc. Transl. 1 (1950) 1–143;
E.B. Dynkin, Am. Math. Soc. Transl. 9 (1962) 328–469.
[17] E.B. Dynkin, Semi-simple subalgebras of semi-simple Lie algebras, Am. Math. Soc. Transl., Ser. 2 6 (1957)
111–244.
[18] E.B. Dynkin, Maximal subgroups of the classical groups, Am. Math. Soc. Transl., Ser. 2 6 (1957) 245–378.
[19] E.B. Dynkin, Selected Papers, Am. Math. Society, 1999.
[20] R. Gilmore, Lie Groups, Lie Algebras, and some of their Applications, Krieger, 1994.
[21] R.N. Cahn, Semisimple Lie Algebras and their Representations, Benjamin–Cummings, 1984.
[22] H. Georgi, Lie Algebras in Particle Physics. From Isospin to Unified Theories, 2nd Edition, in: Perseus
Books, Vol. 54, 1999.
[23] R. Slansky, Phys. Rep. 79 (1981) 1.
[24] A. Hebecker, J. March-Russell, Nucl. Phys. B 625 (2002) 128, hep-ph/0107039.
[25] N. Haba, M. Harada, Y. Hosotani, Y. Kawamura, Nucl. Phys. B 657 (2003) 169, hep-ph/0212035.
[26] L.J. Hall, Y. Nomura, hep-ph/0212134;
M. Quiros, hep-ph/0302189.
[27] H.D. Fegan, An Introduction to Compact Lie Groups, World Scientific, Singapore, 1991.
[28] K.R. Dienes, J. March-Russell, Nucl. Phys. B 479 (1996) 113, hep-th/9604112.
[29] Y. Katsuki, Y. Kawamura, T. Kobayashi, N. Ohtsubo, K. Tanioka, Prog. Theor. Phys. 82 (1989) 171;
Y. Katsuki, Y. Kawamura, T. Kobayashi, N. Ohtsubo, Y. Ono, K. Tanioka, Nucl. Phys. B 341 (1990) 611.
[30] M. Golubitsky, B. Rothschild, Primitive subalgebras of exceptional Lie algebras, Pacific J. Math. 39 (2)
(1971) 371–393.
[31] T. Asaka, W. Buchmüller, L. Covi, Phys. Lett. B 540 (2002) 295, hep-ph/0204358.
[32] N. Haba, Y. Shimizu, hep-ph/0212166.
[33] L.E. Ibanez, H.P. Nilles, F. Quevedo, Phys. Lett. B 192 (1987) 332.
[35] E.P. Wigner, Group Theory, Academic Press, 1949.
[36] S. Groot Nibbelink, hep-th/0305139.
[37] Y. Nomura, D.R. Smith, N. Weiner, Nucl. Phys. B 613 (2001) 147, hep-ph/0104041.
[38] A. Hebecker, Nucl. Phys. B 632 (2002) 101, hep-ph/0112230.
[39] T.J. Li, Nucl. Phys. B 633 (2002) 83, hep-th/0112255.
[40] Y. Hosotani, Phys. Lett. B 126 (1983) 309;
Y. Hosotani, Ann. Phys. 190 (1989) 233.
[41] P. Candelas, G.T. Horowitz, A. Strominger, E. Witten, Nucl. Phys. B 258 (1985) 46.
[42] N. Marcus, A. Sagnotti, W. Siegel, Nucl. Phys. B 224 (1983) 159.
[43] Y. Imamura, T. Watari, T. Yanagida, Phys. Rev. D 64 (2001) 065023, hep-ph/0103251.
[44] T. Asaka, W. Buchmüller, L. Covi, Nucl. Phys. B 648 (2003) 231, hep-ph/0209144.
[45] F. Gmeiner, S. Groot Nibbelink, H.-P. Nilles, M. Olechowski, M.G. Walter, Nucl. Phys. B 648 (2003) 35,
hep-th/0208146.
[46] G. von Gersdorff, M. Quiros, hep-th/0305024.
[47] A.E. Faraggi, talk at 4th Int. Conf. on Phys. Beyond the Standard Model, Lake Tahoe, 1994, hep-ph/9501288.
[48] R. Dermisek, A. Mafi, Phys. Rev. D 65 (2002) 055002, hep-ph/0108139;
H.D. Kim, S. Raby, JHEP 0301 (2003) 056, hep-ph/0212348.
[49] J.E. Kim, hep-th/0301177;
K.S. Choi, J.E. Kim, hep-ph/0305002.
[50] K.S. Choi, K. Hwang, J.E. Kim, hep-th/0304243.
T-duality, and the K-theoretic partition function

of type IIA superstring theory
Gregory Moore a , Natalia Saulina b
a Department of Physics, Rutgers University, Piscataway, NJ 08855-0849, USA
b Department of Physics, Princeton University, Princeton, NJ 08544, USA
Received 10 June 2003; received in revised form 10 July 2003; accepted 30 July 2003
Abstract
We study the partition function of type IIA string theory on 10-manifolds of the form T 2 × X,
where X is 8-dimensional, compact, and spin. We pay particular attention to the effects of the
topological phases in the supergravity action implied by the K-theoretic formulation of RR fields,
and we use these to check the T-duality invariance of the partition function. We find that the partition
function is only T-duality invariant when we take into account the T-duality anomalies in the RR
sector, the fermionic path integral (including 4-fermi interaction terms), and 1-loop corrections
including worldsheet instantons. We comment on applications of our computation to speculations
about the role of the Romans mass in M-theory. We also discuss some issues which arise when one
attempts to extend these considerations to checking the full U-duality invariance of the theory.
PACS: 11.25.-w
1. Introduction and summary
Duality symmetries, such as the U-duality symmetry of toroidally compactified M-

theory, have been of central importance in the definition of string theory and M-theory.
Topologically nontrivial effects associated with the RR sector have also played a crucial
role in defining the theory. It is currently believed that RR fieldstrengths (and their D-
brane charge sources) are classified topologically using K-theory [1–8]. Unfortunately,
this classification is not U-duality invariant. Finding a U-duality invariant formulation of
E-mail address: saulina@princeton.edu (N. Saulina).
doi:10.1016/j.nuclphysb.2003.07.028
28 G. Moore, N. Saulina / Nuclear Physics B 670 (2003) 27–89
M-theory which at the same time naturally incorporates the K-theoretic formulation of RR
fields remains an outstanding open problem.
With this problem as motivation, the present paper investigates the interplay between
the K-theoretic formulation of RR fields and the T-duality group, an important subgroup of
the full U-duality group. While T-duality invariance of the theory was one of the guiding
principles in the definition of the K-theoretic theta function [4,7] we will see that the full
implementation of T-duality invariance of the low energy effective action of type II string
theory is in fact surprisingly subtle, even on backgrounds as simple as T 2 × X, where T 2
is a two-dimensional torus, and X is an 8-dimensional compact spin manifold. We will
show that, in fact, in the RR sector there is a T-duality anomaly. This anomaly is cancelled
by a compensating anomaly from fermion determinants together with quantum corrections
to the 8D effective action. A by-product of our computation is a complete analysis of the
1-loop determinants of IIA supergravity on X × T 2 .
As an application of our discussion, we reexamine a proposal of Hull [9] for interpreting
the Romans mass of IIA supergravity in the framework of M-theory. We will show that,
while the interpretation cannot hold at the level of classical field theory, it might well hold
as a quantum-mechanical equivalence. In Section 10 we comment on some of the issues
which arise in extending our computation to a fully U-duality invariant partition function.
This includes the role of twisted K-theory in formulating the partition sum.
This paper is long and technical. Therefore we have attempted to write a readable
summary of our results in the remainder of the introduction.
1.1. The effective eight-dimensional supergravity, and its partition function
Previous studies of the partition function in type II string theory [4,7] considered the
limit of a large 10-manifold. One chose a family of Riemannian metrics g = t 2 g0 with
t → ∞ and g0 fixed. Simultaneously, one took the string coupling to zero. The focus of
these works was on the sum over classical field configurations of the RR fields. In this
paper we consider the limit where only 8 of the dimensions are large. The metric has the
form
ds 2 = dsT2 2 + t 2 dsX
2
, (1.1)
where dsT2 2 is flat when pulled back to T 2 . The background dilaton gstring
2 = e2ξ is constant.
We will work in the limit
t → ∞,
e−2ξ := e−2φ V → ∞, (1.2)
where V is the volume of T2and φ is the 10-dimensional dilaton. Finally—and this
is important—until Section 10 we assume the background NSNS 3-form flux, H , is
is a globally well-defined harmonic
identically zero. In particular, the 2-form potential, B,
form on X × T 2 .
As is well known, the background data for the toroidal compactification (1.1) include a
pair of points (τ, ρ) ∈ H ×H where H is the upper half complex plane. τ is the Teichmuller
parameter of the torus and ρ := B0 + iV , where B0 dσ 8 ∧ dσ 9 is an harmonic 2-form
G. Moore, N. Saulina / Nuclear Physics B 670 (2003) 27–89 29
on T 2 . While we work in the limit (1.2), within this approximation we work with exact
expressions in the geometrical data (τ, ρ). In this way we go beyond [7].
It is extremely well known that the low energy effective 8D supergravity theory obtained
by Kaluza–Klein reduction of type II supergravity on T 2 has a “U-duality symmetry”
which is classically SL(3, R) × SL(2, R), and is broken to D := SL(3, Z) × SL(2, Z) by
quantum effects [10–14]. These are symmetries of the equations of motion and not of
the action. (The implementation of these symmetries at the level of the action involves
a Legendre transformation of the fields.) What is perhaps less well known is that the
K-theoretic formulation of RR fields leads to an extra term in the supergravity action
which is nonvanishing in the presence of nontrivial flux configurations. Indeed, the proper
formulation of this term is unknown for arbitrary flux configurations with [H 3 ] = 0, but for
topologically trivial NSNS flux the extra term is known [7] and is recalled in Eqs. (1.14)
and (1.15). This term breaks naive duality invariance of the classical supergravity theory
already for the T-duality subgroup of the U-duality group, and makes the discussion of
T-duality nontrivial.
Let us now summarize the fields and T-duality transformation laws in the conventional
description of the eight-dimensional effective supergravity theory on X. The T-duality
group is DT = SL(2, Z)τ × SL(2, Z)ρ . The theory has the following bosonic fields. From
the NSNS sector there is a scalar t, characterizing the size of X, a unit volume metric
gMN , a 2-form potential1 B(2) , with fieldstrength H(3) , and a dilaton ξ , all of which are
invariant under DT . In addition, there is a multiplet of 1-form potentials Amα
(1) transforming
in the (2, 2) of DT . Finally, the pair of scalars (τ, ρ), transform under (γ1 , γ2 ) ∈ DT as
(τ, ρ) → (γ1 · τ, γ2 · ρ) where γ · is the action by a fractional linear transformation. We
therefore call the factors SL(2, Z)τ , SL(2, Z)ρ , respectively.
α , p = 0, 2, α =
The fieldstrengths from the RR sector include a 0-form and a 2-form, g(p)
1, 2 transforming in the (1, 2) of DT , and a 1-form and 3-form g(p)m , p = 1, 3, m = 8, 9
transforming in the (2 , 1) of DT . Finally there is a 4-form fieldstrength g(4) on X. This
field does not transform locally under T-duality, rather its equation of motion mixes with
its Bianchi identity [14]. The fermionic partners are described in Section 7.
The real part of the standard bosonic supergravity action takes the form
(8D)
3
Re Sboson = SNSNS + Sp (g(p) ) + S4 (g(4) ). (1.3)
p=0
In the action (1.3) all of the terms except for the last term are manifestly T-duality
invariant. The detailed forms of the actions are

1
SNSNS = e−2ξ t 6 R(g) + 4 dξ ∧ ∗dξ + 28t −2 dt ∧ ∗dt
2π
X
1 1 dτ ∧ ∗dτ 1 dρ ∧ ∗dρ
+ t 2 H(3) ∧ ∗H(3) + t 6 + t6
2 2 (Im τ )2 2 (Im ρ)2

1 nβ
+ t 4 gmn Gαβ Fmα
(2) ∧ ∗F (2) , (1.4)
2
1 We will always indicate by the subscript (p) the degree p of a differential p-form on X.
where ∗ stands for the Hodge dual with the metric gMN , we also denote
1 nβ
(2) = dA(1) ,
Fmα H(3) = dB(2) − !mn Eαβ Amα
mα
(1) F(2) , (1.5)
2
!mn and Eαβ are invariant antisymmetric tensors for SL(2, Z)τ and SL(2, Z)ρ , respectively,
gmn = M(τ ), g mn = M(τ )−1 , Gαβ = M(ρ) (1.6)

and finally

1 1 Re z
M(z) := . (1.7)
Im z Re z |z|2
The real part of the RR sector action is given by

3

8 β
Sp (g(p) ) = π t Gαβ g(0)
α
∧ ∗g(0) + t 6 g mn g(1)m ∧ ∗g(1)n
p=0 X
β
+ t 4 Gαβ g(2)
α
∧ ∗g(2) + t 2 g mn g(3)m ∧ ∗g(3)n (1.8)
together with

S4 (g(4) ) = π Im(ρ)g(4) ∧ ∗g(4) . (1.9)
X
1.2. The semiclassical expansion
The vevs of the two fields t and e−2ξ (the 8-dimensional length scale of X and the
inverse-square 8D string coupling) define semiclassical expansions when they become
large. We will expand around a solution of the equations of motion on X. To leading
order in our expansion this means X admits a Ricci-flat metric2 gMN . We also have
constant scalars t, ξ, τ, ρ, and Fmα(2) = 0, H(3) = 0, so the background action SNSNS is
zero. Finally, we expand around a classical field configuration for the RR fluxes, and to
leading order these fluxes g(p) are harmonic forms. Nonzero fluxes contribute terms to the
partition function going like O(e−t
8−2p
).
Let us consider the leading order contribution to the partition function. There are several
sources of contributions even at leading order, but, since we are interested in questions of
T-duality, most of these can be neglected.3 The volume of X suppresses the contribution
of fluxes g(p) , p = 0, 1, 2, 3, and to leading order in the t → ∞ expansion these can be set
to their classical values. Note, however, that neither the string coupling, nor the volume of
2 Almost nothing in what follows relies on the Ricci-flatness of the metric. We avoid using this condition since
a T-duality anomaly on non-Ricci-flat manifolds would signal an important inconsistency in formulating string
theory on manifolds of topology X × T 2 .
3 In particular we are neglecting determinants of KK and string modes, and perturbative corrections O(g 2
string ).
These are all T-duality invariant. The backreaction of nonzero RR fluxes on the NSNS action simply renormalizes
V to Veff , where ρ = B0 + iVeff is the variable on which SL(2, Z)ρ acts by fractional linear transformations.
X, suppress the action for g(4) , and thus we must work in a fully quantum-mechanical way
with this field. This is just as well, since (not coincidentally) this is the term in the action
which is not manifestly T-duality invariant. Fortunately, in our approximation, g(4) is a
free, nonchiral field and hence quantization is straightforward (after the K-theory subtleties
are taken into account). Including subleading terms in the expansion parameter t involves
(among other things) summing over the RR fluxes g(p) , p = 0, 1, 2, 3.
Finally, in order to be consistent with our approximation scheme we must allow
the possibility of flat potentials in the background.4 These contribute nontrivially to
the partition function through important phases and accordingly, we will generalize our
background to include these. The real part of the action for the flat configurations
vanishes, of course, and hence in the physical partition function one must integrate over
these flat configurations. In the RR sector the flat potentials are thought to be classified
by K 1 (X10 ; U (1)) [5]. Since we assume NSNS fieldstrengths are zero it follows from
Eqs. (10.9), (10.10) that the connected component makes no contribution to the phase so
we can ignore it. (The group of components of K 1 (X10 ; U (1)) is accounted for in the sum
over cohomology classes â below.)5 The space of flat NSNS potentials is
4
H 2 X; U (1) × H 1 X; U (1) .
In this paper we will work only with the identity component of this torus. Accordingly, we
will identify the space of flat NSNS potentials with the torus
1
H2 (X) H (X) 4
× , (1.10)
HZ2 (X) HZ1 (X)
p
where Hp (X) is a space of harmonic p-forms on X and HZ (X) is the lattice of integrally
normalized harmonic p-forms on X. The first factor is for B(2) and the second factor for
(1) transforming in the (2, 2) of DT .
the fields Amα
Putting all these ingredients together the partition function we wish to study can be
schematically written as

Z(t, gMN , ξ, τ, ρ) = dµflat Det · e−Scl + · · · , (1.11)
flat potentials RR fluxes
where dµflat is a T-duality invariant measure on the flat potentials, Det is a product of
1-loop determinants and Scl is the classical action. Now, to investigate T-duality it is
convenient to denote by F the collection of all fields occurring in (1.11) which transform
locally and linearly under DT . These include the flat NSNS potentials above as well as
the classical fluxes g(p) , p = 0, . . . , 3. We introduce a measure [dF ] on F which includes
integration over the flat potentials and summation over the fluxes for p = 0, 1, 2, 3. This
measure is T-duality invariant, and we can write

Z(t, gMN , ξ, τ, ρ) = [dF ] Z(F ; t, gMN , ξ, τ, ρ). (1.12)
4 By “flat” we mean the DeRham representative of the relevant fieldstrength is zero.

5 If treated as differential forms, RR zero modes do contribute to the overall dependence of the partition sum
on t˜ = te−ξ/3 . See Eq. (7.39). In the K-theoretic treatment they also give a factor of |Ktors
0 (X × T 2 )|.
The invariance of (1.12) under the subgroup SL(2, Z)τ of the T-duality group is essentially
trivial. The relevant actions and determinants are all based on SL(2, Z)τ -invariant
differential operators. The invariance of the theory under SL(2, Z)ρ is, however, much
more nontrivial. Therefore, we simplify notation and just write Z(F , ρ) for the integrand
of (1.12). Now, checking T-duality invariance is reduced to checking the invariance of
Z(F , ρ). This function is constructed from:
(a) the K-theoretic sum over RR fluxes of g(4) in the presence of F ;

(b) the integration over the Fermi zeromodes in the presence of g(4) and F ;
(c) the inclusion of 1-loop determinants, including determinants of the 8D supergravity
fields and the quantum corrections due to worldsheet instantons.
In the following subsections we sketch how each of these elements enters Z(F , ρ).
Briefly, the K-theoretic sum over RR fluxes g(4) leads to a theta function Θ(F , ρ). This
function turns out to transform anomalously under T-duality. The integration over the
fermion zeromodes corrects this to a function Θ(F , ρ). This function still transforms
anomalously. The inclusion of 1-loop effects, including the string 1-loop effects finally
cancels the anomaly.
1.3. The K-theoretic RR partition function
In order to write explicit formulae for the quantities in (1.12) we must turn to the K-
theoretic formulation of RR fields. In practical terms the K-theoretic formulation alters
the standard formulation of supergravity in two ways: first it restricts the allowed flux
configurations through a “Dirac quantization condition” on the fluxes. Second, it changes
the supergravity action by the addition of important topological terms in the action.6
In more detail, the K-theoretic Dirac quantization condition states that the DeRham
class of the total RR fieldstrength [G/(2π)] is related to a K-theory class x ∈ K 0 (X10 ) via

G
= ch(x) A. (1.13)
2π
The topological terms in the action can be described as follows. On a general 10-manifold
this term involves the mod-two index of a Dirac operator and cannot even be written as a
traditional local term in the supergravity action [4,5,7]. In the case of zero NSNS fluxes,
the general expression for the phase in the supergravity theory is
Im(S10D ) = −2πΦ, Φ = Φ1 + Φ2 , (1.14)
where e2πiΦ2
is the mod-two index and Φ1 is given by the explicit expression

1 G2 5 1 G2 3 G4 p1 G0
Φ1 = − + + 1+
15 2π 6 2π 2π 12 8π
X10
6 It also alters the overall normalization of the bosonic determinants by changing the nature of the gauge group
for RR potentials, but we will not discuss this in the present paper.

G2 p1 G4 A8 G0 G0 p1 2
− + (1 + + , (1.15)
2π 48 2π 2 2π 4π 48
is expressed in terms
where G2j , j = 0, 1, 2, are RR fluxes on X10 , p1 = p1 (T X10 ) and A
of the Pontryagin classes of X10 as

= 1 − 1 p1 + 1 7p12 − 4p2 .
A (1.16)
24 5760
In the case that we reduce to 8 dimensions, taking our manifold to be of the form X × T 2
with the choice of supersymmetric spin structure on T 2 the above considerations simplify
and can be made much more concrete.
Consider first the Dirac quantization condition. We reduce RR fieldstrengths as7
G0
= g(0)
2
,
2π
G2
= g(0)
1
dσ 8 ∧ dσ 9 + g(1)m ∧ dσ m + g(2)
2
,
2π
G4
= g(4) + g(3)m ∧ dσ m + g(2)
1
∧ dσ 8 ∧ dσ 9 , (1.17)
2π
where σ m , m = 8, 9, are coordinates on T 2 . In the K-theoretic formulation of flux
α ,g
quantization the fieldstrengths g(4) , g(3)m , g(2) α
(1)m , g(0) are related to certain integral
cohomology classes which we denote as

e
a ∈ H 4 (X, Z), fm ∈ H 3 (X, Z) ⊗ Z2 , eα = ∈ H 2 (X, Z) ⊗ Z2 ,
e

n1
γm ∈ H 1 (X, Z) ⊗ Z2 , nα = ∈ H 0 (X, Z) ⊗ Z2 . (1.18)
n0
The explicit relation between these classes and the g(p) is somewhat complicated and
given in Eq. (4.3). The K-theoretic Dirac quantization condition leaves all integral classes
in (1.18) unconstrained except for fm . One finds that Sq 3 (fm ) = 0. As explained in
Sections 3.3 and 5.2 “turning on” flat NSNS potentials corresponds to acting on the K-
theory torus by an automorphism changing the holonomies of the flat connection on the
torus. In concrete terms, turning on flat potentials modifies the reduction formulae (1.17)
according to Eqs. (5.15) to (5.18).
Now let us consider the phase. It turns out that on 10-folds of the form X × T 2 the phase
e 2πiΦ2 arising from the mod 2 index may be expressed in concrete terms as
exp[2πiΦ2 ]

= exp iπ g(3)8 ∪ Sq 2 (g(3)9) + g(3)8 ∪ Sq 2 (g(3)8) + g(3)9 ∪ Sq 2 (g(3)9)
X
7 Beware of notation! The subscript (p) indicates form degree, while the other sub- and superscripts on g
(p)
2 is the second component of a doublet g α of
indicate DT transformation properties. Thus, for example, g(0) (0)
0-forms.
2
2
g(0) 1 2 2
+ iπ g(0)+ g(4) +
A8 p1 − g(2)
48 2
X

1 2 p1
× g(2) − g(0) g(2) + g(1)8 g(1)9 +
1 2
2

p 2 2 3 2 2 mn
+ 1 + g(1)8g(1)9 g(2) − g(2) ! g(1)m g(3)n . (1.19)
8
This expression is cohomological although it is still unconventional in supergravity theory
since it involves the mod-two valued Steenrod squares, denoted Sq 2 (g(3) ), in the first line.
The above topological term (1.14) is deduced from the K-theory theta function ΘK
defined in [4,5,7], and reviewed below. As explained above, it is convenient to fix the fields
F . We can define a function Θ(F , ρ) by writing ΘK as a sum

ΘK = e−SB (F ) Θ(F , ρ). (1.20)
The sum is over all integral classes except a. That is, we sum over nα , γm , eα , fm subject
to the constraint on Sq 3 fm . The action SB (F ) is the manifestly T-duality invariant action
for the fluxes given in (1.8). ΘK is a function of gMN , ρ, τ and the flat background NSNS
fluxes. As we have mentioned, turning on flat potentials corresponds, in the K-theoretic
interpretation, to acting by automorphisms of the K-theory group K 0 (X) ⊗ R. These
automorphisms act naturally on the theta function. We give concrete formulae for this
action by showing how the inclusion of nonzero flat NSNS fields B0 , B(2) , Amα (1) modifies
the phase Φ. The explicit formula is in Eqs. (5.20)–(5.24).
Since the K-theoretic constraint Sq 3 a = 0, a ∈ H 4 (X, Z) is automatically satisfied
on spin 8-folds X it turns out that Θ(F ) is, essentially, a Siegel–Narain theta function
for the lattice H 4 (X; Z). More precisely, there is a quadratic form on H 4 (X; R) given
by Q = Im(ρ)H I − i Re(ρ)I where H is the action of Hodge ∗ and I is the integral
intersection pairing on H 4 (X, Z). Then

F) α̃
Θ(F , ρ) = e i2π2Φ(
Θ (Q). (1.21)
β̃
α̃
Here Θ (Q) is the Siegel–Narain theta function with characteristics. The characteristics
β̃
) in
are written explicitly in Eqs. (5.10), (5.20), and (5.21). Finally, the prefactor 2Φ(F
(1.21) is defined in (5.23) and (5.24).
1.4. T-duality transformations
One of the more subtle aspects of the K-theoretic formulation of RR fluxes, is that
the very formulation of the action depends crucially on a choice of polarization of the
K-theory lattice K(X10 ) with respect to the quadratic form defined by the index. In the
above discussion we have chosen the “standard polarization” for IIA theory, i.e., Γ2 is the
sublattice of K(X10 ) with vanishing G4 , G2 , G0 . Γ1 is then a complementary Lagrangian
sublattice such that K(X10 ) = Γ1 + Γ2 . The standard polarization is distinguished for
any large 10-manifold in the following sense. When the metric of X10 is scaled up

→ t ĝM the action X10 ĝ |G2p | of the Type IIA RR 2p-form scales as t
ĝM 2 2 10−4p .
N N
This allows the sensible approximation of first summing only over G4 , with G2 = G0 = 0,
then including G2 with G0 = 0, and finally summing over all classical fluxes G4 , G2 , G0 .
In the case of X10 = T 2 × X with the metric (1.1) the standard polarization is no longer
distinguished. Various equally good choices are related by the action of the T-duality
group DT on ΓK := K(X × T 2 ).8 In Section 4 we explain how the duality group DT
acts as a subgroup of symplectic transformations on the K-theory lattice and we give an
explicit embedding DT ⊂ Sp(2N, Z), where 2N = rank(ΓK ). As explained in Section 4.2,
since DT acts symplectically, the function Θ(F , ρ) must transform under T-duality as
Θ(γ · F , γ · ρ) = j (γ , ρ)Θ(F , ρ) where j (γ , ρ) is a standard transformation factor for
modular forms. Nevertheless, this transformation law leaves open the possibility of a T-
duality anomaly through a multiplier system in j (γ , ρ). In order to investigate this potential
anomaly more closely we must choose an explicit duality frame and perform the relevant
modular transformations.
We find that, in fact, the function Θ(F , ρ) does transform as a modular form with a
nontrivial “multiplier system” under SL(2, Z)ρ . That is, using the standard generators T , S
of SL(2, Z)ρ we have:
Θ(T · F , ρ + 1) = µ(T )Θ(F , ρ),

1 + 1 −
Θ(S · F , −1/ρ) = µ(S)(−iρ) 2 b4 (i ρ̄) 2 b4 Θ(F , ρ), (1.22)
where T · F , S · F denotes the linear action of DT on the fluxes. Here is the b4+ , b4−
dimension of the space of self-dual and anti-self-dual harmonic forms on X and the
multiplier system is

iπ
µ(T ) = exp 2
λ ,
4
X

iπ
µ(S) = exp 2
λ , (1.23)
2
X
where λ is the integral characteristic class of the spin bundle on X. (So, 2λ = p1 ). The
multiplier system is indeed nontrivial on certain 8-manifolds. As an example, on all Calabi–
Yau 4-folds we have the relation

1 8 − 4 + 1 χ
λ2 = 62 A (1.24)
4 12
X X
and hence µ is nontrivial if χ is not divisible by 12. In particular, a homogeneous

polynomial of degree 6 in P 5 , has χ = 2610. See, e.g., [15].
In more physical language, the “multiplier system” signals a potential T-duality
anomaly. Such an anomaly would spell disaster for the theory since the T-duality group
8 There is also a polarization on manifolds of the type S 1 × X (in our case X = S 1 × X ) where the measure
9 9
is purely real and the imaginary part of the action is an integral multiple of iπ (without flat NSNS potentials).
However, this polarization does not lead to a good long-distance approximation scheme.
should be regarded as a gauge symmetry of M-theory. Accordingly, we turn to the

remaining functional integrals in the supergravity theory. We will find that the anomalies
cancel, of course, but this cancellation is surprisingly intricate.
1.5. Inclusion of 1-loop effects
We first turn to the 1-loop functional determinants of the quantum fluctuations of the
bosonic fields. We show that these are all manifestly T-duality invariant functions of F
except for the quantum fluctuations of g(4) . The full bosonic 1-loop determinant DetB is
given in Eq. (6.20). The net effect of including the bosonic determinants is thus to replace
e−SB (F ) Θ(F , ρ) → ZB (F , ρ) := DetB e−SB (F ) Θ(F , ρ) (1.25)
Inclusion of this determinant alters the modular weight so that ZB (F , ρ) transforms with
weight ( 14 (χ + σ ), 14 (χ − σ )), in close analogy to the theory of abelian gauge potentials
on a 4-manifold [16,17]. Here χ, σ are the Euler character and signature of the 8-fold X.
The multiplier system (1.23) is left unchanged.
Now let us consider modifications from the fermionic path integral. Recall that we
may always regard a modular form as a section of a line bundle over the modular curve
H/SL(2, Z)ρ . On general grounds, we expect the fermionic path integral to provide a
trivializing line bundle. The gravitino and dilatino in the 8D theory transform as modular
forms under the T-duality group DT with half-integral weights and consequently they too
are subject to potential T-duality anomalies.
The inclusion of the fermions modifies the bosonic partition function in two ways:
through zeromodes and through determinants. The fermion action in the 8D supergravity
has the form
(8)
SFermi = Skinetic + Sfermi-flux + S4-fermi , (1.26)
where kinetic terms Skinetic as well as fermion-flux couplings Sfermi-flux are quadratic in
fermions and S4-fermi denotes the four-fermion coupling. Skinetic is T-duality invariant
but Sfermi-flux and S4-fermi contain some non-invariant terms. The non-invariant fermion
zeromode couplings are collected together in the form

S (zm)ninv
= 4π Im ρg(4) ∧ ∗Y(4) + 2π Im ρY(4) ∧ ∗Y(4) , (1.27)
X
where the harmonic 4-form Y(4) is bilinear in the fermion zeromodes. The explicit
expression for Y(4) can be found in Eqs. (7.21) and (7.41).
The inclusion of the integral over the fermionic zeromodes of Skinetic modifies the
partition function by replacing the expression Θ(F , ρ) in (1.21) by

F) α̂

Θ(F , ρ) = dµF e (zm) i2π 2Φ(
Θ (Q). (1.28)
β̂
Here

α̂
Θ (Q)
β̂
is a super-theta function for a super-abelian variety based on the K-theory theta function.
β̂ differ from α̃,
(This is explained in Appendix F.) In particular, the characteristics α̂, β̃
by expressions bilinear in the fermion zeromodes. Similarly, the prefactor 2Φ differs from
(zm)
2Φ by an expression quartic in the fermion zeromodes. Finally, dµF is a T-duality
invariant measure for the finite-dimensional integral over fermion and ghost zeromodes. It
includes the T-duality invariant term e−S
(zm)inv
from the action.
Including the one-loop fermionic determinants of the nonzero modes we finally arrive
at
ZB+F (F , ρ) := DetB DetF e−SB (F ) Θ(F

, ρ). (1.29)
The formula we derive for (1.29) allows a relatively straightforward check of the T-
duality transformation laws and we find:
ZB+F (T · F , ρ + 1) = µ(T )ZB+F (F , ρ),

ZB+F (S · F , −1/ρ)

X (p2 −λ X (p2 −λ
2) 2)
= (−iρ)1/4χ+1/8 (i ρ̄)1/4χ−1/8 ZB+F (F , ρ). (1.30)
Perhaps surprisingly, the fermion determinants have not completely trivialized the
RR contribution to the path integral measure. However, there is one final ingredient we
must take into account: in the low energy supergravity there are quantum corrections
which contribute to leading order in the t → ∞ and ξ → −∞ limit. From the string
worldsheet viewpoint these consist of a 1-loop term in the α expansion together with
worldsheet instanton
corrections. From the M-theory viewpoint we must include the one-
loop correction C3 X8 in M-theory together with the effect [18] of membrane instantons.
The net effect is to modify the action by the quantum correction

1 1
Squant = χ + p2 − λ2 log η(ρ)
2 4
X

1 1
+ χ− p2 − λ2 log η(−ρ̄) , (1.31)
2 4
X
where η(ρ) is the Dedekind function. The final combination
Z(F , ρ) = e−Squant ZB+F (F , ρ) (1.32)

is the fully T-duality invariant low energy partition function.
1.6. Applications
As a by-product of the above results we will make some comments on the open problem
of the relation of M-theory to massive IIA string theory. In [9] C. Hull made an interesting
suggestion for an 11-dimensional interpretation of certain backgrounds in the Romans
theory. One version of Hull’s proposal states that massive IIA string theory on T 2 × X
is equivalent to M-theory on a certain 3-manifold, the nilmanifold.
In Section 9 we review Hull’s proposal. For reasons explained there we are motivated
to introduce a modification of Hull’s proposal, in which one does not try to set up
a 1–1 correspondence between M-theory geometries and massive IIA geometries, but
nevertheless, the physical partition function Z(F , ρ) of the massive IIA theory can be
identified with a certain sum over M-theory geometries involving the nilmanifold. The
detailed proposal can be found in Section 9.3.
1.7. U-duality and M-theory
In the final section of the paper we comment on some of the issues which arise in trying
to extend these considerations to writing the fully U-duality-invariant partition function.
We summarize briefly the M-theory partition function on X × T 3 , we comment on the
SL(2, Z)ρ duality invariance, and we make some preliminary remarks on how one can see
K-theory theta functions for twisted K-theory from the M-theory formulation.
2. Review of T-duality invariance in the standard formulation of type IIA

supergravity
We start by reviewing bosonic part of the standard 10D IIA supergravity action [19].
Fermions will be incorporated into the discussion in Section 7.
2.1. Bosonic action of the standard 10D IIA supergravity
The 10D NSNS fields are the dilaton φ, 2-form potential B 2 and string frame metric
ĝM
N , where
M,
N = 0, . . . , 9. The 10D RR fieldstrengths are the 4-form G4 , 2-form G2
and 0-form G0 .
We measure all dimensionful fields in units of 11D Planck length lp and set k11 = π, so

1 √ 1
(10)
Sbos = e−2φ g10 R(ĝ) + 4 dφ ∧ ∗ˆ dφ + H 3 ∧ ˆ
∗ H 3
2π 2
X10

+
1 4 ∧ ∗ˆ G
G 4 + i B
∧ G
4 ∧ G
4 + G 2 + √g10 G20 ,
2 ∧ ∗ˆ G (2.1)
4π
X10
where ∗ˆ stands for the 10D Hodge duality operator. The fields in (2.1) are defined as
2 = G2 + B
G 2 G0 , 4 = G4 + B
G 2 G2 + 1 B 2 G0 ,
2 B 3 = d B
H 2 .
2
We explain the relation between our fields and those of [19] in Appendix B.
2.2. Reduction of IIA supergravity on a torus
We now recall some basic facts about the reduction of the bosonic part of the 10D

action on T 2 . Let us consider X10 = T 2 × X and split coordinates as XM = (x M , σ m ),
where M = 0, . . . , 7, m = 8, 9.
The standard ansatz for the reduction of the 10D metric has the form
2
ds10 = t 2 gMN dx M dx N + V gmn ωm ⊗ ωn , (2.2)
where gmn is defined in (1.6), t 2 gMN is 8D metric, det gMN = 1. V is the volume of T 2
and ωm = dσ m + Am (1) . The other bosonic fields of the 8D theory are listed below.
α
(1) g(0) α
, g(2) , α = 1, 2, g(1)m , g(3)m , m = 8, 9, and g(4) are defined from9
G0
= g(0)
2
,
2π
2 1
G 1
= g(0) + g(0)
2
B0 !mn ωm ωn + g(1)m ωm + g(2) 2
,
2π 2
4
G 1 1

= g(4) + g(3)m ωm + B0 g(2)
2
+ g(2) !mn ωm ωn . (2.3)
2π 2
(2) The 8D dilaton ξ is defined by
e−2ξ = e−2φ V . (2.4)
(3) B(2) , B(1)m , B0 are obtained from the KK reduction of the NSNS 2-form potential in
the following way
2 = 1 B0 !mn ωm ωn + B(1)m ωm + B(2) + 1 Am B(1)m .

B (2.5)
2 2 (1)
Now, the real part of the 8D bosonic action obtained by the above reduction is
(8D)
3
Re Sboson = SNS + Sp (g(p) ) + S4 (g(4) ), (2.6)
p=0
where

1 −2ξ
1
SNS = e t 6 R(g) + 4 dξ ∧ ∗dξ + 28t −2 dt ∧ ∗dt + t 2 H(3) ∧ ∗H(3)
2π 2

1 6 dτ ∧ ∗dτ 1 6 dρ ∧ ∗dρ 1 4
+ t + t + t g G
mn αβ F mα
∧ ∗F nβ
, (2.7)
2 (Im τ )2 2 (Im ρ)2 2
where Gαβ is defined in (1.6) and Am
(1) and B(1)m are combined into 1-form as a collection
of
mn
! B(1)n
Amα = . (2.8)
(1) Am(1)
9 ! 89 = 1, ! = 1.
89
Also, we denote10
1 nβ
H(3) = dB(2) − !mn Eαβ A(1)mα F(2), (2.9)
2
3

8 β
Sp (g(p) ) = π t Gαβ g(0)
α
∧ ∗g(0) + t 6 g mn g(1)m ∧ ∗g(1)n
p=0 X
β
+ t 4 Gαβ g(2)
α
∧ ∗g(2) + t 2 g mn g(3)m ∧ ∗g(3)n . (2.10)
Finally we have

S4 (g(4) ) = π Im(ρ)g(4) ∧ ∗g(4) . (2.11)
X

It is convenient to introduce the notation SB (F ) = 3p=0 Sp (g(p) ) for the value of the
actions evaluated on a background flux field configuration. SB (F ) will enter the partition
sum ZB+F (F , τ, ρ) in Eq. (8.1).
2.3. T-duality action on 8D bosonic fields
The T-duality group of the 8D effective theory obtained by reduction on T 2 is known to

be DT = SL(2, Z)τ × SL(2, Z)ρ , where the first factor is mapping class group of T 2 which
acts on τ
aτ + b
τ→ (2.12)
cτ + d
and the second factor acts on ρ = B0 + iV
αρ + β
ρ→ . (2.13)
γρ + δ
Let us denote generators of SL(2, Z)ρ by
S : ρ → −1/ρ, T :ρ → ρ + 1
and generators of SL(2, Z)τ by

S : τ → −1/τ, T : τ → τ + 1.
We now recall how T-duality acts on the remaining bosonic fields of the 8D theory [14].
First, ξ, t, gMN are T-duality invariant. Next, there is the collection of fields F mentioned
in the introduction. These transform linearly under T-duality. They include the NS potential
B(2) , which is T-duality invariant, as well as Amα
(1) , which transform in the (2, 2). The other
components of F are the RR fieldstrengths g(0) α , g α , α = 1, 2, which transform in the
(2)
(1, 2) of DT and g(1)m , g(3)m , m = 8, 9, which transform in the (2 , 1) of DT .
Finally, the field g(4) is singled out among all the other fields since according to the
conventional supergravity [14] SL(2, Z)ρ mixes g(4) with its Hodge dual ∗g(4) and hence
10 E = 1, E = −1.
12 21
g(4) does not have a local transformation. More concretely,

− Re ρg(4) + i Im ρ ∗ g(4)
(2.14)
g(4)
transforms in the (1, 2) of DT . Due to this nontrivial transformation the classical bosonic
(8D)
8D action Sboson is not manifestly invariant under SL(2, Z)ρ .
3. Review of the K-theory theta function
In this section we review the basic flux quantization law of RR fields and the definition
of the K-theory theta function. We follow closely the treatment in [4,5,7].
3.1. K-theoretic formulation of RR fluxes
As found in [1–4] RR fields in IIA superstring theory are classified topologically by an

2 = 0 is
element x ∈ K 0 (X10 ). The relation for B

10
G ch x,
= A G= Gj , (3.1)
2π
j =0
where ch is the total Chern character and A is expressed in terms of the Pontryagin classes
as

= 1 − 1 p1 + 1 7p12 − 4p2 .
A (3.2)
24 5760
In (3.1), the right-hand side refers to the harmonic differential form in the specified real
cohomology class. The quantization of the RR background fluxes is understood in the sense
that they are derived from an element of K 0 (X10 ).
3.2. Definition of the K-theory theta function
Let us recall the general construction of a K-theory theta function, which serves as the
RR partition function in Type IIA. One starts with the lattice ΓK = K 0 (X10 )/K 0 (X10 )tors .
This lattice is endowed with an integer-valued unimodular antisymmetric form by the
formula
ω(x, y) = I (x ⊗ ȳ), (3.3)

where for any z ∈ K 0 (X10 ), I (z) is the index of the Dirac operator with values in z.
Given a metric on X10 , one can define a metric on ΓK

G(x) ∗ˆ G(y)
g(x, y) = ∧ , (3.4)
2π 2π
X10
where ∗ˆ is the 10D Hodge duality operator.

Let us consider the torus T = (ΓK ⊗Z R)/ΓK . The quantities ω and g can be interpreted
as a symplectic form and a metric, respectively, on T. To turn T into a Kähler manifold
one defines the complex structure J on T as
g(x, y) = ω(J x, y). (3.5)
Now, if it is possible to find a complex line bundle L over T with c1 (L) = ω, then T
becomes a “principally polarized abelian variety”. L has, up to a constant multiple, a
unique11 holomorphic section which is the contribution of the sum over fluxes to the RR
partition function.
As explained in detail in [20], holomorphic line bundles L over T with constant
curvature ω are in one–one correspondence with U (1)-valued functions Ω on ΓK such
that
Ω(x + y) = Ω(x)Ω(y)(−1)ω(x,y). (3.6)
For weakly coupled type II superstrings one can take Ω to be valued in Z2 . Motivated by
T-duality, and the requirements of anomaly cancellation on D-branes [5], Witten proposed
that the natural Z2 -valued function Ω for the RR partition function is given by a mod-two
index [4]. For any x ∈ K 0 (X10 ), x ⊗ x̄ ∈ KO(X10 ) lies in the real K-theory group on X10 ,
and for any v ∈ KO(X10 ), there is a well-defined mod 2 index q(v) [21]. We take
Ω(x) = (−1)j (x), (3.7)
where j (x) = q(x ⊗ x̄).
As explained in [4,5,7] there is an anomaly in the theory unless Ω(x) is identically 1 on
the torsion subgroup of K(X10 ). In the absence of this anomaly it descends to a function
on ΓK = K 0 (X10 )/K 0 (X10 )tors and can be used to define a line bundle L and hence the
RR partition function.
To define the theta function one must choose a decomposition of ΓK as a sum Γ1 ⊕ Γ2 ,
where Γ1 and Γ2 are “maximal Lagrangian” sublattices. ω establishes a duality between
Γ1 and Γ2 , and therefore there exists θK ∈ Γ1 /2Γ1 such that
Ω(y) = (−1)ω(θK ,y) , ∀y ∈ Γ2 . (3.8)
Following [7] we choose the standard polarization: the sublattice Γ2std is defined as the set
of x with vanishing G0 , G2 , G4 . This choice implies that G0 , G2 , G4 are considered as
independent variables. This is a distinguished choice for every large 10-manifold in the
sense that it allows for a good large volume semiclassical approximation scheme on any
10-manifold (see Section 5).
It was demonstrated in [7] that Γ1std in the standard polarization consists of K-theory
classes of the form x = n0 1 + x(c1 , c2 ). 1 is a trivial complex line bundle and x(c1 , c2 ) is
defined for c1 ∈ H 2 (X10 , Z) and c2 ∈ H 4 (X10 , Z) with Sq 3 c2 = 0, as

1
ch x(c1 , c2 ) = c1 + − c2 + c12 + · · · . (3.9)
2
11 The uniqueness follows from the index theorem on T using unimodularity of ω and the fact that for any
complex line bundle M over T with positive curvature we have H i (T; M) = 0, i > 0.
The higher Chern classes indicated by · · · are such that x(c1 , c2 ) is in a maximal
Lagrangian sublattice Γ1std complementary to Γ2std . Then, θK for the standard polarization
can be chosen to satisfy
ch0 (θK ) = 0, ch1 (θK ) = 0, ch2 (θK ) = −λ + 2â0 , I (θK ) = 0, (3.10)

where λ = 1
2 p1 and â0 is a fixed element of H 4 (X10 , Z) such that

∀ĉ ∈ L f (ĉ) = ĉ ∪ Sq 2 â0 , (3.11)
X10
where

L = ĉ ∈ Htors
4 4
(X10 , Z)/2Htors (X10 , Z), Sq 3 (ĉ) = 0
and f (â) stands for the mod 2 index of the Dirac operator coupled to an E8 bundle on the
11D manifold X10 × S 1 with the characteristic class â ∈ H 4 (X10 , Z) and supersymmetric
spin structure on the S 1 . (We will show in Section 5.1 that for X10 = X ×T 2 in fact â0 = 0.)
The K-theory theta function in the standard polarization is
1
ΘK = eiu eiπτK (x+ 2 θK ) Ω(x), (3.12)
x∈Γ1

where u = − π4 X10 ch2 (θK ) ch3 (θK ) and the explicit form of the period matrix τK is given
by

1 1
Re τK x + θK = (G0 G10 − G2 G8 + G4 G6 ), (3.13)
2 (2π)2
X10
2
1 1
Im τK x + θK = G2p ∧ ∗ˆ G2p . (3.14)
2 (2π)2
p=0 X10
The RR fields which enter (3.13), (3.14) are:

1 1
G0 x + θK = n0 ,
2π 2

1 1
G2 x + θK = ê,
2π 2

1 1 1 1
G4 x + θK = â + ê2 − (1 + n0 /12)λ, (3.15)
2π 2 2 2
where we denote ê = c1 (x), â = −c2 (x) + â0 .
From (3.12) and (3.13), (3.14) the following topological term was found in [7] to be the
K-theoretic corrections to the 10D IIA supergravity action:

n
e2πiΦ(n0 ,ê,â) = exp −2πin0 ê A Ω(1) 0 e2πiΦ(ê,â) , (3.16)
8
X10
e2πiΦ(ê,â) = (−1)f (â0 ) (−1)f (â)

5
ê ê3 â 11ê3λ êâλ êλ2 1
× exp 2πi + − − + − êA8 . (3.17)
60 6 144 24 48 2
X10
3 ] = 0
3.3. Turning on the NSNS 2-form flux with [H
In the presence of an H -flux we expect K-theory to be replaced by twisted K-theory

KH classifying bundles of algebras with nontrivial Dixmier–Douady class. The Morita
equivalence class of the relevant algebras only depends on the cohomology class of H ,
but this does not mean that the choice of “connection” that is, the choice of B field is
irrelevant to formulating the K-theory theta function. Indeed, when [H ] = 0, the choice

of trivialization B in H = d B changes the action in supergravity and “turning on” this
field in supergravity corresponds to acting with an automorphism on the K-theory torus. In
this section we describe this change explicitly. See [22,23] for recent mathematical results
relevant to this issue.
2 ∈ H 2 (X10 , R). We normalize B
Let us turn on B 2 so that it is defined mod H 2 (X10 , Z)
under global tensorfield gauge transformation. By Morita equivalence, the RR fields are
still classified topologically by x ∈ K 0 (X10 ). The standard coupling to the D-branes
implies that the cohomology class of the RR field is

G(x)

= eB2 ch(x) A. (3.18)
2π
Let us define

G(x)

:= e−B2 ch(x̄) A. (3.19)
2π
The bilinear form on ΓK = K 0 (X10 )/K 0 (X10 )tors is still given by the index

1 ∧ G(y) = I (x ⊗ ȳ)
ω(x, y) = G(x) (3.20)
(2π)2
X10
while the metric on ΓK is modified to be

1 ∧ ∗ˆ G(y)

g̃(x, y) = G(x) (3.21)
(2π)2
X10
and the Z2 valued function Ω(x) is unchanged. If we continue to use the standard
polarization then θK ∈ Γ1 /2Γ1 is unchanged as well.
The net effect to modify (3.12) is that the period matrix τK should be substituted for
τ
K = τK (G → G).

2 ) = eiu eiπ τ
1
K (x+ 2 θK )
ΘK (B Ω(x). (3.22)
x∈Γ1
Note, that the constant phase eiu in front of the sum remains the same as in (3.12).
The imaginary part of the 10D Type IIA supergravity action now becomes Im(S10D ) =
where
−2π Φ,

1 2 2 1 3 2

Φ =Φ+ B2 G4 + B2 G2 G4 + B G + G0 G4
8π 2 3 2 2

1 4 1 5 2
+ B2 G0 G2 + B2 G0 , (3.23)
4 20
Φ is defined in (3.16), (3.17) and G2p (x + 12 θK ), p = 0, 1, 2, are given in (3.15).

From (3.23) we find that corrections to Φ depending on B 2 coincide with the imaginary
part of the standard supergravity action (see, for example, [12]).
Note, that G defined in (3.19) is a gauge invariant field if the global tensorfield gauge
transformation
2 → B
B 2 + f2 , f2 ∈ H 2 (X10 , Z) (3.24)
also acts on K 0 (X10 ) as
x → L(−f2 ) ⊗ x, x ∈ K 0 (X10 ), (3.25)
where the line bundle L(−f2 ) has c1 (L(−f2 )) = −f2 .

Thus, according to (3.25) a tensorfield gauge transformation acts as an automorphism
of ΓK , preserving the symplectic form ω. Eq. (3.25) acts on theta function (1.12) by
multiplication by a constant phase:

ΘK (B2 + f2 ) = e iπ/4 2 ).
f2 (λ − 2â0)2 ΘK (B (3.26)
X10
4. Action of T-duality in K-theory
In this section we consider X10 = T 2 × X and describe the action of T-duality on the
K-theory variables.
As we have mentioned, the standard polarization is distinguished for any large 10-
→ t ĝM
manifold in thefollowing sense. When the metric of X10 is scaled up ĝM 2
N N

the action X10 ĝ |G2p | of the Type IIA RR 2p-form scales as t
2 10−4p . This allows the
successive approximation of keeping only G4 whose periods have the smallest action, then
including G2 and finally keeping all G4 , G2 , G0 .
In the case of X10 = T 2 × X with the metric (1.1), the standard polarization is no longer
distinguished. Various equally good choices are related by the action of the T-duality group
DT on ΓK = K 0 (T 2 × X)/Ktors0 (T 2 × X).
We argue below that DT can be considered as a subgroup of Sp(2N, Z), where N

denotes the complex dimension of the K-theory torus T = K 0 (T 2 × X) ⊗Z R/ΓK and
Sp(2N, Z) stands for the group of symplectic transformations of the lattice ΓK .
4.1. Background RR fluxes in terms of integral classes on X
To describe the action of DT on K-theory variables, we will write RR fields in terms

of integral classes on X. Let us start from the standard polarization12 and write a general
element of Γ1std as

x = n0 1 + L n1 e0 + e + γm dσ m − 1 + x e0 e + a + hm dσ m + ∆, (4.1)

where e0 = dσ 8 ∧ dσ 9 , so that T 2 e0 = 1. L(ê) is a line bundle with c1 (L) = ê ∈
H 2 (X10 ; Z), 1 is a trivial line bundle, and for any â ∈ H 4 (X10 ; Z), x(â) is a K-theory
lift (if it exists). In (4.1) ∆ puts x into the Lagrangian lattice Γ1std and we also introduce
the notations:
a ∈ H 4 (X; Z), e, e ∈ H 2 (X; Z),
hm ∈ H 3 (X; Z), γm ∈ H 1 (X; Z), m = 8, 9. (4.2)
The RR fields entering (3.13), (3.14) are given by

1 1
G0 x + θK = n0 ,
2π 2

1 1
G2 x + θK = n1 e0 + e + γm dσ m ,
2π 2

1 1 1 1
G4 x + θK = a + e2 + e0 e + fm dσ m − (1 + n0 /12)λ, (4.3)
2π 2 2 2
where
e = n1 e + e − γ1 γ2 , fm = hm + am + eγm . (4.4)
Note that (3.13) is in fact only a function of these variables, by the Lagrangian property.
From the 10D constraint Sq 3 â = Sq 3 â0 , valid in the case [H 3 ] = 0, we find the
constraints on the integral cohomology classes: Sq fm = Sq am , m = 8, 9. We will show
3 3
that actually Sq 3 fm = 0, m = 8, 9 (see comment below (5.8)).
4.2. The embedding DT ⊂ Sp(2N, Z)
From the transformation rules of the RR fields under the T-duality group [24] we find
that fm and γm transform in the (2 , 1) of DT and we can form a representation (1, 2) out
of n0 , n1 and e, e in the following way:

n1 e
nα = , eα = . (4.5)
n0 e
We would like to reformulate the transformation rules for RR fields in terms of the
action on ΓK .13 The action of SL(2, Z)τ on ΓK is via standard pullback under topologically
nontrivial diffeomorphisms. The action of SL(2, Z)ρ is more novel.
12 Γ std and Γ std are defined on page 19.

1 2
13 Some discussion of T-duality in the K-theoretic context can be found in [25].
We will explain the action of the two generators S, T of SL(2, Z)ρ separately. To begin,
the action of T on ΓK is a particular case of the global gauge transformation (3.24), (3.25)
with f2 = e0 and for this reason T ∈ Sp(2N, Z). The action of T preserves the standard
polarization since it maps Γ2std → Γ2std :

G2p y ⊗ L(−e0 ) = 0, ∀y ∈ Γ2std , p = 0, 1, 2. (4.6)
The action of the generator S on ΓK is more interesting. By the Kunneth theorem we can
decompose

K 0 X × T 2 = K 0 (X) ⊗ K 0 T 2 ⊕ K 1 (X) ⊗ K 1 T 2 . (4.7)
Both K 0 (T 2 ) = Z ⊕ Z and K 1 (T 2 ) = Z ⊕ Z have natural symplectic bases on which S acts
as the standard symplectic operator iσ2 . For K 0 (T 2 ) we choose basis 1 and L(e0 ) − 1, and
for K 1 (T 2 ) we denote the basis as ζ m , m = 8, 9. We now have a Lagrangian decomposition
of ΓK = Γ1 ⊕ Γ2 :
Γ1 = K 0 (X) ⊗ 1 ⊕ K 1 (X) ⊗ ζ 8 ,

Γ2 = K 0 (X) ⊗ L(e0 ) − 1 ⊕ K 1 (X) ⊗ ζ 9 (4.8)
on which the T-duality generator S acts simply. However, the decomposition (4.8) is
not compatible with the standard polarization, and hence the action of S in the standard
polarization appears complicated. We now give an explicit description of the action of S in
the standard polarization.
Let us write a generic element y ∈ Γ2std as

y = x(ã) ⊗ L(e0 ) − 1 + z1 + z2 + z3 ⊗ L(e0 ) − 1 , ã ∈ H 4 (X, Z). (4.9)
In (4.9) z1 , z2 , z3 are such that
G G G
(z1 ) = jm dσ m , (z2 ) = k, (z3 ) = k , (4.10)
2π 2π 2π
where jm ∈ H 5 (X, R) ⊕ H 7 (X, R), k, k ∈ H 6 (X, R) ⊕ H 8 (X, R). According to the
transformation rules of RR fields [24] S acts on y as

S : y → y , y = x(ã) + z1 + z3 − z2 ⊗ L(e0 ) − 1 . (4.11)
From (4.11) we find that the image Γ2 := S(Γ2std) differs from Γ2std .14
Since we have an embedding DT ⊂ Sp(2N, Z), we can deduce the existence of well-
defined transformation laws under DT of the function Θ(F , ρ), related by (1.20) to the
K-theory theta function ΘK . This follows from the fact that ΘK is an holomorphic section
of the line bundle L over the K-theory torus with c1 (L) = ω. Since L is not affected by
symplectic transformations, and has a one-dimensional space of holomorphic sections, it
follows that under T-duality transformations ΘK can at most be multiplied by a constant.
Nevertheless, this leaves open the possibility of a T-duality anomaly, as indeed takes place.
14 In following [24] we have actually combined the transformation S with the transformation
S from SL(2, Z)τ .
This is a more convenient basis for checking the invariance of the theory.
To conclude this section we show how the multiplier system of (1.22), (1.23) is related
to the standard 8th roots of unity appearing in theta function transformation laws. Let us
recall the general transformation rule under Sp(2N, Z) for the theta function θ [m](τ ) of
m
a principally polarized lattice Λ = Λ1 + Λ2 of rank 2N. Here m = m ∈ R
2N are the
characteristics and the period matrix τ ∈ MN (C), τ = τ is a quadratic form on Λ1 .

T
It was found in [26] that under symplectic transformations

Aτ + B
σ ·τ = , σ ∈ Sp(2N, Z) (4.12)
Cτ + D
the general θ [m](τ ) transforms as
ϑ[σ · m](σ · τ ) = κ(σ )e2πiφ(m,σ ) det(Cτ + D)1/2 ϑ[m](τ ), (4.13)

where

−1 1 CT D d
σ · m = mσ + ,
2 (AT B)d
1
φ(m, σ ) = − m T DB T m − 2m T BC T m + m T CAT m
2
1
+ m T D − m T C AT B d ,
2
where (A)d denotes a vector constructed out of diagonal elements of matrix A.
The factor κ(σ ) in (4.13) has quite nontrivial properties [26]. In particular κ 2 (σ ) is a
character of Γ (1, 2) ⊂ Sp(2N, Z), where
T
σ ∈ Γ (1, 2) iff AT B d ∈ 2Z, C D d ∈ 2Z. (4.14)
One can easily check that SL(2, Z)ρ ⊂ Γ (1, 2) by writing out explicit representations
σ (S) and σ (T ) in Sp(2N, Z). We give σ (S) and σ (T ) in Appendix A.
Using the explicit expressions for σ (S) and σ (T ) as well as the definition of τK (3.13),
(3.14) we find that in (4.13)
1/2 + 1 −
det C(S)τK + D(S) = eiπ/4b4 (−iρ)1/2b4 (i ρ̄) 2 b4 , φ m, σ (S) = 0, (4.15)
1/2
det C(T )τK + D(T ) = 1, φ m, σ (T ) = 0. (4.16)
Now comparing (4.13) and the explicit formulae (5.31) for the transformation laws of
Θ(F , ρ) derived in the next section we find the relation between κ(σ ) and the multiplier
system µ(S), µ(T )
π
κ(S)ei 4 b4 = µ(S), κ(T ) = µ(T ). (4.17)
5. Θ(F , ρ) as a modular form
In this section we derive an explicit expression for Θ(F , ρ) using its relation (1.20) to
the K-theory theta function ΘK and we check that Θ(F , ρ) transforms under the T-duality
group DT as a modular form.
5.1. Zero NSNS fields
We first assume that all NSNS background fields are zero. In this case Θ(F , ρ), defined
in (1.20) is given by an expression of the form

Θ(F , ρ) = ei2πΦ(a,F ) e−π X Im(ρ)g(4)∧∗g(4) , (5.1)
a∈H 4 (X,Z)
where the imaginary part of the 8D effective action 2πΦ(a, F ) is derived as follows. We
substitute
â = a + e0 e + hm dσ m , ê = e + n1 e0 + γm dσ m (5.2)
into the definition (3.16) of ei2πΦ(n0 ,ê,â) .
We need to evaluate f (a + e0 e + hm dσ m ). We use the bilinear identity from [7]

f (u + v) = f (u) + f (v) + uSq 2 v, ∀u, v ∈ H 4 (X10 ; Z) (5.3)
X10
to find

f a + e0 e + hm dσ m = f (a + e0 e ) + f hm dσ m . (5.4)
Let us consider f (hm dσ m )
first. Again using the bilinear identity we obtain

f hm dσ m = f h8 dσ 8 + f h9 dσ 9 + h8 Sq 2 (h9 ). (5.5)
X
From (5.3) it follows that m = 8, 9, are linear functions of h ∈ H 3 (X, Z).

f (hdσ m ),
Moreover, from the diffeomorphism invariance of the mod-two index we see that
f (hdσ 8 ) = f (hdσ 8 + Qhdσ 9 ), for any integer Q and, using the bilinear identity once more
we find that f (hdσ m ) = r(h), m = 8, 9, where

r(h) = hSq 2 h, h ∈ H 3 (X, Z) (5.6)
X
is a spin-cobordism invariant Z2 -valued function. In fact, r(h) is a nontrivial invariant since
for X = SU(3) and h = x3 the generator of H 3 (SU(3), Z) we have r(h) = 1. In conclusion:

f hm dσ m = h8 Sq 2 h8 + h9 Sq 2 h9 + h8 Sq 2 (h9 ) . (5.7)
X
Now we consider f (a + e0 e ):

f (e0 e + a) = f (a) + f (e0 e ) + e0 e Sq 2 a
X10

1 1
= (a)2 − (e )2 λ + (e )2 a = 2
aλ + (e ) a − λ . (5.8)
2 2
X X
This uses the bilinear identity (5.3), the reduction of the mod-two index along T 2 , and the
formula Eq. (8.40) for f (u ∪ v) from [7].
We can now evaluate â0 defined in (3.11). The kernel of Sq 3 is given by those elements
a + e0 e + hm dσ m such that h8 ∪ h8 = h9 ∪ h9 = 0. If we add the condition that the
element is a torsion class then f (a + e0 e ) = 0 and we need only evaluate (5.7). Now,
since Sq 3 (hm ) = hm ∪ hm = 0 it follows that Sq 2 (hm ) has an integral lift. Using again the
condition that hm is torsion we find that the right-hand side of (5.7) is zero. It follows that
â0 = 0.
We can now evaluate the phase. Using (4.4) we reexpress (5.5) as

f hm dσ m = f8 Sq 2 (f9 ) + f8 Sq 2 (f8 ) + f9 Sq 2 (f9 )
X

+ e2 (γ9 f8 − γ8 f9 ) + e3 γ8 γ9 . (5.9)
Taking into account (5.9) and (5.8) we find the total phase Φ(a, F ) in (5.1) is given by

Φ(a, F ) = 2Φ + (a + α)β, (5.10)
X
where the characteristics are defined as:
1 1 1
α = (e)2 + (1 − n0 /12)λ + (e + e)! mn γm γn ,
2 2 2
1 2 1 1
β = (e ) + (1 − n1 /12)λ + (e − e)! mn γm γn (5.11)
2 2 2
and we recall that e = n1 e + e − 12 ! mn γm γn . Note that for convenience we have made a
shift of the summation variable in (5.1) a → a + λ + 12 (e + e )! mn γm γn .
The prefactor 2Φ is given by

exp[2πi2Φ] = exp πi f8 Sq (f9 ) + f8 Sq (f8 ) + f9 Sq (f9 ) ,
2 2 2
(5.12)
X
1 1 1
exp 2πi − (e e)2 − e eλ + e3 e
4 24 6
X
1 1 1
− e2 λ + n0 λ(e )2 + (1 + n0 /12)λ2
4 48 4
2
1
+ (n0 − n1 )A 8 − 1 n0 n1 A8 + λ λ
+ ! mn γm fn
2 2 24 24

1 2

+ n0 (e − e)λ − 12e e − 4eλ − 4e ! γm γn .
3 mn
48
In deriving 2Φ we have used
2
( A )8 = 1 A 8 − λ .
2 24
Also, in bringing 2Φ to the form (5.12) we have used the congruences

1 3 1 1
(e ) e + e e3 + (e )2 e2 − λe e ∈ Z, (5.13)
6 4 12
X

(e e)2 ∈ 2Z, e eλ ∈ 2Z, (5.14)
X X
which follow from the index theorem on X:

1 4 1
e − λe2 ∈ Z, ∀e ∈ H 2 (X, Z). (5.15)
24 24
X
5.2. Including flat NSNS potentials
Let us now take into account globally defined NSNS fields:
2 = 1 B0 !mn ωm ωn + B(1)m ωm + B(2) + 1 Am B(1)m , Am

B
2 2 (1) (1)
and recall that Am(1) and B(1)m are combined into the (2, 2) of DT as in (1.5).
We define a gauge invariant fieldstrength G = eB2 G as in (3.18) where G are given in
(4.3) and we expand G(x + θK ) as
1
2
0
G 1
x + θK = g(0) 2
,
2π 2
2
G 1

1 1
x + θK = g(0) + g(0)
2
B0 !mn ωm ωn + g(1)m ωm + g(2) 2
,
2π 2 2
4
G 1

1 1
x + θK = g(4) + g(3)m ωm + B0 g(2) 2
+ g(2) !mn ωm ωn . (5.16)
2π 2 2
The first effect of including flat NSNS fields is to modify the fields which enter SB (F ).
These fields g(0)α ,g α
(1)m , g(2) , g(3)m are now linear combinations of the integral classes
γm , fm , e , n defined in (4.2), (4.4) with coefficients constructed from Amα
α α
(1) and B(2) :

n1
α
g(0) = , g(1)m = γm + ξ(1)m ,
n0

1
g(2) = e + A(1) γm + ξ(1)m + B(2) g(0)
α α mα α
, (5.17)
2
1 1 pα nβ
g(3)m = fm + B(2) g(1)m + λ(3)m + k(3)m + !mn Eαβ A(1) ξ(1)p A(1), (5.18)
2 6
where we denote
nβ nβ
ξ(1)m = !mn Eαβ g(0)
α
A(1) , λ(3)m = !mn Eαβ eα A(1) ,
pα nβ
k(3)m = !mn Eαβ A(1) γp A(1) . (5.19)
The other effect of including flat NSNS fields is to shift the characteristics and the
prefactor of Θ(F , ρ). Now Θ(F , ρ) has the form

Θ(F , ρ) = e 2πi2Φ
exp −π Im(ρ)g(4) ∧ ∗g(4)
a∈H 4 (X,Z) X

+ iπ Re(ρ)g(4) ∧ g(4) + 2πig(4)β̃ , (5.20)
where [g(4) ] = a + α̃, a ∈ H 4 (X, Z), and the shifted characteristics α̃, β̃ are
α̃ = α + ϕ 2 , β̃ = β + ϕ 1 , (5.21)
where α, β are defined in terms of integral classes n0 , n1 , γm , eα in (5.11), while ϕ α
transform in the (1, 2) of DT . Explicitly,

1 1 1
ϕ α = Amα
(1) f m + λ(3)m + k (3)m + B(2) e α
+ A (1) γm +
mα
ξ(1)m
2 6 2
1
+ B(2) B(2) g(0)α
− ζ(4)g(0)
α
, (5.22)
2
where ξ(1)m , λ(3)m , k(3)m are given in (5.19) and we also denote
1 n β n β m β m β
ζ(4) = Eβ β Eβ β A 1 1 !n n A 2 3 A 1 2 !m1 m2 A(1)2 4 . (5.23)
64 1 2 3 4 (1) 1 2 (1) (1)
The shifted prefactor 2Φ in (5.20) is given by

1 1

2Φ = 2Φ − β ∧ ϕ + ϕ ∧ ϕ + (2Φ)inv ,
2 2
(5.24)
2
X
where 2Φ is defined in terms of integral classes n0 , n1 γm , eα , fm in (5.12) and (2Φ)inv is
the part of the phase which is manifestly invariant under the T-duality group DT . Explicitly,

1 1 1 1
(2Φ)inv = B(2) 3
Eαβ g(0)
α β
e − ! mn γm γn − ! mn ξ(1)m γn − ! mn ξ(1)m ξ(1)n
12 6 4 8
X

1 1
+ B(2) 2
− ! mn ξ(1)m fn − ! mn λ(3)m γn
4 2
X

3 mn 1 mn
− ! λ(3)m ξ(1)n − ! k(3)m ξ(1)n
8 24

1 1 1
+ B(2) − ! mn fm fn − ! mn λ(3)m fn − ! mn λ(3)m λ(3)n
2 2 4
X

1 1 1 β
− ! mn λ(3)m k(3)n + ξ(1)m q(5) m
+ ζ(4) Eαβ eα g(0) + ζ(4)! mn γm γn
6 12 2

1
+ m
λ(3)m q(5) + ζ(4) ! mn γm fn , (5.25)
12
X
pα mβ
m
where q(5) = Eαβ A(1) fp A(1) .
5.3. Derivation of T-duality transformations
Let us study transformations of Θ(F , ρ) defined in (5.20) under DT . First, we note that
Θ(F , ρ) is invariant under SL(2, Z)τ . Next, we consider the action of the generator S. For
any function h(F ) of fluxes F , we denote

S h(F ) := h(S · F )
and
δS [h] := S[h] − h,
where S · F denotes the linear action on fluxes. To check the transformation under
S we need to do a Poisson resummation on the self-dual lattice H 4 (X, Z). The basic
transformation law is:

θ −φ
ϑ (0|−1/τ ) = (−iτ )1/2e2πiθφ ϑ (0|τ ) (5.26)
φ θ
and its generalization to self-dual lattices (4.13).
After the Poisson resummation and a shift of summation variable a → a + e2 + λ we
find that Θ(F , ρ) transforms under S as

1 + 1 −

Θ(S · F , −1/ρ) = e2πi X S[α̃]S[β̃]+δS [2Φ] (−iρ) 2 b4 (i ρ̄) 2 b4 Θ(F , ρ). (5.27)
Now using the definitions of α̃, β̃ (5.21), (5.22) and 2Φ (5.24) as well as the
transformation rules for F , we find after some tedious algebra
2
λ

δS [2Φ] = − S[α̃]S[β̃] + + Z. (5.28)
4
X X
We conclude that the generator S acts as

1 + 1 −
λ2 /2
Θ(S · F , −1/ρ) = eiπ X (−iρ) 2 b4 (i ρ̄) 2 b4 Θ(F , ρ). (5.29)
To check how Θ(F , ρ) transforms under the generator T we use its relation (1.20) to
the K-theory theta function ΘK as well as the transformation of ΘK under global gauge
transformation B 2 → B
2 + f2 (3.26) where the action of the generator T corresponds to
f2 = e0 . In this way we find from (3.26) that
2 /4
Θ(T · F , ρ + 1) = eiπ Xλ Θ(F , ρ). (5.30)
5.4. Summary of T-duality transformation laws
Below we summarize the transformation laws of the function Θ(F , ρ) under the
generators of T-duality group DT .
Θ(F , ρ) is invariant under SL(2, Z)τ :
Θ(T · F , ρ) = Θ(F , ρ),

Θ(
S · F , ρ) = Θ(F , ρ). (5.31)
Θ(F , ρ) transforms as a modular form with a nontrivial “multiplier system” under
SL(2, Z)ρ . That is, using the standard generators T , S of SL(2, Z)ρ we have
Θ(T · F , ρ + 1) = µ(T )Θ(F , ρ),

1 + 1 −
Θ(S · F , −1/ρ) = µ(S)(−iρ) 2 b4 (i ρ̄) 2 b4 Θ(F , ρ), (5.32)
where T · F , S · F denotes the linear action of DT on the fluxes. Here is the b4+ , b4−
dimension of the space of self-dual and anti-self-dual harmonic forms on X and the
multiplier system is

iπ
µ(T ) = exp λ2 ,
4
X

iπ
µ(S) = exp λ2 , (5.33)
2
X
where p1 = p1 (T X). These define the “T-duality anomaly of RR fields”.
6. The bosonic determinants
In this section we compute bosonic quantum determinants around the background

specified in Section 2.
Let us factorize bosonic quantum determinants as: DetB = DRR DNS , where DRR (DNS )
denotes the contribution from RR (NSNS) fields.
6.1. Quantum determinants DRR for RR fields
Quantum determinants DRR for RR fields have the form

4
DRR = ZRR,p , (6.1)
p=1
where ZRR,p is the quantum determinant for g(p) . First, we present the contribution ZRR,4
arising from the fluctuation dC(3) of g(4) . From (1.9) we find the kinetic term for C(3)
S3,cl = π Im(ρ)(dC(3) , dC(3) ), (6.2)

where ( , ) denotes the standard inner product on the space of p-forms on X, constructed
with the background metric gMN .
We use the standard procedure [27,28] for path-integration over p-forms, which can
be summarized as follows. Starting from the classical action for the p-form Sp,cl =
α(dC(p) , dC(p) ) one constructs the quantum action as15

p
1
m+1
k
Sp,qu = α(C(p) , ∆p C(p) ) + α m+1 u(p−m) , ∆p−m uk(p−m) , (6.3)
m=1 k=1
where uk(p−m) , k = 1, . . . , m + 1, m = 1, . . . , p, are ghosts of alternating statistics. For

example, uk(p−1) , k = 1, 2, are fermions, uk(p−2) , k = 1, 2, 3, are bosons, etc. In (6.3) ∆p is
the Laplacian acting on p-forms and constructed with gMN .16
To computeZRR,4 we apply (6.3) for p = 3, α = π Im(ρ) and use the measure [DCp ]
normalized as [DCp ]e−(Cp ,Cp ) = 1:

ZRR,4 = (α)− 2 (B3 −B2 +B1 −B0 )
1

det ∆3 −1/2 det ∆2 det ∆1 −3/2 det ∆0 2
× , (6.4)
V3 V2 V1 V0
where det ∆p is the determinant of nonzero modes of the Laplacian acting on p-forms.
Bp = Bp − bp , where Bp denotes the (infinite) number of eigen-p-forms and bp and Vp
p p
are the dimension and the determinant of the metric of the harmonic torus Tharm = Hp /HZ .
The appearance of Vp in (6.4) is due to the appropriate treatment of zeromodes and is
explained in Appendix E.
The determinants det ∆p together with the infinite powers depending on Bp , here and
below, require regularization and renormalization, of course. These can be handled using,
for example, the techniques of [29]. In particular the expression
q(Im ρ) := (Im ρ)− 2 (B3 −B2 +B1 −B0 )

1
(6.5)

is a local counterterm of the form e−π Im ρ X (uλ +vp2 ) , where
2
the numbers u, v depend on
the regularization. From now on we will assume that π Im ρ X (uλ2 + vp2 ) is included into
the 1-loop action:

2 iπ
S1-loop = π Im ρ uλ + vp2 + Re ρ p2 − λ2 . (6.6)
24
X X
In Section 8 we will show that T-duality invariance determines u and v uniquely.
Next, we consider the contributions to DRR from dC(2)m , dC(1) α , dC (0)m which are
α , g
the fluctuations for g(3)m , g(2) , respectively. Let us also make field redefinition of
(1)m
the quantum fields C (0)m , m = 8, 9, to fields C(0)m , m = 8, 9, which have well defined
transformation properties under the full U-duality group17
√ ξ 1 (0)9 .
C(0)8 = τ2 e C(0)8 , C(0)9 = √ eξ C (6.7)
τ2
15 Factors α 1/(m+1) should be understood as a mnemonic rule to keep track of the dependence on α which
follows from the analysis of various cancellations between ghosts and gauge-fixing fields.
16 ∆ = dd † + d † d.
17 For some discussion of U-duality see Section 10.
From (2.10) we find classical action quadratic in the above fluctuations:

α β
S0,cl = π t˜ 6 g mn C(0)m , d † dC(0)n , S1,cl = πt 4 Gαβ C(1) , d † dC(1) ,

S2,cl = πt 2 g mn C(2)m , d † dC(2)n ,
where t˜ = te−ξ/3 is U-duality invariant, and g 88 = τ12 g 88 , g 99 = τ2 g 99 , g 89 = g 89 . Now,
using (6.3) with a = π t˜ 6 g mn , πt 4 Gαβ , πt 2 g mn and p = 0, 1, 2 correspondingly we find:

−B det ∆0 −1
ZRR,1 = π t˜ 6 0 , (6.8)
V0

4 B −B det ∆1 −1 det ∆0 2
ZRR,2 = πt 0 1 , (6.9)
V1 V0

−B +B −B det ∆2 −1 det ∆1 2 det ∆0 −3
ZRR,3 = πt 2 2 1 0 . (6.10)
V2 V1 V0
In computing (6.8)–(6.10) we also used that detm,n g mn = 1, detm,n g mn = 1 and
detα,β Gαβ = 1.
Collecting together (6.4) and (6.8)–(6.10) we find that DRR has the form

det 23 −1/2 det 21 −1/2
DRR = rRR (t, ρ) , (6.11)
V3 V1
where

rRR (t, ρ) = (eξ )2B0 (Im ρ) 2 (b3 −b2 +b1 −b0 ) t −2B2 −2B1 −4B0
1

× (π)− 2 (B0 +B1 +B2 +B3 )
1
and we recall that q(Im ρ) was included into S1-loop .

We have computed the quantum determinants DRR treating RR fluctuations as
differential forms. It would be more natural if these determinants had a K-theoretic
formulation. This might be an interesting application to physics of differential K-theory.
6.2. Quantum determinants for NSNS fields
Let us first consider fluctuations damα mα

(1) and db(2) of the NSNS field F(2) and H(3) . From
(2.7) we find the quadratic action for fluctuations:
1 −2ξ
4 † nβ

Scl = e t gmn Gαβ amα
(1) , d da(1) + t b(2) , d db(2) .
2 †
(6.12)
4π
Now, again using (6.3) we find
4
t −2ξ 2(B0 −B1 ) det ∆1 −2 det ∆0 4
ZNS,2 = e (6.13)
4π V1 V0
and
1 (B −B −B ) −1/2 −3/2
t 2 −2ξ 2 1 2 0 det ∆2 det ∆1 det ∆0
ZNS,3 = e . (6.14)
4π V2 V1 V0
Let us now consider fluctuations of scalars: δξ, δτ, δρ. From (2.7) we write the action
quadratic in these fluctuations:

1 M 1 M
Sscal = β 8∂ M δξ ∂M δξ + ∂ δτ ∂M δ τ̄ + ∂ δρ∂M δ ρ̄ , (6.15)
(τ2 )2 (ρ2 )2
X
1 −2ξ 6
where β = 4π e t .
Now using the scalar measures defined as
δρ∧∗δρ̄ δτ ∧∗δτ̄
− −
[Dδρ][Dδ ρ̄]e X (Im ρ)2 = 1, [Dδτ ][Dδ τ̄ ]e X (Im τ )2 = 1, (6.16)

[Dδξ ]e−8 X δξ ∧∗δξ = 1 (6.17)
we find the quantum determinants for the NSNS scalars ZNS,0 :

−5/2
5 det ∆0
ZNS,0 = β − 2 B0 . (6.18)
V0
Finally, we consider the fluctuation hMN of the metric t 2 gMN . Recall that we work in
the limit e−ξ → ∞ so that in computing the quantum determinant for the metric we drop
couplings to RR background fluxes.
From (2.7) we find the quadratic action

Smetr = β (DN hMP )P MP QS D N hQS + hMP RMNP Q hNQ
X
2
1
− D hMN − DN h
M
, (6.19)
2
where h = g MN hMN and
1 1
P MP QS = g MQ g P S − g MP g QS .
2 4
In (6.19) RMNP Q is the Riemann tensor of the Ricci-flat18 background metric gMN .
The covariant derivative DM is performed with the background metric, and indices are
raised and lowered with this metric.
Following standard procedure [30,31] we first insert the gauge fixing condition into the
path-integral δ(κN − (D M hMN − 12 DN h)). Then, we insert the unit

1 = det(β11 ) Dκ(1) e−β(κ(1) ,κ(1) ) (6.20)
and integrate over κ(1) in the path-integral. This procedure brings the kinetic term for the
fluctuation hMN to the form

QS QS Q S
β hMP P MP NR KNR hQS , KNR = −δN δR DL D L + 2RN Q R S . (6.21)
X
18 If the background metric is not Ricci-flat there are terms involving the Ricci-tensor in (6.19) as well as in
(6.22).
Gauge fixing also introduces fermionic ghosts k(1) , l(1) with the action
Sgh = β 1/2(l(1) , ∆1 k(1) ). (6.22)

Using the measure [DhMN ]e− X hMN P
MNP Q h
P Q = 1 we obtain the result for the quantum
determinant Zmetr of the metric
det ∆1
Zmetr = (β)− 2 (NK −B1 ) [det K]−1/2
1
, (6.23)
V1
where det K is a regularized determinant of nonzero modes of the operator K defined in
(6.21) and NK = N − n , where N stands for the dimension (infinite) of the space of
K K K
the second rank symmetric tensors and nK is the number of zeromodes of the operator K.
We will explain how we regularize det K shortly.
Combining all NSNS determinants together we find:
−1/2
−1/2 det ∆2
DNS = rNS (t, ξ )[det K] , (6.24)
V2
where

rNS (t, ξ ) = (4π) 2 NK +B0 +B1 + 2 B2 (eξ )NK +B2 +2B1 +2B0 t −3NK −B2 −4B1 −8B0 .
1 1
(6.25)
Finally, from (6.11) and (6.24) we find the full expression for bosonic determinants
DetB = Q(t, gMN )(Im ρ) 2 (b3 −b2 +b1 −b0 ) ,

1
(6.26)
where Q is a function only of the T-duality invariant variables gMN , t and ξ. Explicitly,
−1/2 −1/2 −1/2
det ∆3 det ∆2 det ∆1
Q(t, gMN ) = rtot [det K]−1/2 , (6.27)
V3 V2 V1
where we regularized det K in a way that eliminates dependence on infinite numbers Bp
and NK so that
rtot = (t˜ )3(nK +b2 +2b1 +4b0 ) , (6.28)

where we recall t˜ = te−ξ/3 .
Now, let us check the transformation laws of DetB under DT . From (6.26) it is obvious
that DetB is manifestly invariant under all generators of DT except generator S.
Using
Im(ρ)
Im(−1/ρ) = (6.29)
ρ ρ̄
we find that under S, DetB transforms as
sB = (ρ ρ̄) 2 (b0 −b1 +b2 −b3 ) .

1
DetB (−1/ρ) = sB DetB (ρ), (6.30)
7. Inclusion of the fermion determinants
In this section we include the effects of the fermionic path integral. We recall the
fermion content in the 10-dimensional and 8-dimensional supergravity theories and derive
their actions. In the presence of nontrivial fluxes these fermionic path integrals are
nonvanishing, even for the supersymmetric spin structure on T 2 .
7.1. Fermions in 8D theory and their T-duality transformations
Let us begin by listing the fermionic content in the 8-dimensional supergravity theory
(this content will be derived from the 10-dimensional theory below).
The fermions in the 8D theory include two gravitinos ψ A , ηA , A = 0, . . . , 7, and spinors
˜ µ̃.19 The relation of these fields to the 10D fields is explained in (7.13), (7.14).
Σ, Λ, l, µ, l,
There are also bosonic spinor ghosts b1 , c1 , Υ2 and b2 , c2 , Υ1 which accompany ψ A and
ηA , respectively.
The fermions and ghosts transform under T-duality generators as follows. The
generators T , T,
S act trivially on fermions and ghosts while the under the generator S
they transform as

ψ A → eiα Γ ψ A , ηA → ηA , Λ → e−iα Γ Λ, Σ → Σ, (7.1)
2iα Γ ˜ iα Γ iα Γ
l →e l, l˜ → e−2iα Γ l, µ→e µ, µ̃ → e µ̃ (7.2)
and ghosts transform as

Υ1 → Υ1 , Υ2 → e−iα Γ Υ2 , (7.3)
iα Γ
{c1 , b1 } → e {c1 , b1 }, {c2 , b2 } → {c2 , b2 }, (7.4)
where α is defined by
1
α = ν + π, i ρ̄ = eiν |ρ| (7.5)
2
and Γ is the 8D chirality matrix.
The above transformation rules for space–time fermions follow from the transformation
rules for the appropriate vertex operators on the worldsheet (as discussed, for example, in
[11]). The only generator of DT acting nontrivially on fermions is S. The components
a , a = 8, 9, of the right-moving NS vertex are rotated by 2α, while the components
VNS
VNS
A are invariant. This follows since S does not act on the left-moving components of
˜
vertex operators. In this way we find the transformation rules for ηA , b2 , c2 , Σ, Υ1 , l, l,
which originate from R ⊗ NS sector. To account for the transformation rules for
ψ A , b1 , c1 , Λ, Υ2 , µ, µ̃ we recall that these fields originate from NS ⊗ R sector and that
the right-moving R vertex VR transforms under S as

S : VR → eiα Γ VR . (7.6)
19 These fields are MW in Lorentzian signature. We supress 16 component spinor indices below.
7.2. 10D fermion action
We start from the part of the 10D IIA supergravity action quadratic in fermions [19].
We work in the string frame.20

√ −2φ 1 ¯ A 1 N
(10)
Sferm = −g10 e ψ̂ AΓ N B DN ψ̂B + Λ Γ DN Λ
2 2

1
N
− √ (∂N φ)ΛΓ Γ ψ̂A
A
2

√
−g10 e GAC ψ̂¯ Γ[EΓAC ΓF] Γ11 ψ̂ F
1 E
−φ
+
16

3 E AC 11 5 AC 11

+ √ ΛΓ Γ Γ ψ̂E + ΛΓ Γ Λ
2 4

√ 1 ¯ B 5 A 21
+ −φ
−g10 e G0 ψ̂ AΓ ψ̂B + √ ΛΓ ψ̂A − ΛΛ
A
8 8 2 32

√
1 ψ̂¯ Γ ΓABCD
E
+ −g10 e−φ G AB C D [E ΓF] ψ̂ F
192

1 E ABCD 3 ABCD
+√ Λ Γ Γ ψ̂E + Λ Γ Λ
2 4

1 √
+ −g10 e−2φ HABC
48

¯ E BC
11 F √ E ABC
× ψ̂ Γ[EΓ A
ΓF] Γ ψ̂ + 2Γ Γ 11
ΛΓ ψ̂E , (7.7)

and ψ̂ A are the Majorana dilatino and gravitino and covariant derivatives act on
where Λ
them as
1 C A
DN ψ̂ A = ∂N ψ̂ A + ωN
A
Bψ̂ + ωN
B
Γ
B
ψ̂ ,
4 BC
1
DN Λ + ωNBCΓ BCΛ.
= ∂N Λ
4
There are also terms quartic in fermions in the action. It turns out that it is important
to take them into account to check the T-duality invariance of partition sum. We recall the
4-fermionic terms in Appendix C.
7.3. Reduction on T 2
To carry out the reduction of the fermionic action to 8D we choose the gauge for the
10D vielbein as
1
tEMA Am a
V 2 em
=
E A M
1 (7.8)
M a
0 V 2 em
20 We explain the relation between our conventions and those of [19] in Appendix B.

(recall a = 8, 9 and A = 0, . . . , 7) and use the following basis of 10D 32 × 32 matrices ΓA ,
ΓA = σ2 ⊗ Γ A , A = 0, . . . , 7,
Γ8 = σ1 ⊗ 116 , Γ9 = σ2 ⊗ Γ, Γ = Γ 0 · · · Γ 7 . (7.9)
Here Γ A are symmetric 8D Dirac matrices, which in Euclidean signature can be all chosen
to be real, and σ1,2,3 are Pauli matrices. In this basis the 10D chirality Γ11 and charge
conjugation matrices C (10) have the form
Γ11 = σ3 ⊗ 116 , C (10) = iσ2 ⊗ 116 . (7.10)

in the
The 8D fermions listed in Section 7.1 are related to 10D fields ψ̂ A and Λ
following way21
A √
ψ 1 A Σ 3 2
A = ψ̂ + Γ Γa ψ̂ ,
A a
= Λ + Γa ψ̂ a , (7.11)
η 6 Λ 4 4
√
l 2 µ̃
= Γa ψ̂ −
a
Λ, = ψ̂8 − Γ89 ψ̂9 . (7.12)
µ 2 l˜
7.4. 8D fermion action

(8)
Now we present the 8D action Squad = Skin + Sfermi-flux quadratic in fermionic
22
fluctuations over the 8D background specified in Section 2.2. The kinetic term is standard

1 1
Skin = e−2ξ t 7 ψ̄A Γ AMB DM ψB + η̄A Γ AMB DM ηB
2 2
X
2 M 2 M 1¯ M
+ ΣΓ DM Σ + ΛΓ DM Λ + lΓ DM l
3 3 4
+ µ̄Γ M DM µ + ¯lΓ
1 1˜ M 1¯ M
DM l˜ + µ̃Γ DM µ̃ . (7.13)
4 4 4
The coupling of fluxes to fermion bilinears is:

π n0 ρ + n1 n0 ρ̄ + n1
Sfermi-flux = e−ξ t 8 √ X(0) − √ X(0) + t 7 g(1)m ∧ ∗X(1)
m
4 Im ρ Im ρ
X

2 ρ + g1 2 ρ̄ + g 1
g(2) g(2)
+t 6
√
(2)
∧ ∗X(2) − √
(2)
∧ ∗X(2) + t 5 g(3)m ∧ ∗X(3)
m
Im ρ Im ρ

4
+ t Im ρ g(4) ∧ ∗[X(4) + X(4) ] , (7.14)
where the harmonic fluxes g(p) , p = 0, . . . , 4, were defined in (5.16). These harmonic fields
(p) constructed out of Fermi bilinears. We now give
couple to differential p-forms X(p) , X
21 Λ and Γa ψ̂ a are mixed to give the 8D “dilatino”, the superpartner of e−2ξ = e−2φ V .
22 In Minkowski signature ψ̄ = ψ † Γ 0 . In Euclidean signature ψ̄ and ψ are treated as independent fields.
A A A A
explicit formulae for X(p) :

√
X(0) = − (−)
A Γ
AB (−)
WB − W (−) Γ AB (−) + i 2 Λ (+) Γ A W(+)
B A A
√ √ √ (−) A (+)
(−) A (+) (+) A (−)
− i 2 WA Γ Λ + i 2 Σ Γ A − i 2 A Γ Σ
i i (+) A (−)
+ l¯ (−) Γ A (+)
A − A Γ l
2 2
i (−) A (+) i (+) A (−)
+ µ̄ Γ WA − WA Γ µ
2 2
Σ (+) − ¯l˜ µ̃(+) + µ̃
1 (+) 1 ¯ (+) ˜ (+)
+ 4Σ Λ − 4Λ
(+) (+) (+)
l , (7.15)
2 2
(X(2) )MN = (−) [A B] (−) (−) [A
A Γ ΓMN Γ WB + WA Γ ΓMN Γ B
B] (−)
√ √
+ i 2Λ (+) ΓMN Γ A W(−) − i 2 W (−) Γ A ΓMN Λ(+)
A A
√ √ (−) A
A (−)
+ i 2 Σ ΓMN Γ A + i 2 A Γ ΓMN Σ (+)
(+)
i i (+)
+ l¯ (−) Γ A ΓMN A +
(+)
ΓMN Γ A l (−)
2 2 A
i i (+)
− µ̄(−) Γ A ΓMN W(+) A − WA ΓMN Γ µ
A (−)
2 2
+ 4Σ (+) ΓMN Λ(+) + 4Λ (+) ΓMN Σ (+)
− ¯l˜ ΓMN µ̃(+) − µ̃

1 (+) 1 ¯ (+)
ΓMN l˜ (+) , (7.16)
2 2
where ψA(±) = 12 (116 ± Γ)ψA , etc., and we use the combinations of 8D fields
√ √
2 2
A = ψA + i ΓA Λ, WA = ηA − i ΓA Σ
3 3
to make the expressions for X(0) , X(2) have more nice coefficients.
The forms X (2) can be obtained from X(0) , X(2) by exchange of 8D chiralities
(0) , X
(−) ↔ (+).
Under the T-duality generator S the above forms transform as
{X(0) , X(2) } → e−iα {X(0), X(2) }, (0) , X

{X (2) } → eiα {X
(0), X
(2) } (7.17)
so that the combinations √Im 1
(n ρ + n1 )X(p) , √Im1 (p) for p = 0, 2 which
(n ρ̄ + n1 )X
ρ 0 ρ 0
appear in the action (7.13) are invariant under S.
Also we have defined the 1-form

m m (−) [A
X(1) M = e+ A Γ ΓM Γ B] W(+) (+) [A B] (−)
B − WA Γ ΓM Γ B
√ √
− i 2Λ (+) ΓM Γ A W(+) + i 2 W (+) Γ A ΓM Λ(+)
A A
√ √
A (−)
− i 2 Σ ΓM Γ A + i 2 A Γ ΓM Σ (−)
(−) (−) A
− ¯l˜ Γ A ΓM A +
i (+) i (+)
ΓM Γ A l˜(+)
(+)
2 2 A
i ¯ (−) A i (−) (−) ΓM Λ(+)

− µ̃ Γ ΓM W(−)
A + WA ΓM Γ µ̃
A (−)
− 4Σ
2 2
1 1
+ 4Λ(+)ΓM Σ (−) − µ̄(+) ΓM l (−) + l¯(−) ΓM µ(+) + e−m
(+) ↔ (−)
2 2
(7.18)
and the 3-form

m
X(3) MNP = e+
m
− (−) [A B] (+) (+) [A
A Γ ΓMNP Γ WB − WA Γ ΓMNP Γ B
B] (−)
√ √
+ i 2Λ (+) ΓMNP Γ A W(+) + i 2 W (+) Γ A ΓMNP Λ(+)
A A
√ √ (−) A
A (−)
− i 2 Σ ΓMNP Γ A − i 2 A Γ ΓMNP Σ (−)
(−)
− ¯l˜ Γ A ΓMNP A −
i (+) i (+)
ΓMNP Γ A l˜ (+)
(+)
2 2 A
i ¯ (−) A (−) i (−)
+ µ̃ Γ ΓMNP WA + W ΓMNP Γ A µ̃(−)
2 2 A
− 4Σ (−) ΓMNP Λ(+) − 4Λ (+) ΓMNP Σ (−)

1 1
+ µ̄(+) ΓMNP l (−) + l¯ (−) ΓMNP µ(+) + e− m
(+) ↔ (−) , (7.19)
2 2
where we denote e± m
= e8m ∓ ie9m .
m
The forms X(1) and X(3)m transform in the 2 of SL(2, Z) . Also from (7.18), (7.19) we
τ
m m
find that X(1) and X(3) are invariant under SL(2, Z)ρ if we accompany the action of the
generator S by the U (1) rotation of eam
m
e± → e±iα e±
m
. (7.20)
Since there are no local Lorentz anomalies, we can make this transformation.
The most important objects in (7.14) are the self-dual23 form X(4) and the anti-self-dual
(4) which couple to the flux g(4) . X(4) is defined by
form X
(X(4) )MNP Q = −i (+) [A B] (+) (+) [A

A Γ ΓMNP Q Γ WB + i WA Γ ΓMNP Q Γ B
B] (+)
√ √
− 2Λ (−) ΓMNP Q Γ A W(+) + 2 W (+) Γ A ΓMNP Q Λ(−)
A A
√ √ (+) A
− 2Σ ΓMNP Q Γ + 2
(−) A (+) A Γ ΓMNP Q Σ (−)
A
1 1 (−)
− l¯ (+) Γ A ΓMNP Q (−)
A + A ΓMNP Q Γ l
A (+)
2 2
1 1 (−)
− µ̄(+) Γ A ΓMNP Q W(−)A + WA ΓMNP Q Γ µ
A (+)
2 2
(−) ΓMNP Q Λ(−) − 4i Λ
+ 4i Σ (−) ΓMNP Q Σ (−)
− ¯l˜ ΓMNP Q µ̃(−) + µ̃

i (−) i ¯ (−)
ΓMNP Q l˜(−) (7.21)
2 2
(4) can be obtained from X(4) by the exchange of 8D chiralities (+) ↔ (−).
and X
1 B B B B .
A1 A2 A3 A4 = − 4! !A1 A2 A3 A4 B1 B2 B3 B4 Γ 1 2 3 4 Γ
23 In our conventions Γ
Under the T-duality generator S these forms transform as

X(4) → eiα X(4) , X (4).
4 → e−iα X (7.22)
We have also checked using Appendix C that the 4-fermion terms in the 8D action can
be written as

S4(8D)
-ferm = S4-ferm + S4-ferm ,

π (4) ∧ ∗X
(4) ].
S4 -ferm = e−2ξ t 8 [X(4) ∧ ∗X(4) + X (7.23)
128
X
While S4-ferm is manifestly invariant under T-duality, we will see that the non-invariant
term S4 -ferm is required for T-duality invariance of the total partition sum Z(F , ρ) of (1.32).
7.5. T-duality invariance of the ghost interactions
The classical 8D action obtained from the reduction of 10D IIA supergravity on T 2
is invariant under local supersymmetry (all 32 components survive the reduction). To
construct the quantum action we have to impose a gauge fixing condition on the gravitino

ψA
ψ̂(8D) :=
ηA
and include ghosts. Since the susy transformation laws involve fluxes, there is a potential
T-duality anomaly from the ghost sector. In fact no such anomaly will occur as we now
demonstrate. There are two generic properties of supergravity theories:
(1) In addition to a pair of Faddeev–Popov ghosts associated to the local susy gauge
transformation ψ̂(8D) → ψ̂(8D)
A + δ ψ̂ A
!ˆ (8D) a “third ghost”, the Nielsen–Kallosh ghost,
appears [32].
(2) Terms quartic in Faddeev–Popov ghosts are required [33].
Let us recall first how the “third ghost” appears. Following the standard procedure we
fix the local susy gauge by inserting δ(f − ΓA ψ̂(8D)
A ) into the path integral. Then we also
insert the unit24

−2ξ 7
1 i
t f¯Df
= i ΓN DN
1= [df ]e 2 Xe , D (7.24)
1 −2ξ 7
det 2 e t D
and integrate over [df ]. (If D has zeromodes this expression is formally 0/0, but (7.27)
below still makes sense.)
As a result we first find that the gravitino kinetic term gets modified to

i

− e−2ξ t 7 ψ̄ A MAB ψ B + η̄A MAB ηB , (7.25)
2
X

¯
24 We use the measure [df ]ei X ff = 1.
where the operator MAB acts on sections of the bundle25 Spin(X) ⊗ T X as

MAB = δAB iΓ M DM − 2iΓA DB , (7.26)
where DA = EA M D . The determinant in (7.24) is expressed as the partition function for
M
the “third ghost” Υ with action

i Υ
SΥ = − D
e−2ξ t 7 Υ . (7.27)
2
X
is a bosonic 32 component spinor, which we decompose into 16 component spinors as
Υ

= Υ1 .
Υ
Υ2
Now we come to the most interesting part of quantum action which involves Faddeev–
Popov ghosts b̂, ĉ
(2) (4)
Sbc = Sbc + Sbc , (7.28)
(2) (4)
where Sbc (Sbc ) denotes the parts of the action quadratic (quartic) in FP ghosts. Let us
discuss the quadratic part first. According to the standard FP procedure we have

¯
Sbc = t 7 e−2ξ b̂ΓA δĉ ψ̂(8D)
(2) A
. (7.29)
X8
We decompose bosonic 32 component spinors b̂, ĉ as

c b1
ĉ = 1 , b̂ = .
c2 b2
We can write the action as a sum of two pieces
(2) (2)0 (2)2
Sbc = Sbc + Sbc .
(2)0 (2)2
Here Sbc does not contain fermionic matter fields while Sbc is quadratic in fermions.
(2)0 (2)2
We now present Sbc and put Sbc in Appendix D:

(2)0 7 −ξ ¯ −ξ 2 8 n0 ρ + n1 gh n0 ρ̄ + n1 gh
Sbc = t e b̂(−i D)ĉ − πe t √ X − √ X
3 Im ρ (0) Im ρ (0)
X8
1 ghm
+ t 7 g(1)m ∧ ∗X(1)
2
2
1 6 g(2)ρ + g(2)
1 2 ρ̄ + g 1
g(2)
+ t Im √
gh
∧ ∗X(2) − √
(2) gh
∧ ∗X(2)
3 Im ρ Im ρ

1 ghm
+ t 5 g(3)m ∧ ∗X(3) , (7.30)
8
25 Spin(X) and T X are spinor and tangent bundles on X.

where we define forms bilinear in FP ghosts as

gh 1
(−) (−) (−) (−) (−) (−)
X(0) = b2 c1 − c1 (−) b2 − b1 c2 + c2 (−) b1 , (7.31)
2
gh 1
(−) (−)
X(2) MN = b2 ΓMN c1(−) + c1 (−) ΓMN b2(−) + b1 ΓMN c2(−)
2

+ c2 (−) ΓMN b1(−) , (7.32)
ghm 1 m (+) (−) (+) (−) (+)
X(1) M = e+ b2 ΓM c1 − c1 (−) ΓM b2 − b1 ΓM c2
2
1 m
+ c2 (+) ΓM b1(−) + e− (+) ↔ (−) , (7.33)
2
ghm 1 m (+) (−) (+) (−) (+)
X(3) MNP = e+ b2 ΓMNP c1 + c1 (−) ΓMNP b2 + b1 ΓMNP c2
2
1 m
+ c2 (+) ΓMNP b1(−) + e− (+) ↔ (−) . (7.34)
2
gh , X
The forms X gh can be obtained from Xgh , Xgh by exchange of 8D chiralities
(0) (2) (0) (2)
(−) ↔ (+). Note, that b̂, ĉ do not couple to the flux g(4) .
Let us now present the part of the quantum 8D action which is quartic in ghosts (as
obtained by following the procedure of [33]):

1 ¯ ABC ¯ 1
(4)
Sbc = e−2ξ t 8 b̂Γ ĉ b̂ΓABC ĉ + b̂¯ ΓA ĉ b̂¯ ΓA ĉ . (7.35)
84 3
The presence of this quartic action is due to the fact that gauge symmetry algebra is open
in supergravity: [δ!ˆ1 , δ!ˆ1 ]ψ̂(8D)
A contains a term proportional to the equation of motion of
A .
ψ̂(8D)
(4) (2)0
The T-duality invariance of Sbc , Sbc and SΥ is manifest and we have also checked that
(2)2
Sbc is T-duality invariant, so we conclude that the part of the 8D quantum action which
contains ghosts is T-duality invariant.
7.6. Computation of the determinants
We can now compute the fermionic quantum determinants including ghosts. Let us
˜ µ, µ̃, b1 , b2 , c1 , c2 , Υ1 , Υ2 and ψA , ηA in the full orthonormal
expand the fields Λ, Σ, l, l,
basis of the operators D̆ = iΓ N DN and M, respectively, where the operator M was
defined in (7.26). Note that since we are assuming that background fluxes are harmonic,
fermionic nonzero modes do not couple to them. Moreover,we can rescale nonzero modes
by a factor of e−ξ t 7/2 so that kinetic terms appear without any dependence on ξ and t,
but four-fermionic terms are suppressed as e2ξ t −6 with respect to the kinetic terms. Since
kinetic terms are manifestly T-duality invariant the integration over nonzero modes will
just give a factor DetF depending only on the Ricci-flat metric gMN and the constants t
and ξ, all of which are T-duality invariant. DetF has the form
DetF = rF (ξ, t) det M, (7.36)

where det M is determinant of the operator M defined in (7.26) regularized in a way that
−n
rF (ξ, t) = const × e−2ξ t 7 M , (7.37)
where nM denotes the number of zeromodes of M.
Note, that determinants of nonzero modes of the fermions Σ, Λ, l, µ, l, ˜ µ̃ and bosons

Υ1 , Υ2 , b1 , b2 , c1 , c2 cancel each other and do not contribute to DetF .
The situation is quite different for zeromodes: the kinetic terms are zero but there is
nonzero coupling to harmonic fluxes, so that if we rescale fermion zeromodes by e− 2 ξ t 2
1
we make both the fermion coupling to g(4) and the fermion quartic terms independent of ξ
and t. We will also rescale ghost zeromodes by e− 2 ξ t 2 and include the factor (e−ξ t 4 )nM
1
which comes from the rescaling of fermion and ghost zeromodes into the definition of
DetF , i.e., we define new rF :
n n
rFnew (ξ, t) := rF (ξ, t) e−ξ t 4 M = const · (t)−3nM eξ M . (7.38)
From (6.28) and (7.38) we find that the full quantum determinants depend on t and ξ in
the following way
nM −nK −b2 −2b1 −4b0
t˜−3 , (7.39)
where we recall that t˜ = te−ξ/3 is the U-duality invariant combination.26 Note that the
dependence on t˜ in (7.39) comes entirely from the volume of the space of zeromodes. The
volume of bosonic zeromodes is blowing up in the limit t˜ → ∞, but the volume of fermion
zeromodes is shrinking. Since (7.39) is an overall factor in the partition sum, it is a question
of a net balance between fermion and boson zeromodes whether the partition sum blows
up or vanishes in the limit t˜ → ∞.
7.7. Integration over the space of fermion zeromodes
We can split the action of the rescaled fermion and ghost zeromodes as
S (zm) = S (zm)inv + S (zm)ninv .

Here the part S (zm)inv is invariant under T-duality and includes all the ghost zeromode
interactions, the coupling of the fermion zeromodes to all RR fluxes except for g(4) and the

invariant part of the 4-fermion zeromode couplings, denoted S4(zm)
-ferm .
S (zm)ninv transforms nontrivially under the generator S of T-duality and can be recast in
the following way:

S (zm)ninv = {4π Im ρg(4) ∧ ∗Y(4) + 2π Im ρY(4) ∧ ∗Y(4) }, (7.40)
X
26 For any Ricci-flat spin 8-manifold the numbers n

M and nK can be expressed in terms of topological
invariants.
where we define the harmonic 4-form Y(4) as
1 1 (zm) (zm)
Y(4) = √ X + X(4) . (7.41)
16 Im ρ (4)
This object transforms under S as
S · Y(4) = − Re ρY(4) + i Im ρ ∗ Y(4) . (7.42)
We now expand the harmonic 4-forms in the basis ωi of H 4 (X, Z)

g(4) = ni + α̃ i ωi , Y(4) = y i ωi , β̃ = β̃ i ωi ,
where the chracteristics α̃, β̃ are given in (5.21). Next, we define

, ρ) = α̂
Θ(F dµ(zm)
F ĥei2π 2Φ Θ (Q), (7.43)
β̂
where the shifted characteristics are defined as α̂ i = α̃ i + y i , β̂ i = β̃ i + S · y i , and

(zm)
dµF denotes the measure of the rescaled fermion and ghost zeromodes. Recall that
Q(ρ) = [H Im ρ − i Re ρ]I. In (7.43) ĥ = e−S
(zm)inv
is a T-duality invariant expression
which depends on τ, ρ, t, gMN as well as fermion and ghost zeromodes. The dependence
on τ, ρ, t, gMN comes entirely from the coupling of the rescaled zeromodes (of fermions
and ghosts) to the fluxes g(p) , p = 0, 1, 2, 3. Finally, we have also defined
1
, ρ, y) := 2Φ
2Φ(F − yI S · y − yI β,
(7.44)
2
was defined in (5.24).
where 2Φ

Θ(F , ρ) is invariant under SL(2, Z)τ and transforms under SL(2, Z)ρ as
1 + 1 −
· F , −1/ρ) = sF µ(S)(−iρ) 2 b4 (i ρ̄) 2 b4 Θ(F
Θ(S , ρ), (7.45)
· F , ρ + 1) = µ(T )Θ(F
Θ(T , ρ). (7.46)
We do Poisson resummation to find (7.45) and the extra phase sF is due to the
transformation27 of dµzm
F
I (M)
= (i)I (M) (−iρ)− 2 I (M) (i ρ̄) 2 I (M) ,
1 1
sF = eiα (7.47)
where I (M) is the index of the operator M defined in (7.26). As in the standard compu-
tation of the chiral anomaly [34], only the zeromodes contribute to the transformation of
fermionic measure. Indeed, the contribution of the bosonic ghosts c1 , b1 , Υ2 to the trans-
˜
formation of the measure cancels that of the contribution of the fermions µ, µ̃, Λ, l, l.
27 Here we use the fact that the 10D fermions are Majorana fermions in Minkowski signature.
8. T-duality invariance
8.1. Transformation laws for ZB+F (F , τ, ρ)
Now we study the transformation laws for
ZB+F (F , τ, ρ) = DetB DetF e−SB (F ) Θ(F

, ρ), (8.1)
, ρ) is defined in (7.43), while DetB and Det are defined in (6.26) and (7.36),
where Θ(F F
(7.38), respectively. We also recall that SB (F ) is the real part of the classical action
evaluated on the background field configuration.
First, we note that ZB+F (F , τ, ρ) is invariant under SL(2, Z)τ . Second, we learn how
ZB+F (F , τ, ρ) transforms under SL(2, Z)ρ by using the transformation rules of DetB
, ρ) (7.45), (7.46). We find:
(6.30) and Θ(F
1 + 1 −
ZB+F (S · F , τ, −1/ρ) = sB sF µ(S)(−iρ) 2 b4 (i ρ̄) 2 b4 ZB+F (F , τ, ρ), (8.2)
ZB+F (T · F , τ, ρ + 1) = µ(T )ZB+F (F , τ, ρ), (8.3)
where sB is taken from the transformation of DB .
Now, using the definition of χ and σ
1 1
b0 − b1 + b2 − b3 + b4± = (χ ± σ ), σ = b4+ − b4− (8.4)
2 4
as well as the index theorem:

I (M) + λ = 248A
2 8
X X
we obtain the final result for the transformation under the generator S
1 1

X (p2 −λ
1 1
X (p2 −λ
2) 2)
ZB+F (S · F , τ, −1/ρ) = (−iρ) 4 χ+ 8 (i ρ̄) 4 χ− 8 ZB+F (F , τ, ρ).
(8.5)
From (8.3) and (8.5) we find that there is a T-duality anomaly.
Let us note in passing that the transformations (8.3), (8.5) are consistent for any 8-
dimensional spin manifold. This can be seen by computing28
π

ZB+F (ST )6 · F , τ, ρ = ei 4 X (7λ −p2 ) ZB+F (F , τ, ρ),
2

ZB+F S 4 · F , τ, ρ = ZB+F (F , τ, ρ) (8.6)
and then noting that the index theorem for 8-dimensional spin manifolds implies

2
7λ − p2 ∈ 1440Z. (8.7)
X
28 The branches for the 8th roots of unity are chosen in such a way that S 2 = (−)FR , where F is a space–time
R
fermion number in right-moving sector of type IIA string.
Incidentally, when X admits a nowhere-vanishing Majorana spinor of ± chirality the

Euler characteristic is given by [35]:

1
χ =± p2 − λ2 (8.8)
2
X
and the transformation rule (8.5) simplifies to
1
ZB+F (S · F , τ, −1/ρ) = (−iρ) 2 χ ZB+F (F , τ, ρ), (8.9)
1
ZB+F (S · F , τ, −1/ρ) = (i ρ̄) 2χ ZB+F (F , τ, ρ) (8.10)
for positive and negative chirality, respectively.
8.2. Including quantum corrections
Now we recall that there is a 1-loop correction to the effective 8D action

2 iπ
S1-loop = π Im ρ uλ + vp2 + Re ρ p2 − λ2 , (8.11)
24
X X

where we recall that π Im ρ X (uλ2 + vp2 ) comes from the regularization of q(Im ρ) in
(6.5) and the numbers u and v depend on the regularization.
We now demonstrate that to construct a T-duality invariant partition function this term
should be replaced with

1 1
Squant = χ + p2 − λ2
log η(ρ)
2 4
X

1 1
+ χ− p2 − λ 2
log η(−ρ̄) , (8.12)
2 4
X
where η(ρ) is Dedekind function. Taking the limit Im ρ → ∞ one can uniquely determine
u = − 24
1
and v = 24
1
in (8.11).
η has the following transformation laws:
1 πi
η(−1/ρ) = (−iρ) 2 η(ρ), η(ρ + 1) = e 12 η(ρ) (8.13)
so that e−Squant transforms as

−Squant
(−1/ρ) = (−iρ)− 4 χ− 8 X (p2 −λ (i ρ̄)− 4 χ+ 8 X (p2 −λ e−Squant (ρ),
1 1 2) 1 1 2)
e (8.14)
π

e−Squant (ρ + 1) = e −i 24 X (p2 −λ ) e−Squant (ρ).
2
(8.15)
Finally, we find that the total partition function
Z(F , ρ) := e−Squant ZB+F (F , ρ) (8.16)
is invariant
Z(T · F , ρ + 1) = Z(F , ρ), (8.17)
Z(S · F , −1/ρ) = Z(F , τ, ρ). (8.18)
This is our main result.
As a consistency check consider (for simplicity) the case when X admits a nowhere-
vanishing spinor of positive chirality and take the limit Im ρ = V → ∞
1
iπ
Squant → ρ+ e2πinmρ χ. (8.19)
12 m
n1 m1
We recognize the multiple cover formula for worldsheet instantons on T 2 from [18].
9. Application: Hull’s proposal for interpreting the Romans mass in the framework
of M-theory
As a by-product of the above results we will make some comments on an interesting

open problem concerning the relation of M-theory to IIA string theory.
It is well known that IIA supergravity admits a massive deformation, leading to the
Romans theory. The proper interpretation of this massive deformation in 11-dimensional
terms is an intriguing open problem. In [9] C. Hull suggested an 11-dimensional
interpretation of certain backgrounds in the Romans theory. His interpretation involved
T-duality in an essential way, and in the light of the above discussion we will make some
comments on his proposal. (For a quite different proposal for interpreting this massive
deformation see [36].)
9.1. Review of the relation of M-theory to IIA supergravity
Naive Kaluza–Klein reduction says that for an appropriate transformation of fields

{gM-theory, CM-theory} → {gIIA , HIIA , φIIA , CIIA } we have
SM-theory = SIIA . (9.1)

One of the main points of [7] was that, in the presence of topologically nontrivial fluxes
equation (9.1) is not true! Indeed, given our current understanding of these fields, there
is not even a 1–1 correspondence between classical M-theory field configurations and
classical IIA field configurations. Rather, certain sums of IIA-theoretic field configurations
were asserted to be equal to certain sums of M-theoretic field configurations. In this sense,
the equivalence of type IIA string theory to M-theory on a circle fibration is a quantum
equivalence.
To be more precise, in [7] it was shown that for product manifolds Y = X10 × S 1 , the
sum over K-theory lifts x(â) of a class â ∈ H 4 (X10 ; Z) is proportional to the sum over
torsion shifts of the M-theory 4-form of Y . We have:
N(−)Arf(q)+f (â0 ) −SIIA

√ e
N2 NK
x(â)
2
= exp −GM-theory (â) (−1)f (â+ĉ) . (9.2)
4 (X ;Z)
ĉ∈Htors 10
The above formula is the main technical result of [7]. We recall that [GM-theory(â)] =
2π(â − 12 λ) and the equivalence class of â is defined to contain M-theory field
configurations with fixed kinetic energy

GM-theory(â)2 = 1 GM-theory(â) ∧ ∗ˆ GM-theory(â),
4π
X10
from which follows that these fields are characterized by â = â + ĉ, ĉ ∈ Htors 4 (X ; Z).
10
Also, in (9.2) NK and N is the order of Ktors (X10 ) and Htors(X10 ; Z) respectively, N2
0 4
stands for the number of elements in the quotient L = L/L , where L = Htors 4 (X ; Z)/
10
4
2Htors(X10 ; Z) and L = {ĉ ∈ L, Sq ĉ = 0}. Finally, Arf(q) is the Arf invariant of the
3
quadratic form q(ĉ) = f (ĉ) + X10 ĉ ∪ Sq 2 â0 on L . The identity (9.2) extends to the case
where Y is a nontrivial circle bundle over X10 [7].
As we have mentioned, we interpret the fact that we must sum over field configurations
in (9.2) as a statement that IIA-theory on X10 and M-theory on Y = X10 × S 1 are really
only quantum-equivalent. This point might seem somewhat tenuous, relying, as it does,
on the fact that the torsion groups in cohomology and K-theory are generally different.
Nevertheless, as we will now show, a precise version of Hull’s proposal again requires
equating sums over IIA and M-theory field configurations. In this case, however, the sums
are over nontorsion cohomology classes, and in this sense the claim that IIA-theory and
M-theory are only quantum equivalent becomes somewhat more dramatic.
9.2. Review of Hull’s proposal
One version of Hull’s proposal states that massive IIA string theory on T 2 × X is
equivalent to M-theory on a certain 3-manifold which is a nontrivial circle bundle over
a torus. The proposal is based on T-duality invariance, which allows one to transform away
G0 at the expense of introducing G2 along the torus, combined with the interpretation of
G2 flux as the first Chern class of a nontrivial M-theory circle bundle [7]. We now describe
this in more detail.
Hull’s proposal is based on the result [12] that dimensional reduction of massive IIA
supergravity with mass m on a circle of radius R (denoted SR1 ), gives the same theory as
Scherk–Schwarz reduction of IIB supergravity on S1/R 1 . The IIB fields are twisted by

1 mθ
g(θ ) = , (9.3)
0 1
1
where the coordinate on S1/R is z = 2π
R θ , θ ∈ [0, 1] and the monodromy is

1 m
g(1)g(0)−1 = ∈ SL(2, Z). (9.4)
0 1
Schematically:

IIAm IIB
= , (9.5)
SR × X9
1 S1/R × X9
1
g(θ)
where X9 is an arbitrary 9-manifold. Note, in particular, that the twist acts on the IIB
axiodil τB = C0 + ie−φB as
τB (θ ) = τB (0) + mθ (9.6)
which implies that the IIB RR field G1 has a nonzero period.
Let us also recall the duality between IIB on a circle and M-theory on T 2 :
IIB M
= 2 , (9.7)
SR1 × S1/R × X T (τM , AM ) × S1/R
1 1 ×X
where the T 2 (τM , AM ) on the M-theory side has complex structure τM = τB (0) and area
AM = eφB /3 (R )−4/3 .
Now, invoking the adiabatic argument we have:

IIB M
= , R) × X
, (9.8)
S1/R × SR × X g(θ)
1 1 B(m; R
where B(m; R , R) is a 3-manifold with metric:
2
2π 1 2
ds =
2
(dθ ) + AM
2
dx + (Re τM + mθ ) dy + Im τM dy , (9.9)
2
R Im τM
where x, y are periodic x ∼ x + 1 and y ∼ y + 1.29
Combining (9.5) with (9.8) we get the basic statement of Hull’s proposal:
IIAm M
= . (9.10)
SR1 × SR × X B(m; R , R) × X
1
We can now see the connection between Hull’s proposal and T-duality. A duality
transformation exchanges G0 for a flux of G2 through the torus. Then we can interpret
the nontrivial flux G2 as the first Chern class of a line bundle in the M-theory setting.
9.3. A modified proposal
In view of what we have discussed in the present paper, the equivalence of classical
actions—when proper account is taken of the various phases of the supergravity action—
cannot be true. This is reflected, for example, in the asymmetry of the phase (5.12) in
exchanging n0 for n1 . However, we follow the lead of (9.2) and therefore modify Hull’s
proposal by identifying sums over certain geometries on the IIA and M-theory side.30
A modified proposal identifies Z(F , ρ, τ ) defined in (8.1), (8.16) with a sum over
M-theory geometries as follows. Recall first that in the 8D theory there
p is a doublet of
(0) = Q q where p, q are
α , arising from G and G . Next, let us factor g
zeroforms g(0) 0 2
29 It is not entirely obvious that the invocation is justified, since for a large M-theory torus the twist is carried
out over a small radius on the IIB side.
30 In making these statements we are including the K-theoretic phase as part of the “classical” action. Since the
phase is formally at 1-loop order it is possible that one could associate it with a 1-loop effect in such a way that
classical equivalence does hold.
relatively prime integers and Q is an integer. Then we take a matrix N ∈ SL(2, Z)ρ

r −s
N= , rp − sq = 1 (9.11)
−q p
such that

Q
N g(0) = . (9.12)
0
This is the T-duality transformation that eliminates Romans flux.
Now, thanks to the invariance of Z(F , τ, ρ) under T-duality transformations (see (8.17),
(8.18) above) we find:

pρ + s
Z(F , ρ) = Z N · F , . (9.13)
qρ + r
By the results of [7] the right-hand side of (9.13), having G0 = 0, does have an
interpretation as a sum over M-theory geometries. The M-theory geometry is indeed a
circle bundle over T 2 ×X defined by c1 = Qe0 +pe −qe +γm dσ m (as in Hull’s proposal),
but in addition it is necessary to sum over E8 bundles on the 11-manifold B × X. While
it is essential to sum over g(4) , all other fluxes F may be treated as classical—that is, they
may be fixed and it is not necessary to sum over them.
Both sides of (9.13) should be regarded as wavefunctions in the quantization of self-dual
fields. For this reason we propose that there is only an intrinsically quantum mechanical
equivalence between IIA theory and M-theory in the presence of G0 .
10. Comments on the U-duality invariant partition function
The present paper has been based on weakly coupled string theory. However, our
motivation was understanding the relationship between K-theory and U-duality. In
generalizing our considerations to the full U-duality group D = SL(3, Z) × SL(2, Z)ρ
of toroidally compactified IIA theory it is necessary to go beyond the weak coupling
expansion. Thus, it is appropriate to start with the M-theory formulation. In the present
section we make a few remarks on the U-duality of the M-theory partition function and
its relation to the K-theory partition functions of type IIA strings. In particular, we will
address the following points:
(a) The invariance of the M-theory partition function under the nongeometrical SL(2, Z)ρ
is not obvious and appears to require surprising properties of η invariants. In
Section 10.2 we state this open problem in precise terms.
(b) We will sketch how one can recover “twisted K-theory theta functions”, at weak
coupling cusps when the H -flux is nonzero in Section 10.3.
We believe that one can clarify the relation between K-theory and U-duality by studying
the behavior of the M-theory partition function at different cusps of the M-theory moduli
space. At a given cusp the summation over fluxes is supported on fluxes which can be
related to K-theory. (See, for example, (9.2).) A U-duality invariant formulation of the
theory must map the equations defining the support at one cusp to those at any other cusp.
This should define the U-duality invariant extension of the K-theory constraints.
10.1. The M-theory partition function
Let us consider the contribution to the M-theory partition function from a background
Y which is a T 3 fibration over X.
= V − 3 t˜ 2 gMN dx M dx N + V 3 g̃mn θ m θ n ,
1 2
2
ds11 (10.1)
where θm = dx m + Am
and
(1) ∈ [0, 1].
xm t˜ 2 gMN
is 8D Einstein metric with det gMN = 1.
g̃mn and V are the shape and the volume of the T 3 fiber. We denote world indices on T 3
by m = (m, 11), m = 8, 9 and M = 0, . . . , 7 as before.
Topologically, one can specify the T 3 fibration over X by a triplet of line bundles Lm
which transform in the representation 3 of SL(3, Z) and have first Chern classes c1 (Lm ) =
F(2)
m , where F m = dθ m . Such a specification is valid up to possible monodromies. These
(2)
are characterized by a homomorphism π1 (X) → SL(3, Z).
On a manifold Y of the type (10.1) we reduce the M-theory 4-form GM-theory as
GM-theory 1 m n
= G(4) + G(3)mθ m + F(2)mn + εmnk B0 F(2)
k
θ θ . (10.2)
2π 2
We also include the flat potential
1
c(0) = B0 εmnk θ m θ n θ k (10.3)
6
in the Kaluza–Klein reduction.31 (We will list the full set of flat potentials in this
background below.)
From the Bianchi identity dGM-theory = 0 we have
dG(4) = F(2)
m
G(3)m , dG(3)m = F(2)
n
F(2)mn ,
dF(2)mn = 0, m
dF(2) =0 (10.4)
which implies that fluxes G(4) and G(3)m are in general not closed forms.32
Let us recall how the various fields transform under D = SL(3, Z) × SL(2, Z)ρ [14].
• t˜, gMN are U-duality invariant.

• SL(2, Z)ρ acts on ρ = B0 + iV ∈ H by fractional linear transformations.
• SL(3, Z) acts on the scalars g̃mn parametrizing SL(3, R)/SO(3) via the mapping class
group of T 3 .
F(2)
m
• Fmα(2) = F m transform in the (3, 2) of D, where F(2) := 2 ε
m 1 mnk
F(2)nk .
(2)
31 ε 11,8,9 = ε
11,8,9 = 1.
n
(3)11 = 0 and F(2) = 0, n = 8, 9, so that all background fluxes are
32 In IIA at weak coupling we assumed G
closed forms.
• G(3)m transform in the (3 , 1) of D.

• G(4) is singled out among all the other fields since according to conventional
supergravity [14] SL(2, Z)ρ mixes G(4) with its Hodge dual ∗G(4) . More concretely,

− Re ρG(4) + i Im ρ ∗ G(4)
(10.5)
G(4)
transforms in the (1, 2) of D. Due to this nontrivial transformation the classical bosonic
8D action is not manifestly invariant under SL(2, Z)ρ . In detail, the action has real part:

Re(S8D ) = π Im ρG(4) ∧ ∗G(4) + t˜ 2 g̃ mn G(3)m ∧ ∗G(3)n

X
nβ
+ t˜ 4 g̃mn Gαβ Fmα
(2) ∧ ∗F(2)

1 1
+ t˜6 R + 28t˜−2 ∂M t˜∂M t˜ + 2 ∂M ρ∂ M ρ̄
2π 2ρ2
X

1
+ g̃ mn g̃ kl ∂M g̃mk ∂ M g̃nl , (10.6)
3
where Gαβ is defined in (1.6), g̃ kl is inverse of g̃mk and R is the Ricci-scalar of the
metric gMN .
The imaginary part of the 8D bosonic action follows from the reduction of the M-theory
phase ΩM (C). This phase is subtle to define in topologically nontrivial field configurations
of the G-field. It may be formulated in two ways. The first formulation was given in [37]. It
uses Stong’s result that the spin-cobordism group Ω11 (K(Z, 4)) = 0 [38]. That is, given a
G
spin 11-manifold Y and a 4-form flux 2π one can always find a bounding spin 12-manifold

Z and an extension G of the flux to Z. In these terms the M-theory phase ΩM (C) is given
as:

2πi
ΩM (C) = ! exp G3 − 2πi G p2 − λ2 . (10.7)
6 48
Z Z
Here ! is the sign of the Rarita–Schwinger determinant. The phase does not depend on
the choice of bounding manifold Z, but does depend on the “trivializing” C-field at the
boundary Y .
A second formulation [7,39,40] proceeds from the observation of [37] that the integrand
of (10.7) may be identified as the index density for a Dirac operator coupled to an E8 vector
bundle. The M-theory 4-form can be formulated in the following terms [7,39,40]. We set:
GM-theory
+ dc,
=G (10.8)
2π
where G = 1 Tr248 F 2 + 1 2 Tr R 2 , F is the curvature of a connection A on an E8
2
60 8π 32π
bundle V on Y and R is the curvature of the metric connection on T Y. GM-theory is a
real differential form, and c ∈ Ω 3 (Y, R)/ΩZ3 (Y ), where ΩZ3 (Y ) are 3-forms with integral
periods. The pair (A, c) is subject to an equivalence relation. In these terms the M-theory
phase is expressed as:

η(DV ) + h(DV ) η(DRS ) + h(DRS )
ΩM (C) = exp 2πi + ω(c), (10.9)
4 8
where DV is the Dirac operator coupled to the connection A, DRS is the Rarita–Schwinger
operator, h(D) is the number of zeromodes of the operator D on Y , and η(D) is the η
invariant of Atiyah–Patodi–Singer. The phase ω(c) is given by

2 1
ω(c) = exp πi
c G + X8 + cdcG + c(dc) 2
. (10.10)
3
Y
10.2. The semiclassical expansion
For large t˜ there is a well-defined semiclassical expansion of the M-theory partition

function, which follows from the appearance of kinetic terms in the action (10.6) scaling as
t˜ 2k for k = 0, 1, 2, 3. In the leading approximation we can fix all the fields except G(4) , but
this last field must be treated quantum mechanically. Note that this semiclassical expansion
can differ from that described in the previous sections because we do not necessarily
require weak string coupling. In the second approximation we treat G(4) and G(3)n as
quantum fields, and so on.
In the leading approximation in addition to the sum over fluxes G(4) we must integrate
over the flat potentials. These include flat connection Am 3
(1) of the T fibration and potentials
coming from the KK reduction of c
1
c = C(3) + C(2)m θ m + C(1)mnθ m θ n + c(0) , (10.11)
2
p
= C(2)m − 12 C(1)pm A(1) and C(3) =C
(3) − C(2)m A(1) , and c(0) is defined
where C(2)m m
in (10.3). C(3) is invariant under U-duality, C(2)m transforms in the (3, 1) of D. We can
combine the flat potentials C(1)mn and Am (1) in a U-duality multiplet of D transforming as
(3, 2) by writing
1 mnk
2ε C(1)nk
A(1) =
mα
. (10.12)
Am(1)
The duality invariance in the leading approximation is straightforward to check. We
keep only G(4) . The flux is quantized by [G(4)] = a − 12 λ, where a ∈ H 4 (X, Z) is the
characteristic class of the E8 bundle and λ is the characteristic class of the spin bundle. We
sum over a ∈ H 4 (X, Z). The 8D action, including the imaginary part is SL(3, Z) invariant.
The imaginary part of the 8D effective action in this case takes a simple form which can
be found from (10.9)

1 2
Im(S8D ) = −π a ∪ λ + B0 a − λ . (10.13)
2
X
The invariance under SL(2, Z)ρ then follows in the same way as in our discussion in the
weak string coupling regime.
Let us now try to go beyond the first approximation. In the second approximation
[G(4)] = a − 12 λ + [Am (1) G(3)m ]. We allow nonzero fluxes G(3)m , but still set to zero the
fieldstrengths F(2) and F(2) . We thus have a family of tori with flat connections. Already in
the second approximation, when we switch on nonzero fluxes G(3)m there does not appear
to be a simple expression for the M-theory phase.
Nevertheless, one can get some information about the M-theory phase from the
requirement of U-duality invariance. We know that SL(3, Z) invariance is again manifest
from the definition of ΩM (C) and Re(S8D ). But the expected SL(2, Z)ρ invariance is
nontrivial. We would simply like to state this precisely. To do that we write M-theory
partition function in the second approximation as

ZM-theory(g̃mn , ρ) := dµflat ZM-theory(g̃mn , G(3)m , ρ), (10.14)
G(3)m
where ZM-theory(g̃mn , G(3)m, ρ) is the partition function with fixed, but nonzero, flux,
G(3)m , dµflat stands for the integration over
2 1
H3 (X) H (X) 3 H (X) 6
× × , (10.15)
HZ3 (X) HZ2 (X) HZ1 (X)
p
where Hp (X) is a space of harmonic p-forms on X and HZ (X) is the lattice of integrally
normalized harmonic p-forms on X. The first factor is for C(3) , the second factor for C(2)m
(1) transforming in the (3, 2) of D. The integration
and the third factor is for the fields Amα
measure dµflat is U-duality invariant.
The summand in (10.14) with fixed G(3)m is given by

ZM-theory(g̃mn , G(3)m, ρ) = Det(G(4) , G(3)m)e−Squant e−Scl , (10.16)
a∈H 4 (X;Z)
where

˜2 g̃ mn G(3)m ∧∗G(3)n )
e−Scl = ΩM (G(4) , G(3)m, B0 )e−π X (Im(ρ)G(4) ∧∗G(4) +t
and Det(G(4), G(3)m ) denotes 1-loop determinants. These depend implicitly on the scalars
ρ, g̃mn , t˜ as well as on the metric gMN . We include 1-loop corrections in Squant (see below).
The M-theory phase ΩM in (10.16) depends on the fieldstrengths G(4) , G(3)m and the
flat potentials, but it is metric-independent, and hence should be a topological invariant.
The dependence of ΩM on flat potentials is explicit from (10.10) for c as in (10.11). For
example dependence of ΩM on B0 has the form

eiπ X B0 G(4) G(4) . (10.17)

It is convenient to include 1-loop corrections X B0 X8 together with effect of membrane
instantons in Squant . The nontrivial question is dependence on G(4) and G(3)m which also
comes from η(DV ) + h(DV ).
The independence of ΩM on the metric on Y = X × T 3 (in the second approximation)
follows from the standard variation formula for η-invariant. To show this let us fix the
connection on the E8 bundle V with curvature F and consider the family of vielbeins e(s)
on Y = X × T 3 parametrized by s ∈ [0, 1] such that the metric on T 3 remains flat and

independent of the coordinates on X. The corresponding family of Riemann tensors R(s)
which is a pullback from X × [0, 1]. Now we can write the
gives an A-roof genus A(s)
standard formula for the change in η-invariant under the variation of vielbein [41]:

η e(1) − η e(0) = j +
ch(V )A(s), (10.18)
Y ×[0,1]
iF
where integer j is a topological invariant of Y × [0, 1] and ch(V ) := 30
1
[Tr248 e 2π ]. In

the second approximation we only switch on G = G(4) + G(3)m dx so that neither
m
ch2 (V ) = −2(G + 1 λ) nor ch4 (V ) = 1 (G

+ 1 λ)2 have a piece ∼ dx 8 dx 9 dx 11 and
2 5 2
integral in (10.18) vanishes.
Now we come to the main point. The requirement of the invariance under the standard
generators S, T of SL(2, Z)ρ
ZM-theory(g̃mn , −1/ρ) = ZM-theory(g̃mn , ρ), (10.19)
ZM-theory(g̃mn , ρ + 1) = ZM-theory (g̃mn , ρ) (10.20)
gives a nontrivial statement about the properties of the function ΩM (G(4) , G(3)m, B0 ).
The sum over fluxes G(3)m ∈ H 3 (X, Z) in (10.14) might be entirely supported by
classes which satisfy a system of SL(3, Z) invariant constraints. These constraints can in
principle be determined by summing over torsion classes once the phase ΩM is known in
sufficiently explicit terms. In the simple case when G(3)m are all 2-torsion classes, one can
act by the generators of SL(3, Z) on the constraint
Sq 3 (G(3)9) + Sq 3 (G(3)11) + G(3)9 ∪ G(3)11 = 0 (10.21)
which follows from [7]. If we assume that this constraint is part of SL(3, Z) invariant
system of constraints then we find
G(3)m ∪ G(3)n = 0, m, n = 8, 9, 11. (10.22)
10.3. Comment on the connection with twisted K-theory
In this section we discuss the behavior of the partition function near a weak-coupling
cusp. There is a twisted version of K-theory which is thought to be related to the
classification of D-brane charges in the presence of nonzero NSNS H -flux [2,42–44]. It is
natural to ask if the contributions to the M-theory partition function ZM-theory (g̃mn , ρ) from
fluxes with nonzero H(3) := G(3)11 ∈ H 3 (X, Z) are related, in the weak string-coupling
cusp, to some kind of twisted K-theory theta function.
The weak-coupling cusp may be described by relating the fields in (10.1) to the fields
in IIA theory. First, the scale t˜ is related to the expansion parameter used in our previous
sections by t˜ 2 = e− 3 ξ t 2 . Next, we parametrize the shape of T 3 as g̃mn = em
2
a e b δ where
n ab
 e−ξ/3 
√ 0 0 1 τ1 C(0)8
τ2
a
em = 0 −ξ/3 √  0 1 C(0)9 . (10.23)
e τ2 0
0 0 e 2ξ/3 0 0 1
We denote frame indices by a = (a, 11), a = 8, 9. The weak coupling cusp may be written
as
R × R 2 × SL(2, R)/SO(2), (10.24)
where the first factor is for the dilaton ξ , the second for C(0)8 , C(0)9,33 and the third for the
modular parameter τ of the IIA torus.
As far as we know, nobody has precisely defined what should be meant by the “KH
theta function”. Since the Chern character has recently been formulated in [22,23], this
should be possible. Nevertheless, even without a precise definition we do expect it to
be a sum over a “Lagrangian” sublattice of KH (X × T 2 ). At the level of DeRham
cohomology, this should be a “maximal Lagrangian” sublattice of ker d3 / Im d3 where
d3 : H ∗ (X10 , Z) → H ∗ (X10 , Z) is the differential d3 (ω) = ω ∧ [H(3)]. Using the filtration
implied by the semiclassical expansion, and working to the approximation of e−t˜ this
2
means that we should first define a sublattice of the cohomology lattice by the set of integral
cohomology classes (a, G(3)8, G(3)9) such that (G(4), G(3)8 , G(3)9) are in the kernel of d3 :
H(3) ∧ G(4) = 0, H(3) ∧ G(3)m = 0, m = 8, 9. (10.25)
Then the theta function should be a sum over the quotient lattice obtained by modding out
by the image of d3
G(3)8 ∼ G(3)8 − pH(3) , G(3)9 ∼ G(3)9 − sH(3),
G(4) ∼ G(4) − ω(1) H(3). (10.26)
Here p, s ∈ Z and ω(1) ∈ H 1 (X, Z).
Thus, our exercise is to describe how a sum over this
quotient lattice emerges from (10.14).
Let us consider the couplings of flat potentials C(1)89 and C(2)m to the fluxes which
follow from (10.10)
mn C
C(1)89 H(3) G(4) 2πi X! (2)m G(3)n H(3)
e2πi X e . (10.27)
Integrating over C(1)89 and C(2)m gives H(3) ∧ G(4) = 0 and ! mn H(3) ∧ G(3)n = 0,
respectively.
Next, we note that, due to the SL(3, Z) invariance of the M-theory action we have
(suppressing many irrelevant variables)

ZM-theory C(0)m , G(3)m − pm H(3) , A11
(1) , G(4) − ω(1) H(3)

= ZM-theory C(0)m + pm , G(3)m , A11
(1) + ω(1) , G(4) . (10.28)
Now we use (10.28) to write the sum over all fluxes G(4) , G(3)m , m = 8, 9, in the kernel
of d3 as

ZH = (1) , G(4) =
ZM-theory C(0)m , G(3)m , A11 W, (10.29)
d3 -kernel Mfund
(0)m transforming in the 2 of SL(2, Z)τ as C(0)8 = eξ √τ2 C

33 These are related to the RR potentials C (0)8 ,
1
C(0)9 = e √τ C
ξ (0)9 .
2
where Mfund stands for the fluxes in the fundamental domain for the image of d3 within
the kernel of d3 and

W= ZM-theory C(0)m + pm , G(3)m , A11
(1) + ω(1) , G(4) . (10.30)
pm ∈Z2 ω(1) ∈H 1 (X,Z)
Now, we can recognize that ZH descends naturally to the quotient of the weak-coupling
cusp

Γ∞ \ R × R 2 × SL(2, R)/SO(2) , (10.31)
∼ Z2
where Γ∞ = is the subgroup of the parabolic group Γ∞ consisting of elements of the
form

1 0 p
Lm = 0
n
1 s , p, s ∈ Z. (10.32)
0 0 1
Written this way, ZH is clearly a sum over a Lagrangian sublattice of the KH (X × T 2 )
lattice. (Recall that we are working in the DeRham theory, with the filtration appropriate
to the second approximation.)
The interesting point that we learn from this exercise is that in formulating the KH
theta function, the weighting factor for the contribution of a class in KH should be given
by (10.30). The dependence of the action on the integers pm and ω(1) ∈ H 1 (X, Z) behaves
like exp[−Q(pm , ω(1) )] where Q is quadratic form. Therefore W is itself already a theta
function. This follows because the dependence on C(0)m and A11 (1) comes entirely from the
real part of the classical action (10.6), since, as we have shown, the phase is independent
of the metric on X × T 3 . The dependence on C(0)m comes from X t˜ 2 g̃ mn G(3)m ∧ ∗G(3)n

and the dependence on A11 (1) from X Im ρG(4) ∧ ∗G(4) , where we recall that [G(4) ] =
a − 12 λ + [Am(1) G(3)m ].
It would be very interesting to see if the function ZH defined in (10.29) is in accord
with a mathematically natural definition of a theta function for twisted K-theory. But we
will leave this for future work.
As an example, let us consider X = SU(3). Let x3 generate H 3 (X, Z). Then fixing
H(3) = kx3 we find that the fundamental domain of the image of d3 within the kernel of d3
is given by
G(3)8 = rx3 , G(3)9 = px3 , 0 r, p k − 1, (10.33)
so that the sum over RR fluxes in (10.29) is finite and it is in this sense that RR fluxes are k-
torsion. This example of X = SU(3) is especially interesting since it is well known [44–46]
that at weak string coupling D-brane charges on SU(3) in the presence of H(3) = kx3 are
classified by twisted K-theory groups of SU(3), and these groups are k-torsion. As argued
in [5], from Gauss’s law it is then natural to expect that RR fluxes are also k-torsion. This
is indeed what we find in (10.33).34 On the other hand, the M-theory sum is indeed a full
34 In fact, from [44] we know the order of the torsion group is actually k or k/2, according to the parity of k.
However, given the crude level at which we are working we do not expect to see that distinction. We expect that
a more accurate account of the phases in the partition function will reproduce this result.
sum over all fluxes. This is in harmony with the result of [47] for brane charges. Clearly,
there is much more to understand here.
Acknowledgements
G.M. would like to thank E. Diaconescu and E. Witten for many important discussions
and correspondence related to these matters both during and since the collaboration leading
to [5,7]. He would also like to thank B. Acharya, C. Hull, N. Lambert, and J. Raben for
useful discussions and correspondence, and C. Hull for some remarks on the manuscript.
G.M. would also like to thank L. Baulieu and B. Pioline at LTPHE, Paris and the
Isaac Newton Institute for hospitality during the completion of this manuscript. G.M. is
supported in part by DOE grant # DE-FG02-96ER40949.
Appendix A. Duality transformations as symplectic transformations
Here we give the explicit expressions for representations of S and T in Sp(2N, Z). Let
us choose the following basis of the lattice Γ
x = (x1 , x 2 ), (A.1)

x 1 = yl 1, yl ⊗ L(e0 ) − 1 , L(es ) − 1 , L(es ) − 1 ⊗ L(e0 ) − 1 ,

L γr dσ m − 1 , x fk dσ m , x(ωi ) , (A.2)

x 2 = x(ωi ) ⊗ L(e0 ) − 1 , x dk dσ m , x wr dσ m , x(us ),

x(us ) ⊗ L(e0 ) − 1 , x(hl ), x(hl ) ⊗ L(e0 ) − 1 , (A.3)
where we introduce
yl ∈ H 0 (X, Z), hl ∈ H 8 (X, Z), l = 1, . . . , b0 ,

γr ∈ H (X, Z),
1
wr ∈ H (X, Z),
7
r = 1, . . . , b1 ,
es ∈ H 2 (X, Z), us ∈ H 6 (X, Z), s = 1, . . . , b2 ,
fk ∈ H (X, Z),
3
dk ∈ H (X, Z),
5
k = 1, . . . , b3 ,
ωi ∈ H 4 (X, Z), i = 1, . . . , b4 ,
where bp is the rank of H p (X, Z) and b3 is the rank of the sublattice of H 3 (X, Z) which
is span by classes f such that Sq 3 f = 0.
In the above basis the generators S and T are represented by

A(S) B(S) A(T ) B(T )
σ (S) = , σ (T ) = , (A.4)
C(S) D(S) C(T ) D(T )
 
1b0
 −1b0 
 
 1b2 
 
A(S) =  −1b2 ,
 
 12b1 
 
12b3
0b4
 
 
 
 
 
B(S) =  , (A.5)
 
 
 
1b4
 
−1b4
 
 
 
 
C(S) =  ,
 
 
 
 
0b4
 12b3 
 
 12b1 
 
D(S) =  1b2 , (A.6)
 
 −1b2 
 
1b0
−1b0
 
1b0
−1
 b0 1b0 
 
 1b2 
 
A(T ) =  −1b2 1b2 , B(T ) = 0N , (A.7)
 
 12b1 
 
12b3
1b4
 
1b4
 12b3 
 
 12b1 
 
C(T ) = 0N , D(T ) =  1b2 . (A.8)
 
 −1b2 1b2 
 
1b0
−1b0 1b0
Appendix B. Supergravity conventions
The 10D fields that we use are related to the fields in [19] as
G4 3φ G2 9φ
√ = e− 4 F4Rom , √ = −e− 4 F2Rom ,
2π 2π
2
B 3φ 15φ
√ = −e 2 B2Rom , m = G0 e 4 ,
2π
φ φ
ψ̂A = e− 8 ψA = e− 8 λ(Rom) ,
(Rom) 1 (Rom)
= e 2 gM
φ
, Λ gM
N N .
√
We also remind that we set k11 = π while in [19] k11 = 2π was assumed.
Appendix C. 4-Fermion terms
Below we collect 4-fermionic terms in D = 10 IIA supergravity action which are

obtained from circle reduction of the D = 11 action of [48]:

π √ 1
S4(10)
-ferm = −g10 e −2φ
− χ̄E ΓABCDEFχF + 12χ̄ [A ΓBC χ D]
2 64
× χ̄[A ΓBC χD]
1 1
+ χ̄E ΓABCEF χF χ̄A ΓB χC + χ̄A ΓA χC χ̄B ΓB χ C
32 4

1 1
− χ̄A ΓB χC χ̄B ΓA χ C − χ̄A ΓB χC χ̄ A ΓB χ C , (C.1)
8 16
where

1
,
χA = ψ̂A + √ ΓAΛ
6 2
√
2 2 11
χ11 = − Γ Λ
3
11).
and A = (A,
(11)
Recall that the graviton EA
M and the gravitino ψA of 11D supergravity are related to
10D fields as [48]:

− 3 Aφ 2φ 2φ
=e
EA ,
EM 11 = e ,
E11 = e CM
E11 ,
3 3
M M
(11) 1
ψA = √ eφ/6 χA .
2π
Appendix D. Quartic couplings of ghosts and fermions
Below we collect terms in the 8D quantum action which are bilinear in FP ghosts and
bilinear in fermions:

π 1 ¯
t 8 e−2ξ (χ̄B ΓA χC + 2χ̄A ΓB χC ) b̂ ΓA ΓBC ĉ
(2)2
Sbc =
2 8
X
1 ¯
+ (χ̄B Γā χC + 2χ̄ā ΓB χC ) b̂Γā ΓBC ĉ
6
2
+ (χ̄A ΓBC χD ) b̂¯ ΓABCD ĉ + (χ̄ā ΓBC χD ) b̂¯ ΓāBCD ĉ
1
6 9
1 ¯ 4
− b̂ ΓA Γ ABCDE
+ Γā Γ āBCDE
ĉ(χ̄B ΓCD χE )
48 3
¯ 4 ¯
− ĉ¯ ΓB χA b̂ΓA χB − ĉ¯ ΓB χā b̂Γā χB
3
1 ¯ 11 ¯ A 4 ¯ ā
+ ĉΓ χ11 b̂Γ χA + b̂Γ χā
4 3
¯ A ā 4 ¯ ā D
+ LAā b̂Γ χ + LāD b̂Γ χ
3
1 ¯ A DE
+ LDE b̂ Γ Γ χA + Γ Γ χā ,
4 ā DE
(D.1)
4 3
where we now split indices as A = (A, ā), A = 0, . . . , 7, ā = (a, 11), a = 8, 9. Nonzero
components of LDE are given by
LAd̄ = −ĉ¯ ΓA χd̄ , La11 = −ĉ¯ Γa χ11 .

(2)2 A
Sbc is obtained by relating 8D gauge field ψ̂(8D) ˆ to 11D gravitino
(gauge parameter !)
(11)
ψA (gauge parameter ! (11) ) as

√ φ 1 √ φ
A
ψ̂(8D) = 2π e− 6 ψA(11) + ΓA Γā ψā(11) , !ˆ = 2π e 6 ! (11).
6
Let us also remind a standard fact that to keep the gauge

11 = 0,
EA m=0
EA
used in reduction from 11D one has to accompany supersymmetry transformations of [48]
with field dependent Lorentz transformations.
(2)2
The last line in the action Sbc originates from such Lorentz transformations.
(2)2
To write out Sbc in terms of 8D fields
A
ψ Σ
A
ψ̂(8D) := , Λ (8D) := ,
ηA Λ

l l˜
θ̂(8D) := , ν̂(8D) :=
µ µ̃
one should substitute

√
1 A 2 A
χ A
= ψ̂(8D)
+ Γ θ̂(8D) +
A
Γ Λ(8D) , A = 0, . . . , 7,
12 6
1 1 √
χ 8 = ν̂(8D) + Γ8 (θ̂(8D) + 2 Λ (8D)),
2 3
1 1 √
χ 9 = Γ89 ν̂(8D) + Γ9 (θ̂(8D) + 2 Λ (8D) ),
2 3
√ √
2 2 11 2
χ11 = − Γ Λ (8D) − θ̂(8D) .
3 4
(2)2
We do not present the final expression but we have checked that Sbc is T-duality
invariant.
Appendix E. Measures for path integrals
Here we explain why det ∆p are divided by Vp in (6.4) This is related to the integration
over zeromodes.
p
i
Introducing a basis a(p) , i = 1, . . . , bp , in HZ let us denote

ij j ij
Vp = a(p) i
∧ ∗a(p), Vp = det Vp . (E.1)
i,j
X
p
Note, that Vp is invariant under the choice of basis in HZ .
To explain integration over fermionic zeromodes let us consider the following path-
integral over fermionic p-forms u and v:
bp bp

Du Dv u v e−(v,∆p u) , (E.2)
i=1 γi j =1γj
where γi , i = 1, . . . , bp , is a basis of Hp (X, Z).

%p %p
In (E.2) we have inserted bi=1 γi u bj =1 γj v, to get nonzero answer, i.e., to saturate
fermion zeromodes.
To perform the integration in (E.2) we expand u and v in an orthonormal basis {ψn } of
eigen-p-forms of ∆p :

u= un ψn , v= vn ψn , (ψn , ψm ) = δn,m . (E.3)
n n
p
i
Let us choose the basis a(p) , i = 1, . . . , bp , of the lattice HZ , dual to the basis γi ∈
Hp (X, Z), i.e.,

j
a(p) = δij .
γi
Then, orthonormal zeromodes are expressed as

j i
i
ψzm = a(p) Wp−1 j , (E.4)
where Wp−1 is the inverse of the vielbein for the metric on HZ : (Vp )ij = (WpT Wp )ij .
p
Now, we integrate (E.2) and obtain

det ∆p
. (E.5)
Vp
In the case of bosonic p-forms u and v we do not need to insert anything to get a
nonzero answer:

det ∆p −1
Du Dv e−(u,∆p v) = , (E.6)
Vp
where (E.6) the integration over bosonic zeromodes was performed

bp
bp
j 1
Duizm Dvzm = j
= Vp . (E.7)
i=1 j =1 (deti,j γi ψzm )2
Appendix F. Super-K-theory theta function
Here we explain why Θ(F , ρ) defined in (1.28) is a super-theta function for a family of
principally polarized super-abelian varieties. To show this we use the results of [49], where
super-theta functions were studied.
A generic complex supertorus is defined as a quotient of the affine superspace with even
coordinates zi , i = 1, . . . , Neven , and odd coordinates ξa , a = 1, . . . , Nodd , by the action of
the abelian group generated by {λi , λi+Neven }
λi : zj → zj + δij , ξa → ξa , (F.1)
λi+Neven : zj → zj + (Ωeven )ij , ξa → ξa + (Ωodd )ia . (F.2)
We will restrict to the special case (Ωodd)ia = 0 relevant for our discussion. Let us
also assume that the reduced torus (obtained from the supertorus by forgetting all odd
coordinates) has a structure of a principally polarized abelian variety and denote its Kähler
form by ω.
It follows from the results of [49], that a complex line bundle L on the supertorus with
c1 (L) = ω has a unique section (up to constant multiple) iff Ωeven
T = Ωeven together with
the positivity of the imaginary part of the reduced matrix. This section is a supertheta
function.
Now we can find a family of principally polarized super-abelian varieties relevant to our
case simply by setting Neven = N and Nodd = Nferm.zm and by defining symmetric Ωeven
as
Re(Ωeven )ij = Re τK (xi , xj ), (F.3)

Im(Ωeven )ij = Im τK (xi , xj )

2

+ G2p (xi ) + G2p (xj ) ∧ ∗ˆ J2p (zm) + δij F (zm), (F.4)
p=0X
10
where xi , i = 1, . . . , N , is a basis of Γ1 . In (F.3) J2p (zm) is a 2p-form on X10 constructed

as a bilinear expression in fermion (and ghosts) zeromodes and F (zm) is a functional
quartic in fermion (and ghosts) zeromodes, both J2p (zm) and F (zm) can in principle be
found from the 10D fermion action (7.10), (D.1) as well as from the ghost action (7.35),
(7.40), (E.1). The modified characteristics α̂, β̂ and prefactor 2Φ(F
) in (1.28) all originate
from the shift of the imaginary part of the period matrix described in (F.4). It would be
very nice if one could formulate this super-abelian variety in a more natural way, without
reference to a Lagrangian splitting of ΓK .
References
[1] R. Minasian, G. Moore, K-theory and Ramond–Ramond charge, JHEP 9711 (1997) 002, hep-th/9710230.
[2] E. Witten, D-branes and K-theory, JHEP 9812 (1998) 019, hep-th/9810188.
[3] K. Olsen, R.J. Szabo, Constructing D-branes from K-theory, hep-th/9907140.
[4] E. Witten, Duality relations among topological effects in string theory, JHEP 0005 (2000) 031, hep-
th/9912086.
[5] G. Moore, E. Witten, Self-duality, Ramond–Ramond fields, and K-theory, JHEP 0005 (2000) 032, hep-
th/9912279.
[6] D.S. Freed, M.J. Hopkins, On Ramond–Ramond fields and K-theory, JHEP 0005 (2000) 044, hep-
th/0002027.
[7] D.E. Diaconescu, G. Moore, E. Witten, E8 gauge theory, and a derivation of K-theory from M-theory, hep-
th/0005090.
[8] D. Freed, Dirac charge quantization and generalized differential cohomology, hep-th/0011220.
[9] C.M. Hull, Massive string theories from M-theory and F-theory, JHEP 9811 (1998) 027, hep-th/9811021.
[10] C. Hull, P. Townsend, Unity of superstring dualities, Nucl. Phys. B 438 (1995) 109–137, hep-th/9410167.
[11] A. Giveon, M. Porrati, E. Rabinovici, Target space duality in string theory, hep-th/9401139.
[12] E. Bergshoeff, M. de Roo, M.B. Green, G. Papadopoulos, P.K. Townsend, Duality of II 7-branes and 8-
branes, hep-th/9601150.
[13] I. Lavrinenko, H. Lu, C. Pope, T. Tran, U-duality as general coordinate transformations, and spacetime
geometry, hep-th/9807006.
[14] M. Cvetic, H. Lu, C.N. Pope, K.S. Stelle, T-duality in the Green–Schwarz formalism, and the mass-
less/massive IIA duality map, Nucl. Phys. B 573 (2000) 149–176, hep-th/9907202.
[15] S. Sethi, C. Vafa, E. Witten, Constraints on low-dimensional string compactifications, Nucl. Phys. B 480
(1996) 213–224, hep-th/9606122.
[16] E. Witten, On S-duality in Abelian gauge theory, hep-th/9505186.
[17] E. Verlinde, Global aspects of electric–magnetic duality, Nucl. Phys. B 455 (1995) 211–228, hep-
th/9506011.
[18] E. Kiritsis, B. Pioline, On R 4 threshhold corrections in IIB string theory and (p, q) string instantons, Nucl.
Phys. B 508 (1997) 509;
B. Pioline, H. Nicolai, J. Plefka, A. Waldron, R 4 couplings, the fundamental membrane and exceptional
theta correspondences, hep-th/0102123.
[19] L. Romans, Massive IIA supergravity in ten dimensions, Phys. Lett. B 169 (1986) 374.
[20] E. Witten, Five-brane effective action in M-theory, J. Geom. Phys. 22 (1997) 103, hep-th/9610234.
[21] M.F. Atiyah, I.M. Singer, The index of elliptic operators: V, Ann. Math. 93 (1971) 139.
[22] V. Mathai, D. Stevenson, Chern character in twisted K-theory: equivariant and holomorphic cases, hep-
th/0201010.
[23] V. Mathai, R.B. Melrose, I.M. Singer, The index of projective families of elliptic operators,
math.DG/0206002.
[24] P. Meessen, T. Ortin, An Sl(2, Z) multiplet of nine-dimensional type II supergravity theories, Nucl. Phys.
B 541 (1999) 195–245, hep-th/9806120.
[25] K. Hori, D-branes, T-duality, and index theory, Adv. Theor. Math. Phys. 3 (1999) 281–342, hep-th/9902102.
[26] J. Igusa, Theta Functions, Springer-Verlag, Berlin, 1972.
[27] W. Siegel, Hidden ghosts, Phys. Lett. B 93 (1980) 170.
[28] J. Gegenborg, G. Kunstatter, The partition function for topological field theories, hep-th/9304016.
[29] N.D. Birrell, P.C.W. Davies, Quantum Fields in Curved Space, Cambridge Univ. Press, Cambridge, 1982.
[30] S. Deser, P. van Nieuwenhuizen, One-loop divergences of quantized Einstein–Maxwell fields, Phys. Rev.
D 10 (1974) 401.
[31] G. ’t Hooft, M. Veltman, One-loop divergences in the theory of gravitation, Ann. Inst. Henri Poincaré 20
(1974) 69.
[32] N. Nielsen, Ghost counting in supergravity, Nucl. Phys. B 140 (1978) 499.
[33] R. Kallosh, Modified Feynman rules in supergravity, Nucl. Phys. B 141 (1978) 141.
[34] K. Fujikawa, Path integral for gauge theories with fermions, Phys. Rev. D 21 (1980) 2848.
[35] C. Isham, C. Pope, N. Warner, Nowhere-vanishing spinors and triality rotations in 8-manifolds, Class.
Quantum Grav. 5 (1988) 1297.
[36] A. Adams, J. Evslin, The loop group of E8 and K-theory from 11d, hep-th/0203218.
[37] E. Witten, On flux quantization in M-theory and the effective action, J. Geom. Phys. 22 (1997) 1, hep-
th/9609122.
spin
[38] R. Stong, Calculation of Ω11 (K(Z, 4)), in: M. Green, D. Gross (Eds.), Unified String Theories, 1985 Santa
Barbara Proceedings, World Scientific, Singapore, 1986.
[39] D.E. Diaconescu, G. Moore, unpublished.
[40] G. Moore, Some comments on branes, G-flux, and K-theory, Int. J. Mod. Phys. A 16 (2001) 936–944,
hep-th/0012007.
[41] M. Atiyah, V. Patodi, I. Singer, Math. Proc. Cambridge Philos. Soc. 77 (1975) 43, 405.
[42] P. Bouwknegt, V. Mathai, D-branes, B-fields and twisted K-theory, JHEP 0003 (2000) 007, hep-th/0002023.
[43] E. Witten, Overview of K-theory applied to strings, hep-th/0007175.
[44] J. Maldacena, G. Moore, N. Seiberg, D-brane instantons and K-theory charges, JHEP 0111 (2001) 062,
hep-th/0108100.
[45] S. Stanciu, An illustrated guide to D-branes in SU(3), hep-th/0111221.
[46] S. Fredenhagen, V. Schomerus, Non-commutative geometry, gluon condensates, and K-theory, JHEP 0004
(2001) 007, hep-th/0012164.
[47] J. Maldacena, G. Moore, N. Seiberg, D-brane charges in five-brane backgrounds, JHEP 0110 (2001) 005,
hep-th/0108152.
[48] E. Cremmer, B. Julia, J. Scherk, Supergravity theory in 11 dimensions, Phys. Lett. B 76 (1978) 409.
[49] J. Rabin, M. Bergvelt, Super-curves, their Jacobians, and super-KP equations, alg-geom/9601012.
Where is the Higgs boson?

A. Aranda a,1 , C. Balázs b , J.L. Díaz-Cruz c
a Department of Physics, Boston University, Boston, MA 02215, USA
b Department of Physics, Florida State University, Tallahassee, FL 32306, USA
c Instituto de Fisica, BUAP, Puebla, Pue. 72570, Mexico
Received 20 December 2002; received in revised form 23 June 2003; accepted 28 July 2003
Abstract
Electroweak precision measurements indicate that the standard model Higgs boson is light and
that it could have already been discovered at LEP 2, or might be found at the Tevatron run 2. In the
context of a TeV−1 size extra-dimensional model, we argue that the Higgs boson production rates
at LEP and the Tevatron are suppressed, while they might be enhanced at the LHC or at CLIC. This
is due to the possible mixing between brane and bulk components of the Higgs boson, that is, the
non-trivial brane-bulk ‘location’ of the lightest Higgs. To parametrize this mixing, we consider two
Higgs doublets, one confined to the usual space dimensions and the other propagating in the bulk.
Calculating the production and decay rates for the lightest Higgs boson, we find that compared to
the standard model (SM), the cross section receives a suppression well below but an enhancement
close to and above the compactification scale Mc . This impacts the discovery of the lightest (SM
like) Higgs
√ boson at colliders. To find a Higgs signal in this model at the Tevatron run 2 or at the
LC with s = 1.5 TeV, a√ higher luminosity would be required than in the SM case. Meanwhile, at
the LHC or at CLIC with s ∼ 3–5 TeV one might find highly enhanced production rates. This will
enable the latter experiments to distinguish between the extra-dimensional and the SM for Mc up to
about 6 TeV.
1. Introduction
The Higgs boson is the missing link connecting the real world with the unified elec-
troweak (EW) gauge group by spontaneously breaking the latter. Precision measurements
E-mail addresses: fefo@cgic.ucol.mx (A. Aranda), balazs@hep.fsu.edu (C. Balázs),

ldiaz@sirio.ifuap.buap.mx (J.L. Díaz-Cruz).
1 Present address: Facultad de Ciencias, Universidad de Colima, México.
doi:10.1016/j.nuclphysb.2003.07.023
A. Aranda et al. / Nuclear Physics B 670 (2003) 90–102 91
of EW observables constrain the Higgs mass below about 200 GeV at 95% C.L. [1–8]
within the standard model (SM). Thus, it is expected that a Higgs particle will be dis-
covered at the run 2 of the Tevatron, provided sufficient luminosity [9]. But it is intrigu-
ing to notice that the EW observables strongly prefer a SM like Higgs with mass below
114.1 GeV [5,7], which is the present lower limit from LEP 2. The data also indicate that
the Higgs boson should have already been discovered [5], and the fact that it was not found
can be interpreted as new physics crucially affecting the Higgs sector [7]. In this work
we put forward a model in which the presently missing signal of the lightest Higgs boson
is due to a suppression of the Higgs production cross section at LEP and the Tevatron.
This suppression arises from the non-trivial ‘location’ of the lightest Higgs boson in a five-
dimensional space. However, the same feature promises enhancement of the Higgs signal
at the CERN Large Hadron Collider (LHC) and possibly at a multi-TeV linear collider
(CLIC).
The idea that our universe could be confined to a higher-dimensional defect has been
revived both in field theory [10] and string contexts [11]. It has been laid on more solid
ground in the context of non-perturbative string analyses [12–14], and applied as a possible
solution to the gauge hierarchy problem [15–18]. Such a solution relies on the existence
of n > 0 additional compact space-like dimensions. In models based on this idea, the
four-dimensional Planck scale MPl becomes an effective quantity and it is related to the
fundamental scale M by the volume of the extra space Vn via the relation MPl 2 = M n+2 V .
n
1/n
If one requires M = O(TeV) then for n = 2, the compactification radius R ∼ Vn is in the
order of a millimeter, but for n = 7 it is less than a fermi, not far from the inverse of a TeV.
It is remarkable to notice that with O(TeV−1 ) size extra dimensions the hierarchy problem
is indeed nullified, since the fundamental scales M ∼ 1/R are close to TeV.
The string arguments of Refs. [12,14] also allow the standard gauge and Higgs sectors
to penetrate the bulk. If the compactification scale Mc = 1/R is higher than O(TeV),
phenomenology does not conflict with this scenario either [19–24]. This makes the inverse
TeV size extra-dimensional models attractive. Alternatively, the hierarchy problem can be
solved with the use a non-factorizable geometry, which has also been proposed in five
dimensions. The introduction of an exponential ‘warp’ factor reduces all mass parameters
of the order of a fundamental MPl of a distant brane to TeV’s on the brane where we
live [25]. In order to explain large hierarchies among energy scales one simply has to
explain small distances along the extra dimension, thus this mechanism requires a Planck
scale size extra dimension [25,26].
These intriguing possibilities have opened a new window for the exploration of physics
beyond the SM [27–29], in particular, the phenomenology of the Higgs sector. In Ref. [30]
it is shown that it is possible to obtain electroweak symmetry breaking in an extra-
dimensional scenario even in the absence of tree-level Higgs self interactions. Also,
in Ref. [31] we find scenarios in which the radion in the Randall–Sundrum model is
contrasted with the SM Higgs boson. Studying several models that lead to a universal
rate suppression of Higgs boson observables, Ref. [32] concluded that the Tevatron and
LHC will have difficulty finding evidence for extra-dimensional effects. Yet another study
of universal extra dimensions [33] conjectures that a suppression of the Higgs rates
occurs [34].
92 A. Aranda et al. / Nuclear Physics B 670 (2003) 90–102
However, just as in the SM and other four-dimensional theories, the Higgs sector
remains the least constrained, since it can live either on the brane or in the bulk, each choice
being phenomenologically consistent. One way to parametrize this freedom is to consider
an extra-dimensional two Higgs doublet model (XD THDM) where one doublet lives in
the bulk while the other is confined to the brane. The lightest Higgs boson state, which
will resemble the SM one, will then be a linear combination of the neutral components of
the two doublets. Constraints from electroweak precision data have been applied to such a
model, and it was found that the compactification scale is larger than a couple of TeVs [35].
In this paper, we study the ability of present and future colliders to find the lightest
Higgs boson in a XD THDM with a single TeV−1 size new dimension. In particular, our
aim is to estimate the minimal size of the compact dimension for which the lightest Higgs
signal is distinguishable from that of the SM (or THDM) at the LHC or at CLIC.
We assume that the SM gauge bosons and one of the Higgs doublets propagate in this
compact dimension. The SM W ± and Z particles are identified with the zero modes of the
five-dimensional gauge boson fields. There is a second Higgs field restricted to the brane
together with all the matter fields of the SM. Although the Higgs spectrum includes two
CP-even states (h, H , with mh < mH ), one CP-odd Higgs (A) and a charged pair (H ± ), in
this work we focus on the lightest Higgs boson h, because most likely this will be the first
Higgs state that future colliders will detect. The CP-even Higgses may be the combinations,
i.e., brane-bulk mixed states, of the two Higgs doublets.
We derive the Lagrangian for Higgs interactions and apply it to calculate the cross
section of the associated production of Higgs with gauge bosons at the LC, as well as
the dominant Higgs decays including the possible contributions from virtual KK states.
Crucial to our approach is the cumulative effect of the virtual gauge boson KK states,
W ±(n) and Z (n) , which contribute to the cross section for the production of the Higgs
associated with the W ± and Z and to the three-body decay h → Vf f¯ (V = W ±(0) , Z (0) ).
The corresponding reactions at hadron colliders are studied as well.
We remark that in this scenario, with a low enough compactification scale, the discovery
of the extra dimension would probably precede the discovery of the lightest Higgs boson.
Gauge bosons propagating in the new dimension would exhibit unambiguous resonances,
for example, in their s-channel production at the LHC. Our focus is on the lightest Higgs
because we investigate how much information the various colliders can give us about the
Higgs sector with a second Higgs doublet in the bulk. While more exotic processes might
provide more useful to this end, we restrict ourselves to V h production (V = W ± or Z)
because our emphasis is that this dominant search channel is suppressed at the Tevatron
(and at LEP), while possibly enhanced at future colliders.
The organization of our paper is as follows. In Section 2, we present the model
that we use to study the brane-bulk mixing of the Higgs boson. Then in Section 3, we
derive formulae for the Higgs production and decays. These include the evaluation of the
contribution from virtual Z (n) KK states to the associated production at linear colliders, i.e.,
e+ e− → hZ (0) , as well as to the three-body decay h → Vf f¯ . Results of our calculation
for the hadron collider case are also included. The discussion of the implications of our
results for present and future colliders appears in Section 4, while the conclusions are
presented in Section 5.
2. Modeling the Higgs location
To describe the brane-bulk Higgs mixing, we work with a five-dimensional (5D)

extension of the SM that contains two Higgs doublets. The SM fermions and one Higgs
doublet (Φu ) live on a 4D boundary, the brane, while the gauge bosons and the second
Higgs doublet (Φd ), are all allowed to propagate in the bulk. The constraints from
electroweak precision data [35] show that the compactification scale can be of O(TeV)
(3–4 TeV at 95% C.L.). The relevant terms of the 5D SU(2) × U (1) gauge and Higgs
Lagrangian are given by
1 a 2 1
L5 = − FMN − (BMN )2 + |DM Φd |2 + |Dµ Φu |2 δ x 5 , (1)
4 4
where the Lorentz indices M and N run from 0 to 4, and µ runs from 0 to 3. The covariant
derivative is given by
Y σa
DM = ∂M + ig5 BM + ig5 AaM . (2)
2 2
Given this definition, the mass dimensions of the fields are: dim(Φd ) = 3/2, dim(Φu ) = 1,
dim(AM ) = 3/2, dim(BM ) = 3/2, and the 5D gauge couplings have a mass dimension of
−1/2. Bulk fields are defined to have even parity under x 5 → −x 5 , and are expanded as
∞ 5
µ 5 1 √ nx
S x ,x = √ S (0) x µ + 2 cos S (n) x µ . (3)
πR R
n=1
This decomposition, together with Eq. (1), guarantees that after compactification we obtain
the usual 4D kinetic terms for all fields.
Spontaneous symmetry breaking (SSB) of EW symmetry occurs when the Higgs
doublets acquire vacuum expectation values (vevs). After SSB the Higgs fields on the brane
can be written as

1 Φu0∗ 1 vu + h cos α + H sin α + i cos βA
Φu = √ = √ , (4)
2 Φu− 2 cos βH −
+
(0) 1 Φd 1 sin βH +
Φd = √ = √ , (5)
2 Φd0 2 vd − h sin α + H cos α + i sin βA
where the neutral CP-even bosons are denoted by h and H , and h is identified with the
lightest Higgs: mh < mH . The mixing angle α is introduced to diagonalize the CP-even
mass matrix. The CP-odd and charged Higgs fields are denoted by A and H ± , and vu and
vd are the vevs of Φu and Φd(0) , respectively. Note that the angles α and β = arctan(vu /vd )
parametrize what we call brane-bulk mixing, or higher-dimensional ‘location’, of the
neutral Higgses.
After performing the KK-mode expansion and identifying the physical states, one
derives the interaction Lagrangian for all the vertices of the neutral and charged Higgses.
In particular, the interactions ZZh and ZZ (n) h, which are necessary to calculate the Higgs
production in association with a Z, as well as the vertices involving the W ± bosons, are
given by the following 4D Lagrangian:

gMZ
L4 ⊃ h sin(β − α) + H cos(β − α) Zµ Z µ
2cW
√ gMZ ∞
+ 2 (h sin β cos α + H sin β sin α) Zµ(n) Z µ
cW
n=1

+ gMW h sin(β − α) + H cos(β − α) Wµ+ W −µ
√ ∞
+ −(n)µ
+ 2 gMW (h sin β cos α + H sin β sin α) Wµ W + Wµ− W +(n)µ .
n=1
(6)
Thus, the vertices hZZ and hW W have the same form as in the usual 4D THDM,
i.e., proportional to sin(β − α). Meanwhile the couplings hZZ (n) and hW ± W ±(n) are
proportional to sin β cos α, vanishing either when β = 0 or α = π/2, i.e., either when
EWSB is driven exclusively by the vev of Φd(0) or when the CP-even Higgs comes
entirely from Φd(0) . Similarly, the couplings of the CP-odd Higgs A and the charged Higgs
resemble the THDM, although new vertices of the type H + W − Z (n) or H + W −(n) Z could
be induced.
On the other hand, because the fermions are confined to the brane, the Higgs-fermion
couplings could take any of the THDM I, II or III versions [36,37]. However, for the
THDM III version the possible FCNC problems √ would be ameliorated, as the bulk-brane
couplings will be suppressed by the factor 1/ 2πR [38]. Thus, for the flavor conserving
couplings one can use the formulae of the widely studied THDM II, which appears, for
instance, in the Higgs hunters guide [39].
3. Extra-dimensional contribution to Higgs production and decays
3.1. Associated production h + Z at linear colliders
In order to study Higgs production at future colliders, first we derive the cross section
for the Bjorken process, namely, for e+ e− → hZ. The total amplitude includes the
contribution of virtual Z = Z (0) and Z (n) states in the s-channel. The sum over all KK
modes can be performed analytically, which considerably simplifies the final expression.
Our result for the cross section is given by

σ e+ e− → hZ = σSM FXD (α, β, s). (7)
Here σSM denotes the SM cross section, given by

G2 M 4 1 |k| (3MZ2 + |k|2 )
σSM = F Z 4sw4 − 2sw2 + √ , (8)
3π 2 s (s − MZ2 )
with
2 1/2
1 s + MZ2 − m2h
|k| = √ − sMZ2 (9)
s 2
being the 3-momentum of the Z boson.

The extra-dimensional contribution is factorized into
2
FXD (α, β, s) = sin(β − α) + 2 cos α sin βFKK (s) . (10)
The function FKK , which arises after summing over all the virtual KK-modes, is given by
∞
s − M2
FKK (s) = 2 Z
= RA(s)π cot RA(s)π − 1, (11)
s − Mn2
n=1
where

Mn = n2 /R 2 + MZ2 , (12)
is the mass of the nth KK level. When neglecting the widths,

Γn = Mn αg vf2 + af2 /3, (13)
of the KK resonances,2 we can write

A(s) = s − MZ2 . (14)

(Here vf and af are the SM vector and axial coupling strength of the vector boson to
fermions.) This is a reasonable approximation, since αg = g 2 /(4π) 1. By neglecting
only MZ αg terms (next to MZ ), we can even include the dominant width effect by setting

A(s) = c s − cMZ2 /c, (15)
where c = (1 + 2iαg ).
Recently it was pointed out that summing over a large number of KK resonances may
jeopardize the unitarity of standard-like extra-dimensional models [40]. Thus, we mention
that the sum in Eq. (11) can also be performed analytically for a finite number of terms
with the result:

N
s − M2
FKK (s) = 2 Z
= RA(s) π cot RA(s)π − HN−RA(s) + HN+RA(s) − 1,
s − Mn2
n=1
(16)
where Hx is the harmonic number function. (The KK width can also be included as
above.) Since the FKK function can be calculated even analytically, we can easily check its
sensitivity to the number of KK levels and the inclusion of KK width. Fig. 1 shows that, for
a typical set of parameters, FKK is reasonably insensitive to these, which ensures that our
later results are robust against cutoff and width effects of KK levels in the relevant energy
range.
These expressions also apply for the process q q̄ → W ± h after changing σSM and MZ
to MW at the appropriate places.
2 The width of the nth KK state is defined as its total decay rate into a SM fermion pair.
√
Fig. 1. The extra-dimensional contribution FXD to e+ e− → Zh, as the function of the center of mass energy s
for various other parameter values.
3.2. Associated production h + Z, h + W ± at hadron colliders
When considering Higgs production at hadron colliders, an expression similar to Eq. (7)
holds at the parton level for the production cross section of the Higgs in association with
a W ± or Z. To obtain the hadronic cross section h1 h2 → hZ, the partonic cross section
must be convoluted with the parton distribution functions (PDFs):

1 1
σ (h1 h2 → Zh) = fq/ h1 (x1 , ŝ)σ (q q̄ → Zh)fq̄/ h2 (x2 , ŝ) dx1 dx2 + q ↔ q̄.
q q̄ 0 0
(17)
Here fq/ hi (x, ŝ) gives the distribution of a parton q in the hadron hi as a function of the
longitudinal momentum fraction x and the factorization scale, which is chosen to be the
partonic center of mass ŝ. In our numeric study we use CTEQ4M PDFs [41]. The sum in
Eq. (17) extends over the light quark flavors q = u, d, s, c.
The large center of mass energy that can be achieved at the LHC also opens up the
possibility to produce a Higgs boson in association with KK states, for instance, hZ (1) ,
which will have a very distinctive signature that could allow ‘direct’ detection of the first
KK modes at the LHC. This possibility is studied elsewhere.
3.3. Higgs decays
For Higgs bosons lying in the intermediate mass range, which is in fact favored by
the analysis of electroweak radiative corrections, the dominant decay is into bb̄ pairs. In
our higher-dimensional model this decay width is given by the formulae of the THDM,
just as that of the other tree-level two-body modes. On the other hand, for the three-
body decays h → W lνl and h → Zl + l − , which can play a relevant role at the Tevatron
and LHC, the corresponding decay width could receive additional contributions from the
virtual KK states. The inclusion of these KK modes leads to the following expression for
the differential decay width:
dΓ g 4 mh (x 2 − 4rw )1/2 2
(h → W l ν̄l ) = fV (x) sin(β − α) + 2 cos α sin βFKK ,
dx 3072π 3 1−x
(18)
1/2
where fV (x) = x 2 − 12rw x + 8rw + 12rw2 , with rw = MW 2 /m2 , and 2r
h w < x < 1 + rw .
The FKK function is given as in Eq. (11), with the replacements s → q 2 = m2h (1 − x) and
MZ → MW . A similar expression can be derived for the decay h → Zl + l − .
To study the effect of the KK modes on the decay h → W l ν̄l , we have evaluated the ratio
of the corresponding decay width in the extra-dimensional scenario over the SM decay
width:
Γ (h → W l ν̄l )XD
RhW W ∗ = . (19)
Γ (h → W l ν̄l )SM
Results for this ratio are shown in Table 1, for several representative sets of parameters
which are chosen as
A: Mc = 2 TeV, α = π/3, B: Mc = 2 TeV, α = π/1.28,
C: Mc = 2 TeV, α = π, D: Mc = 5 TeV, α = π,
while β = π/4 remains fixed. We observe that significant deviations from the SM can
appear, although this effect is largely due to the difference of the Higgs couplings in the
THDM and the SM.
On the other hand, the loop induced decays h → γ γ , Zγ also receive contributions
from the W (n) KK modes. But since the coupling hW W (n) that appears in the loop is
proportional to MW , rather than MW (n) , the contribution of the KK states will decouple,
as there are no mass factors that could cancel the ones in the numerator. Thus, the KK
contribution can be neglected for the decay widths of the loop induced decays.
We conclude that in the intermediate Higgs mass range the decay h → bb̄ will continue
to dominate, even more than in the SM case for some values of parameters. In the limit
Mc O(1) TeV, one recovers the SM pattern for the Higgs decays.
Table 1
The ratio RhW W ∗ , introduced in Eq. (19), for several sets of parameters A, B, C, D (as defined in the text), and
with β = π/4
mh (GeV) RhW W ∗ (set A) RhW W ∗ (set B) RhW W ∗ (set C) RhW W ∗ (set D)
130 6.0 × 10−2 0.999 0.51 0.50
140 6.0 × 10−2 0.998 0.51 0.50
150 6.1 × 10−2 0.996 0.50 0.50
160 6.1 × 10−2 0.993 0.50 0.50
4. Implications for Higgs searches at future colliders
4.1. The LC case
Because of its simplicity, first we discuss Higgs production at a linear collider. The
present bound on the compactification scale is Mc 3.8 TeV [42], for the cases when the
Higgs field is either in the bulk or confined to the brane. For the general case of “mixing”
one obtains similar bounds [43].
The results for the e+ e− → hZ cross section, after the inclusion of the virtual Z KK
contribution, are shown in Fig.√2. This plot shows the cross section as a function of the
center of mass energy (200 < s < 4000 GeV), for mh = 120 GeV and a value of the
compactification scale Mc = 4 TeV. We plot the SM cross section σSM , the THDM cross
section σTHDM , as well as the extra-dimensional cross section σXD . Three different pairs
of α and β were chosen to compute σTHDM and σXD . They are given by
a: β = π/4, α = −π/4, b: β = π/4, α = 0,
c: β = π/4, α = 0.241π.
Choice a corresponds to a case in which σTHDM = σSM ; b corresponds to a case where the
effect of the mixing due to α in the term involving the KK sum is maximum; c corresponds
√
to values of α and β for which σXD /σTHDM is larger than 1 above s = 2 TeV.
The first frame of Fig. 2 shows well that the shape of the cross section is determined by
the product of FKK , as shown in Fig. 1, and the SM cross section.
We can see that in all three cases the cross section of the XD model is always smaller
√
than that of the SM at s = 500, 1000 GeV, and that in order to obtain a larger cross
√
section, one needs energies greater than s ∼ 2 TeV. This is understood from the fact
that the heavier KK modes, through their propagators, interfere destructively with the SM
amplitude thus reducing the cross section. We mention that the three cases presented in
Fig. 2 are only representative, and that one can find broad regions of parameter space in
which σXD > σTHDM .
Moreover, as Fig. 2 shows, once the center of mass energy approaches the threshold for
the production of the first KK state, the cross section starts growing. For instance, with
√
Mc = 4 TeV, σSM σXD for s 2 TeV. However, one would need higher energies in
order to have a cross section larger than that of the SM, which may only be possible at
CLIC [45].
According to current studies, when the cross section is 4% larger than the SM cross
section, with the estimated precision that could be obtained at the LC [44], it may be
possible to distinguish between the SM and XD Higgs scenarios. We can see that this
might be possible at CLIC for Mc = 3–4 TeV for a broad range of parameter values. It is
interesting to note that such deviations in the cross section from the SM prediction arise
even when the couplings of the Higgs to the gauge bosons are indistinguishable from the
SM couplings.
Fig. 2. SM, THDM and XD cross sections for e+ e− → hZ. Each plot corresponds to a different set of values for
α and β all with mh = 120 GeV and with a compactification scale Mc = 4 TeV.
4.2. Implications for the Tevatron
After the productive but unsuccessful Higgs search at LEP 2, the run 2 of the Tevatron
continues the search until the LHC starts operating. The luminosity that is required to
achieve a 5 or 3σ discovery, or a 95% C.L. exclusion limit, was presented by the run 2
Higgs working group [9]. For instance, with mh = 120 GeV, the corresponding numbers
are about 20, 6 and 2 fb−1 , respectively.
Assuming Mc 2 TeV, the inclusion of the KK modes decreases the hW ± and hZ
associated production cross section at the Tevatron. Depending on the actual values of Mc ,
α and β, the suppression in the parton level cross section may be anything between 1 and
99 percent. For example, if Mc is a few TeV then the s-channel process q q̄ → W ± h can
receive a considerable suppression, as it can be inferred from Fig. 2. This is so unless
the KK contribution itself is suppressed by cos α and/or sin β in Eq. (7), in which case
the presented model has little relevance. Thus, as a general prediction of this model, we
Fig. 3. Higgs production cross section in association with a Z boson at the LHC as a function of the
compactification scale for selected values of the mixing parameters.
conclude that more than the above listed luminosity is required to find a light Higgs boson.
This slims the chances of the Tevatron to find the Higgs of this model.
4.3. Higgs production at the LHC
The Higgs discovery potential in this model is more promising at the LHC. We illustrate
this in Fig. 3, showing the pp → hZ differential cross section as the function of the hZ
invariant mass MhZ . The typical resonance structure displayed by Figs. 1 and 2 is preserved
by the hadronic cross section. The resonance peak is well pronounced when MhZ ∼ Mc .
This leads to a large enhancement over the SM (or THDM) cross section.
The singularity at Mc = MhZ is regulated by the width of the KK mode, which is
included in our calculation as given by Eq. (13). Thus, Fig. 3 gives a reliable prediction
of the XD cross section even in the peak regions. Depending on the particular values of
α and β the enhancement is more or less pronounced. For an optimistic set α = 0 and
β = π/2, the XD production cross section is considerably enhanced compared to the SM at
MhZ = Mc . This enhancement may be detectable up to about Mc = 6 TeV. We estimate that
with 100 fb−1 for Mc = 6 TeV there are about 20 hZ events in the bins around MhZ ∼ Mc .
As Fig. 3 shows, in the SM less than one event is expected in the same MhZ range. It
is needless to say that similar results hold for pp → hW ± , which further enhances the
discovery prospects.
Based on these results, we conclude that in the Bjorken process alone the reach of the
LHC may extend to about Mc = 6 TeV, depending on the values of α and β. Finally, we
note that the XD contribution to the running of the gauge couplings is important when the
effective center of mass energy of the collider is close to 1/R [23,24]. Since not
√ included
in this work, this contribution is expected to change our results somewhat for ŝ Mc .
5. Conclusions
In this work, we studied the extent to which present and future colliders can probe the
brane-bulk location of the Higgs boson in a model with a TeV−1 size extra dimension. In
this model one Higgs doublet is located on the brane while another one propagates in the
bulk. We found that the virtual KK states of the gauge bosons contribute
√ to the associated
production of the Higgs with W ±(n) and Z (n) , and at low energies ( s < 1/R GeV) the
cross
√ section is suppressed compared to the SM case. Meanwhile at higher energies, i.e.,
at s ∼ 1/R, the cross section can receive an enhancement that has important effects on
the discovery of the Higgs at future colliders.
Analysing compactification scales in the range of 2–8 TeV, we √concluded that to find a
Higgs signal in this model the Tevatron run 2 and the LC with s = 500–1500 GeV are
required to have
√ a luminosity higher than in the SM case. Meanwhile, the LHC and possibly
CLIC with s ∼ 3–5 TeV might have a greater potential to find and study a Higgs signal.
Depending on the model parameters, these colliders may be able to distinguish between
the extra-dimensional and the SM for compactification scales up to about 6 TeV. If this
model is relevant for weak scale physics, the LHC should see large enhancements in the
associated production rates. Thus, not finding the Higgs at the Tevatron may be good news
for the XD Higgs search at the LHC.
Acknowledgements
A.A. was supported by the US Department of Energy under grant DE-FG02-

91ER40676. C.B. was also supported by the DOE, under contract number DE-FG02-
97ER41022. J.L.D.-C. was supported by CONACYT and SNI (México).
References
[1] T. Takeuchi, W. Loinaz, N. Okamura, L.C. Wijewardhana, hep-ph/0304203.

[2] M.S. Chanowitz, hep-ph/0304199.
[3] P. Langacker, hep-ph/0304186.
[4] J. Erler, hep-ph/0212272.
[5] P. Langacker, J. Phys. G 29 (2003) 1.
[6] S. Villa, hep-ph/0209359.
[7] M.S. Chanowitz, Phys. Rev. D 66 (2002) 073002.
[8] H.J. He, Y.P. Kuang, C.P. Yuan, B. Zhang, hep-ph/0211229.
[9] M. Carena, et al., hep-ph/0010338.
[10] V.A. Rubakov, M.E. Shaposhnikov, Phys. Lett. B 125 (1983) 136.
[11] I. Antoniadis, Phys. Lett. B 246 (1990) 377.
[12] J. Polchinski, Phys. Rev. Lett. 75 (1995) 4724.
[13] J.D. Lykken, Phys. Rev. D 54 (1996) 3693.
[14] P. Horava, E. Witten, Nucl. Phys. B 475 (1996) 94.

[15] I. Antoniadis, S. Dimopoulos, G.R. Dvali, Nucl. Phys. B 516 (1998) 70.
[16] N. Arkani-Hamed, S. Dimopoulos, G.R. Dvali, Phys. Lett. B 429 (1998) 263.
[17] I. Antoniadis, N. Arkani-Hamed, S. Dimopoulos, G.R. Dvali, Phys. Lett. B 436 (1998) 257.
[18] N. Arkani-Hamed, S. Dimopoulos, J. March-Russell, Phys. Rev. D 63 (2001) 064020.
[19] A. Pomarol, M. Quiros, Phys. Lett. B 438 (1998) 255.
[20] I. Antoniadis, S. Dimopoulos, A. Pomarol, M. Quiros, Nucl. Phys. B 544 (1999) 503.
[21] A. Delgado, A. Pomarol, M. Quiros, Phys. Rev. D 60 (1999) 095008.
[22] C.D. Carone, Phys. Rev. D 61 (2000) 015008.
[23] K.R. Dienes, E. Dudas, T. Gherghetta, Phys. Lett. B 436 (1998) 55.
[24] K.R. Dienes, E. Dudas, T. Gherghetta, Nucl. Phys. B 537 (1999) 47.
[25] L. Randall, R. Sundrum, Phys. Rev. Lett. 83 (1999) 3370.
[26] N. Arkani-Hamed, S. Dimopoulos, G.R. Dvali, N. Kaloper, Phys. Rev. Lett. 84 (2000) 586.
[27] G.F. Giudice, R. Rattazzi, J.D. Wells, Nucl. Phys. B 544 (1999) 3.
[28] T. Han, J.D. Lykken, R. Zhang, Phys. Rev. D 59 (1999) 105006.
[29] For a list of experimental bounds see:
Y. Uehara, Mod. Phys. Lett. A 17 (2002) 1551;
N. Arkani-Hamed, S. Dimopoulos, G. Dvali, Phys. Rev. D 59 (1999) 086004;
E. Mirabelli, M. Perelstein, M. Peskin, Phys. Rev. Lett. 82 (1999) 2236;
C. Balazs, et al., Phys. Rev. Lett. 83 (1999) 2112;
J.L. Hewett, Phys. Rev. Lett. 82 (1999) 4765;
T.G. Rizzo, Phys. Rev. D 59 (1999) 115010;
K. Aghase, N.G. Deshpande, hep-ph/9902263;
L.J. Hall, D. Smith, Phys. Rev. D 60 (1999) 085008;
S. Cullen, M. Perelstein, Phys. Rev. Lett. 83 (1999) 268;
P. Nath, Y. Yamada, M. Yamaguchi, Phys. Lett. B 466 (1999) 100;
P. Nath, M. Yamaguchi, Phys. Rev. D 60 (1999) 116006;
M.L. Graesser, Phys. Rev. D 61 (2000) 074019.
[30] B. Grzadkowski, J.F. Gunion, Phys. Lett. B 473 (2000) 50.
[31] G.F. Giudice, R. Rattazzi, J.D. Wells, Nucl. Phys. B 595 (2001) 250;
D. Dominici, B. Grzadkowski, J.F. Gunion, M. Toharia, hep-ph/0206197;
J.L. Hewett, T.G. Rizzo, hep-ph/0202155.
[32] J.D. Wells, hep-ph/0205328.
[33] T. Appelquist, B.A. Dobrescu, Phys. Lett. B 516 (2001) 85.
[34] F.J. Petriello, JHEP 0205 (2002) 003.
[35] M. Masip, A. Pomarol, Phys. Rev. D 60 (1999) 096005.
[36] J.L. Diaz-Cruz, G. Lopez Castro, Phys. Lett. B 301 (1993) 405.
[37] J.L. Diaz-Cruz, J.J. Toscano, Phys. Rev. D 62 (2000) 116005.
[38] Y. Sakamura, hep-ph/9912511.
[39] J.F. Gunion, H.E. Haber, G.L. Kane, S. Dawson, The Higgs Hunter’s Guide, SCIPP-89/13.
[40] R.S. Chivukula, D.A. Dicus, H.J. He, S. Nandi, Phys. Lett. B 562 (2003) 109, hep-ph/0302263.
[41] H.L. Lai, et al., Phys. Rev. D 55 (1997) 1280.
[42] T.G. Rizzo, J.D. Wells, Phys. Rev. D 61 (2000) 016007, hep-ph/9906234.
[43] A. Muck, A. Pilaftsis, R. Ruckl, hep-ph/0209371.
[44] M. Battaglia, K. Desch, hep-ph/0101165.
[45] R.W. Assmann, et al., SLAC-REPRINT-2000-096.
Four-point correlators of BPS operators

in N = 4 SYM at order g 4
G. Arutyunov a , S. Penati b , A. Santambrogio c , E. Sokatchev d
a Max-Planck-Institut für Gravitationsphysik, Albert-Einstein-Institut, Am Mühlenberg 1,
D-14476 Golm, Germany
b Dipartimento di Fisica, Università degli studi di Milano-Bicocca and INFN, Sezione di Milano,
piazza della Scienza 3, I-20126 Milano, Italy
c Dipartimento di Fisica, Università degli studi di Milano and INFN, Sezione di Milano,
via Celoria 16, I-20133 Milano, Italy
d Laboratoire d’Annecy-le-Vieux de Physique Théorique 1 LAPTH, B.P. 110, F-74941 Annecy-le-Vieux, France
Abstract
We study the large N degeneracy in the structure of the four-point amplitudes of 12 -BPS operators
of arbitrary weight k in perturbative N = 4 SYM theory. At one loop (order g 2 ) this degeneracy
manifests itself in a smaller number of independent conformal invariant functions describing the
amplitude, compared to AdS5 supergravity results. To study this phenomenon at the two-loop level
(order g 4 ) we consider a particular N = 2 hypermultiplet projection of the general N = 4 amplitude.
Using the formalism of N = 2 harmonic superspace we then explicitly compute this four-point
correlator at two loops and identify the corresponding conformal invariant functions.
In the cases of 12 -BPS operators of weight k = 3 and k = 4 the one-loop large N degeneracy is
lifted by the two-loop corrections. However, for weight k > 4 the degeneracy is still there at the
two-loop level. This behavior suggests that for a given weight k the degeneracy will be removed if
perturbative corrections of sufficiently high order are taken into account. These results are in accord
with the AdS/CFT duality conjecture.
PACS: 03.70.+k; 11.15.-q; 11.10.-z; 11.30.Pb; 11.30.Rd
Keywords: AdS/CFT; SYM theory; Anomalous dimensions; Superspace
E-mail address: agleb@aei-potsdam.mpg.de (G. Arutyunov).

1 UMR 5108 associée à l’Université de Savoie.
doi:10.1016/j.nuclphysb.2003.07.027
104 G. Arutyunov et al. / Nuclear Physics B 670 (2003) 103–147
1. Introduction
In recent years the supersymmetric quantum conformal field theories in dimensions

higher than two have undergone a renaissance due to the discovery of the AdS/CFT
correspondence [1]. In the first place this concerns N = 4 super-Yang–Mills (SYM) theory
in four dimensions whose conjectured dual is type IIB (superstring) supergravity on an
anti-de Sitter background.
On the string theory side the present investigations are focused either on low-energy
supergravity or on semiclassical string solutions carrying nonvanishing global charges of
the superconformal group. In the latter case a direct comparison of the corresponding string
energies to the spectrum of scaling dimensions of certain composite operators in N = 4
SYM is possible [2–4].
According to the duality conjecture, AdS supergravity is dual to the gauge theory taken
at an infinite value of the ’t Hooft coupling λ and in the large N limit, where N is the rank
of the gauge group (S)U(N ). In this limit the gauge theory reveals the following peculiar
behavior:
• the sector of protected operators is enlarged;

• the spectrum of anomalous dimensions degenerates;
• the symmetry is enhanced.
The first property is a simple consequence of the rather general fact that in the large N
limit the anomalous dimension of a multi-trace operator made of single-trace constituents
is equal to the sum of their individual anomalous dimensions. In this sense, in the large N
limit the protected operators form a ring.2
The second property is more subtle and it is reminiscent of the free theory pattern, where
usually several operators with the same canonical dimension coexist. It is well known that
protected (for instance, 12 -BPS) operators with the same quantum numbers can be realized
in the gauge theory both as single- and multi-trace operators. Thus, multiplying protected
operators with the same quantum numbers by an unprotected operator, we obtain different
operators with the same dimension in the planar limit, hence the degenerate spectrum.
Such a mechanism is essentially due to the conformal supersymmetry responsible for the
protection and, again, due to the large N factorization property of the Feynman graphs.
However, this is not the whole story. As was recently shown in Ref. [5], there exist, for
instance, single-trace operators (which cannot be obtained as the product of a protected and
an unprotected operators) whose anomalous dimensions coincide up to two loops. This is
evidence for an extra symmetry in the planar limit which is responsible for this degeneracy
of the perturbative spectrum. The symmetry enhancement [5] is an interesting property of
the gauge theory related, for instance, to the integrability of the spin chain Hamiltonian [6].
The latter can be identified with the one-loop dilatation operator in N = 4 SYM.
2 When computing the two-point function of a multi-trace operator, one can readily see that the leading 1/N
contribution is due to the Feynman graphs which contribute separately to the two-point functions of the individual
single-trace constituents.
G. Arutyunov et al. / Nuclear Physics B 670 (2003) 103–147 105
Recently, another interesting peculiarity of the large N limit of N = 4 SYM was

found [7]. We recall that in regard to the supergravity approximation of the gauge theory,
the most interesting object to study are the four-point correlation functions of 12 -BPS
operators (for reviews see, e.g., [8,9]). These operators are dual to the Kaluza–Klein states
of the compactified type IIB supergravity. They form isolated series of superconformal
representations and their scaling dimension as well as their two- and three-point correlation
functions are protected from receiving quantum corrections. In contrast, their four-point
correlators undergo perturbative renormalization.
The four-point amplitude of 12 -BPS operators with arbitrary weights (dimensions)
kp , p = 1, 2, 3, 4, is specified by a number of conformal invariant functions depending
on the two conformal cross-ratios. There is a general procedure based on the field-
theoretic insertion formula which allows one to determine the maximal number of a priori
independent such functions appearing in the quantum part of the amplitude [10,11].
The explicit amplitudes corresponding to the cases where all kp ≡ k are equal to 2, 3
or 4 have been computed in the supergravity approximation by using the effective five-
dimensional description of type IIB, and also in perturbation theory at one loop (order
g 2 ) and in the planar limit. It was then observed in Ref. [7] that the perturbative amplitudes
exhibit a certain degeneracy in comparison to their supergravity counterparts. Starting from
k = 4, the number of conformal invariant functions describing the amplitude at one loop is
smaller than in the supergravity regime. We can formulate this fact as another interesting
property of the large N limit:
• the perturbative four-point amplitudes of 12 -BPS operators degenerate.
Thus, not only the two-point functions (the spectrum) but also the higher-point perturbative
amplitudes exhibit simplified features in the large N limit.
The aim of the present paper is to the study this degeneracy in more detail. Recall that
for the first time the degenerate situation occurs for k = 4. In the supergravity regime
one finds two different conformal invariant functions, precisely the number predicted on
general grounds. However, at one loop the two functions describing the large N amplitude
coincide. It should be noted that the perturbative computation has been performed for 12 -
BPS operators realized as single-trace operators. Generically, the 12 -BPS operators can be
realized as mixtures of single- and multi-trace operators with arbitrary mixing coefficients.
Therefore, two natural questions can be asked about this large N degeneracy:
1. Does the operator mixing influence the degeneracy?

2. Is the degeneracy lifted when the higher-loop corrections are taken into account?
We study the first question by using the k = 4 amplitude as the simplest example. The
1
2 -BPS operator of dimension 4 is a mixture of one single- and one double-trace operator,
the latter is made of two single-trace constituents of dimension 2. Firstly, we show that
the free amplitude involving mixed operators coincides with that for single
√ traces in the
large N limit, provided that the mixing parameter scales faster than 1/ N . As discussed
in Ref. [7], the supergravity-induced amplitude admits a unique splitting into a “constant”
and an “interacting” parts. The former precisely matches the free amplitude computed for
single-trace operators. In the AdS/CFT language the interacting part must comprise all
loop corrections to the free amplitude. Clearly, in order not to affect this agreement with
the constant part of the supergravity amplitude,
√ in field theory we should allow only mixing
which is more heavily suppressed than 1/ N . Secondly, we argue that with such a mixing
the contribution of the double-trace operators at one loop is suppressed in the large N limit
as well, so the degeneracy is not affected by the operator mixing.
We then investigate the issue of the higher-loop corrections by extending the perturba-
tive treatment to two loops (order g 4 ). To illustrate the more general picture, we consider
the amplitude involving 12 -BPS operators of equal but arbitrary weight k. By performing
an explicit diagrammatic computation we find that at two loops the degeneracy occurring
in the case k = 4 is lifted: two different conformal invariant functions emerge. However,
for amplitudes involving operators of weight k > 4 the degeneracy persists even at the
two-loop level. These observations indicate that for a fixed weight k the degeneracy is re-
moved when loop corrections of sufficiently high order (depending on k) are taken into
account. Since the supergravity-induced amplitudes comprise all perturbative corrections,
it is natural to expect them to realize the maximal possible number of different conformal
invariants.
In fact, the perturbative degeneracy phenomenon has a rather simple origin. At a given
order of perturbation theory we have only a finite number of topologically different planar
interacting graphs. Whether all of them are indeed present depends on the weight k of the
four composite operators for which the amplitude is computed. It is clear that increasing
k makes all possible interacting topologies appear. Eventually, we reach the point of
saturation beyond which increasing the weight amounts to decorating the interacting
graphs with free propagators. In the large N limit this does not generate new conformal
invariants. On the contrary, fixing k and increasing the order of perturbation theory
“awakes” new interacting topologies and hence generates new conformal invariants, thus
lifting the degeneracy.
One could also ask if there is any relationship between the degeneracy of the four-
point amplitudes and that of the spectrum observed in Ref. [5]. Any conformal invariant
function can be expanded over an infinite basis of conformal partial wave amplitudes, each
of which represents the contributions of an individual primary operator. The degeneracy
of the four-point functions is therefore related to some kind of degeneracy of the spectrum
and of the OPE coefficients. It would be interesting to find out if the degeneracy of the
spectrum is also removed when higher loop corrections are taken into account. Indeed,
in the lattice picture [6] the L-loop dilatation matrix induces the mixing of the L + 1
neighboring elementary fields constituting a composite operator, i.e., with L growing the
effective interaction is spreading over the lattice and can subsequently increase its size
which is equal to the dimension of the operator. When this happens, the degeneracy might
be lifted in a way similar to that for the four-point amplitudes.
The organization of the paper is as follows. In Section 2 we describe the structure of
the N = 4 four-point amplitude of 12 -BPS operators determined by the field-theoretic
insertion procedure and discuss its “pure” N = 2 hypermultiplet projection. It involves
a subset of k − 2 conformal invariant functions (or one such function for k = 2) from the
original N = 4 amplitude. For k > 4 this subset is not complete but it is sufficiently large
to illustrate the degeneracy problem. The advantage of the hypermultiplet projection is that
we can use the quantum formalism of N = 2 off-shell harmonic superspace [12] which
is particularly efficient at two loops.3 For the sake of clarity, at the end of Section 2 we
present and discuss our main results, whose derivation is explained in detail in Sections 4
and 5. In Section 3 we discuss the influence of operator mixing on the degeneracy problem.
In Section 4 we summarize the N = 2 insertion procedure and the Feynman diagram
tools necessary to perform the two-loop computation of the hypermultiplet projection.
In Section 5 a diagrammatic calculation is presented and the corresponding conformal
invariants are identified. Finally, in Section 6 we use the operator product expansion and
the knowledge of the one-loop anomalous dimensions of certain operators to get an insights
into the structure of the k = 3 and k = 4 amplitudes, independently of the diagrammatics.
In particular, for k = 3 we are able to completely reconstruct the two-loop amplitude from
these OPE considerations. Appendix A clarifies the technical tool of harmonic analyticity
which we systematically use in order to drastically reduce the number of relevant Feynman
graphs. Appendix B contains the details of the graph calculation for k = 3 and 4 whose
generalization for arbitrary k is given in Section 5.
2. Generalities and main result
The lowest component of a 12 -BPS multiplet in N = 4 SYM is a real scalar field of

dimension k transforming in the irrep [0, k, 0] of the R symmetry group SO(6) ∼ SU(4).
In terms of the elementary fields it can be realized, e.g., as a single-trace operator

Tr φ {a1 · · · φ an } . (2.1)
Here φ a , a = 1, . . . , 6 are the N = 4 SYM scalars and { , } denotes traceless symmetriza-

tion. A convenient way to handle the SO(6) indices is to project the operator (2.1) onto the
highest weight state of the irrep [0, k, 0]. This can be done with the help of a complex null
vector ua (ua ua = 0) carrying U(1) charge, which can be viewed as a harmonic variable
parametrizing the coset space SO(6)/SO(2) × SO(4) (see, e.g., Ref. [10] for details):

O(k) = ua1 · · · uak Tr φ a1 · · · φ an . (2.2)
We start by summarizing the general properties of the four-point amplitude of 12 -BPS

operators in the N = 4 SYM theory. It is sufficient to restrict our discussion to the case
where all the operators involved have equal weights k. On the basis of conformal and
R symmetry alone the general form of the four-point amplitude can be parametrized as
follows:
(k)
k|k|k|k = amnl (s, t)X m Y n Z l , (2.3)
m+n+l=k
3 Correlation functions of 1 -BPS operators at order g 4 have also been computed in the N = 1 superspace
2
formalism [13,14].
Fig. 1. The propagator structure corresponding to the monomial X m Y n Z l .
where m, n, l are nonnegative integers. This expression is a polynomial in the three

elementary propagator (Wick) contractions of the four points
(12)(34) (13)(24) (14)(23)

X= 2 x2
, Y= 2 x2
, Z= 2 x2
. (2.4)
x12 34 x13 24 x14 23
Here (12) = (21) ≡ ua1 ua2 , etc., denote the SO(6)-invariant contractions of the harmonic
variables at two different points. Every monomial in Eq. (2.3) corresponds to a certain
propagator structure built out of the propagators of the elementary fields (see Fig. 1). We
refer to Eq. (2.3) as to the representation of the amplitude in the propagator basis. The
(k)
coefficients amnl (s, t) are arbitrary functions of the two independent conformal invariant
cross-ratios
2 x2
x12 2 x2
x14
s= 34
2 x2
, t= 23
2 x2
.
x13 24 x13 24
Superconformal symmetry puts additional kinematical restrictions on the coefficient

functions in (2.3) which take the form of differential equations (superconformal Ward
identities) [15,16]. Further, dynamical constraints on the amplitude are obtained through
the procedure of inserting the N = 2 SYM action. According to the “partial non-
renormalization theorem” of Ref. [17], the quantum N = 2 amplitude takes a factorized
form (see Section 4 for details). Recently this insertion procedure has been generalized
to the N = 4 case [7,10]. The factorized form of the quantum corrections to the N = 4
amplitude is
(k)
k|k|k|kquant = RN =4 Fmnl (s, t)X m Y n Z l , (2.5)
m+n+l=k−2
where RN =4 is a universal polynomial prefactor carrying weight 2 at each point.

Explicitly,
RN =4 = sX 2 + Y 2 + tZ 2 + (s − t − 1)YZ
+ (1 − s − t)X Z + (t − s − 1)X Y. (2.6)
All the dynamical information is thus encoded in the conformally invariant coefficient
(k)
functions Fmnl (s, t) which are the subject of our subsequent investigation.4 Substituting
the universal prefactor into Eq. (2.5) and comparing it with Eq. (2.3), we read off the
(k)
following expression for the quantum corrections to the coefficients amnl :
(k) (k) (k) (k)
amnl (s, t) = sFm−2,n,l + Fm,n−2,l + tFm,n,l−2
(k) (k)
+ (s − t − 1)Fm,n−1,l−1 + (1 − s − t)Fm−1,n,l−1
(k)
+ (t − s − 1)Fm−1,n−1,l . (2.7)
(k)
If any of the indices of Fmnl in this equation becomes negative, then the corresponding
term is absent.
Since all the operators involved are identical, the four-point amplitude is invariant under
the symmetric group S4 which permutes the points 1, . . . , 4. Only the subgroup S3 acts
(k)
nontrivially on the cross-ratios s and t and, as a result, on the coefficient functions Fmnl .
It is generated by the two independent crossing transformations, e.g., 1 ↔ 2 and 1 ↔ 3.
(k)
If all the indices m, n, l of Fmnl are different, one can use the action of S3 to order them,
(k)
e.g., m > n > l. The function Fmnl with m > n > l does not satisfy any crossing symmetry
relation. It can be taken as a representative of the corresponding crossing equivalence class
(k)
consisting of six elements. If any two indices of Fmnl coincide, then the function satisfies
an additional crossing symmetry relation (and the corresponding crossing equivalence
class consists of three elements). Finally, in the case where all three indices coincide, the
function transforms into itself under the whole group S3 (and the corresponding crossing
equivalence class consists of a single element).
For sufficiently small values of k the number of conformal invariants F representing the
independent crossing equivalence classes is easy to work out (the general formula is given
in Ref. [11]). For instance, both for k = 2, 3 the quantum part of the amplitude is given by
a single function F , while already for k = 4 two functions are needed.
The functions F can be explicitly computed both in perturbation theory and in the
strong coupling regime by using the AdS/CFT duality. In Ref. [7] it was observed that
in the large N limit the perturbative four-point amplitudes of the 12 -BPS operators exhibit
a certain degenerate behavior in comparison to their supergravity partners: the number
of independent conformal invariants describing the one-loop (order λ ∼ g 2 ) perturbative
amplitude is smaller than that in the supergravity regime. In what follows we study this
degeneracy phenomenon at the two-loop (order λ2 ∼ g 4 ) level by explicitly working out
the corresponding four-point amplitude.
Computing four-point correlation functions at two loops is of course much more
involved than at one loop, due to the rapidly increasing number of Feynman graphs. A
very efficient technique is provided by the N = 2 harmonic superspace formalism. This
approach, combined with the N = 2 insertion procedure, allows us to drastically reduce
the number of relevant graph topologies to be computed.
4 To make the reader familiar with the notation we stress that the upper index in F (k) (s, t) denotes the weight
mnl
of the 12 -BPS operators. For the sake of clarity we sometimes separate the lower indices by commas.
In order to apply the N = 2 harmonic superspace formalism we need to decompose

the 12 -BPS operators into their N = 2 hypermultiplet (HM) and SYM constituents. This
is done by first decomposing the N = 4 field-strength superfield W = ua φ a (x) + · · · into
the Grassmann analytic HM superfield q + = u+ i φ (x) + · · · (and its harmonic conjugate
i 5
+ + i
q̃ = ui φ̄ (x) + · · ·) and the chiral field strength W = w(x) + · · · (and its antichiral
conjugate W = w̄(x) + · · ·). Here the six real scalars φ a (x) forming an SO(6) vector
have been split into an SU(2) doublet φ i (x), i = 1, 2, and a complex singlet w(x). The
SO(6) vector harmonics ua have been replaced by SU(2) fundamental harmonics u± i
(u− = (u+ )∗ , u+ ij −
i uj = 1), parametrizing the sphere S ∼ SU(2)/U(1).
2
In the sequel we will be interested in the “pure” HM projection in which we put only
q̃’s at points 1 and 4 and only q’s at points 2 and 3. As explained in Ref. [7], taking
this projection means considering only the terms with l = 0 in Eq. (2.3) (in Fig. 1 they
correspond to the graphs without diagonals). Further, the propagator structures (2.4) are
replaced by
[12][43] [13][42]
X →X= 2 x2
, Y →Y = 2 x2
, (2.8)
x12 34 x13 24
where, e.g.,
[12] = −[21] ≡ (u1 )+ ij +

i (u2 )j (2.9)
is the SU(2) invariant contraction of the two sets of harmonic variables at points 1 and 2.
Under this projection Eq. (2.5) reduces to

k−2
Q(k) ≡ q̃ k q k q k q̃ k = RN =2
(k)
Fm,k−m−2,0 (s, t)Xm Y k−m−2 , (2.10)
m=0
where
RN =2 = sX2 + (t − s − 1)XY + Y 2 . (2.11)

The factorized expression (2.10) with the universal prefactor (2.11) is precisely what
one finds by the direct application of the field-theoretic insertion procedure to the four-
point amplitude in the N = 2 theory (see Section 4).
(k)
According to the previous discussion of the crossing symmetry, we can take Fm,k−m−2,0
with
1
m− ≡ (k − 2) m k − 2 ≡ m+ for k even, (2.12)
2
1
(k − 1) m k − 2 ≡ m+ for k odd, (2.13)
2
as representatives of the different crossing equivalence classes of the coefficient functions
of the four-point amplitude.
5 Here q̃ means the usual complex conjugation for the field together with an antipodal reflection on the sphere
S 2 ∼ SU(2)/U(1) for the harmonic variable.
It should be pointed out that for k > 4 the whole N = 4 amplitude cannot be restored
(k)
from its pure N = 2 projection6 and consequently, the functions Fm,k−m−2,0 above do not
form a complete set of a priori independent conformal invariants needed to parametrize
the amplitude. However, the subset of conformal invariants emerging in the pure HM
projection is sufficiently large to illustrate the large N degeneracy problem at the two-loop
level.
Among the chosen set of independent coefficient functions only one (for k odd) or
two (for k even) admit a pair of coincident indices. The first such function F+(k) (s, t) ≡
Fk−2,0,0(s, t) is obtained by setting m = m+ , so it exists for any value of k.7 It has the
additional symmetry property
1
F+(k) (s, t) = F+(k) (s/t, 1/t). (2.14)
t
(k)
The second function Fm− ,m− ,0 is obtained by setting m = m− , so it exists only for k even.
In this case we have
1
Fm(k)− ,m− ,0 (s, t) = Fm(k)− ,m− ,0 (1/s, t/s). (2.15)
s
It is more convenient to introduce the functions
(k)
Fm(k) (s, t) ≡ Fm,k−m−2,0 (t, s), m = m+ (2.16)
(k) (k)
since for m = m− the function F− (s, t) ≡ Fm− ,m− ,0 (t, s) has the same crossing symmetry
(k)
as F+ :
(k) 1 (k)
F± (s, t) = F± (s/t, 1/t). (2.17)
t
The functions Fm(k) with m− < m < m+ do not a priori obey any crossing symmetry
relations. Finally, we note that the case k = 2 is exceptional as all the three indices of
F (2) vanish. This function transforms covariantly under the whole group S3 .
In summary, the pure HM projection is parametrized by the chain of conformally
invariant functions Fm(k) whose “boundaries” F±(k) are subject to the crossing symmetry
(k)
relation (2.17). The function F− exists only for k even.
Concluding this section, we give the main result of our two-loop calculation, the details
of which are presented in Sections 4 and 5. In the large N limit we find

(k) λ2 k 2 1 (1) 2 1
F− (s, t) = 2 s Φ (s, t) + Φ (2) (t/s, 1/s) = Fm(k) (s, t),
N 4 4 s

(k)
2
λ k 12 2 1
F+ (s, t) = 2 (t + 1) Φ (1) (s, t) + Φ (2) (s, t) + Φ (2) (s/t, 1/t) , (2.18)
N 4 4 t
6 To this end we would also need some of the “mixed” projections of the type q̃ k−p q p . Obtaining such
projections involves some amount of linear algebra; in addition, the corresponding two-loop graphs are more
complicated. Therefore, here we prefer to restrict ourselves to the “pure” sector.
7 One could instead take F (k) (k) (k)
0,k−2,0 which obeys F0,k−2,0 (s, t) = F0,k−2,0 (t, s). The relation between the
(k) (k)
two functions is Fk−2,0,0 (s, t) = (1/s)F0,k−2,0 (1/s, t/s).
where the range of m is given by Eqs. (2.12) and (2.13), in particular, for k odd the function
F−(k) is absent. In Eqs. (2.18) λ = g 2 N /4π 2 is the ’t Hooft coupling and the functions Φ (1)
and Φ (2) are the so-called one- and two-loop scalar box conformal integrals

d4 x5 iπ 2
2 x2 x2 x2
≡ − 2 x2
Φ (1) (s, t), (2.19)
x15 25 35 45 x13 24
2 d4 x d4 x
x13 5 6 (iπ 2 )2 (2)
2 x2 x2 x2 x2 x2 x2
≡ 2 x2
Φ (s, t). (2.20)
x15 25 35 56 16 36 46 x13 24
We see that in the large N limit all but one conformal invariants describing the pure HM
projection are equal. It is only the “boundary” function F+(k) which is distinctly different
from the others.
Having found the pure HM part of the four-point amplitude of the 12 -BPS operators,
we can now understand what happens to the large N degeneracy at the two-loop level.
According to Ref. [7], at one loop and in the large N limit all conformal invariants
emerging in the pure HM projection coincide:
λ k 2 (1)
F±(k) (s, t) = Fm(k) (s, t) = − Φ (s, t). (2.21)
N2 2
(3)
This degeneracy manifests itself already in the case k = 3: the single function, F+ ,
(3)
describing the one-loop amplitude obeys the extra crossing symmetry relation F+ (s, t) =
(3)
F+ (t, s) which is not required on general grounds. Just as in the one-loop case, the two-
loop function F+(3) from Eq. (2.18) obeys Eq. (2.17) imposed by the symmetry of the
amplitude, but the one-loop “bonus” crossing symmetry relation does not hold anymore.
We therefore conclude that this symmetry of the k = 3 one-loop amplitude is destroyed by
the two-loop quantum corrections.
Another type of degeneracy appears in the case k = 4. On the general grounds of
superconformal and crossing symmetry two a priori independent conformal invariants are
needed in this case, while at one loop both of them appear to coincide in the large N
limit [7]. According to Eqs. (2.18), at the two-loop level we indeed find two conformally-
invariant functions F±(4) which are distinctly different. Again we observe that the one-loop
degeneracy is lifted when the two-loop corrections are switched on.
However, we also see that both types of large N degeneracy mentioned above still
hold at the two-loop level for operators of higher weight, k > 4. Indeed, for k > 4 the
(k) (k)
functions Fm coincide (they are equal to F− for k odd) and have the same bonus crossing
(k)
symmetry (2.17) as F− .
Finally, we mention that in the special case k = 2 we find

λ2 1 2
F (2) (s, t) = 2 (s + t + 1) Φ (1) (s, t)
N 4

1 1
+ Φ (2) (t/s, 1/s) + Φ (2) (s, t) + Φ (2) (s/t, 1/t) . (2.22)
s t
This result has been previously obtained in Refs. [14,18].
3. Mixing and large N degeneracy
In field theory 12 -BPS operators can be realized as mixtures of single- and multi-trace
operators. Here we investigate whether the operator mixing can affect the degeneracy
problem of four-point correlation functions. As an explicit example we consider the case
of weight k = 4 operators.
As discussed in Ref. [7], for 12 -BPS operators of weight k 4 the complete N = 4 four-
point amplitude can be reconstructed from the “pure” N = 2 hypermultiplet projection,
Eq. (2.10). In particular, in the case k = 4 we can write down
4
(4)
Q(4) = am,4−m,0 (s, t)Xm Y 4−m . (3.1)
m=0
This projection involves five of the fifteen coefficient functions of the original N = 4
(4) (4) (4) (4)
amplitude, which belong to the three crossing equivalence classes (a040 , a400 ), (a130 , a310 )
(4)
and a220 .
In order to study possible mixing effects on Q(4) we compute the coefficients a (4) at
tree level. There, generically, the 12 -BPS operator is of the form
O ∼ S + aD, (3.2)
where we have defined

S ≡ Tr q 4 , D ≡ Tr q 2 Tr q 2 , (3.3)
q being the N = 2 Grassmann analytic HM superfield and a an N -dependent mixing
parameter.
We fix the normalization coefficient for O by requiring its two-point function to be
normalized canonically. Computing separately the two-point functions SS, SD and
DD, in the case of SU(N) color group we find (since we are interested in the color
combinatorics we do not exhibit the space–time dependence of the correlators)

SS = 4 N 2 − 1 P1 ,

SD = 8 N 2 − 1 P2 ,

DD = 8 N 2 − 1 P3 , (3.4)
where, for convenience, we have introduced
1 4 1 2
P1 = 2
N − 6N 2 + 18 , P2 = 2N − 3 , P3 = N 2 + 1 . (3.5)
N N
Therefore, the canonically normalized operator reads
−1/2
O = 4 N 2 − 1 P1 + 4aP2 + 2a 2 P3 (S + aD). (3.6)
The tree level contributions to the coefficients a (4) are given by diagrams with
propagator structures as in Fig. 1, with l = 0, n = 4 − m, m = 0, . . . , 4. The coefficient
(4) (4)
functions a040 and a400 correspond to disconnected diagrams built as product of 2-point
(4)
Fig. 2. Building blocks for the coefficient a310 .
function diagrams. Therefore, given the above normalization, they are trivially equal to
1. In order to determine the other coefficients we need to compute the actual four-point
amplitudes for the mixed operator O. This can be achieved by evaluating separately the
four-point functions SSSS, DDDD and SSSD, SSDD, SDDD as well as the
amplitudes obtained from the latter by permuting S and D.
Since it is enough to compute one coefficient function within each crossing equivalence
(4)
class, we concentrate, for instance, on the coefficient function a310 . It is convenient to first
evaluate the building blocks of Fig. 2.
These building blocks can be obtained from the tree diagrams of the 2-point amplitudes
(3.4) by cutting a line. This operation simply amounts to inserting a δa1 b1 for the line
cut without affecting the color structure. Therefore, the three building blocks are still
proportional to the polynomials given in Eq. (3.5). Explicitly,
SS → 24 P1 δa1 b1 ,
SS → 25 P2 δa1 b1 ,
SS → 25 P3 δa1 b1 . (3.7)
These building blocks are then combined to realize all possible configurations of single-
and double-trace operators. The result is
(4) (4)
a310 = a130
28 (N 2 − 1)(P12 + 16a 2P22 + 4a 4 P32 + 8aP1P2 + 16a 3P2 P3 + 4a 2P1 P3 )
=
24 (N 2 − 1)2 (P1 + 4aP2 + 2a 2P3 )2
2 (N − 1)(P1 + 4aP2 + 2a 2 P3 )2
8 2 16
= 4 2 = 2 . (3.8)
2 (N − 1) (P1 + 4aP2 + 2a P3 )
2 2 2 N −1
Being independent of the mixing parameter a, it gives the finite N result for the tree level
(4)
part of the coefficient function a310 , independently of the presence of the double-trace
operator. In the large N limit this expression agrees with the result of Ref. [7]. We note
that the independence of a is due to the particular structure (3.7) of the building blocks.
(k)
For k 4 we observe the same phenomenon in the coefficients a(k−1)10 .
(4)
Next we concentrate on a220 . Again, the calculation can be carried out by first
computing the building blocks depicted in Fig. 3.
Introducing the concise notation Tr(Ta Tb · · · Tc Td ) ≡ (ab · · · cd) for the trace of the
color generators, the algebraic expressions for the building blocks of Fig. 3 can be written
(4)
Fig. 3. Building blocks for the coefficient a220 .
as
(N 2 − 9)
SS → 4 (a1 a2 b1 b2 ) + (a2 a1 b1 b2 ) + (a1 a2 b2 b1 ) + (a2 a1 b2 b1 )
N
2N 2 + 9
+8 2
(a1 a2 )(b1 b2 ) + 4(a1 b1 )(a2 b2 ) + 4(a1 b2 )(a2 b1 ),
N
SD → 23 (a1 a2 b1 b2 ) + (a2 a1 b1 b2 ) + (a1 a2 b2 b1 ) + (a2 a1 b2 b1 )

(2N 2 − 3)
+ (a1 b1 a2 b2 ) + (a1 b2 a2 b1 ) + (a1 a2 )(b1 b2 ) ,
N
2
DD → 2 N + 3 (a1 a2 )(b1 b2 ) + 2(a1 b1 )(a2 b2 ) + 2(a1 b2 )(a2 b1 ) .
3
All possible configurations of the four-point amplitudes of S and D are obtained by

combining the building blocks of Fig. 3 in all possible ways. The last contraction between
two blocks produces an extra factor of 4. Pulling out the common factor 28 (N 2 − 1), we
find

15 2 729 1215
SSSS → N − N + 135 − 2 +
4
≡ P4 ,
2 N N4
1
SSSD → 3 10N 6 − 54N 4 + 216N 2 − 405 ≡ P5 ,
N
1
SSDD → 2 4N 6 + 6N 4 − 63N 2 + 135 ≡ P6 ,
N
1 6
SDDD → 2N + 5N 4 + 18N 2 − 45 ≡ P7 ,
N

DDDD → N 6 + 5N 4 + 19N 2 + 15 ≡ P8 . (3.9)
(4)
The final answer for the coefficient a220 is then
(4) 16(P4 + 4aP5 + 6a 2P6 + 4a 3 P7 + a 4 P8 )

a220 = . (3.10)
(N 2 − 1)(P1 + 4aP2 + 2a 2P3 )2
We notice that by setting a = 0 (i.e., by neglecting the double-trace contributions) we
obtain the finite N result for the free amplitude involving only the single-trace operator:

(4) single 16 N 4 − 15
2 N + 135 − N 2 + N 4
2 729 1215
a220 = 2 (3.11)
(N 2 − 1) N 2 − 6 + N182
Fig. 4. The leading large N contributions to the four-point amplitudes SSSS and DDDD at order λ.
In the large N limit this expression agrees with that in Ref. [7].
Next we analyze the large N limit of the general expression (3.10). Taking the large N
limit of Pi we obtain

(4) 16 N 2 + 40(Na) + 24(Na)2 + 8(Na)3 + (Na)4
a220 = 2 . (3.12)
N 2 + 8(Na) + 2(Na)2
If in the large N limit the mixing parameter is chosen to be of order one, then the double-
(4)
trace part of O provides a dominant contribution and one finds a220 ∼ 1.
We recall that according to our general treatment in Section 2 the coefficient function
(4)
a220(s, t) admits a unique splitting on the nontrivial (“quantum”) function of the conformal
(4)
cross-ratios constructed out of F± and an integration (“free field”) constant a220
(4) (4)
a220(s, t) = a220 + F+ (s, t) + F+ (1/s, t/s) + (t − s − 1)F− (t, s). (3.13)
(4)
The constant part of a220 (s, t) was computed in the supergravity regime [7] with the result
(4) sugra
a220 = 16/N 2 , i.e., it scales as 1/N 2 . Therefore, the AdS/CFT correspondence does
√
not allow the mixing parameter a to be of order one. If we pick a = κ/ N + · · ·, then it
(4)
will affect the coefficient a220 at order 1/N 2 :
(4) 16
a220 = 2
1 + κ4 . (3.14)
N
(4)
However, if we want a220 to match the large N value computed from supergravity exactly,
√
we ought to require κ = 0, i.e., the mixing parameter has to scale faster than 1/ N .
Moreover, we note that if a is expanded in the odd powers of 1/N , then the free amplitude
has a ’t Hooft type expansion in powers of 1/N 2 . Therefore, the simplest assumption for
the mixing coefficient is a ∼ 1/N + · · ·.
Let us now briefly discuss the influence of the mixing on the large N degeneracy
g2
occurring at the one-loop level (order λ ≡ 4π ). As shown in Ref. [7], harmonic analyticity,
i.e., the superconformal kinematics, allows one to reduce the one-loop calculation of the
four-point amplitude just to a single Feynman graph given in Fig. 4. This statement holds
irrespectively of the intrinsic trace structure of the BPS operators and therefore it can be
applied to the operator O.
(4)
The various contributions to the coefficient a220 from the amplitude at order λ are
then obtained by specifying the operator (S or D) at each corner of the graph in Fig. 4.
Using a double line notation for the propagators in there, the leading large N behavior is
easily identified: both the purely single-trace part SSSS and the purely double-trace part
DDDD grow as N 7 . This is in contradistinction to the free field case, where SSSS ∼
N 6 while DDDD ∼ N 8 (see Eq. (3.9)). Therefore, at one loop the contribution of the
double-trace operator appears to be more suppressed than in the free case, if compared to
that of the single trace. The double-trace contribution can compete with the single-trace
one if the mixing parameter is of order one. This is however
√ ruled out by the free field
considerations above which set a to scale faster than 1/ N .
In conclusion, we see that if one requires the large N free field theory amplitude to
coincide with the one for single-trace operators (and therefore with the constant part of the
supergravity induced amplitude), then the mixing produces no effect on the leading large
N one-loop amplitude and, as a consequence, on its degeneracy. Such a behavior follows
the supergravity pattern: the redefinitions of the supergravity fields (corresponding to the
field-theoretic mixing) produce no effect on the four-point amplitude, provided that it is of
the regular (i.e., not of the extremal or sub-extremal) type [19].
4. The N = 2 insertion procedure and Feynman rules
We would like to find the quantum corrections to the lowest component (at θ1,2,3,4 = 0)
of the four-point correlator Qk and identify the corresponding conformal invariants F in
Eq. (2.10). Already in the one-loop case, but even more so at two loops, the most efficient
technique, in our opinion, is the insertion procedure in N = 2 harmonic superspace [17,
20–22].
The basic ingredients of the N = 4 SYM theory in terms of N = 2 superfields are
the hypermultiplet q + (x, θ + , θ̄ + , u) and the N = 2 SYM field strength W (x, θ ). The
distinctive feature of the former is that it is a Grassmann analytic superfield (i.e., it depends
only on the harmonic U(1) projections θ + , θ̄ + ) while the latter is a chiral superfield.
One of the advantages of the N = 2 formulation of the theory compared to N = 1 is
that the HM composite operators like Tr(q + )k need no covariantization (no presence of the
gauge superfield at the external points of the amplitude). Further, the HM matter interacts
with the gauge sector only through a single cubic vertex. The true non-Abelian nature of
the theory is encoded in the gauge self-interactions (as well as in the ghost sector, but we
do not see it at the level g 4 ).
The quantum corrections to the four-point correlator Qk at order g 4 can be obtained
by a double insertion of the N = 2 SYM action

1
SN =2 SYM = d4 x d4 θ L, L = 2 Tr W 2 . (4.1)
4g
More precisely,

1
(k)
Q g4
= d4 x5 d4 θ5 d4 x6 d4 θ6 Q(k) L(5)L(6) tree . (4.2)
2
Here Q(k) |L(5)|L(6)tree is a new, six-point correlator calculated at “tree level” (Born
approximation). The corresponding Feynman graphs are obtained by drawing the usual
two-loop (order g 4 ) graphs for Q(k) and then inserting the linearized N = 2 SYM
Lagrangian L into the gluon lines twice, each time into a different gluon line.8
The most important property of the new six-point amplitude is that it can be written in
the factorized form
(k)
Q L|L = Θ × A(k−2)(x, u), (4.3)
where Θ is a particular nilpotent six-point superconformal covariant carrying harmonic
U(1) charge 2 at points 1 to 4, and R charge 2 at points 5 and 6 (the chiral field strength W
has R charge 1 while the HMs carry harmonic U(1) but no R charge). Since the left-handed
θα are the only superspace coordinates with positive R charge 1/2, we conclude that the
expansion of Θ must start with four such θ ’s. The rest of the six-point correlator (4.3) is
given by the function A(k−2) (x, u) which carries U(1) charge k − 2 at points 1 to 4 and can
be expanded in the propagator basis as follows:

k−2
A (k−2)
(x, u) = Xm Y k−m−2 Am (x). (4.4)
m=0
The coefficients Am (x) are conformally covariant functions of the six space–time points.
They have vanishing U(1) charge and are thus harmonic independent.
The structure of the nilpotent covariant Θ is determined by superconformal symmetry
combined with the Grassmann analytic (or 12 -BPS) nature of the first four points and the
chiral nature of the last two points. The explicit form of Θ (for θ̄r+ = 0, r = 1, . . . , 4) has
been worked out in Ref. [22]:
4
x2 x2 x2 x2 x4 x4
Θ = r=1 4 r5 r6 12 34 13 24
x56 RN =2

× [12]2[34]2τ14 τ23 + [14]2[23]2τ12 τ34

+ [12][23][34][41][τ13τ24 − τ12 τ34 − τ14 τ23 ] . (4.5)
Here
τrs ≡ 4(ρr ρs )(σr σs ) + ρr2 σs2 + ρs2 σr2 (4.6)

and
−1 −1
ρr = θr+ − θ5i u+
ri xr5 , σr = θr+ − θ6i u+
ri xr6 , r = 1, . . . , 4, (4.7)
8 This insertion procedure is based on the formula for the second derivative of the four-point correlator with
respect to the gauge coupling constant:

∂ 2 (k) 2 1
2
Q = 4 d4 x5 d4 θ5 Q(k) L + 4 d4 x5 d4 θ5 d4 x6 d4 θ6 Q(k) L|L.
∂g g g
If both insertions are made into the same gluon line, this means inserting the chiral-to-chiral propagator
W (5)W (6) ∼ δ(5, 6) into that gluon line. By performing the chiral superspace integration over point 6, one
can show that the five-point correlator arising in this way precisely cancels against the single-insertion term in
this formula.
are Q-supersymmetry-invariant combinations of the analytic θr+ with the chiral θ5,6 i
+
projected with the SU(2) harmonics (ur )i . The harmonic contractions [12, 34], etc., and
the polynomial RN =2 have been defined in Eqs. (2.9) and (2.11), respectively.
The aim of our two-loop calculation is to determine the factor A(k−2)(x, u) in the six-
point covariant (4.3) and then to substitute everything in the insertion formula (4.2). Since
we are only interested in the lowest component of the four-point correlator Qk , we can
set all the external θ ’s to zero, θr+ = θ̄r+ = 0, r = 1, . . . , 4. In this case Θ is rather simple:
2 x2 x4 x4
x12
Θθ + =0 = θ54 θ64 34 13 24
4
RN =2 . (4.8)
x56
Consequently, the six-point correlator becomes
x2 x2 x4 x4
Q(k) L|Lθ + =0 = θ54 θ64 12 344 13 24 RN =2 A(k−2)(x, u). (4.9)
x56
Further, substituting this into the double-insertion formula (4.2) and performing the trivial
chiral integrations over θ5,6 , we obtain the two-loop correlator
4
(k) d x5 d4 x6 (k−2)
Q θ + =0 = x12 2 2 4 4
x34 x13 x24 RN =2 4
A (x, u). (4.10)
x56
Rewriting the amplitude in the form (2.10), we can read off the following expression for
the conformally invariant coefficient functions F (k) at two loops:
4
(k) d x5 d4 x6
Fm,k−m−2,0 (s, t) = x12 x34 x13x24
2 2 4 4
4
Am (x). (4.11)
x56
Now, the practical question is how to compute A(k−2) (x, u) (and hence Am (x) and
F (s, t)) from the corresponding set of two-loop Feynman diagrams. It turns out that instead
of setting θ + = 0, as required in the final expression (4.10), it is much more convenient to
do the computations with θ5,6 = 0. The knowledge of the structure of Θ (4.5) allows us to
easily switch from one of these superconformal frames to the other.9 In the new frame Θ
becomes even simpler,

4
+ 2
Θθ5,6 =0 = θr , (4.12)
r=1
so that Eq. (4.9) is replaced by

4
+ 2 (k−2)
Q(k) L|Lθ5,6 =0 = θr A (x, u). (4.13)
r=1
Then it is clear that in working out the expressions for the various Feynman graphs we
can concentrate only on the terms with the maximal number of external θ ’s. In particular,
9 By “superconformal frames” we mean that Θ can be cast into one of these forms by means of a finite
superconformal transformations.
(a)
(b)
Fig. 5. Building blocks of the Feynman graphs.
at order g 4 this choice removes all graphs which contain non-Abelian interaction vertices.
For example, the Y-shaped gluon subgraph in Fig. 5(a) vanishes because it has two chiral
ends at the insertion points 5 and 6 and one analytic end (the gluon without insertion);
after setting θ5,6 = 0 we are left with too few left-handed θ ’s at the analytic gluon end to
supply the required R charge 2. Similarly, the TTT block in Fig. 5(a) has three chiral ends
(in fact, only two, points 6 and 7 should be identified) and two analytic ends; once again,
the analytic θ ’s cannot provide the required R charge 3.
As a result of all these simplifications our task is reduced to listing all tree level Feynman
graphs made out of the two building blocks T and TT in Fig. 5(b). Although these blocks
contain interaction vertices and hence integrals, the latter are easily computed using the
harmonic superspace Feynman rules and they produce very simple rational space–time
functions [20,22]:
2gfabc − 2
T125 ≡ q̃ + (1)W (5)q + (2) = 2
[21 ]ρ1 + [12− ]ρ22 − 2(ρ1 ρ2 ) , (4.14)
4
(2π) x12
4g 2 fabc fcde − − 2 2
T T1562 ≡ q̃ + (1)W (5)W (6)q + (2) = − 2
[1 2 ]ρ1 σ2 , (4.15)
(2π)6 x12
where ρ, σ have been defined in (4.7) and [12− ] = (u1 )+ ij −
i (u2 )j (cf. (2.9)).
Notice the characteristic presence of negative-charged harmonics in both expressions
(4.14) and (4.15). This has to do with the important issue of harmonic analyticity [15]. The
point is that the free HM satisfies the massless field equation

D ++ q + x, θ + , θ̄ + , u = 0, (4.16)
where
∂ ∂
D ++ = u+i −i
− 4iθ + σ µ θ̄ + µ (4.17)
∂u ∂x
is the supersymmetrized harmonic derivative (the raising operator of the group SU(2)
realized on the charges ± of the harmonics). This equation can also be viewed as a Cauchy–
Riemann condition on the harmonic coset SU(2)/U(1), hence the name “harmonic
analyticity”. Yet another interpretation of Eq. (4.16) is that it defines q + as the highest
weight state of an SU(2) irrep of charge 1 (isodoublet). When the HMs interact with the
gauge sector, the harmonic derivative in (4.16) is modified by a gauge connection. Still,
the gauge invariant composite operators like Tr(q + )k satisfy the same equation with a flat
harmonic derivative. So, such operators correspond to the highest weight state of an SU(2)
irrep of charge k.
As the simplest example of harmonic analyticity in action, consider the HM two-point
function (propagator). It has to satisfy the Green’s function equation

D1++ q̃ + (1)q + (2) = δ(1, 2), (4.18)
so here harmonic analyticity holds up to contact terms. The solution to this equation is (for
simplicity we set θ2 = 0)
[12]
q̃ + (1)q + (2) = − 2 . (4.19)
x12 + 4i [1[12]2] θ1+ σ θ̄1+
This can be checked by expanding the denominator in θ1+ θ̄1+ , using the relation
1 (1/x12
2 ) ∼ δ(x ) and the naive harmonic differentiation rules
12
D1++ θ1+ = 0, D1++ [1− 2] = [12], D1++ [12] = 0, (4.20)

as well as the rule
1
D1++ = δ(u1 , u2 ) (4.21)
[12]
which follows from the relation ∂/∂ z̄(1/z) ∼ δ(z). In the context of our four-point
investigations we always set all θ̄ = 0. Therefore here we never encounter harmonic
singularities of the type (4.21) and can safely apply the naive rules (4.20). Then harmonic
analyticity simply means polynomial dependence in u+ and no dependence in u− . This is
the case of the propagator (4.19) which now becomes
[12]
q̃ + (1)q + (2) θ̄=0 = 2 . (4.22)
x12
The same must hold for any n-point correlator of 12 -BPS operators: they must be
polynomials in the harmonic variables u+ at each point.
Clearly, the expressions for the building blocks T (4.14) and TT (4.15) are not harmonic
analytic because of the presence of 1− and 2− . This, however, is not a problem: The
various building blocks or even complete Feynman graphs are not expected to be harmonic
analytic, much like they are not conformal covariants. It is only the sum of all graphs that
has these properties (see examples in Appendix A). In what follows we shall often profit
from the expected harmonic analyticity of the final result to greatly simplify our graph
calculations.
5. Two-loop four-point amplitudes
In this section we work in the frame θ5,6 = 0 where the six-point nilpotent covariant
takes the form
k−2
(k) 4
+ 2

Q L|Lθ5,6 =0 = θr m k−m−2
X Y Am (x) . (5.1)
r=1 m=0
Here the coefficients Am (x) are not all independent, as the crossing symmetry 2 ↔ 3 of
Eq. (2.10) implies that Am (x) is equal to Ak−m−2 (x) with x2 and x3 interchanged. In
particular, for k even, A(k−2)/2 is invariant under this crossing transformation. By using
Eq. (4.11) one can see that these facts are in accord with the corresponding transformation
(k)
property of Fm,k−m−2,0 (s, t). Therefore, we need compute only k−2 2 + 1 coefficient
k−1
functions if k is even and 2 if k is odd.
At two loops the correlator (5.1) is given by the sum of a number of Feynman graphs
containing four interaction vertices of HM with gluons and two L insertions. Let us discuss
the general structure of these diagrams. According to the rules discussed in Section 4,
they are made out of the building blocks T and TT. The same argument based on θ
counting allows us to discard other sets of vanishing diagrams. Take, for instance, the
graphs involving a HM loop with the insertion of two gluon lines connecting the HM
propagators. Such a loop produces the expression (T T )2 ∼ ρ 4 σ 4 = 0 (see (4.15)). For the
same reason any graph containing a corner with all the incoming or outgoing HM lines
being free vanishes, since this implies concentrating too many θ ’s at some other corner.
Finally, disconnected graphs containing subgraphs in the form of the one- or two-loop
corrections to the two-point function of protected 1/2 BPS operators vanish as well.
With these rules in mind and observing that at order g 4 the maximal number of HM lines
involved in interactions with gluons is four, we can now draw all possible configurations.
The planar interaction topologies relevant to the calculation of the functions Am (x) are the
same as the ones shown in Fig. 12 for the k = 4 case (thin lines).10 The complete graphs
are then obtained by combining an interaction topology from Fig. 12 (thin lines) with the
matching free HM frame from Fig. 6(a). Each resulting topology is labeled by a pair of
integers (p, q) (0 p 2m, 0 q 2(k − m − 2)) which fixes the particular structure of
the free HM lines. This procedure automatically takes into account the possible crossing
symmetry transformations of the end points.
Evaluating the explicit expressions for each topology amounts to multiplying the
building blocks (4.14) and (4.15) and using a simple Fierz identity for the two-component
spinors ρ and σ . In this way the contributions of the individual graphs are always expressed
in terms of the variables ρr2 , σr2 and τrs .
Since the building blocks contain negatively charged harmonic variables, each indi-
vidual graph involves some nonanalytic harmonic dependence. The final sum (5.1) must
however be harmonic analytic and moreover, the coefficients Am are harmonic indepen-
dent. Thus, all the terms in the graphs with nonanalytic dependence are in fact spurious
10 Beyond those of Fig. 12 there are also other interacting topologies which, however, play only an auxiliary
role as it will become clear from our procedure of identifying the diagrammatic contributions to Am (x).
(a) (b)
Fig. 6. (a) The free HM frame (p, q) associated with the function Am (x). (b) The HM frame common to all
graphs contributing to Am (x).
and should cancel out. Due to the huge number of diagrams the evaluation of each diagram
and the proof of the actual cancellation of nonanalytic terms can be rather intricate (see
Appendix A for a sample calculation).
However, the knowledge that the final result has to be analytic allows us to skip the
actual computation of all the diagrams and directly identify a minimal set of graphs
relevant to restoring the correct harmonic analytic structure of the final sum. This can be
achieved by a suitable procedure of identifying the harmonics. In Appendix B we describe
this procedure in detail when applied to operators of weight k = 3, 4. The corresponding
minimal sets of relevant graphs for the k = 3 case are given in Fig. 11, whereas for k = 4
they are listed in Figs. 13 and 14.
For generic k, the procedure of identifying the harmonics goes along the same lines
as in Appendix B. In this way, we are led to select the minimal set of relevant diagrams
which contribute to the function Am . They are drawn in Fig. 7 where the labels (p, q) are
explicitly indicated for each graph.
We note that the topologies of interaction lines appearing in the diagrams of Fig. 7 are
the same as the ones for k = 3, 4 cases (see Figs. 11–14 in Appendix B). Let us explain
why this happens by first looking at the case 1 m k − 3. We observe that the diagrams
in Fig. 7 all contain the common free HM frame depicted in Fig. 6(b). This frame provides
the harmonic factor
m−1 k−m−3
Xm−1 Y k−m−3 → [12][43] [13][42] (5.2)
which can be pulled out from all the graphs. It is easy to see that what remains forms the
set of graphs contributing to the function A1 (x) for the amplitude of weight 4 operators
(see Fig. 14). Therefore, we can exploit the results for k = 4 to select the minimal set of
interacting subdiagrams which survive harmonic identification. The particular cases m = 0
and its crossing symmetry partner m = k − 2 can be treated in the same way. For m = 0
the graphs a3 and a4 in Fig. 7 are absent and pulling out a common factor Y k−4 we are led
to the set of graphs for the function A0 of the k = 4 case (see Fig. 13).
The important conclusion we reach at this point is that the two-loop calculation for
arbitrary k is reduced to that for k = 4. The reason for this is quite clear: since at two
loops at most four HM lines are involved in the interactions, for a sufficiently large k the
majority of the HM lines play the role of “spectators”.
Let us explain why the diagrams in Fig. 7 with the labels (p, q) indicated do indeed
provide the contributions to Am . Recall that the graphs which contribute to Am are the
Fig. 7. Relevant topologies of the interacting subgraphs of the planar graphs contributing to the function Am .
Every subgraph has to be multiplied by the free contribution from Fig. 6(a).
ones from which we can eventually extract the propagator factor Xm Y k−m−2 . We have
already pulled out a common factor Xm−1 Y k−m−3 . Leaving aside for the moment the
graphs of the type 7a we see that all the other diagrams contain an additional factor
XY ∼ [12][43][13][42]. Pulling it out we restore the required overall factor Xm Y k−m−2 .
What remains from the graph is a chargeless combination of harmonics which may contain
constant harmonic independent terms, but also nonanalytic terms. However, once we have
pulled out the required overall factor, the rest has to be harmonic independent and can be
computed by identifying all the harmonics as explained in Appendix B.
We still have to take care of the graphs which contain interaction subgraphs of type
7a. For instance, for the interaction topology 7a1 besides the common frame (5.2) we can
also pull out an extra factor [12][43] ∼ X. This leaves behind the free factor [13]2 which
is different from the expected [13][42] ∼ Y . The missing factor Y is restored only after
summing up 7a1 with all of its analyticity partners. The partners of the graph 7a1 needed
to restore analyticity, e.g., at points 1 and 3 are shown in Fig. 8. In Appendix A we analyze
in detail how analyticity is achieved in this case and how such a long calculation can be
efficiently bypassed by the procedure of identifying the harmonics.
The reader may wonder why the graph 8d, which has the topology 7b, has been included
in the set 8 where it clearly plays the auxiliary role of an analyticity partner of the graph
8a. In fact, this graph plays a dual role: it becomes the principal graph contributing to the
coefficient function Am+1 (x). Indeed, pulling out the common factor Xm−1 Y k−m−3 and
Fig. 8. The graph 8a (topology 7a1 ) and its partners restoring analyticity at points 1 and 3 (all the graphs have the
same bottom part). The common free HM frame has been pulled out as the prefactor Xm−1 Y k−m−3 .
the extra factor X2 contained in the the graph 8d, we find

τ13 − [13][1−3− ] ρ12 σ32 + ρ32 σ12 τ24 − [24][2−4− ] ρ42 σ22 + ρ22 σ42 . (5.3)
This expression is chargeless, so in it we can identify all the harmonics, after which we are
left with the analytic term τ13 τ24 . We can say that the analytic part of Eq. (5.3) contributes
to the function Am+1 , while the nonanalytic part combines with the other graphs in Fig. 8
into a harmonic analytic expression which gives a contribution to Am .
Having identified the relevant graphs, we now proceed to computing the six-point
correlator. The Grassmann and space–time structures of the graphs in Fig. 7 are
τ24 (ρ12 σ32 + ρ32 σ12 )
Ga1 = T125T346 + (5 ↔ 6) T245T246 = 4 x2 x2
,
x24 12 34
τ13 (ρ22 σ42 + ρ42 σ22 )
Ga2 = T125T346 + (5 ↔ 6) T135T136 = 4 x2 x2
,
x13 12 34
τ34 (ρ12 σ22 + ρ22 σ12 )
Ga3 = T135T246 + (5 ↔ 6) T345T346 = 2 x2 x4
,
x13 24 34
τ12 (ρ32 σ42 + ρ42 σ32 )
Ga4 = T135T246 + (5 ↔ 6) T125T126 = 4 x2 x2
,
x12 13 24
τ13 τ24
Gb = T135 T136 T245T246 + (5 ↔ 6) = 2 4 4 ,
x13 x24
τ12 τ34
Gb = T125 T126T345 T346 + (5 ↔ 6) = 2 4 4 ,
x12x34
τ23 (ρ12 σ42 + ρ42 σ12 )
Gc1 = T125T135 T246 T346 + (5 ↔ 6) = 2 x2 x2 x2
,
x13 24 12 34
τ14 (ρ22 σ32 + ρ32 σ22 )
Gc2 = T125T245 T136 T346 + (5 ↔ 6) = 2 x2 x2 x2
,
x13 24 12 34
and
Gd = T125T345 T136 T246 + (5 ↔ 6)

1
= 2 2 2 2 τ14 τ23 − τ13 τ24 − τ12 τ34 − τ23 ρ12 σ42 + ρ42 σ12
x12 x34x13 x24

− τ14 ρ32 σ22 + ρ22 σ32 + τ13 ρ42 σ22 + ρ22 σ42 + τ24 ρ12 σ32 + ρ32 σ12

+ τ12 ρ42 σ32 + ρ32 σ42 + τ34 ρ12 σ22 + ρ22 σ12 .
Collecting everything we arrive at the following complete harmonic analytic expression
for the six-point correlator:
1
k−2
Q(k) L|L ∼ 2 x2 x2 x2
Xm Y k−m−2
x12 34 13 24 m=0

2 b
× Cm d
τ14 τ23 + 2sCm b
− Cm
d
τ13 τ24 + Cm − Cm d
τ12 τ34
s
c 2 2 c 2 2
+ Cm m 23 ρ1 σ4 + ρ4 σ1 + Cm − Cm τ14 ρ3 σ2 + ρ2 σ3
1 − Cd τ 2 2 2 d 2 2
d 2 2 d 2 2
+ Cm − Cm a2
τ13 ρ4 σ2 + ρ22 σ42 + Cm − Cm
a1
τ24 ρ1 σ3 + ρ32 σ12

d 2 2 d 2 2
+ Cm − Cm τ12 ρ4 σ3 + ρ3 σ4 + Cm − Cm τ34 ρ1 σ2 + ρ2 σ1 ,
a4 2 2 a3 2 2
(5.4)
where Cm denote the combinatorial factors associated to the various graphs.
The expression (5.4) has been derived for 1 m k − 3. As was already mentioned
above, the cases m = 0 (and its crossing partner m = k − 2) are a bit special, as for m = 0
the graphs a3 and a4 are absent, while for m = k − 2 we do not have a1 and a2 . Otherwise,
the derivation of the relevant contributions for these cases follows the same steps as above.
Thus, the formula (5.4) remains valid for m = 0 and for m = k − 2 as well, provided we
impose the following formal requirement on the combinatorial coefficients
a
C0 3 = C0a4 = Ck−2
a1 a2
= Ck−2 = 0.
Before we proceed, let us briefly discuss the general symmetry properties of the
combinatorial factors. We note that for the graphs a1 and a2 the range of indices is
0 m k − 3, while for a3 and a4 it is 1 m k − 2. For a generic m the combinatorial
a1 a2 a1 a2
coefficients Cm and Cm corresponding to the graphs a1 and a2 are equal, Cm = Cm ≡ Cma
a3 a4 c1
(though their analytic structures are different). Analogously, Cm = Cm ≡ Cm and Cm =
a
c2
Cm = Cmc . We also have the following additional relations

a
Cm = Ck−m−2
a
, b
Cm = Ck−m−2
b
, (5.5)
which in the special case m = m− reduce to

a
Cm −
= Cm
a
−
, b
Cm −
= Cm
b
−
. (5.6)
Finally, to recast the six-point amplitude (5.4) in the form (5.1) we can use the relations
θr+2 θr+2 2 x2
xrs
ρr2 = 2
, σr2 = 2
, τrs = θr+2 θs+2 56
2 x2 x2 x2
(5.7)
xr5 xr6 xr5 r6 s5 s6
valid for θ5,6 = 0. The very last step is to perform the space–time integration over the
insertion points 5 and 6 in Eq. (4.11). This introduces the functions Φ (1) and Φ (2) defined
in (2.19) and (2.20) and their point permutations. Note that Φ (1) is totally symmetric under
permutations of the external points.
With these definitions at hand, from Eqs. (5.1) and (4.11) we can identify

(k) 1 d b b (1) 2
Fm,k−m−2,0 (s, t) = : C t + 2Cm − Cm
d
s + 2Cm − Cm d
Φ (s, t)
4 m
d
a 1 (2)
d (2)
+ Cm − Cm Φ (t/s, 1/s) + Cm − Cm a
Φ (s, t)
s
c
d 1 (2)
+ Cm − Cm Φ (s/t, 1/t) . (5.8)
t
Here : is an overall normalization coefficient independent of m, which is related to the
normalization of the 12 -BPS operators.
The final result (5.8) essentially depends on the values of the combinatorial coefficients
Cm . The particular crossing symmetry properties (2.17) of the functions F± provide
some additional relations among the corresponding combinatorial coefficients. Indeed, for
m = m+ we find

c
Cm +
= 2Cm
b
+
= 2Cm
d
+
,
but there are no new restrictions on Cm− .
We remark that the case k = 2 is special as there are no graphs of type 6a and the higher
crossing symmetry requires 2C0b = C0c = 2C0d .
The general expressions for the combinatorial coefficients Cm valid for finite N are
fqpe frt e fcse fnle
a
Cm = (a1 · · · am am+1 · · · ak−1 p)
2(m!) (k − 1 − m)!(k − 3 − m)!
2
× (a1 · · · am bm+1 · · · bk−3 qrn)(c1 · · · cm am+1 · · · ak−1 c)

× (c1 · · · cm bm+1 · · · bk−3 tsl),
2Cmb
= (a1 · · · am am+1 · · · ak−2 ps)(a1 · · · am bm+1 · · · bk−2 rn)
2[(k − m − 2)!m!]2
× (c1 · · · cm am+1 · · · ak−2 qc)(c1 · · · cm bm+1 · · · bk−2 tl),
c
Cm = (a1 · · · am am+1 · · · ak−2 pt)(a1 · · · am bm+1 · · · bk−2 qn)
2[(k − m − 2)!m!]2
× (c1 · · · cm am+1 · · · ak−2 rc)(c1 · · · cm bm+1 · · · bk−2 sl),
d
Cm = (a1 · · · am am+1 · · · ak−2 cp)(a1 · · · am bm+1 · · · bk−2 lq)
2[(k − m − 2)!m!]2
× (c1 · · · cm am+1 · · · ak−2 sr)(c1 · · · cm bm+1 · · · bk−2 nt),
where
(a1 · · · ak ) ≡ Tr(t(a1 · · · tak ) )
is the symmetrized (without 1/k!) trace of k generators of the color group. In these
formulae the combinatorial factors are necessary to prevent the overcounting of the HM
lines induced by symmetrization.
To compute these coefficients in the large N limit we use the conventions [tp , tq ] =
ifpqe te and fabe fabe = 2Nδee as well as the following fusion and splitting rules
(a1 · · · al A)(a1 · · · al B) ≈ (l + 1)2 l!N l−1 Tr(AB),

Tr[al · · · a1 a1 · · · al A] ≈ N l Tr A.
In the large N limit we find
a
Cm = 2Cm
b
= Cm
d
= k 4 N 2k , c
Cm = 2k 4 N 2k . (5.9)
In this formula the range of the index m is 0 m k − 3 for Cm a , 0 m k − 2 for C d ,
m
1 m k − 2 for Cm b and 0 m k − 2 for Cm . The case of C0b deserves some special
c
attention. Looking at the graph 6b for m = 0, we realize that the gluon lines can “slide”
along the HM loops, thus giving rise to two equivalent planar diagrams. Hence, the factor
b with m = 0:
C0b is twice as big as the other Cm
C0b = k 4 N 2k . (5.10)
So, the combinatorial coefficients have the remarkable property that they do not depend
on m, except for the single “anomalous” coefficient Cm b which changes its value only for
m = 0.
Substituting these combinatorial coefficients into Eq. (5.8), choosing canonical normal-
ization for the 12 -BPS operators and recalling the definition (2.16), we obtain our main
result (2.18).
6. Two-loop four-point amplitudes from the OPE
Here we show that the four-point two-loop amplitude for 12 -BPS operators of weight
3 can be completely reconstructed just on the basis of the expected operator product
expansion structure and the knowledge of the one-loop anomalous dimensions for certain
operators of twist 2 and 4. Analogously, we obtain some restrictions on the structure of the
amplitude for weight 4 operators. The discussion in this section provides an independent
check on our diagrammatic computation.
6.1. Weight 3
The diagrammatic treatment clearly shows that at two loops the four-point amplitude is
made out of two basic conformal integrals, (2.19) and (2.20). Therefore, the most general
Ansatz compatible with the symmetry requirement F (s, t) = (1/t)F (s/t, 1/t) is

λ2 1 2
F (s, t) = 2 <s + η(t + 1) Φ (1) (s, t)
N 4

µ 1
+ Φ (2) (t/s, 1/s) + ν Φ (2) (s, t) + Φ (2) (s/t, 1/t) , (6.1)
s t
i.e., each function depends on four unknown coefficients <, η, µ and ν.
Already at the level of the free field theory one can see that the OPE O(k) O(k) of two
1
2 -BPS operators exhibits a heredity property: if some superconformal primary operator
emerges in the spectrum of this OPE then it must also appear in the spectrum of O(n) O(n)
with n > k. Therefore, knowing the spectrum of the operator product O(2) O(2) (see
Refs. [23,25]), we can use it to obtain restrictions on the four-point amplitude underlying
the OPE of the 12 -BPS operators with higher weights. From the technical point of view, we
need to build the conformal partial wave amplitude (CPWA) expansion of the conformal
integrals entering Eq. (6.1) and then to match the corresponding short-distance expansion
of the four-point amplitude with perturbative CPWA contributions of operators with known
anomalous dimensions. In this way one can fix the undetermined numerical coefficients in
Eq. (6.1).
The CPWA expansions of the conformal integrals in Eq. (6.1), corresponding to taking
a short-distance limit x1 → x2 , x3 → x4 , have already been obtained in Ref. [23] (cf.
Section 3 in [23]). They are formulated in terms of the conformal cross-ratios v = s/t
and Y = (t − 1)/t. Under the assumption that the conformal dimension ∆ of an operator
is of the form ∆ = ∆0 + γ , where ∆0 and γ are the canonical and the anomalous parts,
respectively, the nonanalytic term v ∆/2 entering the CPWA of this operator gives rise to
perturbative logarithms
1 1
v ∆/2 = 1 + γ ln v + γ 2 ln2 v + · · · .
2 8
Therefore, the information about the one-loop (order λ) anomalous dimensions is reflected
in the coefficients of the ln2 v terms in the perturbative CPWA expansions and this is
how we can use it to derive the form of the two-loop (order λ2 ) amplitude. The explicit
calculations are not particularly instructive and therefore in the following we mainly restrict
ourselves to the formulation of the results obtained.
6.1.1. Operators of twist 2

The R-symmetry singlet of the OPE O(2) O(2) contains an infinite tower of twist 2
operators of increasing spin, the lowest member being the Konishi scalar of canonical
dimension 2 and of one-loop anomalous dimension 3λ. The requirement of reproducing
this field in the OPE derived from Eq. (6.1) fixes two of the four unknown coefficients,
namely,
32
η=ν = . (6.2)
4
Given relation (6.2), the whole tower of twist 2 operators is then correctly reproduced by
the four-point amplitude. However, the operators of twist 2 produce no restrictions on the
coefficients < and ν.
6.1.2. Operators of twist 4 in the singlet channel

To get further restrictions on the four-point amplitude we consider the superconformal
primary operators of twist 4 in the singlet R-symmetry channel. The simplest example
are the four quadrilinear dimension 4 operators Σ1,2 , Σ± studied in Ref. [24] (see also
[5]). These operators diagonalize the matrix of the dilatation operator at one loop with the
following result for the large N one-loop anomalous dimensions11

1 √
γ1 = 0, γ2 = 6λ, 13 ± 41 λ.
γΣ± = (6.3)
4
By using the explicit form of the (canonically normalized) operators Σ1,2 , Σ± established
in Ref. [24] we then compute their free three-point functions with the canonically
normalized 12 -BPS operator O(3) in the large N limit and find
1/2
1 3 18
A± = O (3)
O (3)
Σ± = ∓ √ , (6.4)
N 5 5 41
while O(3) O(3) Σ1,2 ∼ O(1/N 2 ). Such a suppressed behavior of the three-point ampli-
tudes involving Σ1,2 is naturally explained by the fact that these fields are double-trace
operators. Thus, among the four operators only Σ± participate in the large N operator
product expansion of the four-point amplitude. The normalizations of their CPWAs are
given by the constants A± .
Combining the free, the one-loop amplitudes [7] and the two-loop ansatz (6.1), we have
worked out the corresponding CPWA expansion and have found the following system of
equations12 encoding the large N contribution of the fields Σ± :
6
A+ + A− = , (6.5)
5N 2
21 λ
A+ γΣ+ + A− γΣ− = , (6.6)
10 N 2
2
81 8 4 λ
A+ γΣ2 + + A− γΣ2 − = + <+ µ . (6.7)
20 3 3 N2
It is easy to check that the first two equations are indeed satisfied with the values of the A’s
and γ ’s found, while the third equation yields
µ
<=− . (6.8)
2
At this stage only one coefficient, e.g., µ, remains undetermined.
6.1.3. Operators of twist 4 in the irrep [0, 2, 0]

The OPE O(2) O(2) has ten different SO(6) channels. Among them only three may
contain unprotected superconformal primary operators: [0, 0, 0], [0, 2, 0] and [1, 0, 1]. As
we have already exploited the known information about the lowest-dimensional operators
in the singlet, now we move to the superconformal primaries of dimension 4 arising13 in
the irrep [0, 2, 0].
11 At fixed λ the anomalous dimension γ vanishes only when N = ∞ [24].

1
12 These equations are obtained after careful disentangling of the contribution of the superconformal
descendants of the twist 2 fields.
13 There is only one 1 -BPS operator of dimension 2 in the irrep [0, 2, 0] and it does not produce any new
2
restrictions on the four-point amplitude.
There are four superconformal primary operators of canonical dimension 4 which might
contribute to the operator product expansion under study. One of them is the protected
semi-short double-trace operator [25]. The other three are long unprotected operators
Θ, Θ± whose one-loop anomalous dimensions were found in Ref. [26] (see also [27]):
1 √
γΘ = 3λ, γΘ± = 5 ± 5 λ. (6.9)
2
The protected operator as well as Θ are double-traces, therefore we do not expect them
to contribute to the four-point amplitude in the large-N limit. Denoting by B± the
normalizations of the CPWAs of the operators Θ± and proceeding in the same manner
as before, we find the following system of equations:
147
B+ + B− = , (6.10)
4N 2
105λ
B+ γΘ+ + B− γΘ− = , (6.11)
2N 2
35λ2
B+ γΘ2 + + B− γΘ2 − = (9 + 2µ). (6.12)
4N 2
The first two equations are used to find
21 √
B± = 2
7∓3 5 . (6.13)
8N
The normalization constants are positive, as they should (the squares of the normalization
coefficients of the three-point functions). Substituting these values in the third equation we
find µ = 0.
Thus, by exploiting the information about the one-loop anomalous dimensions of some
superconformal primaries, we have completely determined the freedom in the most general
ansatz (6.1):
32
< = µ = 0, η=ν= . (6.14)
4
One can easily see that with these values the amplitude (6.1) indeed coincides with the
function F+(3) found in the previous section.
Finally, we remark that in Ref. [5] the two-loop anomalous dimensions for the operators
Θ± were obtained
√
17 ± 5 5 2
γ± = − λ . (6.15)
8
Working out the logarithmic terms in the two-loop OPE implied by the four-point
amplitude for k = 3 we have checked that Eq. (6.15) is indeed compatible with our findings
provided the one-loop corrections to the normalization constants are

(3) (3) 2 21 √ 1 √
B± = O O Θ± = 7∓3 5 1− 35 ± 3 5 λ . (6.16)
8N 2 20
6.2. Weight 4
Now we obtain some restrictions on the four-point two-loop amplitude for the 12 -BPS
operators of weight 4 from the knowledge of the one-loop anomalous dimensions for the
same operators of twist 2 and 4 that were already discussed in the previous section.
(4)
For k = 4 we expect two functions F± with the symmetry property
1
F±(4) (s, t) = F±(4)(s/t, 1/t).
t
The general ansatzes compatible with these symmetries are

λ2 1 2
F±(4) (s, t) = 2
<± s + η± (t + 1) Φ (1) (s, t)
N 4

µ± (2) 1
+ Φ (t/s, 1/s) + ν± Φ (2) (s, t) + Φ (2) (s/t, 1/t) ,
s t
(6.17)
i.e., each function depends on four unknown coefficients <, η, µ and ν.
6.2.1. Operators of twist 2

This time the Konishi field (as well as its higher-spin cousins) imposes the following
relations
42
η+ = ν+ = = 4, (6.18)
4
while the other coefficients remain invisible. Clearly, these values of η+ , ν+ coincide with
those in Eq. (2.18) obtained by explicit diagrammatic computations.
6.2.2. Operators of twist 4 in the singlet channel

A calculation similar to the case k = 3 produces the following system (after disentan-
gling the contributions of the superconformal descendants of the twist 2 fields):
32
A+ + A− = ,
15N 2
56 λ
A+ γΣ+ + A− γΣ− = ,
15 N 2
2
2 6 λ
A+ γΣ2 + + A− γΣ2 − = 3<+ − + η− + 2µ− + 2ν− + <− , (6.19)
3 5 N2
where γΣ± are given by Eqs. (6.3). Solving this system we find
η− = 12 − 3<+ − 2µ− − 2ν− − <− . (6.20)
This completes our consideration of the singlet channel.

6.2.3. Operators of twist 4 in the irrep [0, 2, 0]

The occurrence of the superconformal primaries Θ± in the double operator product
expansion implied by the four-point amplitude leads to the following system of equations:
2744
B+ + B− = ,
25N 2
784λ
B+ γΘ+ + B− γΘ− = ,
5N 2
98λ2
B+ γΘ2 + + B− γΘ2 − = (η− + 2µ− + 2ν− + <− ). (6.21)
5N 2
It allows us to find
η− = 12 − 2µ− − 2ν− − <− . (6.22)
Comparing the last formula to Eq. (6.20) gives <+ = 0, which coincides with the
corresponding value of the coefficient <+ in Eq. (2.18).
This exhausts the predictive power of the method. The remaining coefficients µ± , ν−
and ρ− can be found only from the diagrammatic computation. If we had the information
about some other superconformal primaries appearing in the OPE for k = 3 or k = 4, we
could use it to further refine our ansatz.
Finally, according to formula (2.18) we have
µ+ = η− = ν− = 0, µ− = <− = 4.
With these values Eq. (6.22) is trivially satisfied.
Acknowledgements
We are grateful to N. Beisert, S. Frolov, S. Kovacs, H. Osborn, J. Plefka, Y. Stanev

and M. Staudacher for many useful discussions. This work has been supported in part by
INFN, MURST and the European Commission RTN program HPRN-CT-2000-00131, in
which G.A. is associated to the University of Bonn, S.P. is associated to the University of
Padova, and A.S. is associated to the University of Torino. The work of G.A. has been also
supported by RFBI grant N02-01-00695.
Appendix A. Harmonic analyticity constraints
In this appendix we develop a set of systematic rules to select the relevant graphs
contributing to a correlation function and find consistency conditions for the combinatorial
coefficients. These rules are based on the general property of harmonic analyticity satisfied
by any correlation function of composite operators in G-analytic harmonic superspace.
Harmonic analyticity for a composite operator Ok = Tr(q̃ n q k−n ) means D ++ Ok = 0,
where D ++ is the harmonic derivative given in (4.17). Correlation functions then satisfy

Di++ Ok (1) · · · Ok (n) = 0, i = 1, . . . , n (A.1)
Fig. 9. Example of an analyticity condition. The arrows point at the propagator hit by D1++ .
(modulo contact terms). At every order in perturbation theory we can implement this
condition directly on Feynman graphs. As a consequence of the identity
←
−
D1++ q̃ + (1)q + (2) = − q̃ + (1)q + (2) D ++
2 = δ(1, 2), (A.2)
where δ(1, 2) is the delta function in the G-analytic harmonic superspace, every time the
derivative hits a HM propagator shrinks it to a point. In particular, when D ++ hits a free
line we have a vanishing contribution (we neglect contact terms), whereas when it hits a
HM line involved in an interaction with a gluon, the effect is to shrink the propagator line
to a point and move the internal gluon vertex to coincide with an external vertex. A simple
example is given in Fig. 9 where analyticity implemented at the vertex 1 of diagram a1
produces the shrunk configuration a2 .
If the same shrunk configuration can be obtained when hitting different graphs (but at
the same point), we can say that all these graphs conspire together to restore harmonic
analyticity at that point. Repeating this procedure, one eventually recovers the whole class
of graphs needed to form a harmonic analytic amplitude.
Again, we exemplify this by taking the simple case of Fig. 9. There, the same shrunk
configuration is obtained both from diagrams a1 and a3 . Since these are the only two
diagrams which give the configuration a2 when hit at the point 1, the sum of their
contributions must result into an expression which is analytic in the 1 variable. In fact,
using Eq. (4.14) their contributions read

a1 → Ca1 [13][42] [21− ]ρ12 + [12− ]ρ22 − 2(ρ1 ρ2 )

× [43− ]ρ32 + [34− ]ρ42 − 2(ρ3 ρ4 ) ,

a3 → Ca3 [12][42] [31− ]ρ12 + [13− ]ρ32 − 2(ρ1 ρ3 )

× [43− ]ρ32 + [34− ]ρ42 − 2(ρ3 ρ4 ) , (A.3)
where Cai are the combinatorial factors (with possible signs included). The two expres-
sions are nonanalytic in the 1 variable because of their nontrivial dependence on u− 1 . The
action of D1++ on the sum of the two contributions must give zero, according to the con-
dition (A.1). Remembering that D1++ only acts on the u− +
1 harmonic converting it into u1 ,
we obtain

(Ca1 + Ca3 )[42][13][21] [43− ]ρ12 ρ32 + [34− ]ρ12 ρ42 = 0 (A.4)
which implies
Ca1 = −Ca3 . (A.5)
Using this condition, the sum of the two contributions in (A.3) reads (we consider only the
part depending on 1− )

[42] [13][21−] − [12][31−] [43− ]ρ12 ρ32 + [34− ]ρ12 ρ42 . (A.6)
The harmonic cyclic identity
[13][21−] − [12][31−] = [23] (A.7)
shows that the nonanalytic contributions from the two diagrams cancel against each other
leaving an analytical result in the 1 variable.
As appears from this simple example, imposing analyticity conditions allows to draw
systematically all the diagrams which at a given order contribute to the analytic final result
for the correlation function and to find nontrivial relations like (A.5) which constrain the
combinatorial coefficients. These relations can be used as a check when the coefficients are
directly computed from the diagrams.
We notice that the final analytic result can be obtained by using a shortcut procedure:
given the sum of the two 1− dependent contributions in (A.3) we make the harmonics
identification 1 ≡ 2. As a consequence, the contribution from a3 vanishes whereas a1 gives

Ca1 [42][13][21−] [43− ]ρ12 ρ32 + [34− ]ρ12 ρ42

⇒ Ca1 [42][23] [43− ]ρ12 ρ32 + [34− ]ρ12 ρ42 ,
1≡2
(A.8)
which is the correct result. The identification of the harmonics we chose is not the only
possible one. We could have made the alternative identification 1 ≡ 3. In this case the
contribution from diagram a1 vanishes whereas from diagram a3 we obtain

Ca3 [12][42][31−] [43− ]ρ12 ρ32 + [34− ]ρ12 ρ42

⇒ −Ca3 [23][42] [43− ]ρ12 ρ32 + [34− ]ρ12 ρ42 ,
1≡2
(A.9)
which again is the correct result.
We now consider the cases of our interest. We start by studying the analyticity
conditions for the set of diagrams of Fig. 8 in the main text. They have in common two free
lines which give the analytic structure [12][43] and a “base” T245T246 :

T245 T246 = [24− ][42−] ρ42 σ22 + ρ22 σ42 + 4(ρ2 ρ4 )(σ2 σ4 )

= τ24 − [24][2−4− ] ρ42 σ22 + ρ22 σ42 . (A.10)
It is easy to see that leaving aside the base as well as the free propagators leads to the
following harmonic and Grassmann structures:

(8a): [13]2[T125 T436 + 5 ↔ 6] = [13]2[21− ][43− ] ρ12 σ32 + ρ32 σ12 ,

(8b): [13][43][T125T136 + 5 ↔ 6] = [13][43][21−][13− ] ρ12 σ32 + ρ32 σ12 ,

(8c): [13][12][T435T136 + 5 ↔ 6] = [13][12][43−][31− ] ρ12 σ32 + ρ32 σ12 ,
(8d): [12][43][T135T136 + 5 ↔ 6]

= 2[12][43] − [13][1− 3− ] ρ12 σ32 + ρ32 σ12 + τ13 . (A.11)
The factor 2 in the graph (8d) is due to the symmetrization 5 ↔ 6. Note that this is the
only graph from this set which gives an analytic contribution proportional to τ13 . The
other Grassmann term ρ12 σ32 + ρ32 σ12 always appears with nonanalytic harmonic factors.
Collecting all such terms we find

[13] Ca [12][43−][31− ] + Cb [13][21−][43− ]

+ Cc [43][21−][13− ] − 2Cd [12][43][1−3− ] , (A.12)
where we have introduced combinatorial factors Ca , . . . , Cd . We want to achieve harmonic
analyticity at points 1 and 3. According to the rule described above, we hit the graphs with
harmonic derivatives D1++ and D3++ .
We begin by applying D1++ to the expression (A.12) to obtain

[13][12] (−Cb − 2Cd )[43][13−] + (Ca + Cc )[43− ][31] = 0. (A.13)
Since the two harmonic structures within the brackets are linearly independent, we have
the following two conditions
Cb = −2Cd , Ca = −Cc . (A.14)
Similarly, hitting (A.12) with D3++
we obtain

[13][43] (Ca + Cb )[21− ][13] + (Cc + 2Cd )[12][31−] = 0 (A.15)
which implies
Cc = −2Cd , Ca = −Cb . (A.16)
The last set of conditions is consistent with (A.14) if Cb = Cc . This was somehow expected
since graphs (8b) and (8c) only differ by a reflection (1 ↔ 3, 2 ↔ 4) which should not
affect the combinatorial factor.
We can solve the constraints (A.14) ending with only one arbitrary coefficient, e.g., Cd .
With this choice for the coefficients the expression (A.12) becomes harmonic analytic with
the help of the identity
[43][21−][13−] + [12][43−][31− ]
− [13][21−][43−] + [12][43][1−3− ] = [42]. (A.17)
Finally, taking all this into account, attaching the “base” and restoring the missing
space–time factors, we find the total contribution of the graphs in Fig. 8 to be
2Cd [12][43] 2 2
Total Fig. 8: 4 x4 x4 x4
− [13][42] ρ1 σ3 + ρ3
2 2
σ1 + [12][43]τ 13
x12 13 24 34

× −[24][2−4− ] ρ42 σ22 + ρ22 σ42 + τ24 . (A.18)
We see that analyticity at points 1 and 3 has indeed been achieved but not at points 2 and 4,
of course.
Fig. 10. Graphs conspiring to harmonic analyticity at points 1 and 4.
As in the previous example, the correct result can be obtained by using harmonics
identification. In fact, pulling out the free line contribution [12][43], we can identify the
harmonics 1 ≡ 2 and 3 ≡ 4. As a consequence, in Eq. (A.12) the contributions from
diagrams a, c, d vanish and we are left with
1≡2,3≡4
Cb [13]2[21− ][43− ] ρ12 σ32 + ρ32 σ12 ⇒ Cb [13]2 ρ12 σ32 + ρ32 σ12 . (A.19)
Since the final result must be analytic, we can find an analytic term which under the above
identification brings to the same expression
1≡2,3≡4
−Cb [13][42] ρ12 σ32 + ρ32 σ12 ⇒ Cb [13]2 ρ12 σ32 + ρ32 σ12 . (A.20)
Therefore, this must be the correct term which appears in the final result and it is indeed
what we found in (A.18).
We notice that using this shortcut we can easily figure out the complete analytic structure
for the sum of diagrams in Fig. 8. In fact, analyticity at points 2 and 4, which should be
restored by adding another set of partner graphs, can rather be obtained by identifying
2 ≡ 4. This identification leaves the single term τ24 in the bottom part of Eq. (A.18) which
becomes completely analytic.
We now consider the more complicated situation of graphs in Fig. 10. By imposing
analyticity at points 1 and 4 we can find consistency conditions between the coefficients
of the graphs (10c) which have been checked independently in Section 6, and the
coefficients (10a) and (10b) which do not have any independent check. These conditions
will necessarily involve new graphs contributing only to nonanalytic structures, necessary
in order to cancel similar terms coming from the graphs (10a), (10b) and (10c).
We start by considering the graph (a) and apply the outlined procedure in order to
achieve harmonic analyticity at points 1 and 4. It is easy to realize that in the large N
limit, i.e., neglecting nonplanar graphs, the set of conspiring graphs is just the one showed
in Fig. 10.
As in the previous example, we write the harmonic and Grassmann factors coming from
these graphs:
(a): [12][13][42][43][T135T126 T425T436 + 5 ↔ 6]

−→ [12][13][42][43]

× [31− ][12− ][24−][43− ] + [13− ][21−][42− ][34− ]

× ρ12 σ22 σ32 ρ42 + σ12 ρ22 ρ32 σ42 ,
(b): [12][13][42][43][T136T125 T425T436 + 5 ↔ 6]
−→ [12][13][42][43]

× [31− ][12− ][24− ][43−] ρ12 σ22 ρ32 σ42 + σ12 ρ22 σ32 ρ42

+ [13− ][21− ][42−][34− ] σ12 σ22 ρ32 ρ42 + ρ12 ρ22 σ32 σ42 ,
(c1 ): [12]2[43]2[T135 T136T425 T426 + 5 ↔ 6]

−→ 2[12]2[43]2 − [13][1−3− ] ρ12 σ32 + ρ32 σ12 + τ13

× −[24][2−4− ] ρ22 σ42 + ρ42 σ22 + τ24 ,
(c2 ): same as (c1 ) with 2 ↔ 3 or 1 ↔ 4,

(d1 ): [12]2[13][42][43]2 (T T )1563(T T )4652 + 5 ↔ 6

−→ [12]2[13][42][43]2[1− 3− ][4− 2− ] ρ12 ρ22 σ32 σ42 + σ12 σ22 ρ32 ρ42 ,
(d2 ): same as (d1 ) with 2 ↔ 3 or 1 ↔ 4,

(e1 ): [12][13][42][43]2 T135 T126(T T )4562 + 5 ↔ 6

−→ [12][13][42][43]2[2− 4− ][13−][21− ] σ12 σ22 ρ32 ρ42 + ρ12 ρ22 σ32 σ42 ,
(e2 ): same as (e1 ) with 2 ↔ 3,
(e3 ): same as (e1 ) with 1 ↔ 4,
(e4 ): same as (e1 ) with 2 ↔ 3 and 1 ↔ 4,
(f1 ): [12][13][43]2[T135T126 T425 T426 + 5 ↔ 6]

−→ [12][13][43]2[13− ][21− ] σ12 ρ32 + ρ12 σ32

× −[24][2−4− ] ρ22 σ42 + ρ42 σ22 + τ24 ,
(f2 ): same as (f1 ) with 2 ↔ 3,

(f3 ): same as (f1 ) with 1 ↔ 4,
(f4 ): same as (f1 ) with 2 ↔ 3 and 1 ↔ 4. (A.21)
We concentrate on the contributions to the three independent structures not containing τ
factors
U ≡ ρ12 ρ22 σ32 σ42 + σ12 σ22 ρ32 ρ42 ,
V ≡ ρ12 σ22 ρ32 σ42 + σ12 ρ22 σ32 ρ42 ,
Z ≡ ρ12 σ22 σ32 ρ42 + σ12 ρ22 ρ32 σ42 . (A.22)
They are:

(a): Ca [12][13][42][43] [31− ][12− ][24−][43− ] + [13− ][21−][42− ][34− ] Z,

(b): Cb [12][13][42][43] [31− ][12− ][24− ][43−]V + [13− ][21−][42− ][34− ]U ,
(c1 ): 2Cc [12]2[43]2[13][1−3− ][24][2−4− ](U + Z),
(c2 ): same as (c1 ) with 2 ↔ 3 or 1 ↔ 4,
(d1 ): Cd [12]2[13][42][43]2[1− 3− ][4− 2− ]U,
(d2 ): same as (d1 ) with 2 ↔ 3 or 1 ↔ 4,
(e1 ): Ce [12][13][42][43]2[2− 4− ][13− ][21− ]U,
(e2 ): same as (e1 ) with 2 ↔ 3,
(e3 ): same as (e1 ) with 1 ↔ 4,
(e4 ): same as (e1 ) with 2 ↔ 3 and 1 ↔ 4,
(f1 ): Cf [12][13][43]2[13− ][21− ][42][2−4− ](U + Z),
(f2 ): same as (f1 ) with 2 ↔ 3,
(f3 ): same as (f1 ) with 1 ↔ 4,
(f4 ): same as (f1 ) with 2 ↔ 3 and 1 ↔ 4. (A.23)
We have assumed that diagrams which differ by the exchange 2 ↔ 3 or/and 1 ↔ 4 have
the same combinatorial factor.
By acting with D1++ and setting to zero the independent structure proportional to U we
get

[12][13][42][43] [12][43][13−][2− 4− ](−2Cc − Cd − Ce − Cf )

+ [12][34−][13− ][42− ](−Cb − Ce − Cf ) = 0. (A.24)
Doing the same for the structure proportional to V gives an equation like (A.24) with
2 ↔ 3, then it will not impose any new condition on the coefficients. The part proportional
to Z gives instead

[12][13][42][43] [12][43][13−][2− 4− ](−2Cc − Cf )

+ [12][34−][13−][42− ](−Ca − Cf ) + 2 ↔ 3 = 0. (A.25)
By now setting to zero the two linearly independent analytic structure appearing in (A.24)
and (A.25) we get the following conditions relating the coefficients of the graphs of Fig. 10
Ca = 2Cc = −Cf , Cb = Ca − Ce = Ca + Cd . (A.26)
We notice that the condition Cf = −2Cc has been already obtained in (A.16) (diagrams
(10c) and (10f) are the same as (8d) and (8c), respectively).
It is now easy to see that there are no extra conditions coming from the terms containing
one τ factor. The same is true if we impose the analyticity condition at point 4. So we
conclude that (A.26) is the full set of conditions imposed by the requirement of partial
analyticity at points 1 and 4 on the graphs in Fig. 10.
Using the relations (A.26) one can check that the total contribution obtained by
summing the expressions in (A.21) ends up to be analytic in the 1 and 3 variables.
Again, as in the previous examples, the final answer can be obtained by suitable
identification of the harmonics.
To summarize, what we have leaned from the examples described above is the
following. On one hand, imposing harmonic analyticity on the final result allows to
find consistency conditions among the coefficients of different sets of graphs. On the
other hand, once we give analyticity for granted, we can bypass the whole procedure
of evaluating all diagrams conspiring to a final analytic result by performing harmonics
identification, so drastically reducing the number of graphs to be computed.
Appendix B. Sample calculations: k = 3 and k = 4
In this appendix we describe in detail the calculation of the Am functions (see Eq. (5.1))
for the cases k = 3 and k = 4. In particular, we show how the procedure of identifying the
harmonics can be used both to unambiguously select the diagrams which contribute to a
given function and to drastically reduce the number of diagrams one needs compute.
B.1. The case k = 3
The expected form of the complete amplitude for k = 3 case is:

(3) 4
+ 2 [13][42] [12][43]

Q L|Lθ5,6 =0 ∼ θr A0 (x) + 2 2 A1 (x) . (B.1)
2 x2
x13 x12 x43
r=1 42
As discussed in the main text, the two functions A0 , A1 are related by crossing symmetry,
and we need compute only one of them. To calculate for example A0 , it is sufficient to
look at contributions proportional to the harmonic structure [13][42] in (B.1). They can be
unambiguously selected by identifying the harmonics 1 ≡ 2 and 3 ≡ 4 in (B.1)

(3) 4
+ 2 [13]2
Q L|Lθ5,6 =0 ∼ θr 2 x2
A 0 (x) . (B.2)
r=1
x13 42
When looking at perturbative contributions at two loops, the harmonics identification

implies the vanishing of all graphs containing:
(i) at least one free line 1 → 2 or 3 ← 4,

(ii) the blocks T T1562 or T T3564 ,
(iii) the blocks T125T125 and alike.
Discarding also the graphs which vanish by simple theta counting (see the general
discussion in Section 5), the relevant surviving diagrams are then shown in Fig. 11. Upon
the above harmonic identification, they already contain the analytic factor [13]2 required
in Eq. (B.2). Since the coefficient function A0 is independent of the harmonics, once we
have pulled [13]2 out, we can go a step further and identify the harmonics 1 ≡ 3 in the rest
of the expression.
As a consequence, the contributions from graphs (11c) and (11d) drop out. In the rest the
harmonic dependence reduces to a common factor and, neglecting the space–time structure,
the final contribution of each diagram is given by its combinatorics
2Ca(11) = 2 · 34 N 6 , Cb(11) = Ce(11) = Cf(11) = 34 N 6 ,

Cg(11)
1
= Cg(11)
2
= 2 · 34 N 6 , 2Ch(11) = 34 N 6 . (B.3)
(11) (11)
The extra factor 2 in front of Ca and Ch is due to the symmetry of the corresponding
diagrams under the exchange 5 ↔ 6.
In this simple example it is already clear how one can exploit the condition of harmonic
analyticity of the final result. It is in fact this condition which allowed us to make the
identification 1 ≡ 3, thus reducing the number of Feynman diagrams to be computed.
In a complete calculation performed without identifying the harmonics, the diagrams
which vanished under this identification would contribute only to cancel nonanalytic terms
coming from the rest of the diagrams.
B.2. The case k = 4
We now consider the more complicated case k = 4. There we have

(4)
Q L|Lθ5,6 =0

4
+ 2 [13][42] 2 [13][42] [12][43]
∼ θr 2 x2
A0 (x) + 2 x2 2 2
A1 (x)
r=1
x13 42 x13 42 x12 x43

[12][43] 2
+ 2 x2
A 2 (x) . (B.4)
x12 43
This correlator is crossing symmetric under the exchange 2 ↔ 3 which relates the
coefficient functions A0 and A2 to each other, while A1 is crossing symmetric and
independent. Therefore, we need to perform the calculation in two independent channels.
To select the relevant graphs which contribute in the two channels we observe that in
the k = 4 case we have 8 HM lines and 2 gluons, which can at most connect 4 HM lines.
Therefore, at least 4 HM lines remain free. If these free lines form a free corner, i.e., they all
come out from the same vertex, we know that such graphs vanish because of theta counting
Fig. 11. Relevant graphs for the k = 3 case.
(see the general discussion in Section 5). It follows that the free lines have to form at least
one disconnected pair.
If the free pair is [13][42], such graphs can contribute to the functions A0 , A1 in
Eq. (B.4) but not to A2 . Then we can write down the contribution of all such graphs with a
free pair [13][42] in the form

[13][42] [13][42] [12][43]
2 x2 2 x2
A0 (x) + 2 2 A1 (x) . (B.5)
x13 42 x13 42 x12 x43
[13][42]
Pulling out the factor 2 x2 is equivalent to removing the corresponding free lines from
x13 42
the graphs, which results in a configuration with k = 3. Then we can go a step further and
identify the harmonics pairwise within the brackets. If we identify 1 ≡ 2, 3 ≡ 4 and the
graph does not vanish, then it contributes to A0 . If instead we identify 1 ≡ 3, 2 ≡ 4 and
the graph does not vanish, then it contributes to A1 . Note that it may happen that the graph
vanishes under each of these identifications, then it should be discarded.
Fig. 12. Relevant graphs for k = 4. The thick lines are free HM propagators whereas thin lines contain vertices of
interaction with gluons.
Alternatively, if the free pair is [12][34], Eq. (B.5) is replaced by

[12][43] [13][42] [12][43]
2 x2 2 x2
A 1 (x) + 2 x2
A 2 (x) . (B.6)
x12 43 x13 42 x12 43
Once again, removing the free lines we obtain a k = 3 configuration in which we identify
the harmonics pairwise within the brackets. If we identify 1 ≡ 2, 3 ≡ 4 and the graph does
not vanish, then it contributes to A1 . If instead we identify 1 ≡ 3, 2 ≡ 4 and the graph does
not vanish, then it contributes to A2 .
We now use this strategy to select the relevant diagrams which eventually contribute to
the coefficient functions. We carefully draw all the diagrams which are nonvanishing for
theta counting and select the ones which survive under one of the harmonics identifications
described above. We are then led to the set of graphs on Fig. 12 where, in order to reduce
the number of diagrams, we have indicated with thick lines all free HM propagators,
whereas thin lines are interacting (they contain an interaction vertex and gluons connect
these vertices in all possible ways). All the graphs have 4 free lines and they have been
Fig. 13. Diagrams contributing to the A0 function for k = 4 case.
organized according to the number of free lines coming out of a single point. We have not
drawn those diagrams that vanish after harmonics identification (the analog of graphs (11c)
and (11d) of the previous example). As already explained, in an exhaustive calculation
done without identifying harmonics these diagrams would conspire to cancel nonanalytic
contributions from the rest.
We now apply the harmonics identification procedure to figure out which diagrams
contribute to which function. Let us analyze in detail the graphs in Fig. 12(a). The first kind
of graphs contain a free pair [13][42], so they belong to the type of Eq. (B.5). According
to the rule above, we pull out the free factor and then we identify the harmonics pairwise.
The identification 1 ≡ 3, 2 ≡ 4 annihilates the graphs, so they cannot contribute to A1 .
On the contrary, the graphs can survive the identification 1 ≡ 2, 3 ≡ 4, so they contribute
to A0 . If we now draw the gluons, we see that only one configuration is allowed and the
graph is reduced to (11e) from the case k = 3. The same argument applies to the second
kind of graphs (12a), reducing them to a single structure like (11f). Similarly, the third
and the fourth graphs (12a) contain a free pair [12][43], so they belong to the type of
Eq. (B.6). According to the rule above, we pull out the free factor and then we identify the
harmonics pairwise. The identification 1 ≡ 2, 3 ≡ 4 annihilates the graphs, so they cannot
contribute to A1 . On the contrary, the graphs survive the identification 1 ≡ 3, 2 ≡ 4, so they
contribute to A2 . If we now draw the gluons, we see that these graphs are reduced to (11e),
(11f) rotated by 90◦ .
Analyzing all the graphs along the same lines we eventually identify the coefficient
functions they contribute to. The complete identification is indicated in Fig. 12. Drawing
the gluon lines brings us to the diagrams in Fig. 13 for A0 and Fig. 14 for A1 .
A couple of comments are now in order. First of all we notice that the configurations
of gluon lines appearing in Figs. 13 and 14 are the same as the ones we have already
Fig. 14. Diagrams contributing to the A1 function for k = 4 case.
selected for k = 3. Moreover, a diagram with a given configuration of interacting lines can
contribute to different functions according to the structure of free propagators which dress
it. In fact, all the allowed configurations of interacting lines are present for both functions.
Now, computing the contribution of each graph is simply a matter of combinatorics. At
large N , neglecting the space–time structure, for the A0 function we obtain
Ca(13) = Cb(13) = 44 N 8 , 2Cc(13) = 44 N 8 , Cd(13) = 44 N 8 ,

(13)
Ce(13)
1
= Ce(13)
2
= 2 · 44 N 8 , 2Cf = 2 · 44 N 8 (B.7)
whereas for A1
(14) (14)
Ca(14) = Cb = Cc(14) = Cd = 44 N 8 ,
2Ce(14) = 44 N 8 , 2Cf(14) = 44 N 8 , Cg(14) = 44 N 8 ,
(14) (14)
Ch1 = Ch2 = 2 · 44 N 8 . (B.8)
It is important to notice that, at large N , the combinatorial factors only depend on

the structure of the interacting lines: diagrams with the same interactions but different
configurations of free propagators have the same coefficient.
References
[1] J. Maldacena, The large N limit of superconformal field theories and supergravity, Adv. Theor. Math. Phys. 2
(1998) 231;
G.G. Gubser, I.R. Klebanov, A.M. Polyakov, Gauge theory correlators from noncritical string theory, Phys.
Lett. B 428 (1998) 105, hep-th/9802109;
E. Witten, Anti-de Sitter space and holography, Adv. Theor. Math. Phys. 2 (1998) 253, hep-th/9802150.
[2] A.M. Polyakov, Gauge fields and space–time, Int. J. Mod. Phys. A 17 (S1) (2002) 119, hep-th/0110196.
[3] D. Berenstein, J.M. Maldacena, H. Nastase, Strings in flat space and pp waves from N = 4 super-Yang–
Mills, JHEP 0204 (2002) 013, hep-th/0202021.
[4] S.S. Gubser, I.R. Klebanov, A.M. Polyakov, A semi-classical limit of the gauge/string correspondence, Nucl.
Phys. B 636 (2002) 99, hep-th/0204051.
[5] N. Beisert, C. Kristjansen, M. Staudacher, The dilatation operator of N = 4 super-Yang–Mills theory, hep-
th/0303060.
[6] J.A. Minahan, K. Zarembo, The Bethe-ansatz for N = 4 super-Yang–Mills, hep-th/0212208.
[7] G. Arutyunov, E. Sokatchev, On a large N degeneracy in N = 4 SYM and the AdS/CFT correspondence,
hep-th/0301058.
[8] O. Aharony, S.S. Gubser, J.M. Maldacena, H. Ooguri, Y. Oz, Large N field theories, string theory and
gravity, Phys. Rep. 323 (2000) 183, hep-th/9905111.
[9] E. D’Hoker, D.Z. Freedman, Supersymmetric gauge theories and the AdS/CFT correspondence, hep-
th/0201253.
[10] G. Arutyunov, F.A. Dolan, H. Osborn, E. Sokatchev, Correlation functions and massive Kaluza–Klein modes
in the AdS/CFT correspondence, hep-th/0212116.
[11] P.J. Heslop, P.S. Howe, Four-point functions in N = 4 SYM, JHEP 0301 (2003) 043, hep-th/0211252.
[12] A. Galperin, E. Ivanov, S. Kalitsyn, V. Ogievetsky, E. Sokatchev, Unconstrained N = 2 matter, Yang–Mills
and supergravity theories in harmonic superspace, Class. Quantum Grav. 1 (1984) 469;
A. Galperin, E. Ivanov, V. Ogievetsky, E. Sokatchev, Harmonic Superspace, Cambridge Univ. Press,
Cambridge, UK, 2001.
[13] S. Penati, A. Santambrogio, D. Zanon, Two-point functions of chiral operators in N = 4 SYM at order g 4 ,
JHEP 9912 (1999) 006, hep-th/9910197;
S. Penati, A. Santambrogio, D. Zanon, More on correlators and contact terms in N = 4 SYM at order g 4 ,
Nucl. Phys. B 593 (2001) 651, hep-th/0005223.
[14] M. Bianchi, S. Kovacs, G. Rossi, Y.S. Stanev, Anomalous dimensions in N = 4 SYM theory at order g 4 ,
Nucl. Phys. B 584 (2000) 216, hep-th/0003203.
[15] B.U. Eden, P.S. Howe, A. Pickering, E. Sokatchev, P.C. West, Four-point functions in N = 2 superconformal
field theories, Nucl. Phys. B 581 (2000) 523, hep-th/0001138.
[16] F.A. Dolan, H. Osborn, Conformal four point functions and the operator product expansion, Nucl. Phys.
B 599 (2001) 459, hep-th/0011040;
F.A. Dolan, H. Osborn, Superconformal symmetry, correlation functions and the operator product expansion,
Nucl. Phys. B 629 (2002) 3, hep-th/0112251.
[17] B. Eden, A.C. Petkou, C. Schubert, E. Sokatchev, Partial non-renormalisation of the stress-tensor four-point
function in N = 4 SYM and AdS/CFT, Nucl. Phys. B 607 (2001) 191, hep-th/0009106.
[18] B. Eden, C. Schubert, E. Sokatchev, Three-loop four-point correlator in N = 4 SYM, Phys. Lett. B 482
(2000) 309, hep-th/0003096.
[19] G. Arutyunov, S. Frolov, On the correspondence between gravity fields and CFT operators, JHEP 0004
(2000) 017, hep-th/0003038.
[20] P.S. Howe, C. Schubert, E. Sokatchev, P.C. West, Explicit construction of nilpotent covariants in N = 4
SYM, Nucl. Phys. B 571 (2000) 71, hep-th/9910011.
[21] B. Eden, C. Schubert, E. Sokatchev, Three-loop four-point correlator in N = 4 SYM, Phys. Lett. B 482
(2000) 309, hep-th/0003096;
B. Eden, C. Schubert, E. Sokatchev, Four-point functions of chiral primary operators in N = 4 SYM, Talk
given at “Quantization, Gauge Theory and Strings”, Moscow, June 5–10, 2000, hep-th/0010005.
[22] B. Eden, C. Schubert, E. Sokatchev, unpublished.
[23] G. Arutyunov, B. Eden, A.C. Petkou, E. Sokatchev, Exceptional non-renormalization properties and OPE
analysis of chiral four-point functions in N = 4 SYM(4), Nucl. Phys. B 620 (2002) 380, hep-th/0103230.
[24] G. Arutyunov, S. Penati, A.C. Petkou, A. Santambrogio, E. Sokatchev, Non-protected operators in N = 4
SYM and multiparticle states of AdS(5) SUGRA, Nucl. Phys. B 643 (2002) 49, hep-th/0206020.
[25] G. Arutyunov, S. Frolov, A.C. Petkou, Operator product expansion of the lowest weight CPOs in N = 4
SYM(4) at strong coupling, Nucl. Phys. B 586 (2000) 547, hep-th/0005182;
G. Arutyunov, S. Frolov, A.C. Petkou, Perturbative and instanton corrections to the OPE of CPOs in N = 4
SYM(4), Nucl. Phys. B 602 (2001) 238, hep-th/0010137.
[26] M. Bianchi, B. Eden, G. Rossi, Y.S. Stanev, On operator mixing in N = 4 SYM, Nucl. Phys. B 646 (2002)
69, hep-th/0205321.
[27] N. Beisert, BMN operators and superconformal symmetry, hep-th/0211032.
The cosmological constant and domain walls in

orientifold field theories and N = 1 gluodynamics
A. Armoni a , M. Shifman a,b
a Theory Division, CERN, CH-1211 Geneva 23, Switzerland
b William I. Fine Theoretical Physics Institute, University of Minnesota, Minneapolis, MN 55455, USA 1
Received 20 March 2003; received in revised form 30 June 2003; accepted 4 August 2003
Abstract
We discuss domain walls and vacuum energy density (cosmological constant) in N = 1
gluodynamics and in non-supersymmetric large N orientifold field theories which have been recently
shown to be planar equivalent (in the boson sector) to N = 1 gluodynamics. A relation between
the vanishing force between two parallel walls and vanishing cosmological constant is pointed out.
This relation may explain why the cosmological constant vanishes in the orientifold field theory at
leading order although the hadronic spectrum of this theory does not contain fermions in the limit
N → ∞. The cancellation is among even and odd parity bosonic contributions, due to NS–NS and
R–R cancellations in the annulus amplitude of the underlying string theory. We use the open–closed
string channel duality to describe interaction between the domain walls which is interpreted as the
exchange of composite “dilatons” and “axions” coupled to the walls. Finally, we study some planar
equivalent pairs in which both theories in the parent–daughter pair are non-supersymmetric.
PACS: 11.15.-q; 11.25.-w; 11.15.Tk
1. Introduction
Domain walls are BPS objects which appear in N = 1 supersymmetric (SUSY)

gluodynamics [1]. If the gauge group is SU(N), there are N distinct discrete vacua labeled
E-mail addresses: adi.armoni@cern.ch (A. Armoni), michael.shifman@cern.ch (M. Shifman).

1 Permanent address.
doi:10.1016/j.nuclphysb.2003.08.004
A. Armoni, M. Shifman / Nuclear Physics B 670 (2003) 148–160 149
by the order parameter, the gluino condensate,

2πk
λλk = NΛ exp i
3
, k = 0, 1, 2, . . . , N − 1. (1)
N
The domain wall W{k,k+1} interpolates between the kth and k + 1 vacua. Moreover, at
N → ∞ two parallel domain walls W{k,k+1} and W{k+1,k+2} are also BPS—there is neither
attraction nor repulsion between them.
It is known that the BPS domain walls in N = 1 gluodynamics present a close parallel
to D-branes in string theory [2]. In particular, a fundamental flux tube can end on the BPS
domain wall, similarly to F1 ending on D-branes in string theory [2] (see also [3–7]). The
purpose of this paper is three-fold. First, we show that this parallel can be further extended.
In string theory the absence of forces between parallel D-branes is due to a cancellation
between the interactions induced by NS–NS and R–R charges. We show how a similar
cancellation works in N = 1 gluodynamics. This parallel yields an important insight
revealing a relation between the vanishing of the cosmological constant and cancellation
of forces. This observation will be used later.
Second, we will extend this parallel to a non-supersymmetric gauge field theory.
Recently, we discussed a non-supersymmetric theory, where exact results on the strong
coupling regime could be obtained [8]. The theory, named “orientifold field theory”, is
a daughter of N = 1 gluodynamics. The parent–daughter relationship is understood in
the sense of [9]. The parent theory is N = 1 gluodynamics with the gauge group U(N).
The daughter theory also has U(N) gauge group, the same gauge coupling as the parent
one, and the fermion sector consisting of one Dirac fermion in the antisymmetric tensor
representation.
The advantage of the orientifold daughter compared to orbifold discussed by Strassler
[9] is the absence of the twisted sector in the former. The non-perturbative planar
equivalence between N = 1 gluodynamics and its orbifold, conjectured by Strassler, was
questioned in the literature (see, e.g., Refs. [10,11]), with the twisted sector of the orbifold
theory being the main suspect. The non-perturbative planar equivalence between N = 1
gluodynamics and its orientifold daughter was shown [8] to be on a much more solid
theoretical footing. We argued that the orientifold gauge theory, at large N , contains N
degenerate vacua, has a bifermion condensate which serves as an order parameter, much in
the same way as the gluino condensate, Eq. (1). Another finding was the vanishing of the
cosmological constant at order N 2 . These results seem to be surprising since the hadronic
spectrum of the orientifold theory is purely bosonic.
The orientifold theory has N discrete degenerate vacua. Hence, one can expect domain
walls. Indeed, the daughter inherits domain walls from its supersymmetric parent. Two
parallel walls of the type W{k,k+1} and W{k+1,k+2} are at indefinite equilibrium. In this
sense they are “BPS”, although the standard definition of “BPS-ness”, through central
charges and supercharges, is certainly not applicable in the non-supersymmetric theory.
In this paper we elaborate on physics of the domain walls in the orientifold theory. We will
show that these walls carry charges similar to NS–NS and R–R charges. In addition, we
will argue that an open–closed string channel duality holds for the analogous field theory
annulus amplitude. Moreover, by exploiting the similarity between string theory and field
150 A. Armoni, M. Shifman / Nuclear Physics B 670 (2003) 148–160
theory we will provide a reason why the cosmological constant of the gauge theory is zero
at order N 2 despite the fact that the hadronic spectrum of the theory contains only bosons.
Finally, in the third part we explain how the parent–daughter relationship (non-
perturbative planar equivalence) between N = 1 gluodynamics and its orientifold can be
extended to include pairs of theories none of which is supersymmetric.
2. The orientifold field theory
This “orientifold field theory” was suggested in Refs. [12,13] in a somewhat different
context. The field content of the orientifold gauge field theory differs from the one of its
parent theory, U(N) SUSY gluodynamics, in that the gluinos are replaced by one massless
Dirac fermions in the rank-two antisymmetric tensor representation of U(N) (denoted by
+ ). The total number of (say) left-handed fermions is thus N(N − 1) in the daughter
theory and N 2 in the parent theory, which agrees to leading order in 1/N . The realization
of the orientifold field theory in string theory is as follows: this theory lives on a brane
configuration of type 0A string theory [12] which consists of NS5-branes, D4-branes and
an orientifold plane—hence the name “orientifold field theory”.
The massless open strings on the brane correspond to the ultraviolet (UV) degrees of
freedom of the field theory: the gauge field and the antisymmetric fermion.
We will assume that our gauge theory has a string theory dual in the spirit of Ref. [14]
(yet to be found, though). It is presumably of the type 0B on a curved background, similarly
to the orientifold field theory analog of N = 4 SYM which is type 0B on AdS5 × RP5
[13]. In this picture the closed strings correspond to the infrared (IR) degrees of freedom,
the glueballs and “quarkonia”. Indeed, the type 0 (closed) strings are purely bosonic, in
agreement with our expectation from the confining orientifold field theory. Moreover, the
bosonic IR spectrum of the gauge theory is even/odd parity degenerate, in accordance with
degeneracies between the NS–NS and R–R towers of the type 0 string.
3. Parallel domain walls versus D-branes
As was mentioned, the gauge theory fundamental flux tubes can end on a BPS domain
wall. Let us assume that for N = 1 gluodynamics/orientifold theory the domain walls
have a realization in terms of Dp-branes (p > 1) of the corresponding type IIB/0B string
theory. Their world volume is 012 + (p − 2) directions transverse to the four-dimensional
space–time 0123. Specific AdS/CFT realizations of domain walls in N = 1 theories are
given in Refs. [5–7,15], mostly in terms of wrapped D5-branes. We deliberately do not
specify which particular branes are used to model the BPS walls, since we do not perform
actual AdS/CFT calculations. D-branes carry the NS–NS charge, as well as the R–R charge
[16]. Moreover, interactions induced by these charges exactly cancel guaranteeing that the
parallel D-branes neither attract nor repel each other.
Let us see how this is realized in N = 1 SYM. We start from two parallel BPS walls
W{k,k+1} and W{k+1,k+2} . Each of them is BPS, with the tension [1]
N
2 sin π .
T{k,k+1} = T{k+1,k+2} = tr λλ (2)
8π 2 N
The tension of the configuration W{k,k+2} is
N 2π
T{k,k+2} = 2
tr λλ2 sin . (3)
8π N
This means that at leading order in N (i.e., N 1 ) two parallel walls W{k,k+1} and W{k+1,k+2}
do not interact. There is no interaction at the level N 0 either. An attraction emerges at the
level N −1 . That the inter-wall interaction potential is O(N −1 ) can be shown on general
kinematic grounds (Ref. [17] presents a detailed discussion, see Eq. (37); see also Ref. [18].
This and other aspects of dynamics of inter-wall separation will be considered in Ref. [19].)
Our task is to understand dynamics of this phenomenon from the field theory side.
Assume that two walls under consideration are separated by a distance Z m−1 where m
is the mass of the lightest composite meson. What is the origin of the force between these
walls?
The interaction is due to the meson exchange in the bulk. Consider the lightest mesons,
scalar and pseudoscalar. The scalar meson σ , the “dilaton”, is coupled to the trace of the
µ
energy–momentum tensor θµ ,
σ µ 3N
θ = σ tr F 2 (4)
f µ 16π 2 f
(see, e.g., Ref. [20]). Here f is a coupling constant scaling as
f ∼ NΛ. (5)
µ
Integrating over the transverse direction and using the fact that θµ translates into mass, we
find that the “dilaton”–wall coupling (per unit area) is
T σ/f, (6)
where T is the tension, see Eq. (2). The σ -wall coupling scales as N 0.
The coupling of the pseudoscalar meson (the “axion” or “η ”; we will denote this field
by η—it will have a realization in terms of the RR 0-form of type IIB/0B) to the wall
is related to the change of the phase of the gluino condensate across the wall. It can be
estimated as2

η N η N ∂α
dz 2 tr F F̃ → dz 2 | tr λλ| , (7)
f 8π f 8π ∂z
where z is the coordinate transversal to the wall, and α is the phase of the order parameter.
large N the absolute value of the order parameter stays intact across
At the wall, while
dz (∂α/∂z) = 2π/N . Thus, the “axion”–wall coupling scales as N 0 . Note that the sign
2 Eq. (7) guarantees, automatically, that the “axion”–wall coupling is saturated inside the wall. Outside the
wall, in the vacuum, α = const, while tr F 2 = tr F F̃ = 0.
of the coupling depends on whether we cross the wall from left to right or from right to
left. This is why the exchange of the dilaton between two parallel BPS walls leads to the
wall attraction while that of the axion leads to repulsion.
Since the “dilaton/axion” coupling to the wall ∼ N 0 , we get no force at the level N 1 for
free. The underlying reason is that the BPS domain wall tension scales not as N 2 , as would
be natural for solitons, but as N 1 —the D-brane type of behavior.
Moreover, we want to say that σ and η give contributions which are exactly equal in
absolute values (at the level N 0 ) but are opposite in signs. This requires the degeneracy
of their masses—which is certainly the case in N = 1 SYM—and their couplings to the
walls (except for the relative sign). Taking the right-hand side of Eq. (7) literally we get
the “axion”–wall coupling in the form,
T η/f + O(1/N), (8)
i.e., the same as in Eq. (6).
At finite N the wall thickness which scales as [4] (NΛ)−1 is finite too. We have
(2 + 1)-dimensional supersymmetry on the wall world volume, which places scalars and
pseudoscalars into distinct (non-degenerate) supermultiplets [21]. At N = ∞ the wall
thickness vanishes and it is natural that in this limit the wall coupling involves the lowest
component of (3 + 1)-dimensional chiral superfield which has the form σ + iη. This
component is coupled to the (1/2, 1/2) central charge Z as Z exp(σ + iη) + h.c.3
Thus, the wall “R–R charge” is due to the axion-like nature of η. The “NS–NS charge” is
due to the wall tension. Both charges are indeed equal and scale as N 0 Λ2 . This guarantees
that at order N 0 there is no force. The σ and η coupling to the wall split at order N −1 .
Needless to say, the very same arguments can be repeated verbatim in the orientifold
theories. In the limit N → ∞ the degeneracy of the σ and η masses holds, and so does the
degeneracy of the wall “NS–NS and R–R charges”. The reason why the couplings to the
wall (associated with the trace of the energy–momentum tensor for the “dilaton” and the
axial charge for the “axion”) are the same is that these charges and couplings are inherited
from the parent supersymmetric theory at N → ∞, see [8]. For what follows it will be
useful to note that the σ exchange alone (before cancellation) generates the interaction
potential between the walls (per unit area)
V
∼ N 0 Λ3 e−mZ . (9)
A
This scaling law is in agreement with the string theory expectations.
3 If the BPS wall in N = 1 gluodynamics becomes an ordinary D-brane at N = ∞, then it must support a
massless U(1) gauge field. The gauge field in 1 + 2 dimensions is equivalent to a (pseudo)scalar field with the
S1 target space [22]. The above argument suggests an answer to a question formulated in Ref. [21]. Namely, the
U(1) gauge field on the world volume of the domain wall was shown to be described by the Lagrangian
1 N
L1+2 = − Fmn F mn + Fmn Ak εmnk + ferm. terms.
4e2 16π
The Chern–Simons term makes the A field massive, mA = N e2 /(4π ), non-degenerate with the translational
modulus. The scaling law of mA depends on that of e2 . We suggest that at N → ∞ the degeneracy is restored,
i.e., e2 ∼ ΛN −2 , so that mA ∼ ΛN −1 → 0 at N → ∞.
Fig. 1. The annulus diagram for the orientifold field theory. The D-branes are domain walls. Closed strings are
bosonic glueballs and open strings are UV degrees of freedom.
Now let us see how this picture is implemented in string theory. From the open string
standpoint the result of the zero force is natural. This is the Casimir force between the
walls. Since the UV degrees of freedom are Bose–Fermi degenerate the vacuum energy
and, hence, the Casimir force is zero. (See Fig. 1.)
In the gauge theory the closed strings are the glueballs of the field theory [23,24]. Let us
consider the large separation case, first. At the lowest level we have a massive scalar and
a pseudoscalar (we assume a mass gap). These two exactly degenerate states correspond
to the “dilaton” and “axion” (RR 0-form of type IIB/0B). They are expected to become
massive when the theory is defined on curved background [23]. The type IIB/0B action
contains the couplings
e−Φ tr F 2 + C tr F F̃ , (10)
where Φ denotes the dilaton and C the R–R zero-form. In addition we have a coupling of
the graviton and the R–R four-form
ηµρ hνλ tr Fµν Fρλ + C µνρλ tr Fµν Fρλ . (11)
In the gauge theory the “graviton” (tensor meson) and the four-form are heavier glueballs
since they carry higher spins than the dilaton and the zero-form. Similarly, the whole tower
of degenerate bosonic hadrons of the “orientifold field theory” should correspond to NS–
NS and R–R fields of type II/0 string theory. This gives us a new picture of why the force
between domain walls is zero in terms of the glueball exchanges: even-parity glueballs lead
to an attractive force between the walls whereas odd-parity glueballs lead to repulsion. The
sum of the two is exactly zero at the leading N 0 order. The 1/N force between the walls is
related to a possible non-vanishing force between parallel D-branes in curved space at the
order O(gst2 ).
We can also exploit the above picture to estimate the potential between a wall and a
anti-wall. This configuration is not BPS and a non-zero force is expected at the leading N 0
order. Ar large separations, the force is controlled by an exchange of the lowest massive
closed strings. These are the dilaton and the 0-form. Now, their contributions add up. We
get an attractive potential as indicated in Eq. (9).
4. Vanishing of the vacuum energy in N = 1 SYM and orientifold theory
One of the surprising results of our previous work [8] is that the N 2 part of the vacuum
energy density vanishes in the “orientifold field theory”. While this result makes sense from
the UV point of view, where we have Bose–Fermi degeneracy, it looks rather “mysterious”
from the IR point of view, since at the level of the composite color-singlet states we
have only bosonic degrees of freedom. Indeed, since at large N we have free bosons, it
is legitimate to sum the bosonic contributions to the vacuum energy density as follows:
1
k2 + Mn2 , (12)
2
n k
where Mn are the hadron masses. The paradox arise since the sum runs over positive
contributions. How can positive contributions sum to zero?
Before we present our solution, we would like to make two comments. First, the sum
(12) is not well defined since the expected Regge trajectory is not bounded from above and
therefore a regularization is needed. Second, in the above sum (12) the N dependence of
each individual mode is N 0 . The expected UV N 2 dependence of the vacuum energy is
hidden in the sum over hadronic modes. Thus, though formally (12) represents the vacuum
energy density of the theory, it is not the most efficient way to calculate it. Below, we
present an alternative way of calculation of the vacuum energy density—which explains
naturally the vanishing result.
Let us consider the contribution to the cosmological constant from the open string
sector. At large N it is dominated by the annulus diagram where each boundary consists
of N D-branes and a summation over the various D-branes is assumed. The Möbius
and Klein-bottle as well as higher-genus amplitudes are suppressed at large N . It is not
surprising that the cosmological constant vanishes, as we have N 2 bosons (NS open
strings) and N 2 fermions (Ramond fermionic open strings).
The annulus diagram has another interpretation. It represents the force between the D-
branes. The force is mediated by bosonic closed strings. In a SUSY setup (the type II
string), D-branes are BPS objects—hence, the zero force. As has been discussed above,
the balance, at large separation, is achieved in this case due to cancellation between the
dilaton, the graviton and the massless R–R forms.
It is interesting that the force between parallel selfdual D-branes is zero also in type-0
string theory [13,25]
A = N 2 (V8 − S8 ) ≡ 0. (13)
This is due to the underlying SUSY on the world sheet. The mechanism is exactly as in the
type II case: the R–R modes cancel the contributions of the NS–NS modes. Note that since
we are interested only in the planar gauge theory, we can restrict ourselves to gst = 0 on
the string theory side. Therefore, higher-genus amplitudes are irrelevant for our discussion.
At this level the relevant type-0 amplitudes, as well as the bosonic spectrum, are identical
to the type II ones, in a not too surprising similarity with the situation in the large N dual
gauge theories (the type-0 string becomes, in a sense, supersymmetric at the tree level). In
particular, the induced dilaton tadpole and cosmological constant are irrelevant, and, thus,
the background inherited from the supersymmetric theory remains intact.
The vanishing of the annulus diagram leads to an explanation of the “mysterious”

vanishing of the cosmological constant in the orientifold field theory: if one views the
hadrons (in the spirit of the AdS/CFT) as closed strings, the degenerate bosonic spectrum is
the reason behind the vanishing result for both wall–wall interaction and the cosmological
constant.
We hasten to add that though the mechanism is similar, there is a difference between the
two cases: the wall–wall interaction involves the force between “D2”-branes (wrapped D5-
branes), whereas the vanishing cosmological constant involves the force between “D3”-
branes. The two sorts of branes are not necessarily the same—it depends on the specific
realization. However, from the bulk point of view, the mechanism is identical. The only
requirement is the degeneracy of the NS–NS and the R–R tower and their couplings to the
branes.
The difference between string theory and field theory is that in string theory the force
between D-branes, from the closed string standpoint, is related to the contribution to
the cosmological constant from the open string sector. However, closed strings and open
strings are independent degrees of freedom, and so one has to add the contribution of the
closed strings to the cosmological constant. In the gauge theory string picture, the closed
strings are simply hadrons made out of the constituent open strings—the gluons and quarks.
Therefore, the value of the cosmological constant can be determined by either ultraviolet
(UV) or infrared (IR) degrees of freedom. It is the same quantity.
Below we will present a purely field-theoretic consideration which will, hopefully, make
transparent the issue of the vanishing of the vacuum energy density E (at level N 2 ) in the
orientifold theory. Usually it is believed that one needs full supersymmetry to guarantee
that E = 0. It turns out that a milder requirement—the degeneracy between scalar and
pseudoscalar glueballs/mesons—does the same job. Of course, in supersymmetric theories
this degeneracy is automatic. The orientifold field theory is the first example where it takes
place (to leading order in 1/N ) without full supersymmetry.
To begin with, we will outline some general relations relevant to E which are valid in
any gauge theory with no mass scale other than the dynamically generated Λ. The vacuum
energy density E is defined through the trace of the energy–momentum tensor,

1 µ 1
E= θµ = DA DΨ θµµ exp(iS),
4 4
3N 2
θµµ = − F , (14)
32π 2
where
F 2 ≡ Fµν
a
F µν,a , F F̃ ≡ Fµν
a
F̃ µν,a . (15)
The second line in Eq. (14) is exact in N = 1 gluodynamics and is valid up to 1/N
corrections in the orientifold theory.
Now, we use an old trick [26] to express the trace of the energy–momentum tensor in
terms of a two-point function. The idea is to vary both sides of Eq. (14) with respect to
1/g 2 invoking the fact that the only dimensional parameter of the theory, Λ, exponentially
depends on 1/g 2 . In this way one obtains [26]

1 µ 1 ν
E = −i d x vac|T
4
θ (x), θν (0) |vacconn
4 µ 4

3N 3N
≡ −i d x 4 2
F (x), 2
F (0) . (16)
128π 2 128π 2 conn
The above expression (16) is formal—it is not well-defined in the ultraviolet. This is the
same divergence that plagues any calculation of E (remember, Eq. (16) is general, it is
not related to supersymmetry). In order to give sense to this relation one needs a particular
regularization. For instance, one can always think of a non-SUSY theory as a SUSY theory
with mass terms for unwanted superpartners (soft SUSY breaking).
In supersymmetric theories SUSY prompts us a natural regularization. Indeed, let us
consider the two-point function of the lowest components of two chiral superfields D 2 W 2 ,

tr D 2 W 2 (x), tr D 2 W 2 (0) . (17)
Supersymmetry Ward identity tells us that this two-point function vanishes identically.
There are two remarkable facts encoded in Eq. (17). First, since D 2 W 2 ∝ (F 2 + iF F̃ ),
the vanishing of (17) is not due to the boson–fermion cancellation but, rather, due to the
cancellation between even/odd parity mesons (glueballs).
Second, Eq. (17) generalizes Eq. (16), so that the expression for the vacuum energy
density takes the form
2
3N
E = −i 2
d 4 x F 2 (x), F 2 (0) − F F̃ (x), F F̃ (0) , (18)
128π
where the connected correlators are understood on the right-hand side. To see that Eq. (18)
is a heir of Eq. (16), please, observe that

3N 3N
0 = −i d 4 x F F̃ (x), F F̃ (0) . (19)
128π 2 128π 2
The vanishing in Eq. (19) is due to the fact that F F̃ is proportional to the divergence of the
axial current aµ both in N = 1 gluodynamics and in the orientifold theory.
It is absolutely clear that this regularization works perfectly in the orientifold theories
(at level N 2 ). Indeed, the part of the two-point function that involves F 2 is saturated at
N → ∞ by the propagators of the glueballs with the even parity. Similarly, the part that
involves F F̃ is saturated by the odd parity glueballs. We then get
λ2n λ2n
E= − , λ2n ∼ N 2 for all n, (20)
Mn2 Mn2
even parity odd parity
µ
where λn are the couplings to Tµ and ∂ µ aµ , respectively, and Mn are the glueballs masses.
Clearly, if the masses and the couplings of the glueballs are even/odd-parity degenerate, as
is the case in N = 1 gluodynamics and in the large-N orientifold field theory, E vanishes.
In summary, in the ultraviolet calculation the Fermi–Bose degeneracy was responsible
for the vanishing of the cosmological constant both in supersymmetric gluodynamics and
in orientifold theory (where the cancellation was at level N 2 ). In dealing with E a certain
regularization procedure is needed. In SUSY it is implicit. In passing from the UV language
to the IR one, we make it explicit through Eq. (18). The range of the applicability of the
latter is wider than just SUSY. It is perfectly applicable in the orientifold theories too.
The expression (20) is in a remarkable agreement with our string theory picture. It shows
that only bosonic glueballs are involved and also that the even and odd parity glueballs
contribute with the opposite signs.
Perhaps the most interesting lesson from this picture is that the cosmological constant
can vanish even though only bosonic IR degrees of freedom are present in the given
gauge theory (at least, to the leading order in N ). In addition, the “correct” calculation
of the cosmological constant in N = 1 SYM, using the IR degrees of freedom, involves a
cancellation among the degenerate bosons and omission of the fermions!
5. Non-supersymmetric parent–daughter pairs
A simple proliferation of the fermion fields in the form of “flavors” leads to non-
supersymmetric parent–daughter pairs. This was first mentioned in Ref. [10] in the context
of Z2 orbifolds. The planar equivalence here holds perturbatively [27,28] but most likely
fails at the non-perturbative level. If we use, as a starting point, N = 1 gluodynamics and
its orientifold, the planar equivalence has solid chances to hold non-perturbatively.
Non-supersymmetric planar-equivalent pairs were mentioned in passing in Ref. [8]. For
instance, gauge theories with one and the same number of Dirac fermions either in the
antisymmetric two-index or symmetric two-index representations are planar equivalent.
Now we would like to discuss in more detail parent–daughter pairs which are obtained
from N = 1 gluodynamics and its orientifold by introducing fermion replicas. Thus, as
previously, the parent and daughter theories share one and the same gauge group, U(N),
and one and the same gauge coupling. The two respective fermion sectors are:
(i) k species of the Weyl fermions in the adjoint, to be denoted as (λA )ij ; and
(ii) k species of the Dirac fermions (Ψ A )[ij ] .
Here i, j are (anti)fundamental indices running i, j = 1, 2, . . . , N while A is the flavor

index running A = 1, 2, . . . , k. Note that k 5. Otherwise we loose asymptotic freedom.
Each Dirac fermion is equivalent to two Weyl fermions,

Ψ[ij ] → η[ij ] , ξ [ij ] .
Of course, since both theories are non-supersymmetric, predictive power is significantly

reduced compared to the case of a SUSY parent. Still, one can benefit from the comparison
of both theories, in particular, the Goldstone meson sectors. Let us start with the case (ii),
k species of Ψ[ij ] . Since the fermion fields are Dirac and in the complex representation of
the gauge group, the theory has the same (non-anomalous) chiral symmetry as QCD with
k flavors, namely, SU(k)L × SU(k)R . Various arguments tell us [29–31] that the pattern of
the chiral symmetry breaking is the same as in QCD too, namely
SU(k)L × SU(k)R → SU(k)V . (21)

√
The only distinction is that in QCD the constant f scales as N Λ while in our case its
scaling law is NΛ. Moreover, the coefficient n in front of the Wess–Zumino–Novikov–
Witten term Γ (see, e.g., Ref. [32]) equals N in QCD and (1/2)N(N − 1) in the case at
hand (see below).
All axial (non-anomalous) currents are spontaneously broken, giving rise to k 2 − 1
Goldstone mesons, “pions”. Some of them—those coupled to the axial currents that can be
elevated from the daughter theory (ii) to the parent theory (i)—persist in the parent theory
(i), where the fermion fields belong to the real representation, with the same coupling to
the corresponding axial currents. This is because of the planar equivalence of two theories.
(Remember, currents with the structure ξ̄ ξ − η̄η cannot be elevated from (ii) to (i).)
It is not difficult to count the number of the axial currents that are elevated from (ii)
to (i): there are (1/2)k(k − 1) off-diagonal currents of the type ξ̄ A ξ B + η̄A ηB (A = B)
plus all k − 1 diagonal axial currents of the type ξ̄ A ξ A + η̄A ηA (no summation over A).
Altogether we get
k(k + 1)
−1
2
Goldstone mesons. This obviously corresponds to the following pattern of the chiral
symmetry breaking:
SU(k) → SO(k), (22)

with the Goldstone mesons in the symmetric two-index representation of SO(k). The
pattern of the chiral symmetry breaking for the quarks belonging to a real representation of
the gauge group indicated in Eq. (22) was advocated many times in the literature [29–31,
33], but no complete proof was ever given.
We conclude this section by a brief comment on the topological properties of the
corresponding chiral Lagrangian and how they match the underlying gauge field theory
expectations. The theory (i) is expected to be confining and support flux tubes—
fundamental color charges cannot be screened. On the other hand, we do not expect stable
baryons with mass growing with N . Composite color-singlet states of gluon and (λA )ij
form baryons with M ∼ N 0 .
At the same time, the theory (ii) does not have baryons with M ∼ N 0 . Here the baryon
masses grow with N . The theory is expected to be confining too, but two flux tubes (each
attached to a color source in the fundamental representation) can be screened by (Ψ A )[ij ] .
As was suggested in Refs. [32,33], at large N one can try to identify baryons with the
Skyrmions supported by the corresponding chiral Lagrangians. Since

π3 SU(k)/O(k) = Z4 at k = 3,

π3 SU(k)/O(k) = Z2 at k 4,

π3 SU(k) = Z at all k (23)
both theories, (i) and (ii), yield Skyrmions with MSkyrme ∼ N 2 , albeit the theory (ii) has a
richer spectrum. The above scaling law, MSkyrme ∼ N 2 , is due to the fact that f scales as
N in the theories under consideration.
Skyrmion statistics is determined by the (quantized) factor in front of the Wess–
Zumino–Novikov–Witten term,
(−1)N(N−1)/2 . (24)
It has half-integer spin provided that N(N − 1)/2 is odd, i.e., N = 4p + 2 or N =
4p + 3 where p is an integer. In both cases one can construct, in the microscopic theory,
interpolating baryon currents with an odd number of constituents scaling as N . Why then
the Skyrmion mass scales as N 2 ? A possible explanation is as follows. For quarks in the
fundamental representation of SU(N) the color wave function is antisymmetric, which
allows them all to be in the S wave in the coordinate space. With antisymmetric two-index
spinor fields the color wave function is symmetric, which would require them to occupy
orbits with angular momentum up to ∼ N . Then the scaling law MSkyrme ∼ N 2 seems
natural.
Since π2 {SU(k)} = 0 the chiral sector of the theory (ii) does not support flux tubes.
Albeit disappointing, such a situation was anticipated by Witten [33] who noted that
topology of the full space of the large N theory need not coincide with topology of its
Goldstone sector.
In light of this remark we can understand the complete failure of the Skyrmion
description of the theory (i). In particular, since π2 {SU(k)/O(k)} = Z2 at k 3 we get
flux tubes in the chiral theory, while we do not expect them in the microscopic theory.
Moreover, stable Skyrmions of the chiral sector should become unstable in the full theory.
6. Conclusions
In this work we tried to further develop a parallel between N = 1 gluodynamics and its
non-supersymmetric orientifold daughter on the one hand, and string/D-brane paradigm,
on the other. We discussed forces between two BPS domain walls in field theory terms
and established contact with the string theory description. In the latter, there is a well-
known cancellation between NS–NS and R–R interactions. In field theory terms this
cancellation manifest itself as follows: the exchange of a composite dilaton coupled to
the domain wall is canceled (at leading order) by that of a composite axion. The string
theory interpretation allows us to establish a one-to-one relation between vanishing force
and vanishing cosmological constant. Thus, the question “who is responsible for the
vanishing cosmological constant in non-supersymmetric orientifold field theory?” gets a
rather unexpected answer—the degeneracy of the even–odd parity composite mesons.
In the last part of the paper we discuss “flavor proliferation” as a device allowing one
to get planar equivalent pairs in which both theories in the parent–daughter pair are non-
supersymmetric, starting from the original pair—N = 1 gluodynamics and its orientifold
daughter. Although we loose predictive power based on supersymmetry of the parent,
some predictions survive. In particular, we compare Goldstone meson sectors, and obtain
consequences for the patterns of the chiral symmetry breaking.
Acknowledgements
We are grateful to I. Brunner, G. Gabadadze, I. Kogan, W. Lerche, A. Loewy,

R. Rabadan, J. Sonnenschein, and G. Veneziano for fruitful discussions. Special thanks
go to A. Gorsky and A. Ritz for insightful communications. The work of M.S. is supported
in part by DOE grant DE-FG02-94ER408.
References
[1] G.R. Dvali, M.A. Shifman, Phys. Lett. B 396 (1997) 64;
G.R. Dvali, M.A. Shifman, Phys. Lett. B 407 (1997) 452, Erratum, hep-th/9612128.
[2] E. Witten, Nucl. Phys. B 507 (1997) 658, hep-th/9706109.
[3] I.I. Kogan, A. Kovner, M.A. Shifman, Phys. Rev. D 57 (1998) 5195, hep-th/9712046.
[4] G.R. Dvali, G. Gabadadze, Z. Kakushadze, Nucl. Phys. B 562 (1999) 158, hep-th/9901032;
G. Gabadadze, M.A. Shifman, Phys. Rev. D 61 (2000) 075014, hep-th/9910050.
[5] J. Polchinski, M.J. Strassler, The string dual of a confining four-dimensional gauge theory, hep-th/0003136.
[6] I.R. Klebanov, M.J. Strassler, JHEP 0008 (2000) 052, hep-th/0007191.
[7] A. Loewy, J. Sonnenschein, JHEP 0108 (2001) 007, hep-th/0103163.
[8] A. Armoni, M. Shifman, G. Veneziano, Exact results in non-supersymmetric large N orientifold field
theories, hep-th/0302163.
[9] M.J. Strassler, On methods for extracting exact non-perturbative results in non-supersymmetric gauge
theories, hep-th/0104032.
[10] A. Gorsky, M. Shifman, Phys. Rev. D 67 (2003) 022003, hep-th/0208073.
[11] D. Tong, JHEP 0303 (2003) 022, hep-th/0212235.
[12] A. Armoni, B. Kol, JHEP 9907 (1999) 011, hep-th/9906081.
[13] C. Angelantonj, A. Armoni, Nucl. Phys. B 578 (2000) 239, hep-th/9912257.
[14] J.M. Maldacena, Adv. Theor. Math. Phys. 2 (1998) 231, Int. J. Theor. Phys. 38 (1999) 1113, hep-th/9711200.
[15] J.M. Maldacena, C. Nuñez, Phys. Rev. Lett. 86 (2001) 588, hep-th/0008001.
[16] J. Polchinski, Phys. Rev. Lett. 75 (1995) 4724, hep-th/9510017.
[17] A. Ritz, (Dis)assembling composite supersymmetric solitons, in: K. Olive, M. Shifman, M. Voloshin (Eds.),
Continuous Advances in QCD 2002, Arkadyfest, World Scientific, Singapore, 2002, pp. 345–368.
[18] A. Ritz, M. Shifman, A. Vainshtein, Phys. Rev. D 66 (2002) 065015, hep-th/0205083.
[19] A. Ritz, M. Shifman, A. Vainshtein, in preparation.
[20] A.A. Migdal, M.A. Shifman, Phys. Lett. B 114 (1982) 445.
[21] B.S. Acharya, C. Vafa, On domain walls of N = 1 supersymmetric Yang–Mills in four dimensions, hep-
th/0103011.
[22] A.M. Polyakov, Nucl. Phys. B 120 (1977) 429.
[23] E. Witten, Adv. Theor. Math. Phys. 2 (1998) 505, hep-th/9803131.
[24] C. Csaki, H. Ooguri, Y. Oz, J. Terning, JHEP 9901 (1999) 017, hep-th/9806021.
[25] I.R. Klebanov, A.A. Tseytlin, Nucl. Phys. B 546 (1999) 155, hep-th/9811035.
[26] V.A. Novikov, M.A. Shifman, A.I. Vainshtein, V.I. Zakharov, Nucl. Phys. B 191 (1981) 301.
[27] M. Bershadsky, Z. Kakushadze, C. Vafa, Nucl. Phys. B 523 (1998) 59, hep-th/9803076.
[28] M. Bershadsky, A. Johansen, Nucl. Phys. B 536 (1998) 141, hep-th/9803249.
[29] S. Dimopoulos, Nucl. Phys. B 168 (1980) 69;
M.E. Peskin, Nucl. Phys. B 175 (1980) 197.
[30] Y.I. Kogan, M.A. Shifman, M.I. Vysotsky, Yad. Fiz. 42 (1985) 504, Sov. J. Nucl. Phys. 42 (1985) 318 (in
English).
[31] J.J. Verbaarschot, Phys. Rev. Lett. 72 (1994) 2531, hep-th/9401059;
A. Smilga, J.J. Verbaarschot, Phys. Rev. D 51 (1995) 829, hep-th/9404031;
M.A. Halasz, J.J. Verbaarschot, Phys. Rev. D 52 (1995) 2563, hep-th/9502096.
Tachyon condensation and universal solutions

in string field theory
Tomohiko Takahashi
Department of Physics, Nara Women’s University, Nara 630-8506, Japan
Received 25 February 2003; accepted 5 August 2003
Abstract
We investigate a non-perturbative vacuum in open string field theory expanded around the analytic
classical solution which has been found in the universal Fock space generated by matter Virasoro
generators and ghost oscillators. We carry out level-truncation analyses up to level (6, 18) in the
theory around one-parameter families of the solution. We observe that the value of the vacuum
energy cancels the D-brane tension as the approximation level is increased, but this non-perturbative
vacuum disappears at the boundary of the parameter space. These results provide strong evidence for
the conjecture that, although the universal solutions are pure gauge in almost all the parameter space,
they are regarded as the tachyon vacuum solution at the boundary.
PACS: 11.25.Sq
1. Introduction
String field theory is a promising approach to investigate non-perturbative aspects of

string theory. As conjectured by Sen [1,2], we can describe the annihilation process of
D-branes by using the condensation of the tachyon in open string field theory, in which
there is a stable vacuum [3] and the energy difference between the stable and unstable
vacua is in precise agreement with a D-brane tension [4–6]. Since these discoveries much
progress has been made in string field theory, particularly in the formulation of vacuum
string field theory [7], where some exact results are obtained [8]. However, despite these
developments, we have not yet found the analytic classical solution, which is eagerly
awaited, corresponding to the tachyon vacuum in open string field theory.
E-mail address: tomo@asuka.phys.nara-wu.ac.jp (T. Takahashi).
doi:10.1016/j.nuclphysb.2003.08.007
162 T. Takahashi / Nuclear Physics B 670 (2003) 161–182
While there are some formal attempts to construct analytic solutions in string field
theory [9–12], strange problems often arise from such formal solutions. For example, if
QL stands for the left-half integration of the BRS current and I is the identity string
field, QL I is a formal solution in purely cubic string field theory [13]. The equation
of motion is given by Ψ ∗ Ψ = 0 and then −QL I is also a solution, around which
the kinetic operator becomes −QB and the theory should describe open strings with a
negative tension. Therefore, ordinary D-branes and negative tension branes are realized
with the same energy density. We find another example in the context of vacuum string
field theory as pointed out in [7]. From similar discussions, it follows that the solutions
leading to pure ghost kinetic operators provide the same energy density as the perturbative
vacuum. Usually, these unreliable results are caused by midpoint singularities in a half
string formulation. In the former example, though the operator Q2L appears in solving
the equation of motion, this operator itself is ill-defined due to a midpoint singularity,
as pointed out, for example, in [10]. We find the same singularity in the latter case.
Fortunately, the analytic solutions found in [9,10] escape from all these problems related
to the midpoint singularity. In addition they have many remarkable features: the solutions
can be expressed by states in the universal Fock space which is spanned by matter Virasoro
generators and ghost oscillators acting on the SL(2, R) invariant vacuum (so we call them
universal solutions). This universality is necessary for the solution corresponding to the
tachyon vacuum [1]. Secondly, open string excitations disappear after the string field
condensation to a specific class of the universal solutions [14,15]. This property is also
indispensable for the tachyon vacuum solution. Consequently, we naturally expect that a
certain kind of the universal solutions is regarded as the tachyon vacuum solution.
Even if the universal solutions are irrelevant to the tachyon vacuum, it should be
emphasized that they have intimate relations to gauge symmetry in string field theory,
which is an underlying principle in the theory and which is much larger symmetry than
existing in the low energy effective theory. We can construct the universal solutions with a
parameter. They are pure gauge solutions in almost all the region of the parameter, but they
become non-trivial solutions at the boundary of the parameter space. Hence, the non-trivial
universal solutions can be regarded as a kind of singular gauge transformations from the
perturbative vacuum [10]. Moreover, the gauge symmetry and the annihilation mechanism
of open strings are inseparable. In the theory around the non-trivial solution, the kinetic
operator is given by the modified BRS charge which has the cohomology with different
ghost numbers from the original cohomology. Therefore, all of on-shell states are reduced
to gauge degrees of freedom in the gauge unfixed theory and then open string excitations
disappear after the condensation [14].
Although the tachyon vacuum solution has already been obtained approximately in the
level truncated theory, we cannot so simply compared the universal solutions with the level
truncated solution, because the gauges of these solutions differ. The energy density of
the universal solutions, then, should be calculated in order to clarify the relation between
these solutions. However, if we try to calculate the potential height directly, there are some
technical problems which are remained to be resolved. To avoid these we should adopt
other approaches at present.
The purpose of this paper is to apply level truncation scheme to investigate the theory
around the universal solutions, and to determine the potential height of the solutions
T. Takahashi / Nuclear Physics B 670 (2003) 161–182 163
indirectly. If a one-parameter family of the solutions can be interpreted as explained above,

the theory around the solutions should describe the perturbative vacuum in the almost
region of the parameter and the tachyon vacuum at the boundary. Hence, we should observe
the situation that in the theory for the almost parameter, there is a non-perturbative vacuum
which gives the same energy density as the D-brane tension, but the non-trivial vacuum
disappears at the endpoint.
In Section 2 we review the universal solutions in string field theory with some new
results, and we explain the difficulty of calculating the potential height. In Section 3 we
analyze the non-perturbative vacuum in the theory around the universal solutions up to level
(6, 18). Our results strongly suggest that the non-trivial universal solution corresponds to
the tachyon vacuum. In Section 4, we give summary and discussions.
2. Classical solutions and potential heights
2.1. Universal solutions
The action of cubic open string field theory is given by [16]

1 1 1
S =− 2 Ψ ∗ QB Ψ + Ψ ∗ Ψ ∗ Ψ . (2.1)
g 2 3
By variation of the action, we find the classical equation of motion
QB Ψ + Ψ ∗ Ψ = 0. (2.2)
One of the analytic solutions with translational invariance has been found as [10]

Ψ0 = QL eh − 1 I − CL (∂h)2 eh I, (2.3)
where I denotes the identity string field. The operators QL and CL are defined by

dw dw
QL (f ) = f (w)JB (w), CL (f ) = f (w)c(w), (2.4)
2πi 2πi
Cleft Cleft
where JB (w) and c(w) are the BRS current and the ghost field, respectively, and Cleft
stands for the contour along the left-half of strings. The function h(w) satisfies h(−1/w) =
h(w) and h(±i) = 0.
The solution (2.3) obeys the equation of motion (2.2) in the following. The anti-
commutation relations of QL and CL are given by
h
QL e − 1 , QL eh − 1 = 2 QB , CL (∂h)2 e2h ,
h

QL e − 1 , CL (∂h)2 eh = QB , CL (∂h)2 e2h − eh , (2.5)
and others are zero. We define similar operators QR (f ) and CR (f ) by replacing the
contour Cleft in (2.4) to Cright corresponding to the right half of strings. Then, for the
star product, these operators satisfy

QR eh − 1 A ∗ B = −(−1)|A| A ∗ QL eh − 1 B ,

CR (∂h)2 eh A ∗ B = −(−1)|A| A ∗ CL (∂h)2 eh B . (2.6)
Through conservation of the BRS current and the ghost field on the identity string field, we
find that

QL eh − 1 + QR eh − 1 I = 0,

CL (∂h)2 eh + CR (∂h)2 eh I = 0. (2.7)
From (2.5), (2.6) and (2.7), it follows that

QB Ψ0 = − QB , CL (∂h)2 eh I,

2
Ψ0 ∗ Ψ0 = QL eh − 1 − CL (∂h)2 eh I = QB , CL (∂h)2 eh I. (2.8)
As a result, the equation of (2.3) is a classical solution in the string field theory.
Though the function h(w) must cancel the midpoint singularity of the ghost field on I to
make the operator CL I well-defined, this cancellation actually occurs due to the previous
two conditions for h(w). Around the midpoint w0 = ±i, the ghost field behaves as [9,10,
17]

1 w0
c(w)I ∼ −c0 + (c1 − c−1 ) I + O (w − w0 )0 , (2.9)
w − w0 2
and then its singularity is a pole at the midpoint. If the function h(w) is analytic around w0 ,
h(w) can be expanded as h(w) = h (w0 )(w − w0 )2 + · · · because h (w) = h (−1/w)/w2 .
Therefore, (∂h)2 eh c(w) becomes regular at the midpoint and then the operator CL is well-
defined on the identity string field.
In the following, let us consider the solution generated by the function

a 1 2
ha (w) = log 1 + w+ , (2.10)
2 w
and we parameterize the solution by a real number a. The function is rewritten as

ha (σ ) = log(1 + 2a cos2 σ ) on the unit circle w = exp(iσ ), and the parameter a is larger
than −1/2 accordingly. In the region a −1/2, the function ha (σ ) can be expanded by
the Fourier series
∞

2 (−1)n
ha (σ ) = − log 1 − Z(a) − 2 Z(a)n cos(2nσ ), (2.11)
n
n=1
√
where Z(a) = (1 + a − 1 + 2a)/a and −1 Z(a) < 1 (−1/2 a < ∞).
Substituting the function into the form (2.3) and expanding it by oscillators, we find the
solution up to level two

8a
|Ψ0 (a) = J1 (a)c1 |0 + + J1 (a) LX−2 c1 |0
3π

8a 8a
+ + J2 (a) c−1 |0 + + 2J1 (a) c0 b−2 c1 |0 + · · · , (2.12)
π 3π
where LX
n denote matter Virasoro generators and J1 (a) and J2 (a) are given by
π/2
dσ 1
J1 (a) = − ha (σ )2 eha (σ ) ,
2π 2 cos σ
−π/2
π/2
dσ 1 + 2 cos(2σ )
J2 (a) = − h (σ )2 eha (σ ) . (2.13)
2π a 2 cos σ
−π/2
Using the Fourier series (2.11), we can carry out these integrations. The results of the
calculations are, for a 0
√
4a 1 1 1 + Z(a)
J1 (a) = − 1− Z(a) + √ log √ , (2.14)
π 2 Z(a) 1 − Z(a)

4a 1 1
J2 (a) = + Z(a) +
π 3 Z(a)
√
1 1 1 + Z(a)
− Z(a) Z(a) + √ log √ , (2.15)
2 Z(a) Z(a) 1 − Z(a)
and for −1/2 a < 0,

4a 1
J1 (a) = − 1+ −Z(a) − √ arctan −Z(a) , (2.16)
π −Z(a)

4a 1 1
J2 (a) = + Z(a) +
π 3 Z(a)

1
+ Z(a) −Z(a) − √ arctan −Z(a) . (2.17)
Z(a) −Z(a)
It is interesting to note that the solution has a well-defined Fock space expression and the
coefficients of its component states have no divergence. This situation is different from
the case of the dilaton condensation in light-cone type string field theories [18,19]. For
instance, the functions J1 (a) and J2 (a) have finite values as depicted in Fig. 1 and then the
coefficients become finite up to level two. Moreover, the Fock space used in the solution
can be spanned by the universal basis, because the solution is made of the BRS current, the
ghost field and the identity string field. This universality is indispensable for the tachyon
solution [1]. Finally, we indicate that the solution is outside Siegel gauge since it contains
states proportional to the ghost zero mode c0 .
Fig. 1. Plots of the functions J1 (a) and J2 (a).
2.2. Physical interpretation
To consider physical meaning of the solution, we expand the string field Ψ by the
solution Ψ0 and the quantum fluctuation Φ as
Ψ = Ψ0 (a) + Φ. (2.18)
Substituting this form into (2.1), the action becomes

1 1 1
S[Ψ ] = S[Ψ0 (a)] − 2 Φ ∗ QB (a)Φ + Φ ∗ Φ ∗ Φ , (2.19)
g 2 3
where the modified BRS charge QB is given by

QB (a) = Q eha − C (∂ha )2 eha , (2.20)
and Q(f ) and C(g) are defined by QL (f ) + QR (f ) and CL (g) + CR (g), respectively. The
first term in the action corresponds to the vacuum energy at the solution, and the second
term represents the action for the quantum fluctuation.
We can show that the solution for a > −1/2 is expressed by a gauge transformation of
the trivial vacuum as

Ψ0 (a) = exp KL (ha )I ∗ QB exp −KL (ha )I , (2.21)
where exp A for a string field A is defined by its series as exp A = I + A + A ∗ A/2! + · · ·,
and the operator KL is defined by

dw 3
KL (f ) = f (w) Jgh (w) − w−1 . (2.22)
2πi 2
Cleft
Consequently, we naturally expect that the theory around the solution for a > −1/2
describes the physics on the perturbative vacuum. Indeed, we can transform the action
for the fluctuation Φ into the original action through the string field redefinition
Φ = eK(ha ) Φ, (2.23)
where K(f ) = KL (f ) + KR (f ), and KR (f ) is the counterpart of KL (f ) related to right

strings. The equivalence of these actions is based on the fact that the original and modified
BRS charges are connected by the similarity transformation
QB (a) = eK(ha ) QB e−K(ha ) . (2.24)

However, since the operators eKL and eK becomes ill-defined at a = −1/2, the solution
cannot be represented by the pure gauge, and the theory around it cannot be connected
to the original theory through the string field redefinition. Consequently, the solution at
a = −1/2 represents a non-trivial vacuum, while the solution is a pure gauge for a > −1/2.
For example, the operator eK is written by the normal ordered form
−1 2
eK(ha ) = 1 − Z(a)2 exp −q̃0 log 1 − Z(a)

∞ ∞
(−1)n (−1)n
× exp − q−2n Z(a) exp −
n n
q2n Z(a) , (2.25)
n n
n=1 n=1
where qn denote the oscillators of the ghost number current and they are written by the
ghost oscillators as
∞
1
q̃0 = (c0 b0 − b0 c0 ) + (c−n bn − b−n cn ),
2
n=1
∞

qn = cn−m bm (n = 0). (2.26)
m=−∞
If we take a = −1/2, the first factor in (2.25) diverges because Z(−1/2) = −1. In order to
find the singularity more explicitly, we write the string field by the oscillator expression as
|Ψ = φ(x)c1 |0 + · · · + β(x)c−1 |0 + γ (x)b−2 c0 c1 |0 + · · · . (2.27)

Using the normal ordered expression (2.25), we find that, through the redefinition of (2.23),
the lowest level component field φ(x) is transformed as
1 Z(a)
φ (x) = φ(x) + −β(x) + 2γ (x) + · · · , (2.28)
1 + Z(a) 1 + Z(a)
where the abbreviation denotes the contribution from the higher level component fields.
Thus, by the string field redefinition, the component fields are transformed into the linear
combination of an infinite number of fields. However, its coefficients diverge at a = −1/2
and then this redefinition is ill-defined. Similarly, the operator eKL has a singularity at
a = −1/2.
To find the physical meaning of the solution at a = −1/2 from a different viewpoint,
we can determine the cohomology of the new BRS charge and the perturbative spectrum
around the solution. As in Ref. [14], the new cohomology appears only in the sector
with different ghost numbers from the original cohomology. Consequently, in the gauge
unfixed theory around the solution, we can solve the equation of motion QB (−1/2)Φ = 0
as Φ = QB (−1/2)φ. Since the theory is invariant under the gauge transformation δΦ =
QB (−1/2)δΛ, all of the on-shell physical states become gauge degrees of freedom in the
theory perturbatively.
Hence, we find that the universal solution at a = −1/2 represents a non-trivial vacuum,
where there is no physical excitation perturbatively. This is the very feature required for
the tachyon vacuum. As discussed above, the solution satisfies universality. Putting these
observations together, we expect that the universal solution corresponds to the tachyon
vacuum itself.
2.3. Potential heights
Formally, we can find that the potential height −S[Ψ0 (a)] is zero for a > −1/2.1
Indeed, the derivative of S[Ψ (a)] with respect to a is given by

d 1 dΨ0 (a)
S[Ψ0 (a)] = − 2 QB Ψ0 (a) + Ψ0 (a) ∗ Ψ0 (a) ∗ = 0, (2.29)
da g da
where we have used the equation of motions for the last equality [20]. Since ha=0 =
∂ha=0 = 0 and then S[Ψ0 (a = 0)] = 0, it turns out that S[Ψ0 (a)] is equal to zero. This
zero potential height can be shown only for a > −1/2, because the solution is ill-defined
for a < −1/2 and it is undifferentiable at a = −1/2. This formal discussion is consistent
with the expectation that the universal solution is a pure gauge for a > −1/2.
However, we cannot calculate the potential height more explicitly beyond the formal
evaluation. Substituting the solution (2.3) into the action (2.1), we find that
1
S[Ψ0 (a)] = − 2
I |CL (∂ha )2 eha QB CL (∂ha )2 eha |I
6g

1 dw dw
=− 2
6g 2πi 2πi
Cleft Cleft
2 2
× ∂ha (w) eha (w) ∂ha (w ) eha (w ) I |c(w)∂c(w )|I , (2.30)
where we have used QB I = 0 and {QB , c} = c∂c. The identity string field is written by the
tensor product of the matter and ghost sectors and the matter sector of the identity string
field is given by [21]
∞
X (−1)n
I = exp − α−n · α−n |0. (2.31)
2n
n=1
Then, I |c∂c|I is an indefinite quantity for any a due to the infinite determinant factor of
the matter sector. Thus, it is difficult to evaluate the potential height because we have not
yet known how it should be regularized.2
1 More precisely, the potential V is divided by the D-brane volume V as V = −S[Ψ ]/V .
D 0 D
2 We are still suffering from the disastrous divergence even if we use usual regularization with the insertion of
e−.L0 [22].
Instead of the exact calculation, we try to use level truncated solutions to evaluate the
potential height. For a = −1/2, the solution (2.12) becomes
2 2 X 2 8
|Ψ0 (−1/2) = c1 |0 + L−2 c1 |0 − c−1 |0 + c0 b−2 c1 |0 + · · · . (2.32)
π 3π 3π 3π
At level zero, the truncated solution is 2/π × c1 |0 and so the function f (T ) defined in
Ref. [4] takes the value

1 2 2 1 2 3
2π −2
+ −0.279. (2.33)
2 π 3 π
This provides 28% of the D-brane tension. Furthermore, we calculate the function f (T ) at
level two and, then, it takes the value ∼ 85, which is far from a stationary point at the level
two potential. For general a, we are faced with such terrible behavior. Although this result
discourages us to perform further calculations, this is a natural result because our solution
is not a solution in the level truncated theory.
Hence, we cannot calculate the vacuum energy of the universal solutions at present, in
order to compare it with the D-brane tension.
3. Level truncation in string field theory with the modified kinetic term
In this section we explore another possibility of clarifying the relation between the
universal solutions and the tachyon vacuum. Instead of the direct calculation of the
potential height, we analyze the non-perturbative vacuum in the theory expanded around
the universal solutions. By using the level truncation scheme in Siegel gauge, we can
find the non-perturbative vacuum without any difficulty, and moreover we come across
the impressive result which supports our conjecture for the universal solutions.
3.1. Conjectures and setup
As in (2.19), the action for the fluctuation Φ is written by

1 1 1
S[Φ] = − 2 Φ ∗ QB (a)Φ + Φ ∗ Φ ∗ Φ , (3.1)
g 2 3
where the modified BRS charge is given by (2.20). Substituting (2.11) into (2.20) and
performing the w integration, we obtain the oscillator expressions of the new BRS charge
a
QB (a) = (1 + a)QB + (Q2 + Q−2 ) + 4aZ(a)c0 − 2aZ(a)2(c2 + c−2 )
2
∞

− 2a 1 − Z(a)2 (−1)n Z(a)n−1 (c2n + c−2n ), (3.2)
n=2
−n−1 . The details of the calculation
where we expand the BRS current as JB (w) = n Qn w
are presented in Appendix A.
Under the Siegel gauge condition b0 Φ = 0, the quadratic term in the action becomes

1 1 1 1
− 2 Φ ∗ QB (a)Φ = − 2 Φ ∗ c0 L(a)Φ, (3.3)
g 2 g 2
where L(a) = {QB (a), b0 }. Using the anti-commutation relation [14]
3
{Qm , bn } = Lm+n + mqm+n + m(m − 1)δm+n,0 , (3.4)
2
we can calculate the operator L(a) as
a
L(a) = (1 + a)L0 + (L2 + L−2 ) + a(q2 − q−2 ) + 4aZ(a). (3.5)
2
Therefore, according to the notations of [1,4], the potential in the new string field theory is
given by the ‘modified’ universal function

1 1
fa (Φ) = 2π 2 Φ, c0 L(a)Φ + Φ, Φ ∗ Φ . (3.6)
2 3
As seen in the previous section, we expect that the solutions Ψ0 (a > −1/2) and Ψ0 (a =
−1/2) are regarded as a pure gauge and the tachyon vacuum, respectively. Consequently,
in the case of a > −1/2, the action of (3.1) describes the perturbative vacuum, and then,
in the potential, there is the stationary point which corresponds to the tachyon vacuum.
On the other hand, the stationary point must vanish at a = −1/2, because the theory has
already stayed on the non-perturbative vacuum. Hence, due to our conjectures, the modified
universal function at the stationary point Φ0 must satisfy

0 (a = −1/2),
fa (Φ0 ) = (3.7)
−1 (a > −1/2).
Let us consider a level truncated expression of the modified universal function. The new
action is invariant under the twist transformation σ → π − σ [4]. Due to this symmetry, we
have only to look for a stationary point where Φ0 contains even level states as well as the
case of the original level truncated analysis. In general, we can write a scalar string field
|Φ by the tensor product of the matter and ghost states as
∞

|Φ = ψi |si , |si = |ηm(i) ⊗ |χg(i). (3.8)
i=0
Our notations for the decomposition of states are almost same as in Ref. [5]. Up to level 6,
the matter and ghost states, |ηi and |χi , are given in Appendix B and the decomposition
of states is in Appendix C.
Using the component fields ψi with zero momentum, we can express the modified
universal function as follows,

1
2 1
fa (ψ) = 2π dij (a)ψi ψj + tij k ψi ψj ψk . (3.9)
2 3
ij ij k
The cubic coefficients tij k does not change from the previous analysis in Ref. [5]. The
quadratic coefficients are calculated as
gh mat gh
dij (a) = (1 + a) level(i) − 1 Amat ij Aij + 4aZ(a) Aij Aij
gh gh
+ aBijmat Aij + aAmat
ij Bij , (3.10)
mat(gh) mat(gh)
where level(i) denotes the level of the state |si ,3 and Aij and Bij are defined by
ij = ηm(i) |ηm(j ) ,
Amat (3.11)
gh
Aij = χg(i) |c0 |χg(j ), (3.12)
Bijmat = ηm(i) |Lmat
−2 |ηm(j ) , (3.13)
gh gh
Bij = χg(i) |c0 L−2 − q−2 |χg(j ). (3.14)
gh
Here, Lmat
n and Ln are the matter and ghost parts of the total Virasoro generators Ln . In
this expression, L2 and q2 are converted to L−2 and −q−2 by the hermitian conjugation.
mat(gh)
A list of the coefficients Aij up to level 6 is given in Appendix B. Up to level 6, the
mat(gh) mat(gh)
coefficients Bij can be calculated by using Aij through the following equations,
1
−2 |η0 = |η1 ,
Lmat
2
1
−2 |η1 = 2|η3 + |η5 ,
Lmat
2
1
−2 |η3 = 3|η6 + |η8 + |η9 ,
Lmat
2
−2 |η4 = 4|η7 + |η10 ,
Lmat
1
−2 |η5 = 4|η9 + |η12 ,
Lmat
2
gh
L−2 − 2q−2 |χ0 = −|χ1 ,
gh
L−2 − 2q−2 |χ1 = 3|χ4 + |χ6 ,
gh
L−2 − 2q−2 |χ4 = 5|χ7 + |χ9 ,
gh
L−2 − 2q−2 |χ5 = 4|χ8 + 2|χ10 + |χ12 ,
gh
L−2 − 2q−2 |χ6 = 3|χ9 + 3|χ11. (3.15)
3.2. Level zero analysis

At level (0, 0) approximation, the component field is tc1 |0 and then the modified
universal function is

1 1
fa (t) = 2π 2 − λ(a)t 2 + K 3 t 3 , (3.16)
2 3
3 We set the level of ground states as level(η ) = 0 and level(χ ) = 0.

0 0
Fig. 2. Plots of the potential minimum in level (0, 0) truncation.

√ √
where λ(a) = 4 1 + 2a − 3(1 + a) and K = 3 3/4. It is easy to see that λ(a) has two
roots
√ √
7+4 7 7−4 7
a+ = = 1.954 . . ., a− = = −0.398 . . ., (3.17)
9 9
then
if a − < a < a + , λ(a) > 0,

− +
if − 1/2 a < a or a > a , λ(a) < 0. (3.18)
Therefore, fa (t) has a local minimum at
−3
K λ(a) (a − a a + ),
t0 = (3.19)
0 (−1/2 a < a − or a > a + ),
and fa (t) at this minimum takes the value

 π2
− 6 λ(a)3 (a − a a + ),
fa (t0 ) = (3.20)
 3K
0 (−1/2 a < a − or a > a + ).
The a dependence of this value is depicted in Fig. 2. Though this behavior is quite different
from the expectation that a does not affect the potential minimum for a > −1/2 as in (3.7),
this a dependence is introduced merely by the level truncation approximation.
The value −0.684 . . . at a = 0 equals to the minimum derived from the previous level
truncation analysis [4], because the kinetic operator L(a) becomes L0 at a = 0 and the
modified universal function agrees with the universal function. This agreement is realized
for any level analysis as seen in (3.5). In addition, it should be noticed that the minimum is
exactly zero at a = −1/2.
3.3. Higher level analysis
We apply the iterative approximation algorithm used by Moeller and Taylor [5] to higher
level calculations. However, there is a slight change in the procedure to find the stable
vacuum. According to Ref. [5], the solution which minimizes the potential of (3.9) has
many branches, and which branch should be chosen is determined by the condition that
the solution becomes the level zero stable vacuum if higher level fields are turned off. As a
result, the branch depends on the sign of the quadratic coefficients dij (a). As an example,
let us see the tachyon field ψ0 = t. Since the coefficient d00 (a) equals to −λ(a) for any
level analysis, the tachyon field which minimizes the potential can be expressed by other
fields in the following,


 −β + β 2 − 4αγ
 (a − a a + ),
t= 2α (3.21)

 −β − β − 4αγ (−1/2 a < a − , a > a + ).
 2
2α
Here, α, β and γ are given by
α = t000 ,

N
β = −λ(a) + t00i ψi , (3.22)
i=1

N
N
γ= d0i (a)ψi + t0ij ψi ψj , (3.23)
i=1 i,j =1
where N is the number of truncated fields. Thus, the branch is determined depending on
the value of a in our analysis.
Let us consider the level two approximation. The level two field is given by
(2)
Φ = t|s0 + ψ1 |s1 + ψ2 |s2
= tc1 |0 + ψ1 (α−1 · α−1 )c1 |0 − ψ2 c−1 |0. (3.24)
From (3.10) and (3.15), we find the quadratic term of the modified universal function as

1 1
2π 2 − λ(a)t 2 + 26 1 + a + 4aZ(a) ψ12 − 1 + a + 4aZ(a) ψ22
2 2

1
+ 13atψ1 + atψ2 . (3.25)
2
The potential minimum is depicted in Fig. 3. We observe that (2, 4) and (2, 6)
approximations lead to almost same results. At a = −1/2, the vacuum expectation
values of component fields and the potential minimum are equal to zero as used (0, 0)
approximation. The vacuum energy are varying slowly in the neighborhood of a = 0, in
which the potential height provide 96% of the D-brane tension. For our conjectures, it is
a desirable fact that the potential minimum changes slowly along the expected vacuum
energy. Thus, even at this level, we can expect that the universal solution is a pure gauge at
least near a = 0, as our conjecture.
Let us progress to the higher level approximation. The potential minimum using level
four and six approximations is depicted in Fig. 4. Like level zero and two cases, the results
of the level (L, 3L) calculation slightly change with the level (L, 2L). In the cases of
Fig. 3. The potential minimum in level two truncations.
Fig. 4. Plots of the potential minimum in level four and six truncations.
both of level four and six, the potential minimum displays a flat region along the expected
vacuum energy. The potential heights and the vacuum expectation values become zero at
a = −1/2 as before. This behavior can be seen in detail in Fig. 5, which magnifies the area
near a = −1/2.
Fig. 5. The enlarged graph of the potential minimum around a = −1/2.
Table 1
Vacuum energy in level truncation scheme for several points of a
Level V /T25
a = −0.2 a = −0.1 a = 0.0 a = 0.1 a = 0.2
(0, 0) −0.233203 −0.462912 −0.684616 −0.866692 −0.995360
(2, 4) −0.777067 −0.908062 −0.948553 −0.955031 −0.956909
(2, 6) −0.854866 −0.944975 −0.959377 −0.955239 −0.955100
(4, 8) −0.965369 −0.986459 −0.986403 −0.985538 −0.985329
(4, 12) −0.988826 −0.990313 −0.987822 −0.986499 −0.985300
(6, 12) −0.993496 −0.995449 −0.994773 −0.994077 −0.993590
(6, 18) −0.996274 −0.996056 −0.995177 −0.994346 −0.993715
In these analyses, it is a remarkable fact that the higher the approximation level is
increased, the wider the flat region grows. Moreover, the potential value in the flat region
approaches the expected value increasingly as the level is raised. At level six, the vacuum
energy becomes almost −1 in the region from −0.2 to 1. In Table 1, we pick out the
values of the potential height for several points of a.4 All of the values are about 99% of
the D-brane tension at level six. For our conjecture, the most important result is that the
stable vacuum disappears at a = −1/2 in every level analysis. These potential behavior to
the parameter a suggests that, if the approximation level approaches infinity, the potential
minimum takes the value of −1 for a > −1/2, but it remains being zero at a = −1/2.
Hence, these results lead us to believe that the conjecture for the universal solutions,
which is expressed by (3.7), should be true, and then the universal solution at a = −1/2
corresponds to the tachyon condensation conjectured by Sen.
4 Of course, our values at a = 0 agree with previous results in Refs. [4–6]. But only the value at level (6, 12)
does not coincide with a result in [5].
3.4. Other universal solutions

We can provide other universal solutions by choosing the function in (2.3) as

a 2
hla (w) = log 1 − (−1)l wl − (−1)l w−l , (3.26)
2
where l = 1, 2, 3, . . . [14]. The case of l = 1 corresponds to the previous example. The
action around the solution has the modified BRS charge
a
QlB (a) = (1 + a)QB + (−1)l (Q2l + Q−2l )
2
+ 4al Z(a)c0 + (−1)l al 2 Z(a)2 (c2l + c−2l )
2
∞

− 2al 2 1 − Z(a)2 (−1)nl Z(a)n−1 (c2nl + c−2nl ). (3.27)
n=2
The kinetic operator in Siegel gauge is given by
a
Ll (a) = Ql (a), b0 = (1 + a)L0 − (−1)l (L2l + L−2l )
2
− (−1)l al(q2l − q−2l ) + 4al 2 Z(a). (3.28)
At level zero, the modified universal function becomes

1 1
fal (t) = 2π 2 − λl (a)t 2 + K 3 t 3 ,
2 3
√ 2
λl (a) = 4l 1 + 2a − 4l − 1 (1 + a).
2
(3.29)
At the local minimum it takes the value

 π
2
− 6 λl (a)3 al− a al+ ,
fa (t0 ) = 3K (3.30)

0 −1/2 a < al− or a > al+ ,
where the branch points al± are given by
√
8l 2 − 1 ± 4l 2 8l 1 − 1
al± = . (3.31)
(4l 2 − 1)2
Let us consider the case of l = 2. The kinetic operator is given by
a
L2 (a) = (1 + a)L0 − (L4 + L−4 ) − 2a(q4 + q−4 ) + 16aZ(a). (3.32)
2
Similarly to the case of l = 1, the quadratic terms up to level six can be calculated by the
following equations,
1
−4 |η0 = |η3 + |η4 ,
Lmat
2
1
−4 |η1 = |η9 + |η6 + |η10 ,
Lmat
2
gh
L−4 − 4q−4 |χ0 = −3|χ4 − 2|χ5 − |χ6 ,
gh
L−4 − 4q−4 |χ1 = 5|χ7 + |χ11 + 2|χ12. (3.33)
Fig. 6. The potential minimum for the universal solution with l = 2.
The branch points are

a2− = −0.258 . . ., a2+ = 0.533 . . . . (3.34)
The results of numerical analysis up to level six are depicted in Fig. 6. At a = −1/2, the
potential minimum remains being zero for any approximation level. From level four, the
flat region begins to spread along the expected vacuum energy. As a different feature from
the previous case, we observe that, if the level is raised, the shape of the curve changes with
the period of level two, in particular near the point that the potential minimum approaches
to zero. On the whole we can see that the curve tends to approach the step function (3.7)
as increasing the level. However, it seems to converge more slowly than the case of l = 1.
4. Summary and discussions
We have studied a non-perturbative vacuum in open string field theory around the
universal solutions by using a level truncation scheme. We have observed that, as increasing
the approximation level, the potential minimum gradually approaches to the step function
which equals to the negative D-brane tension for almost parameters and becomes zero at
the boundary. These results strongly support our conjecture that the universal solution are
pure gauge for the almost parameter region, but the non-trivial solution at the boundary
corresponds to the tachyon vacuum solution.
Though this physical interpretation for the universal solutions is plausible, we can
construct them as many as functions used in Eq. (2.3). At least, the non-trivial solutions are
countable by the natural number as in Eq. (3.26). On the other hand the tachyon vacuum
should be unique in string theory. In order for them to be consistent, these universal
solutions must be changed by gauge transformations. Though this gauge equivalence
remains to be proved, this may be probably possible by the gauge transformation related
to the operators Kn = Ln − (−1)n L−n as discussed in [14]. In addition, the equivalence
between our solutions in this paper is consistent with the conjecture proposed by Drukker
[23,24] which concerns the order of zero of the function with the number of D-branes.
Finally, two problems at least are remained to be solved in order to prove the
correspondence between the universal solutions and the tachyon vacuum. First, we should
calculate the potential height of the universal solutions exactly instead of these numerical
analysis. Secondly, we should find closed strings in the theory around the solutions. In
[24] Drukker discussed how closed strings should be appeared in the theory, but we should
show them more explicitly, for example, as closed string poles in amplitudes.
Acknowledgements
We would like to thank H. Hata, M. Kanemune, I. Kishimoto and T. Kugo for valuable
discussions and comments.
Appendix A. Mode expansion of the modified BRS charge
The modified BRS charge QB (a) is defined in (2.20) and the function ha (w) is given
by (2.10). We can easily evaluate the first term of (2.20) as

h a 1 2 a
Q e =Q 1+
a
w+ = Q 1 + a + w2 + w−2
2 w 2
a
= (1 + a)QB + (Q2 + Q−2 ). (A.1)
2
From (2.10), we can find
2
∂ha (w) eha (w) = aw−1 ∂ha (w) w2 − w−2 . (A.2)
If we differentiate the Fourier series of (2.11), we get
∞

−1
∂ha (w) = −4iw (−1)n Z(a)n sin 2nσ, (A.3)
n=1
where w = exp(iσ ). From these equations, it follows that

2
∂ha (w) eha (w)
∞

= 8aw−2 (−1)n Z(a)n sin 2nσ sin 2σ
n=1
∞

−2
= 8aw −Z(a) sin 2σ +
2 n n
(−1) Z(a) sin 2nσ sin 2σ . (A.4)
n=2
Substituting (A.4) into (2.20), we can evaluate the second term of (2.20) as
∞
π
dσ −inσ 2
C (∂ha ) e 2 ha
= −8aZ(a) cn e sin 2σ
n=−∞
2π
−π
∞
∞
π
dσ −inσ
+ 8a cn (−1)m Z(a)m e sin 2mσ sin 2σ. (A.5)
n=−∞
2π
m=2 −π
We can calculate the integrations in this equation and the results are
1
π 
2 (n = 0),
dσ −inσ 2
e sin 2σ = − 1
(n = ±4), (A.6)
2π 
 4
−π 0 (otherwise),
1
π 
4 (n = ±2(m − 1)),
dσ −inσ
e sin 2mσ sin 2σ = − 1
(n = ±2(m + 1)), (A.7)
2π 
 4
−π 0 (otherwise).
Substituting (A.6) and (A.7) into (A.5), we find that

1 1
C (∂ha )2 eha = −8aZ(a) c0 − (c4 + c−4 )
2 4
∞
1
+ 8a (−1)m Z(a)m (c2(m−1) + c−2(m−1) )
4
m=2

1
− (c2(m+1) + c−2(m+1) )
4
= −4aZ(a)c0 + 2aZ(a)2(c2 + c−2 )
∞

+ 2a 1 − Z(a)2 (−1)n Z(a)n−1 (c2n + c−2n ). (A.8)
n=2
Finally, from (A.1) and (A.8), we obtain the oscillator expression of the modified BRS
charge as in (3.2).
Appendix B. Table of matter and ghost states at levels 6
Table 2 describe the matter and ghost states which contribute to scalar fields at levels
gh
6. The inner products Amat
ij and Aij are defined by Eqs. (3.11) and (3.12).
Table 2
State Inner products
|η0 |0 Amat
00 = 1
|η1 (α−1 · α−1 )|0 Amat
11 = 52
|η2 (α−1 · α−2 )|0 Amat
22 = 52
|η3 (α−1 · α−3 )|0 Amat
33 = 78
|η4 (α−2 · α−2 )|0 Amat
44 = 208
|η5 (α−1 · α−1 )(α−1 · α−1 )|0 Amat
55 = 5824
|η6 (α−1 · α−5 )|0 Amat
66 = 130
|η7 (α−2 · α−4 )|0 Amat
77 = 208
|η8 (α−3 · α−3 )|0 Amat
88 = 468
|η9 (α−1 · α−3 )(α−1 · α−1 )|0 Amat
99 = 4368
|η10 (α−2 · α−2 )(α−1 · α−1 )|0 Amat mat
1010 = 10816, A1011 = 416
|η11 (α−1 · α−2 )(α−1 · α−2 )|0 Amat mat
1111 = 5616, A1110 = 416
|η12 (α−1 · α−1 )(α−1 · α−1 )(α−1 · α−1 )|0 Amat
1212 = 1048320
gh
|χ0 |1 A0,0 = 1
gh
|χ1 b−1 c−1 |1 A1,1 = − 1
gh
|χ2 b−1 c−2 |1 A2,3 = − 1
gh
|χ3 b−2 c−1 |1 A3,2 = − 1
gh
|χ4 b−1 c−3 |1 A4,6 = − 1
gh
|χ5 b−2 c−2 |1 A5,5 = − 1
gh
|χ6 b−3 c−1 |1 A6,4 = − 1
gh
|χ7 b−1 c−5 |1 A7,11 = − 1
gh
|χ8 b−2 c−4 |1 A8,10 = − 1
gh
|χ9 b−3 c−3 |1 A9,9 = − 1
gh
|χ10 b−4 c−2 |1 A10,8 = − 1
gh
|χ11 b−5 c−1 |1 A11,7 = − 1
gh
|χ12 b−2 b−1 c−2 c−1 |1 A12,12 = 1
Appendix C. Table of scalar states at levels 6
Table 3 lists scalar states at levels 6. ψi (6,18) denote vacuum expectation values of
the scalar states ψi in level truncation calculations at (6, 18).
Table 3
ψi State ψi (6,18)
a = −0.50 a = −0.40 a = −0.25 a = 0.00 a = 0.50 a = 1.00
ψ0 |η0 ⊗ |χ0 0.00000 0.49637 0.56739 0.54793 0.48059 0.42623
ψ1 |η1 ⊗ |χ0 0.00000 0.07446 0.05809 0.02857 0.00038 −0.01221
ψ2 |η0 ⊗ |χ1 0.00000 −0.31083 −0.30955 −0.21181 −0.10466 −0.04878
ψ3 |η3 ⊗ |χ0 0.00000 0.01000 −0.00008 −0.00573 −0.00524 −0.00271
ψ4 |η4 ⊗ |χ0 0.00000 −0.00351 −0.00355 −0.00255 −0.00152 −0.00100
ψ5 |η5 ⊗ |χ0 0.00000 0.00426 0.00176 −0.00016 −0.00055 −0.00018
ψ6 |η1 ⊗ |χ1 0.00000 −0.02664 −0.01077 0.00370 0.00817 0.00629
ψ7 |η0 ⊗ |χ4 0.00000 −0.05489 0.00881 0.05739 0.06430 0.04973
ψ8 |η0 ⊗ |χ5 0.00000 0.04502 0.04922 0.03406 0.01876 0.01125
ψ9 |η0 ⊗ |χ6 0.00000 −0.01830 0.00294 0.01913 0.02143 0.01658
ψ10 |η6 ⊗ |χ0 0.00000 0.00273 0.00178 0.00175 0.00180 0.00138
ψ11 |η7 ⊗ |χ0 0.00000 0.00049 0.00128 0.00146 0.00108 0.00075
ψ12 |η8 ⊗ |χ0 0.00000 0.00118 0.00084 0.00072 0.00066 0.00049
ψ13 |η9 ⊗ |χ0 0.00000 0.00111 0.00025 0.00014 0.00036 0.00031
ψ14 |η10 ⊗ |χ0 0.00000 −0.00015 −0.00003 0.00008 0.00009 0.00007
ψ15 |η11 ⊗ |χ0 0.00000 0.00002 0.00001 0.00001 0.00001 0.00000
ψ16 |η12 ⊗ |χ0 0.00000 0.00012 0.00003 0.00000 0.00001 0.00001
ψ17 |η3 ⊗ |χ0 0.00000 −0.00501 −0.00232 −0.00132 −0.00167 −0.0015
ψ18 |η4 ⊗ |χ1 0.00000 −0.00059 −0.00074 −0.00058 −0.00034 −0.00022
ψ19 |η5 ⊗ |χ1 0.00000 −0.00096 −0.00021 −0.00004 −0.00025 −0.00027
ψ20 |η2 ⊗ |χ2 0.00000 −0.00012 −0.00008 −0.00006 −0.00005 −0.00003
ψ21 |η2 ⊗ |χ3 0.00000 −0.00006 −0.00004 −0.00003 −0.00003 −0.00002
ψ22 |η1 ⊗ |χ4 0.00000 −0.00506 −0.00131 −0.00168 −0.00376 −0.00371
ψ23 |η1 ⊗ |χ5 0.00000 0.00126 −0.00038 −0.00142 −0.00123 −0.00084
ψ24 |η1 ⊗ |χ6 0.00000 −0.00169 −0.00044 −0.00056 −0.00125 −0.00124
ψ25 |η0 ⊗ |χ7 0.00000 −0.04011 −0.03204 −0.03001 −0.03071 −0.02579
ψ26 |η0 ⊗ |χ8 0.00000 −0.00536 −0.01649 −0.01875 −0.01288 −0.00819
ψ27 |η0 ⊗ |χ9 0.00000 −0.01448 −0.01167 −0.01142 −0.01206 −0.01018
ψ28 |η0 ⊗ |χ10 0.00000 −0.00268 −0.00824 −0.00938 −0.00644 −0.00409
ψ29 |η0 ⊗ |χ11 0.00000 −0.00802 −0.00641 −0.00600 −0.00614 −0.00516
ψ30 |η0 ⊗ |χ12 0.00000 −0.00765 −0.01029 −0.00773 −0.00413 −0.00243
References
[1] A. Sen, Universality of the tachyon potential, J. High Energy Phys. 9912 (1999) 027, hep-th/9911116.
[2] A. Sen, Descent relations among bosonic D-branes, Int. J. Mod. Phys. A 14 (1999) 4061, hep-th/9902105.
[3] V.A. Kostelecky, S. Samuel, On a nonperturbative vacuum for the open bosonic string, Nucl. Phys. B 336
(1990) 236.
[4] A. Sen, B. Zwiebach, Tachyon condensation in string field theory, J. High Energy Phys. 0003 (2000) 002,
hep-th/9912249.
[5] N. Moeller, W. Taylor, Level truncation and the tachyon in open bosonic string field theory, Nucl. Phys.
B 538 (2000) 105, hep-th/0002237.
[6] D. Gaiotto, L. Rastelli, Experimental string field theory, hep-th/0211012.
[7] L. Rastelli, A. Sen, B. Zwiebach, String field theory around the tachyon vacuum, Adv. Theor. Math. Phys. 5
(2002) 353, hep-th/0012251.
[8] L. Rastelli, A. Sen, B. Zwiebach, Boundary CFT construction of D-branes in vacuum string field theory, J.
High Energy Phys. 11 (2001) 045, hep-th/0105168;
L. Rastelli, A. Sen, B. Zwiebach, Star algebra spectroscopy, J. High Energy Phys. 0203 (2002) 029, hep-
th/0111281;
H. Hata, T. Kawano, Open string states around a classical solution in vacuum string field theory, J. High
Energy Phys. 0111 (2001) 038, hep-th/0108150;
H. Hata, S. Moriyama, Observables as twist anomaly in vacuum string field theory, J. High Energy
Phys. 0201 (2002) 042, hep-th/0111034;
H. Hata, S. Moriyama, Reexamining classical solution and tachyon mode in vacuum string field theory,
Nucl. Phys. B 651 (2003) 3, hep-th/0206208;
H. Hata, H. Kogetsu, Higher level open string states from vacuum string field theory, J. High Energy
Phys. 0209 (2002) 027, hep-th/0208067;
H. Hata, S. Moriyama, S. Teraguchi, Exact results on twist anomaly, J. High Energy Phys. 0202 (2002) 036,
hep-th/0201177;
K. Okuyama, Ghost kinetic operator of vacuum string field theory, J. High Energy Phys. 0201 (2002) 027;
K. Okuyama, Siegel gauge in vacuum string field theory, J. High Energy Phys. 0201 (2002) 043, hep-
th/0111087;
K. Okuyama, Ratio of tensions from vacuum string field theory, J. High Energy Phys. 0203 (2002) 050,
hep-th/0201136;
T. Okuda, The equality of solutions in vacuum string field theory, Nucl. Phys. B 641 (2002) 393, hep-
th/0201149;
Y. Okawa, Open string states and D-brane tension from vacuum string field theory, J. High Energy
Phys. 0207 (2002) 003, hep-th/0204012.
[9] T. Takahashi, S. Tanimoto, Wilson lines and classical solutions in cubic open string field theory, Prog. Theor.
Phys. 106 (2001) 863, hep-th/0107046.
[10] T. Takahashi, S. Tanimoto, Marginal and scalar solutions in open cubic string field theory, J. High Energy
Phys. 03 (2002) 033, hep-th/0202133.
[11] I. Kishimoto, K. Ohmori, CFT description of identity string field: toward derivation of the VSFT action, J.
High Energy Phys. 05 (2002) 036, hep-th/0112169.
[12] J. Klusoň, Exact solutions of open bosonic string field theory, J. High Energy Phys. 04 (2002) 043, hep-
th/0202045;
J. Klusoň, Marginal deformations in the open bosonic string field theory for N D0-branes, Class. Quantum
Grav. 20 (2003) 827, hep-th/0203089;
J. Klusoň, New solution of the open bosonic string field theory, hep-th/0205294;
J. Klusoň, Time dependent solution in open bosonic string field theory, hep-th/0208028;
J. Klusoň, Exact solutions in open bosonic string field theory and marginal deformation in CFT, hep-
th/0209255.
[13] G.T. Horowitz, J. Lykken, R. Rohm, Purely cubic action for string field theory, Phys. Rev. Lett. 57 (1986)
283.
[14] I. Kishimoto, T. Takahashi, Open string field theory around universal solutions, Prog. Theor. Phys. 108
(2002) 591, hep-th/0205257.
[15] T. Takahashi, S. Zeze, Gauge fixing and scattering amplitudes in string field theory expanded around
universal solutions, Prog. Theor. Phys. 110 (2003) 159–177, hep-th/0304261.
[16] E. Witten, Non-commutative geometry and string field theory, Nucl. Phys. B 268 (1986) 253.
[17] M. Schnabl, Wedge states in string field theory, hep-th/0201095.
[18] T. Yoneya, Phys. Lett. B 197 (1987) 76.
[19] H. Hata, Y. Nagoshi, Prog. Theor. Phys. 80 (1988) 1088.
[20] T. Kugo, B. Zwiebach, Target space duality as a symmetry of string field theory, Prog. Theor. Phys. 87
(1992) 801, hep-th/9201040.
[21] D. Gross, A. Javichi, Operator formalism of interacting string field theory (I), Nucl. Phys. B 283 (1987) 1;
D. Gross, A. Javichi, Operator formalism of interacting string field theory (II), Nucl. Phys. B 287 (1987)
225.
[22] I. Kishimoto, T. Takahashi, unpublished.
[23] N. Drukker, Closed string amplitudes from gauge fixed string field theory, Phys. Rev. D 67 (2003) 126004,
hep-th/0207266.
[24] N. Drukker, On different for the vacuum of bosonic string field theory, J. High Energy Phys. 08 (2003) 017,
hep-th/0301079.
Wilson lines corrections to gauge couplings

from a field theory approach
D.M. Ghilencea
DAMTP, CMS, University of Cambridge, Wilberforce Road, Cambridge, CB3 0WA, United Kingdom
Received 13 May 2003; accepted 8 August 2003
Abstract
Using an effective field theory approach, we address the effects on the gauge couplings of one
and two additional compact dimensions in the presence of a constant background (gauge) field. Such
background fields are a generic presence in models with extra dimensions and can be employed
for gauge symmetry breaking mechanisms in the context of 4D N = 1 supersymmetric models. The
structure of the ultraviolet (UV) and infrared (IR) divergences that the gauge couplings develop in the
presence of Wilson line vev’s is investigated. One-loop radiative corrections to the gauge couplings
due to overlapping effects of the compact dimensions and Wilson line vev’s are computed for generic
4D N = 1 models. Values of Wilson lines vev’s corresponding to points (in the “moduli” space) of
enhanced gauge symmetry cannot be smoothly reached perturbatively from those corresponding to
the broken phase. The one-loop corrections are compared to their (heterotic) string counterpart in the
“field theory” limit α → 0 to show remarkably similar results when no massless states are present
in a Kaluza–Klein tower. An additional correction to the gauge coupling exists in the effective field
theory approach when for specific Wilson lines vev’s massless Kaluza–Klein states are present. This
correction is not recoverable by the limit α → 0 of the (infrared regularised) string because the
infrared regularisation limit and the limit α → 0 of the string result do not commute.
 2003 Published by Elsevier B.V.
1. Introduction
The phenomenological and theoretical implications of the physics of “large” extra

dimensions has recently attracted an increased research interest in the context of effective
field theory approaches. String theory is ultimately thought to provide a complete and fully
consistent description of the high energy physics. Nevertheless, effective field theory (EFT)
E-mail address: dmg29@damtp.cam.ac.uk (D.M. Ghilencea).
0550-3213/$ – see front matter  2003 Published by Elsevier B.V.

doi:10.1016/j.nuclphysb.2003.08.011
184 D.M. Ghilencea / Nuclear Physics B 670 (2003) 183–220
approaches are able to describe accurately many aspects of the physics of extra dimensions,
without relying on the string picture.
In this work we adopt an effective field theory approach to investigate the corrections
to the gauge couplings in 4D N = 1 supersymmetric models with one and two extra
dimensions, in the presence of a constant background gauge field. The extra dimensions
may be compactified on manifolds or orbifolds thereof [1] providing the possibility to
construct chiral models. The compactification chosen and the value of the background
gauge field (so-called Wilson lines’ vev’s [2,3]) have implications for the one-loop
corrections to the 4D gauge couplings and for the amount of gauge symmetry present.
The interplay of these two effects on the 4D gauge couplings will be discussed in this work
for the class of models considered. Let us first present this problem in detail.
Additional compact dimensions can induce significant changes to the 4D gauge
couplings. The one-loop corrected coupling in an orbifold compactification to a 4D N = 1
supersymmetric model is
4π 4π bi Ms
= + ln + Ω̃i + · · · , i: gauge group index. (1)
gi2 (Q) gi2 (Ms ) 2π Q
Q is a low energy scale above the supersymmetry breaking scale, Ms is the ultraviolet
scale or in the case of string theory, the string scale. In the EFT approach gi (Ms )
is the tree level (“bare”) coupling while in the (heterotic) string gi (Ms ) is actually a
gauge group independent function of the so-called S and T moduli, invariant under
SL(2, Z)T × SL(2, Z)U × Z2T ↔U [4]. This ensures that the string coupling is a well-defined
expansion parameter, invariant under this symmetry of the string. Similar considerations
may apply to other string models [5]. Further, the logarithmic term in (1) is due to the
(infrared and ultraviolet regularised) contribution of the light (“massless”) states charged
under the gauge group. In a realistic model these states are N = 1 multiplets and account
for the spectrum of the Minimal Supersymmetric Standard Model (MSSM) or similar
models. In various compactifications to 4D such states can arise as the 4D Kaluza–Klein
“zero” (or massless) modes of the initial (higher-dimensional) fields after compactification.
If the 4D N = 1 string orbifold models that we consider in this paper (for a review see
[8]) have an N = 2 sector of states (“bulk”) (e.g., Z4 orbifold) other one-loop corrections
Ω̃i exist [9]. Such orbifolds can also have N = 4 sectors, but these do not affect the
couplings due to the higher amount of supersymmetry. In an EFT description Ω̃i accounts
for the sum of individual one-loop corrections due to massive Kaluza–Klein modes (non-
zero levels) associated with the compactification and charged under the gauge group. Such
states which contribute are organised as N = 2 multiplets [7,9]. In the heterotic string
picture Ω̃i includes [9] in addition the effect of the so-called winding states associated
with the extra dimensions and symmetries of the string (e.g., modular invariance) and
which have no EFT description. Other string constructions [5] bring tadpole cancellation
constraints on Ω̃i , again with no clear EFT equivalent, or may relate Ω̃i to the free energy
of compactification [10]. Despite such additional string effects, the EFT results and the field
theory limit (i.e., infinite string scale) of some string calculations can lead to somewhat
“similar” results for Ω̃i . An example is the power-like dependence of the couplings on the
scale [6]. However, understanding the exact relationship between such approaches requires
a careful investigation.
D.M. Ghilencea / Nuclear Physics B 670 (2003) 183–220 185
A step towards clarifying this relationship is the analysis in [11] where Ω̃i was
computed on pure EFT grounds for 4D N = 1 orbifold compactifications with an N = 2
sector (this is effectively a “bulk” as a two-dimensional torus, while any completely
untwisted N = 4 sector of such orbifold does not affect Ω̃i ). This calculation was done
by summing (infinitely many) one-loop corrections due to associated massive Kaluza–
Klein states. The result agrees with the limit α → 0 (α ≈ 1/Ms2 ) of the heterotic string
result [9] due to massive Kaluza–Klein and winding modes (in this limit string effects such
as winding modes effects were shown to be suppressed1). However, differences emerge
between the EFT and string results [12] when one includes the effect of the massless states
of the theory (in this particular case these were Kaluza–Klein states of level “zero”2). The
one-loop correction of the massless states is infrared (IR) divergent both in the EFT and
in string case. The differences mentioned between the EFT and the limit α → 0 of the
string result are caused by the infrared regularisation of the string [12]. When the string
IR regulator is removed, this regularisation discards α dependent terms (divergent for
α → 0) which multiply the regulator. These terms become relevant in the (field theory)
limit α → 0 which does not commute with the IR regularisation of the string. Such terms
are present in the final EFT correction to gi2 [12]. In the models we address we will obtain
such terms whenever massless Kaluza–Klein modes are present.
The discussion so far did not exhaust all possible one-loop effects in Ω̃i that one
encounters in string or EFT models with extra dimensions. Such models usually have a
larger amount of gauge symmetry than the Standard Model does. A mechanism to reduce it
is then required in a realistic model. Introducing the usual Higgs mechanism is not always
the most economical approach. For multiply-connected manifolds a symmetry breaking
mechanism exists known as the Hosotani mechanism [2] or Wilson-line symmetry
breaking [3]. Explicit models of this type are known at the string level [13–16]. In this
mechanism a constant background gauge field vev (of higher-dimensional components of
the gauge fields) controls the amount of gauge symmetry left after compactification. It also
affects the free energy of compactification [17] (also [10]) and may “shift” the 4D Kaluza–
Klein mass spectrum of the initial, higher-dimensional fields. As a result of this change
new radiative corrections to the gauge couplings are expected.
The gauge symmetry breaking by Wilson lines is spontaneous and provides a viable
approach to model building (see [18] for the relation to orbifold breaking). The effect of the
Wilson lines may be reexpressed as a “twist” in the boundary conditions for the initial fields
(with respect to the compact dimensions) and which is removed by the limit of vanishing
Wilson lines vev’s. Since the breaking is spontaneous, the UV behaviour (encoded in the
couplings gi2 ) of the 4D models obtained after compactification should not be worsened
by non-zero Wilson lines vev’s. To check this we evaluate the 4D Kaluza–Klein masses in
the presence of non-zero Wilson lines vev’s, to investigate the radiative corrections to gi2
and their dependence (continuity) on these vev’s. We show that these corrections have a
UV scale dependence similar to that when Wilson lines have vanishing vev’s. As for the
1 Even in this limit winding modes still play an UV role [11] in fixing the numerical coefficient of the UV
leading term of Ω̃i when α → 0. At the EFT level this coefficient is regulator dependent.
2 As we will discuss, such massless states may also appear from non-zero Kaluza–Klein levels if non-vanishing
background fields (Wilson lines vev’s) exist.
IR behaviour a regularisation is needed when for specific Wilson line vev’s some Kaluza–
Klein modes of non-zero level may become massless and induce a gauge symmetry change.
Our intention is to present a general EFT method to compute radiative corrections to the
4D gauge couplings due to massive Kaluza–Klein states in the presence of Wilson lines.
The framework is that of 4D N = 1 supersymmetric models with one and two compact
dimensions, which correspond at a string level to 4D N = 1 orbifolds with an N = 2 sector
of massive states, in the presence of Wilson line background [20]. The method evaluates
the UV and IR behaviour of the couplings and their one-loop correction in function of
the Wilson line vev’s and may easily be applied to specific models. Our results apply if the
radii of compactification are “large” (in units of UV cut-off), without any reference to string
theory. The EFT one-loop correction is compared to its (heterotic) string counterpart in the
limit α → 0 to find remarkably similar results. The correction may affect the unification
of gauge couplings in MSSM-like models derived from the heterotic string [19]. Wilson
lines corrections to the gauge couplings were not computed previously in a field theory
approach. They were studied in the heterotic string in [17,20]. See [21] for compactification
on G2 manifolds.
The paper is organised as follows. In Section 2 we discuss the Wilson line breaking
of the gauge symmetry and its effects on the masses of 4D Kaluza–Klein modes of the
initial higher-dimensional fields. We then address the effects of the Kaluza–Klein states
and Wilson lines vev’s on the gauge couplings (Section 3). The conclusions are given in
Section 4. Appendices A and B provide extensive technical details for computing general
Kaluza–Klein integrals used in the text (in DR and proper-time regularisation) and these
results can be used for other applications as well.
2. Wilson line effects and extra dimensions
2.1. Definition of the models and Wilson line symmetry breaking
To begin with we review the gauge symmetry breaking by Wilson lines and reexpress
it in terms of boundary conditions for higher-dimensional fields. The class of EFT models
considered in this work is that of 4D N = 1 supersymmetric orbifold models with gauge
symmetry group G larger than the Standard Model (SM) group (e.g., SU(5)). The models
are assumed to have in addition to an N = 1 spectrum an N = 2 sector of massive states
(“bulk”) associated with the extra dimensions ym compactified on a circle (m = 1) or a two-
dimensional torus T 2 (m = 1, 2) and are regarded as the “field theory” limit (Ms → ∞)
of a string compactification. A string embedding of such models is a compactification on
T 6 /P with T 6 a six-dimensional torus and P the point group, subgroup of SU(3) [22] (e.g.,
Z4 ). The action of elements of P on the six compact dimensions can leave one complex
plane (ym ) unrotated (giving the N = 2 sector), rotate all three complex planes (the N = 1
sector), and rotate none of them (the N = 4 sector). Therefore the string compactification
has in addition to the N = 2 and N = 1 sectors, an N = 4 sector as well. The N = 1
sector gives the usual (MSSM-like) logarithmic corrections to the gauge couplings while
the N = 4 sector does not affect them. There remains the N = 2 sector (“bulk”) of the
unrotated plane (ym ) compactified on T 2 or a circle (if one dimension has radius set equal
to 1/Ms ). Finally, a constant background (gauge) field may be present. This gives the string
embedding of our EFT models. It also justifies our considering of EFT models with extra
dimensions ym compactified on a circle or a two-torus (corresponding to the N = 2 sector)
rather than on orbifolds thereof. From now on we use an EFT approach and always refer
to this sector only. From the 4D perspective towers of Kaluza–Klein states are present
associated with the compact dimension(s) ym and which correspond to initial, higher-
dimensional fields charged under G. In the string picture these modes build up together
with the winding modes, N = 2 multiplets of 4D N = 1 orbifolds with non-zero Wilson
line vev’s [17,20].
First, the group G can be broken spontaneously by those Wilson lines vev’s which
“survive” (i.e., commute with) any orbifold action on the fields (see [23] for examples
at the EFT level). A Wilson line operator is defined as

Wi = exp i dym AIym T I , I = 1, . . . , rk G (2)
γi
with a sum over I and m understood; γi labels the contour(s) of integration over the ith
compact dimension(s) (cycle). For two extra dimensions Ay1 and Ay2 should commute,
otherwise the 4D effective action would contain terms Tr F 2 ∝ Tr[Ay1 , Ay2 ]2 from the field
strength in 6D [24]. For phenomenological purposes, we would like to avoid such terms,
thus Ay1 and Ay2 will lie in the Cartan sub-algebra of the Lie algebra of G (of rank rk G). In
(2) TI stands for a generator of this sub-algebra. The gauge symmetry left unbroken by the
Wilson lines vev’s is that whose generators commute with all Wi . Using the commutators
in the Weyl–Cartan basis [25]
[TI , TJ ] = 0, [TI , Eα ] = αI Eα ,
I, J = 1, . . . , rk G; α = 1 + rk G, . . . , dim G (3)
one shows3
[Wi , TI ] = 0, i = 1, 2;

1
[Wi , Eα ] = 0 ⇐⇒ ρi,α ≡− dym AIym αI = 0 mod n,
2π
γi
n ∈ Z, i = 1, 2, (4)
where a sum over I and m is understood; α has components αI , I = 1, rk G and denotes
the root associated with the generator Eα . The first relation in (4) shows that the rank (rk)
of the group is not changed, while the second relation controls the amount of symmetry
breaking, through the vev’s AIym , I = 1, rk G of the Wilson lines in various directions in
the root space. If the constraint in the rhs of the second line of Eq. (4) is not respected, some
gauge fields “outside” the Cartan sub-algebra become massive4 and the gauge symmetry G
3 We also use that exp(iu T )E exp(−iu T ) = E exp(iu α ), where a sum over I is understood.
I I α I I α I I
4 Examples of such fields would be for the SU(5) case the so-called X, Y fields.
is reduced.5 We denote by G ∗ this remaining symmetry, generated by TI (I = 1, . . . , rk G)

and those Eα with vanishing ρi,α , i = 1, 2 (if no such Eα existed, then G would be broken
to a product of U (1)’s). Additional (model-dependent) constraints may apply to ρi,α in
the string case which depend on the embedding of the point group in the gauge group
E8 × E8 . To keep the EFT approach general we do not impose such constraints but keep
ρi,α as parameters throughout the calculation. In the final EFT result such constraints can
then easily be implemented. Further insight into the symmetry breaking G → G ∗ is gained
by using a y-dependent gauge transformation as we discuss separately for one and two
extra dimensions.
2.2. Effects on the 4D Kaluza–Klein mass spectrum: one extra dimension
Consider the 5D fields Aµ̃ (x, y) and Φ(x, y) in the adjoint and fundamental represen-
tations of G respectively, with x ∈ M 4 and index µ̃ = {µ, y}, µ = 0, 3. One has
Aµ̃ (x, y + 2πR) = QAµ̃ (x)Q† , Φ(x, y + 2πR) = QΦ(x, y), (5)
where R is the radius of the extra dimension y and Q is some global transformation. For
our purpose one can actually set Q = 1 as the conclusions below will not depend on this. In
the following we assume Ay constant (position independent) and attempt to “gauge away”
the field Ay using a y-dependent transformation U (y) (or U (y)Q−1 if Q is included). The
new fields are
Ay (x, y) = 0 if U (y) ≡ e−iyAy ,
Aµ (x, y) = U (y)Aµ (x, y)U −1 (y) = AIµ (x, y)TI + Aαµ (x, y)Eα e−iyAy αI ,
I
Φ (x, y) = U (y)Φ(x, y), (6)
where a sum over I and α is understood. Since the generators TI , Eα form a linear
independent set, the second equation says that the fields AIµ in the Cartan sub-algebra
do not “feel” the “background” field Ay and are invariant under U (y). However, fields
outside the Cartan algebra (Aαµ ) are transformed and the same applies to the field in the
fundamental representation. The initial condition Eq. (5) is then changed into Ay = 0 and
AµI (x, y + 2πR) = AµI (x, y),
Aµα (x, y + 2πR) = e−i2πRAy αI Aµα (x, y),

I
Φλ (x, y + 2πR) = e−i2πRAy λI Φλ (x, y),

I
(7)
where λI are the weights (eigenvalues of TI ) and Φλ denotes a component of the

multiplet Φ. The new fields Aµα (x, y) and Φλ (x, y) satisfy modified (“twisted”) boundary
conditions which induce non-zero mass shifts for their 4D Kaluza–Klein modes.
5 If the Wilson line is not in the Cartan sub-algebra, the rank can also be reduced.
A solution to Eq. (7) is

2
Aµα (x, y) ∼ e−iyAy αI
I
eiyn/R Aαµ,n (x), Mn2 (α) = n − R AIy αI /R 2 ,
n∈Z
2
−iyAIy λI
Φλ (x, y) ∼ e eiyn/R φn,λ (x), Mn2 (λ) = n − R AIy λI /R 2 , (8)
n∈Z
where the fields depending on x only are 4D Kaluza–Klein modes for vanishing
background Ay and λ, α denote the weights/roots of corresponding fields. Using the Klein–
Gordon equation6 one finds the mass of the 4D Kaluza–Klein modes written in the rhs of
(8). The gauge symmetry G is reduced (to G ∗ ) since the mass of the zero-modes of some of
the gauge fields Aµα may become non-zero, while for AµI they remain massless. Kaluza–
Klein levels of Aµα are shifted by (non-zero) ρα ≡ −RAIy αI which equals the value of ρα
introduced in Eq. (4). Note that the “twist” of the boundary conditions in (7) is removed by
the formal limit of vanishing Wilson line vev’s (AIy → 0). Eq. (8) is used in Section 3.1
to compute one-loop corrections to the 4D gauge coupling(s) of G ∗ .
2.3. Effects on the 4D Kaluza–Klein mass spectrum: two extra dimensions
The results of the previous section can be extended to the case with two extra dimensions
y1,2 , each compactified on a one-cycle γi , i = 1, 2 (Fig. 1). Consider now the 6D fields
Fig. 1. A two-dimensional toroidal compactification constructed by identifying the opposite sites, with angle θ
between the two cycles γ1,2 . The points (y1 , y2 ) and (y1 + 2π R1 , y2 ) are identified; the same applies to (y1 , y2 )
and (y1 + 2π R2 cos θ, y2 + 2π R2 sin θ ). γ1,2 are defined along R1 and R2 directions respectively and Ay1,2
along the orthogonal dimensions y1,2 .
6 (D D µ + ∂ ∂ y )Ψ (x, y) = 0, with Ψ to denote A or Φ 5D fields.

µ y
Aµ̃ (x, ym ) and Φ(x, ym ), in the adjoint and fundamental representations of G respectively,
with the index µ̃ = µ, y1 , y2 , x ∈ M 4 , m = 1, 2 (µ = 0, 3). Assuming constant Ay1,2 , one
computes the Wilson lines Wi of Eq. (2) corresponding to each γi . Using definition (4) for
ρi,α one can show

ρ1,α = −R1 αI AIy1 , ρ2,α = −R2 αI AIy1 cos θ + AIy2 sin θ . (9)
If either of the quantities in the rhs of (9) is non-integer/non-zero for some set of α, the
symmetry G is broken. Note the dependence on the angle θ .
To compute the dependence of the 4D Kaluza–Klein masses on ρi,α we proceed as
follows. First, one has the periodicity conditions (Fig. 1)
Ψ (x; y1 + 2πR2 cos θ, y2 + 2πR2 sin θ ) = Ψ (x; y1, y2 ),

Ψ (x; y1 + 2πR1 , y2 ) = Ψ (x; y1, y2 ); Ψ : Aµ̃ or Φ, (10)
where Ψ stands for any of the fields Aµ̃ or Φ. These conditions are valid for an arbitrary
two-dimensional toroidal compactification (for an orthogonal torus θ = π/2 and the
periodicity conditions on (y1 , y2 ) “decouple” to leave one such condition in each compact
direction). As in the previous section, a y-dependent transformation V is introduced to
“gauge away” the fields Ay1 , Ay2 . One finds that the new (transformed) fields must satisfy
Ay1,2 (x; y1, y2 ) = 0 if V (y1 , y2 ) = e−iy1 Ay1 −iy2 Ay2 ,

Aµ (x, y1 , y2 ) = V Aµ (x, y1 , y2 )V −1

−i y1 AIy +y2 AIy α I
= AIµ (x, y1 , y2 )TI + Aαµ (x, y1 , y2 )Eα e 1 2 ,

Φ (x; y1, y2 ) = V Φ(x; y1, y2 ) (11)
with a summation over I and α understood. With TI , Eα linear independent one finds
that the components AIµ are invariant under V, AµI (x, y1, y2 ) = AIµ (x, y1 , y2 ) while
Aαµ (x, y1 , y2 ) may not necessarily be so. The initial periodicity conditions (10) are changed
into
−i2πR1 AIy αI
Aµα (x; y1 + 2πR1 , y2 ) = e 1 Aµα (x; y1, y2 ),
Aµα (x; y1 + 2πR2 cos θ, y2 + 2πR2 sin θ )

−i2πR2 AIy cos θ+AIy sin θ αI
=e 1 2 Aµα (x; y1, y2 ),
Φλ (x; y1 + 2πR2 cos θ, y2 + 2πR2 sin θ )

−i2πR2 AIy cos θ+AIy sin θ λI
=e 1 2 Φλ (x; y1, y2 ),
−i2πR1 AIy λI
Φλ (x; y1 + 2πR1 , y2 ) = e 1 Φλ (x; y1, y2 ), (12)
where Φλ denotes the component λ of the multiplet Φ and where we used that TI Φλ =
λI Φλ . A solution to these equations has the structure

−i y1 AIy +iy2 AIy σI

Ψσ (x; y1, y2 ) ∼ e 1 2 ψn1 ,n2 ,σ (x)un1 ,n2 (y1 , y2 ),
n1,2 ∈Z

n1 y2 n2 y2
with un1 ,n2 (y1 , y2 ) = exp i y1 − +i ,
R1 tan θ R2 sin θ
Ψα ≡ Aµα and Ψλ ≡ Φλ ; σ = α, λ, (13)
where the field Ψ is a notation for either Aµα Φλ
or while σI denotes the roots αI or
the weights λI , respectively. un1 ,n2 (y1 , y2 ) is an eigenfunction of the Laplacian ∂y21 + ∂y22
in the absence of the background gauge field because this was already “gauged away”
in (11) (to derive un1 ,n2 see for example [26]). If θ = π/2, un1 ,n2 (y1 , y2 ) is a product of
one-dimensional eigenfunctions for each yi .
Using the Klein–Gordon equation in 6D for massless fields and the above mode
expansion, one finds the mass of the Kaluza–Klein modes of 4D fields in the adjoint and
fundamental representations
2 2
n1 n2 n1
Mn1 ,n2 (σ ) =
2
− Ay1 σI +
I
− − Ay2 σI
I
R1 R2 sin θ R1 tan θ
2
1 1 eiθ

≡ 2 (n2 + ρ2,σ ) − (n1 + ρ1,σ , σ = α or σ = λ.
) (14)
sin θ R 2 R
1
It turns out that ρi,σ , i = 1, 2 (σ = α, λ) introduced in the last step in Eq. (14) have values
equal to those found in Eq. (9) using the definition of Eq. (4). Here σ = α (σ = λ) for
the adjoint (fundamental) representation (we will also use the notation vσ ≡ σI AIy1 ,
wσ ≡ σI AIy1 (sum over I )).
If either ρ1,α , ρ2,α are non-zero for a fixed α, there is no massless “zero mode”
(n1 , n2 ) = (0, 0) boson in the associated Kaluza–Klein tower. The initial symmetry G is
broken to a sub-group G ∗ generated by T I and those Eα for which Aµα has a massless 4D
zero-mode. This is controlled by the choice of AIyi vev’s in the root space of initial G. The
scale where G is broken to G ∗ depends on the potential developed by Ay fields and is thus
model dependent (see [27] for an example). Further, in specific cases ρ1,σ and ρ2,σ may be
simultaneously integers for a fixed σ = α or λ and according to (14) (if σ = α) there exists
a massless Kaluza–Klein boson of non-zero level (n1 , n2 ) = (0, 0). Consequently a gauge
symmetry enhancement (beyond G ∗ ) takes place, enabled by additional corresponding Eα .
These results agree with the previous findings in Section 2.1.
One may regard the overall effect of symmetry breaking as a shift of the Kaluza–Klein
levels ni to “effective” levels ni,eff = ni + ρi,α of non-integer values. The shift is due to
the geometry of compactification, Eq. (10) in the presence of constant background fields
whose effect was replaced by “twisted” boundary conditions, Eq. (12). Eq. (14) will be
used in Section 3.2 to compute the radiative corrections to the 4D gauge coupling(s) of the
group G ∗ .
3. Wilson line corrections to 4D gauge couplings
For realistic models G ∗ must include the Standard Model group (we will use i = 1, 2, 3
to label its component groups); this can happen if G is SU(5). In general the group G may
be larger and G ∗ contains additional group factors (not discussed). The coupling gi of a
group factor i of G ∗ is

4π 4π
= + ΩiT ,
gi2 one-loop gi2 tree-level
∞
1 dt −πt Mψ2 /µ2
ΩiT ≡ β̃i (ψ) e . (15)
4π t reg
ψ 0
ΩiT sums all one-loop corrections to the coupling gi , induced by the states ψ of mass
Mψ . ΩiT is given by the Coleman–Weinberg formal equation in the rhs of Eq. (15) (for a
discussion see [7]). In (15) µ2 is a finite, non-zero mass parameter introduced to enforce a
dimensionless equation; the subscript “reg” expresses that a regularisation of the integral
is required.
ΩiT receives corrections from the 4D massless and massive fields charged under the
group factor i of G ∗ . For our purpose we need not specify the 4D massless or Kaluza–
Klein zero level spectrum, which depends on further details of the model considered. We
restrict ourselves to computing the structure of the corrections to ΩiT from the 4D massive
sector and we sum over all Kaluza–Klein towers of states charged under the group i of G ∗ .
These are:
(1) 4D Kaluza–Klein towers of states whose levels are not shifted (i.e., ρk,σ = 0, k = 1, 2)
and have massless zero-modes. An example is (if σ = α) that of Kaluza–Klein states
associated with the “unbroken” generators Eα (of the group i of G ∗ ), and
(2) 4D Kaluza–Klein towers of states of levels shifted by the amount ρk,σ , k = 1, 2.
An example is (if σ = α) that of Kaluza–Klein states associated with the “broken”
generators Eα (like X, Y gauge bosons and their Kaluza–Klein tower for the SU(5)
breaking to the SM group).
For the beta functions βi one has (after suppressing the subscript i) that β̃(σ ) =
kr (σI σ I )/rk G ∗ for σ belonging to representation r; kr = {−11/3, 2/3, 1/3} for ad-
representations, Weyl fermion and scalar, respectively. The Dynkin index T (r) =
joint
( σ σI σ I )r /(rk G ∗ ) where the sum is over all weights/roots σ belonging to representa-
tion r, each occurring
a number of times equal to its multiplicity [25]. With the defini-
tion bi (r) ≡ σ β̃i (σ ) for the weights σ belonging to r one has bi = −(11/3)Ti (A) +
(2/3)Ti (R) + (1/3)Ti (S), to account for the adjoint, Weyl fermion in representation R and
scalar in representation S. Massive N = 1 Kaluza–Klein states can be organised as N = 2
hypermultiplets with bi = 2Ti (R) and N = 2 vector supermultiplets with bi = −2Ti (A).
3.1. One extra dimension and Wilson line corrections
For the case of one additional compact dimension ΩiT can be written as

ΩiT = Ωi (σ ),
r σ =λ,α
∞
1 dt −πt Mm2 (σ )/µ2 −πχt
Ωi (σ ) ≡ β̃i (σ ) e e . (16)
4π t
m∈Z ξ
Ωi (σ ) is the contribution of a tower of Kaluza–Klein modes associated with a state σ
“shifted” by ρ(σ ) real, with σ = λ, α the weight/root belonging to the representation r.
The sum over m runs over all integers, representing Kaluza–Klein levels of mass Mm (σ )
given by Eq. (8)

Mm (σ ) = (m + ρσ )2 /R 2 , ρσ = −R AIy σI ≡ −Rvσ ; σ = α, λ, (17)
with σ = α, (λ) for the adjoint (fundamental) representation. In Eq. (16) a regularisation of
the integral was performed. Since ΩiT is UV-divergent (t → 0) an UV regulator ξ → 0 was
introduced as the lower limit of the integral. For the special case when there are massless
states in the Kaluza–Klein tower, the integral is also IR-divergent (t → ∞) and an IR
regulator χ → 0 is introduced. If no massless states exist in the Kaluza–Klein tower, one
formally sets χ = 0. For other regularisations and their relationship with that employed
here see Appendix A.4 of this work and Appendices B and C of [11]. From Eq. (16) the
relation between the UV/IR regulators and their associated mass scales can be inferred to
be of type Λ2 ∝ µ2 /ξ for the UV scale and Q2 ∝ χµ2 for the IR scale.
As usually done when computing the one-loop corrections in 4D compactified models
[7,9] we isolate in Eq. (16) the contribution of the “zero” modes (whose existence is model
dependent, they may be projected out by the initial orbifolding) from that of the non-
zero level modes, which is general and is computed in string case. To keep track of this
separation we relabel by βi (σ ) (β̄i (σ )) the one loop beta function of “zero” (non-zero)
modes, respectively. In the following the dependence of ρ, β̄, β and Mm on σ = α, λ is not
written explicitly. From (16) we have
βi 0 β̄i
Ωi = J + J (18)
4π 4π
with the notation
∞
dt −πt M 2 /µ2 −πχt
1
J ≡
0
e 0 e = Γ 0, πνξ ρ 2 + χ/ν , ν≡ ,
t (Rµ)2
ξ
dt
∞
e−πt Mm /µ e−πχt
2 2
J≡
t
m∈Z ξ
2e−πξ χ

= √ − Γ 0, πνξ ρ 2 + χ/ν + 2π(χ/ν)1/2 Erf πξ χ

νξ

2
− ln2 sin π ρ + i(χ/ν)1/2 , (19)
with the functions Γ [0, x] and Erf[x] defined in Appendix A, Eq. (A.8). A “prime” on the
sum over “m” indicates that the sum is over all integers m with m = 0. To evaluate J the
results of Appendix A.1, Eq. (A.1) were used. Eq. (19) is valid if
1 µ2
νξ 1 or 2
. (20)
R ξ
Eqs. (16)–(20) give the most general result for the radiative correction to gauge couplings.
In the limit of “removing” the regulators dependence (ξ → 0) the Γ functions contributions
in J 0 , J can be approximated by logarithms while the Erf function contribution vanishes.
The presence of the regulator χ ensures that the result (19) applies whether or not there are
massless states in the Kaluza–Klein tower.7
The mass parameter µ combines with the (dimensionless) regulators ξ and χ to
introduce the following associated mass scales

2 µ2
Q ≡ πe χµ χ→0 ,
2 γ
Λ ≡
2
. (21)
ξ ξ →0
Q is therefore the low(est) energy scale, and Λ is the high(est) energy (UV) scale of the
theory. With this notation the one-loop correction is
βi Λ2 β̄i
ρ = 0: Ωi = ln 2 − ln 4πe−γ (ΛR)2 e−2ΛR , (22)

4π Q 4π
βi Λ 2 β̄i
ρ ∈ Z∗ : Ωi = ln 2 − ln 4πe−γ (ΛR)2 e−2ΛR

4π Q 4π

β̄i − βi πeγ (ρ/R)2
+ ln 1 + , (23)
4π Q2
βi Λ2 β̄i
ρ∈/ Z: Ωi = ln 2 − ln 4πe−γ (ΛR)2 e−2ΛR

4π Q 4π
γ
β̄i sin(πρ) 2 βi πe (ρ/R)2
− ln − ln , (24)
4π πρ 4π Q2
where8 according to Eq. (17), ρ/R is a vacuum expectation value in a direction in the
weight/root space. These equations are valid if the following conditions are respected:
1
ρ = 0: Q2 Λ2 , (25)
R2
1 ρ2
ρ ∈ Z∗ : Q2 2 Λ2 and Λ2 , (26)
R R2
1 ρ2
ρ∈/ Z: 2
Λ2 and Λ2 . (27)
R R2
Eqs. (22)–(27) provide the threshold correction to the gauge couplings at one-loop level
due to a tower of Kaluza–Klein states in the presence/absence of a constant background
gauge field.
The first logarithmic term in Eqs. (22)–(24) stands for the contribution of the “zero”
modes from the high scale Λ to some low energy scale Q. The second term in these
equations proportional to β̄i contains the linear contribution (in Λ) to the gauge couplings
7 Special care is needed when removing the regulators χ → 0, ξ → 0 as these limits do not always commute
[12].
8 We denoted by γ the Euler constant, γ = 0.577216 . . . .
and is due to the Kaluza–Klein states of the extra dimension, in the absence of Wilson lines
[11]. Eq. (22) gives the correction for the case of vanishing Wilson line vev’s (ρ = 0).
For the case ρ ∈ Z∗ the last term in Eq. (23) gives an additional contribution due to the
Wilson line vev’s. Note that this correction is proportional to beta functions differences of
zero and non-zero levels which may actually vanish in specific cases. The correction also
depends on the low energy scale Q brought in by the need for an IR regulator when ρ ∈ Z∗ .
Consequently, the low energy physics represented by Q is related to Wilson line effects,
even though the latter may take place at a very high energy scale or may have a large vev
(ρ) compared to Q. Formally, if one sets ρ = 0 the case of Eq. (22) is recovered. Finally,
the constraint (26), ρ/R Λ may be “relaxed” into ρ/R Λ or ρ/R ≈ Λ if one uses
Eqs. (18)–(20) instead of (23) and (26).
The result for the case with ρ ∈/ Z does not depend on Q (the dependence displayed in
(24) cancels out). For Ωi (ρ ∈/ Z) the limit of reaching an integer or vanishing ρ is not finite,
see the last two terms in (24). Indeed, the one loop correction has an infrared divergence
at integer and vanishing values of ρ, which are “moduli” points where new massless states
appear. If these correspond to vector superfields (σ = α) the symmetry G ∗ is enlarged.
Therefore, these “moduli” points cannot be smoothly reached perturbatively by taking the
limit of integer ρ.
We are now able to write the most general threshold correction to the gauge couplings
by combining the contributions Ωi for all possible values for ρσ . Making the dependence
on σ manifest, one has from (16) and (22) to (24)

ΩiT = Ωi (ρσ = 0) + Ωi ρσ ∈ Z∗ + Ωi (ρσ ∈ / Z) (28)

r σ =α,λ
with the remark that if the condition for the argument of any term Ωi is not respected,
that term in the sum should be set to zero. Conditions (25)–(27) should be considered
accordingly.
3.2. Two extra dimensions and Wilson line corrections
The effects on the gauge couplings of the 4D Kaluza–Klein states in the presence of a
background (gauge) field are similar to the one-dimensional case. Eq. (15) becomes

ΩiT ≡ Ωi (σ ),
r σ =α,λ
∞
1 dt −πt Mm2 ,m (σ )/µ2 −πχt
Ωi (σ ) ≡ β̃i (σ ) e 1 2 e (29)
4π t
m1,2 ∈Z ξ
with a summation over all weights/roots belonging to representation r. As in the one-

dimensional case an UV regulator ξ (ξ → 0) is introduced since the integral is divergent
at t → 0 and the associated UV scale is then proportional to µ2 /ξ . For special values of
Kaluza–Klein levels (m1 , m2 ) and Wilson lines vev’s (see Eq. (14)), Mm1 ,m2 may vanish,
and the integral becomes infrared divergent (t → ∞). An IR regulator χ is introduced
χ → 0 for these particular cases, with associated infrared scale Q2 ∝ µ2 χ . The regulator
plays the role of a 6D mass term in the Klein–Gordon equation which shifts the value of
Mm1 ,m2 in Eq. (14). The cases with Mm1 ,m2 = 0 for Kaluza–Klein bosonic states of non-
zero level are important since they signal an enlargement of the gauge symmetry G ∗ . The
mass of the Kaluza–Klein states given in (14) can be written as
µ2 2
2
Mm 1 ,m2
(σ ) = m2 + ρ2,σ − U (m1 + ρ1,σ ) , σ = α, λ, (30)
T 2 U2
with the notation
R2 iθ
U ≡ U1 + iU2 = e (U2 > 0); T (µ) ≡ iT2 (µ) = iµ2 R1 R2 sin θ. (31)
R1
The structure of the mass formula (30) is very general and also applies to string
compactifications. At the string level U and T have correspondents in the so-called moduli
fields related to the complex structure and (imaginary part of) the Kähler structure of the
two-torus, respectively.9 The notation in Eq. (31) is introduced to facilitate a comparison
with the string results where this notation is standard. ρ1,2 are related to the Wilson line
vev’s, see Eqs. (9) and (14). The mass (30) equals that encountered in (heterotic) string
compactification for zero winding modes. This is expected, since an effective field theory
approach corresponds to the case of an infinite string scale when winding modes are
infinitely heavy and their effects are suppressed.
To compute ΩiT of (29) we isolate the contribution of (0, 0) mode from that of the rest
of the Kaluza–Klein modes (this is allowed after the regularisation of the integral in (29)).
This separation is needed for two reasons. First, the exact final spectrum of zero level or
massless modes depends on further details of the models considered; in particular some
(0, 0) modes may not “survive” the initial orbifold projections. Second, string calculations
of Ωi [9,20] that we want to compare with only compute the effects of the non-zero levels.
To keep track of this mode separation we relabel β̃i by β̄i (βi ) for the Kaluza–Klein levels
(m1 , m2 ) = (0, 0), ((m1 , m2 ) = (0, 0)), respectively.
The analysis below considers separately Case 1 of vanishing Wilson lines vev’s ρk,σ =
0, k = 1, 2 (σ fixed) as a reference for when we evaluate Case 2 of non-vanishing Wilson
vev’s. Case 2(B) when the Wilson line breaks the gauge symmetry is the most interesting
for phenomenology.
3.2.1. Case 1. Wilson line background absent

In this case there are vanishing vev’s of the Wilson lines ρk,σ = 0, k = 1, 2 for a
particular set of σ = α, λ, with corresponding Eα as unbroken generators. If this is true
for all α there is no breaking of the initial symmetry G. For simplicity we assume this is
indeed the case, otherwise the discussion refers to the unbroken part G ∗ of G only and the
effect of Kaluza–Klein towers “unshifted” (by ρk,α ). Therefore, the problem is that of one-
loop corrections to the gauge couplings due to the N = 2 sector (of a 4D N = 1 orbifold)
compactified on a two-torus (no Wilson lines vev’s) and well known in the heterotic string
9 In the heterotic string T is expressed in string units (α ∝ 1/M 2 ). To be exact, it is T ∗ = T /ξ that we later
S 2 2
refer to and which is expressed in UV cut-off scale units that plays effectively the role that T2 does in string
theory.
[9]. At the effective field theory level this calculation was performed in [11,12], reviewed
here for later reference. In Eq. (29) the mass of the Kaluza–Klein states does not depend
on σ and the sum of β̃i (σ ) over α, λ, r (for the group G!) may be performed before the
integral itself. We denote this overall sum by bi , b̄i for (0, 0) and for (m1 , m2 ) = (0, 0)
modes, respectively. The total correction ΩiT has then the structure
bi (1) b̄i (1)

ΩiT = J + J , (32)
4π 0 4π
with
∞ ∞
dt −πt M0,0
2 /µ2 dt −πχt
J0(1) ≡ e e−πχt = e = Γ [0, πξ χ],
t t
ξ ξ
∞ ∞
dt −πt Mm2 ,m /µ2 −πχt dt − T πUt |m2 −U m1 |2 −πχt
J (1)
≡ e 1 2 e = e 2 2
m1,2
t m
t
ξ 1,2 ξ

T2 T 4 ξ
= − ln 4πe−γ e− ξ U2 η(U ) + πχT2 ln 4πe−γ U2
2
, (33)
ξ T2
and with

χ min U2 /T2 , 1/(T2 U2 ) ; max U2 /T2 , 1/(T2 U2 ) 1/ξ. (34)
A “prime” on a double sum stands for a sum over all integers (m1 , m2 ) except the mode
(0, 0) considered by J0(1) . An infrared regulator was introduced in ΩiT before its splitting
(1)
into J0 , J (1) , because M0,0 = 0 according to (30). The correction J (1) was evaluated in
detail in Ref. [12]. Condition (34) ensures that higher order corrections in the regulators
χ and ξ vanish in the limit of removing them, χ, ξ → 0. From (33) one finds the overall
correction to the gauge couplings due to the towers of Kaluza–Klein states associated with
two dimensions (and vanishing Wilson line vev’s)

bi Λ2 b̄i ∗ 4
ΩiT = ln 2 − ln 4πe−γ e−T2 T2∗ U2 η(U )
4π Q 4π

b̄i 2 −γ −γ U2
+ Q R1 R2 sin θ e ln 4πe . (35)
4π T2∗
We introduced the notation

T2 µ2
T2∗ ≡ = Λ2 R1 R2 sin θ and Λ2 ≡ , Q2 ≡ πeγ µ2 χ χ→0 ,
ξ ξ →0 ξ ξ →0
(36)
which will be used in the remaining sections. Eq. (34) becomes

1 1 1 1
Q min , ; max , Λ, (37)
R1 R2 sin θ R1 R2 sin θ
which implies Λ2 R1 R2 sin θ 1, i.e., a large area of compactification (in UV scale units).
Note that (R2 sin θ ) plays the role of an “effective” radius of the extra dimension. Eq. (36)
clarifies the link between the UV/IR regulators and the mass scales Λ and Q which emerge
as corresponding UV and IR cut-off scales respectively.
The first term in ΩiT is the usual one-loop logarithmic correction which accounts for
the effects of (0, 0) modes, from the high scale Λ to the low energy scale Q. The second
term accounts for the effects of the massive Kaluza–Klein states associated with the two
extra dimensions and shows the usual power-like dependence [6] on the UV scale since
ΩiT ∼ T2∗ ∼ Λ2 . Its structure can be compared with that of the corresponding (heterotic)
string result in the limit of an infinite string scale or α → 0, when the additional effects of
the winding states of the string are minimised.10 The string result in this limit agrees with
that of field theory. For a detailed discussion on the second term in ΩiT and its link with
the heterotic string see [11].
There remains the last term in ΩiT , i.e., Q2 R1 R2 sin θ ln U2 /T2∗ ∼ χ ln ξ whose origin
was analysed in detail in [12]. Here we review briefly its significance. This correction arises
from the UV divergent contribution J (1) of the massive momentum states in the presence
of the infrared regulator χ . The latter was required by the (IR-divergent) contribution of
the massless mode M0,0 . Therefore, the term χ ln ξ establishes an (infrared) link between
the massive and massless sectors. The term χ ln ξ or the last term in ΩiT that it induces
must be kept in the final correction to the gauge couplings because the limits of removing
the regulators χ → 0 and ξ → 0 do not commute.
The presence of the last term in ΩiT shows that even though Kaluza–Klein states may
have a very large mass, of the order of the compactification scales, their overall contribution
is still proportional to a much lower scale Q, where they may actually be decoupled. The
contribution of this term may be large since U2 T2∗ (equivalently 1/R1 Λ). However,
the coefficient in front may be small Q2 R1 R2 sin θ 1 for our result to be accurate, see
conditions in Eq. (37). We conclude that the overall effect of the infinite tower of Kaluza–
Klein states on the gauge couplings cannot be split into massless and massive modes only,
and a combined effect of these mass sectors through infrared effects is present.
The last term in ΩiT has no equivalent at the string level [12]. To understand why this
is so, note that at the string level, the UV regulator ξ ∝ 1/Λ2 has a “correspondent” in
α ∝ 1/Ms2 with the limit ξ → 0 to correspond to an infinite string scale or α → 0. String
calculations of ΩiT require an infrared regularisation,11 so χ also has a string infrared
regulator correspondent that we denote @, with @ → 0. Therefore, a string equivalent of
the last term in ΩiT would be @ ln α . Such term does appear in string calculations during
the string infrared regularisation (see [12] and Appendix A of [28]). However, in string
calculations α = 0 and consequently @ ln α → 0 in the final, infrared regularised string
result when @ → 0. The subsequent “field theory” limit α → 0 of this string result will
then miss the last term in ΩiT . The origin of this discrepancy is that in specific cases (such
10 Such additional effects are (related to) world-sheet instanton effects, vanish if α → 0 and have no EFT
description.
11 For various infrared regularisations of the string see Appendices of [9,28,29].
as two extra dimensions) the infrared regularisation of the (world-sheet integral of the)
string and its “field theory” limit α → 0 do not commute.
The conclusion is that the UV behaviour found using effective field theory methods for
a 4D N = 1 orbifold compactification with N = 2 sub-sector (two-torus) is not necessarily
that of its (infrared regularised) string embedding in the limit α → 0. This issue is closely
related to the infinite number of states in the Kaluza–Klein tower that one sums over, which
bring about a “non-decoupling” of the UV effects from the low energy (Q) sector. This
concludes our review of the radiative corrections to the couplings for vanishing Wilson
lines vev’s. For more details on this issue see [12].
3.2.2. Case 2. Wilson lines background present

In this case there is a non-zero Wilson lines background ρi,σ , i = 1, 2, σ = α, λ, that can
in some cases (σ = α) affect the symmetry G ∗ . We distinguish two general possibilities for
Ωi (σ ) of Eq. (29)
(A) ρ1,σ ∈ Z, ρ2,σ ∈ Z (σ fixed), and

(B) ρ1,σ ∈
/ Z and ρ2,σ real (or ρ2,σ ∈
/ Z and ρ1,σ real).
Case 2(A). Computing Ωi for ρ1,σ ∈ Z, ρ2,σ ∈ Z

In this case there exists a non-zero Wilson lines vev, with ρi,σ , i = 1, 2, simultaneously
integers, for a fixed σ = α or λ. Therefore, for a specific value of the Kaluza–Klein levels,
Mm1 ,m2 (σ ) vanishes and if this corresponds to a bosonic state α, an enlargement of the
gauge symmetry G ∗ takes place. When Mm1 ,m2 (σ ) vanishes an infrared regulator χ is
required for the corresponding tower of Kaluza–Klein states which contributes to Ωi (σ ).
This is given by (below the σ dependence will not be written explicitly)
βi (2) β̄i (2)
Ωi = J + J , (38)
4π 0 4π
where we isolated the contribution J0(2) of the (0, 0) mode with beta function βi , from that
of non-zero modes J (2) with beta function β̄i and
∞ ∞
dt −πt M0,0
2 dt − T πUt |ρ2 −Uρ1 |2 −πχt
e−πχt =
(2)
J0 ≡ e e 2 2
t t
ξ ξ

= Γ 0, πξ |ρ2 − Uρ1 |2 /(T2 U2 ) + χ ,

dt −πt M 2 dt − π t |m +ρ −U (m +ρ )|2 −πχt
∞ ∞
2
m1 ,m2 /µ −πχt
J ≡
(2)
e e = e T2 U2 2 2 1 1
m
t m
t
1,2 ξ 1,2 ξ

=J (1)
+ Γ [0, πξ χ] − Γ 0, πξ |ρ2 − Uρ1 |2 /(T2 U2 ) + χ (39)
with

χ min U2 /T2 , 1/(T2 U2 ) max U2 /T2 , 1/(T2 U2 ) 1/ξ. (40)
For a simple evaluation of J (2) one adds and subtracts under its integral the exponential
evaluated for (m1 , m2 ) = (0, 0), then shifts the summation variables m1,2 by the integers
ρ1,2 respectively, and finally isolates from the sum the “new” (0, 0) mode. One then
recovers an integral equal to J (1) of Eq. (33) plus two additional terms, giving the result
(39).
Using the notation introduced in Eq. (36) we obtain from (38) and (39) the result
βi Λ2 β̄i ∗ 4
Ωi = ln 2 − ln 4πe−γ e−T2 T2∗ U2 η(U )

4π Q 4π

β̄i 2 −γ −γ U2
+ Q R1 R2 sin θ e ln 4πe
4π T2∗

β̄i − βi πeγ |ρ2 − Uρ1 |2
+ ln 1 + (41)
4π Q2 (R2 sin θ )2
with

1 1 1 1 |ρ2 − Uρ1 |
Q min , ; max , , Λ. (42)
R1 R2 sin θ R1 R2 sin θ R2 sin θ
Except its last term, the result for Ωi is similar to that for vanishing Wilson lines discussed
in Case 1. The first, second and third terms in (41) account for the effects of the (0, 0)
modes, massive Kaluza–Klein modes alone and the “mixed” contribution proportional to
Q2 , respectively. However, there exists an additional term proportional to (β̄i − βi ), not
present in Case 1, and this is due to the effects of the Wilson lines alone. This term vanishes
if one formally sets ρ1 = ρ2 = 0 when Case 1 is recovered. Note the similarity of the last
term in Ωi to that of the last term in (23) of the one extra dimension case. The Wilson line
contribution can also be written as

β̄i − βi πeγ (vσ2 + wσ2 )
ln 1 + ,
4π Q2

vσ ≡ σI AIy1 , wσ ≡ σI AIy2 , σ = α, λ. (43)
where vσ and wσ denote a vev in a direction in the root (α) or weight (λ) space. Note that
βi and β̄i depend on σ . This is relevant when the sum over σ of Eq. (29) is performed to
compute total ΩiT .
The correction due to Wilson lines vev’s shows how the couplings change when non-
zero level Kaluza–Klein modes become massless. The correction due to the Wilson lines in
Eq. (41) depends on their vev’s and on the low scale Q, but the ultraviolet behaviour of the
models is not changed. Indeed, the T2∗ ∼ Λ2 dependence of Ωi is identical to that of the
case with vanishing Wilson lines vev’s addressed in the previous section. Regarding the Q
dependence of Ωi we remark the following. The gauge couplings are changed at the low
scale Q as a result of (possibly) large Wilson lines vev’s ρi and of their near-cancellations
against large Kaluza–Klein levels (m1 , m2 ) corresponding to a large momentum in the
compact directions m1,2 /R1,2 and giving Mm1 ,m2 ≈ 0.
Case 2(B). Computing Ωi for ρ1,σ ∈

/ Z, ρ2,σ real (or ρ2,σ ∈
/ Z, ρ1,σ real)
In the following we assume ρ1,σ ∈ / Z and ρ2,σ real. (Appendix B extends this analysis
to ρ2,σ ∈/ Z, ρ1,σ real.) Phenomenologically this is the most interesting case we discuss.
There are non-zero Wilson line vev’s and Mm1 ,m2 (σ ) = 0 for a fixed set of σ = λ, α. The
corresponding generators Eα are broken, the symmetry is (reduced to) G ∗ and Kaluza–
Klein towers are “shifted” (by ρi,σ ). Since there is no massless state for any integers
m1 , m2 , in Eq. (29) there is no need for an IR regulator χ for the corresponding correction
Ωi (σ ) of the Kaluza–Klein tower. In the following the σ dependence is not shown
explicitly. After Poisson resummation (A.2) the integrand in (29) becomes
−πt Mm
2 /µ2
e 1 ,m2
m1 ,m2
− T πUt |m2 +ρ2 −Uρ1 |2
− T πUt |m2 +ρ2 −U (m1 +ρ1 )|2
= e 2 2 + e 2 2
m2 m1 m2
1/2
− T πUt |m2 +ρ2 −Uρ1 |2 T 2 U2 −πt U2 (m1 +ρ1 )2
= e 2 2 + e T 2
m2
t m1
1/2
T 2 U2 −π m̃22
T2 U2 U2
t −πt T2 (m1 +ρ1 ) +2πi m̃2 (ρ2 −U1 (ρ1 +m1 ))
2
+ e , (44)
t m1 m̃2
with a prime on the double sum in the lhs to indicate the mode (m1 , m2 ) = (0, 0) is not
included. Since ρ1 is non-integer the above three contributions can be integrated separately
over (ξ, ∞).
In Ωi of (29) we isolate the contribution of (0, 0) Kaluza–Klein levels (J0(3) ) from that
of levels with (m1 , m2 ) = (0, 0) and denoted J (3) , with an obvious notation for the beta
functions coefficients:
βi (3) β̄i (3)

Ωi = J + J , (45)
4π 0 4π
and
∞
dt −πt M 2 /µ2
1
J0(3) ≡ e 0,0 = Γ 0, πξ τ |ρ2 − Uρ1 |2 , τ≡ ,
t T 2 U2
ξ
∞
dt −πt Mm2 ,m /µ2
J (3)
≡ e 1 2 ≡ L1 + L2 + L3 (m1 , m2 ) = (0, 0). (46)
t m ,m
ξ 1 2
L1 , L2 and L3 denote the following integrals of the three contributions given in the rhs of
Eq. (44)
∞
dt − T πUt |m2 +ρ2 −Uρ1 |2
L1 ≡ e 2 2
t m
ξ 2
2

= √ e−πτ ξρ1 U2 + 2π|ρ1 |U2 Erf |ρ1 |U2 πτ ξ − Γ 0, πτ ξ |ρ2 − Uρ1 |2

2 2
τξ
2
− ln2 sin π(ρ2 − Uρ1 ) ,
∞
dt −πt
U2
T2 (m1 +ρ1 )
2
L2 ≡ (T2 U2 ) 1/2
e
t 3/2 m1
ξ
2
1
= − √ e−πτ ξρ1 U2 − 2π|ρ1 |U2 Erf |ρ1 |U2 πτ ξ +
2 2
τξ τ ξ U2

1
+ 2πU2 |ρ1 | + + ∆2ρ1 − ∆ρ1 ,
6
∞
dt −π m̃22 T2tU2 −πt UT 2 (m1 +ρ1 )2 +2πi m̃2 (ρ2 −U1 (ρ1 +m1 ))
L3 ≡ (T2 U2 )1/2 e 2
t 3/2 m
ξ 1 m̃2

2 1

= ln 2 sin π(ρ2 − Uρ1 ) − 2πU2 |ρ1 | − ∆ρ1 +
6

ϑ1 (∆ρ2 − U ∆ρ1 |U ) 2
− ln ,
(47)
η(U )
where ∆y denotes the positive definite fractional part of y defined as y = [y] + ∆y ,
0 < ∆y < 1, with [y] an integer number. ϑ1 (z|τ ) and η(U ) are special functions defined in
Appendix A, Eq. (A.32).
The above integrals are evaluated in detail in Appendix A, see Eqs. (A.1), (A.13)
and (A.28), respectively. While evaluating them one introduces errors @i which are higher
order corrections in the regulator ξ , see Eqs. (A.7), (A.27), (A.37). Imposing that these
corrections vanish gives the constraint
(T2 U2 )
max 1/U22 , U22 (48)
ξ
which must be respected for Eqs. (47) to hold true. Using the results of Eqs. (45), (46),
(47) and the notation introduced in Eq. (36), we find the result for the radiative correction
to the gauge couplings
βi Λ2 (R2 sin θ )2 β̄i ∗ 4
Ωi = ln γ − ln 4πe−γ e−T2 T2∗ U2 η(U )

4π πe |ρ2 − Uρ1 | 2 4π

β̄i ϑ1 (∆ρ2 − U ∆ρ1 |U ) 2
− ln
4π η(U )3
β̄i 2
+ ln 2π(ρ2 − Uρ1 ) + 2π|U2 |∆2ρ1 , (49)

4π
valid if

1 1 |ρ1 | |ρ2 − Uρ1 |
max , , , Λ. (50)
R1 R2 sin θ R1 R2 sin θ
The first two conditions in (50) follow from (48). The last two conditions originate from
imposing that the argument of the Γ functions present in Eqs. (46), (47) be small enough, to
approximate these functions with the familiar logarithm12 of one-loop radiative corrections
(these conditions can thus be relaxed). They state that the vev’s of the Wilson lines in any
directions in the root space be smaller than Λ: vσ2 , wσ2 Λ2 , (with definition (43)).
The first term in (49) is a correction due to (0, 0) modes. The second term is due to
non-zero modes and shows that the leading UV behaviour (i.e., T2∗ dependence) of the
correction Ωi is not changed from the case of vanishing vev of the Wilson line. Indeed,
this term has a power-like dependence Ωi ∼ T2∗ = Λ2 R1 R2 sin θ and a logarithmic one
ln T2∗ ∝ ln Λ similar to those of Case 1. Since the Wilson lines introduce a spontaneous
gauge symmetry breaking, this confirms the expectation that the UV behaviour of the
couplings not be changed by their acquiring of non-zero vev’s. The last two terms in (49)
are due to the Wilson lines effects alone. These terms depend on U and ρ1,2 or rather its
fractional part ∆ρ1,2 , but no additional UV scale dependence is present.
There is one notable difference between Ωi of (49) and Ωi of Case 1 or Case 2(A). This
is the absence in (49) of the term
2
Q R1 R2 sin θ ln 4πe−γ U2 /T2∗ . (51)

This term was present in Cases 1, 2(A) due to massless Kaluza–Klein states and it was
induced by the contribution of massive Kaluza–Klein states in the presence of the infrared
regulator required by the massless modes. Such term is not present in (49) since no
massless states exist in this case.
The result of Eq. (49) can be rewritten in a more compact form
βi Λ2 (R2 sin θ )2
Ωi = ln γ
4π πe |ρ2 − Uρ1 |2

β̄i ∗ T2∗ U2 e−γ ϑ1 (∆ρ2 − U ∆ρ1 |U ) 2

+ T − ln − ln + 2πU2 ∆ρ1 .
2
4π 2 π|ρ2 − Uρ1 |2 η(U )
(52)
This is the main result of the paper and gives the general correction to the gauge couplings
in the presence of background gauge fields of vev’s which break the gauge symmetry to G ∗ .
We note that the results (49), (52) were evaluated for ρ1 non-integer, while ρ2 was kept
arbitrary. Further, taking in (49) the formal limit ρ1 integer (∆ρ1 = 0) with ρ2 non-integer
gives a finite result. Detailed calculations in Appendix B show that this limit does provide
the correct result. In conclusion one can extend the validity of Eqs. (49), (52) to all cases
with either ρ1 or ρ2 non-integer and real values of the other, respectively. This is partly
expected given the somewhat symmetric role that ρ1 and ρ2 play.
12 We use −Γ [0, z] = γ + ln z + · · ·, 0 < z 1, and γ = 0.577216 . . . is the Euler constant.

In the formal limit when both ρ1 and ρ2 are integers (for a given σ ) the radiative
correction Ωi of (49), (52) becomes logarithmically divergent in the IR region and a
regulator is required. In such situations to evaluate Ωi one should apply the approach
of Case 2(A). This shows that Ωi as a function of ρ1,2 is not continuous at points in the
“moduli” space of (ρ1 , ρ2 ) with (ρ1 , ρ2 ) = (m, n) with m, n ∈ Z when massless Kaluza–
Klein states may appear. Recalling that ρi are σ dependent, such states would signal for
σ = α an enhancement of the gauge symmetry beyond G ∗ . This discontinuity is also due
to the term in Eq. (51) present for ρ1,2 ∈ Z. To conclude, values of the Wilson lines vev’s
corresponding to ρi integers cannot be smoothly reached perturbatively from those with
ρi non-integers.
We can now address the link of the result (52) for the one-loop corrections in the
presence of Wilson lines with (heterotic) string theory. When doing so we only refer to
the term in (52) proportional to β̄i and due to non-zero modes, and which is evaluated at
the string level. The result (52) has a string counterpart in the context of heterotic (0, 2)
string compactifications [20] with N = 2 sector and Wilson line background which breaks
the initial gauge symmetry. To compare the two results one must consider the string result
in the limit of large compactification area in string units, or equivalently the limit α → 0.
The string correction to the gauge coupling due to non-zero levels of Kaluza–Klein and
winding modes is in this limit [20]

β̄i,H 2π 4 1 ϑ1 (B|U ) 2
Ωi,H =
T2 − ln(T2 U2 ) − ln η(U ) − ln , (53)
4π 5 5 η(U ) H
where the index H stresses that the notation used in this equation is that of the heterotic
string, to be distinguished from that used in our field theory approach. T2 of (53) is
the imaginary part of the Kähler structure measured in α units and has a field theory
counterpart in T2∗ measured in UV cut-off length unit and defined in (31), (36). Also
B = µ − U µ is a mass shift induced by µ, µ which are the two Wilson lines vev’s at
the string level in a particular direction in the root space of the group considered there.
µ, µ have field theory counterparts in ρ1,2 present in Eqs. (4), (9), (30).
Comparing (52), (53) one notices that the power-like and logarithmic behaviour in T2∗
and T2 respectively is similar, up to the numerical coefficient which in the effective field
theory approach is regularisation dependent (T2∗ depends on UV regulator ξ ) and cannot be
fixed on field theory grounds only [11]. The presence of the odd elliptic theta function ϑ1
is similar in (52) and (53). However, its argument depends at the field theory level on ∆ρ1,2
which is the Wilson lines vev’s ρ1,2 modulo an integer, as actually expected from Eq. (4).
Further, the term ln |η(U )|4 appearing in (53) and not present in (52) is not a discrepancy
of the two approaches. This term is due at the string level to including the contribution
µ = µ = 0. At the field theory level this requires one consider the case ρi,α = 0, not
included in (52) and which does bring in such a term as shown in Case 1 Eqs. (35). The only
difference between the two approaches is the existence in (52) of the term ln |ρ2 − Uρ1 |,
absent in (53). With this exception, the effective field theory and string calculation in the
limit α → 0 lead to remarkably close results, despite their entirely different approaches to
computing Ωi .
We conclude with a reminder that the most general correction to the gauge couplings in
the presence of Wilson lines is a sum of the results of type found in Cases 1 and 2(B)
(ignoring the very special case of integer Wilson line vev’s ρk,α ). According to the
discussion in Section 3 and Eq. (29) the total correction is a sum of Eq. (49) and Eq. (35).
Eq. (49) is associated with “broken” generators of initial G while Eq. (35) is associated
with “unbroken” generators Eα of G ∗ (with appropriate beta functions). Note that βi , ρi ,
∆ρi have all a dependence on σ = α, λ not shown explicitly in this section.
4. Conclusions
In general models with additional compact dimensions have a larger amount of gauge
symmetry than the SM does and a mechanism to break it is required. A natural procedure to
achieve this is that of Wilson line breaking in which components of the higher-dimensional
gauge fields develop vacuum expectation values in some directions in the root space. It
is interesting to note that the effect of the background (gauge) field may be reexpressed
as a “twist” of the boundary conditions of the initial fields (with respect to the compact
dimensions) and which is removed by the formal limit of vanishing Wilson lines vev’s.
A consequence of this symmetry breaking is that after compactification (part of) the 4D
Kaluza–Klein mass spectrum of the initial fields is changed and the levels are shifted by
values proportional to the compactification radii and the Wilson lines vev’s.
We considered generic 4D N = 1 models with one and two extra dimensions
compactified on a circle and two-torus respectively, in the presence of a constant
background (gauge) field. Such models are “field theory” (α → 0) limits of 4D N = 1
orbifold compactification of the heterotic string with an N = 2 sector (“bulk”) in the
presence of Wilson lines. For these models we evaluated the structure of the overall one-
loop correction to the 4D gauge couplings including Wilson line effects.
The results depend significantly on the directions (in the root space) of the vev’s of the
Wilson lines which in a realistic model are expected to be fixed by a dynamical (possibly
non-perturbative) mechanism. We computed the corrections to the gauge couplings for
cases when the initial gauge symmetry is broken to a subgroup G ∗ and for the special case
when a non-zero level Kaluza–Klein state of the tower becomes massless, leading to an
enhancement of the gauge symmetry.
The calculation required a careful analysis of the UV and IR behaviour of the gauge
couplings. The Wilson line corrections were identified and it was observed that the UV
behaviour of the models considered is not worsened by their non-zero vev’s. The couplings
(regarded as functions of the Wilson line vev’s ρk,α ) have a discontinuity (infrared
divergence) at all “moduli” points where Kaluza–Klein states of non-zero levels become
massless. As a result, values of the Wilson lines vev’s corresponding to (ρ1,σ , ρ2,σ ) integers
(σ fixed) cannot be smoothly reached perturbatively from those with (ρ1,σ , ρ2,σ ) non-
integers.
The results obtained were compared with their heterotic string counterpart. When no
massless state is present in a Kaluza–Klein tower (Case 2(B)), the one-loop correction of
the effective field theory approach has strong similarities with that of the heterotic string in
the presence of Wilson lines and in the limit α → 0 (when the effects of winding modes
are negligible). This finding is remarkable given the different approach of the two methods
and shows that effective field theories can indeed yield very reliable results in the region
of large compactification radii (in units of UV cut-off length). In both cases the results
can be written as a sum over elliptic theta functions of genus one (at the string level the
general correction is further associated with the topology of a genus two Riemann surface).
When a massless state is present in a Kaluza–Klein tower an infrared regulator is needed
for evaluating the contribution to the gauge couplings (Cases 1, 2(A)). For two compact
dimensions this has as effect the presence of a correction which cannot be recovered by the
(infrared regularised) string calculations available, in the limit α → 0. This is ultimately
caused by the infrared regularisation of the string which does not commute with the (field
theory) limit α → 0.
The techniques developed in this work can easily be applied to specific models. The
results obtained may be used for the study of the unification of the gauge couplings in
MSSM-like models derived from a grand unified model with Wilson line gauge symmetry
breaking. The results can also be applied to models with extra dimensions with a structure
of the 4D Kaluza–Klein masses similar to the general one considered in this paper, even
when this structure is not induced by a Wilson line background, but by orbifolding, etc.
It would be interesting to know if the results of this work can be extended to regions of
small radii of compactification (in units of UV cut-off length) and if the agreement with
the corresponding string results can be maintained.
Acknowledgements
The author thanks Graham Ross and Fernando Quevedo for many discussions on related
topics. This work was supported by a research fellowship from the part of PPARC, United
Kingdom. Part of this work was performed during a visit at CERN supported by the RTN-
European Program HPRN-CT-2000-00148, “Physics Across the Present Energy Frontier:
Probing the Origin of Mass”.
Appendix A. Evaluation of Kaluza–Klein integrals

In the following a “primed” sum m f (m) stands for a sum over all non-zero, integer
numbers, m ∈ Z − {0}. Similarly, m,n f (m, n) is a sum over all integers (m, n) with
(m, n) = (0, 0).
A.1. Computing R1
We compute the integral (ξ > 0, δ > 0)

∞
dt −πt [(m+ρ)2 +δ]
R1 [ξ, ρ, δ] ≡ e (A.1)
t m
ξ
for ξ 1, ρ ∈ R. With this notation, the integral encountered in the text in Eq. (19) will be
given by J = R1 [ξ/(Rµ)2 , ρ, χ(Rµ)2 ] while L1 in Eq. (47) is L1 = R1 [ξ/(T2 U2 ), ρ2 −
U1 ρ1 , U22 ρ12 ].
To compute R1 we use the Poisson resummation formula

1 −πA−1 ñ2 +2iπ ñσ
e−πA(n+σ ) = √
2
e . (A.2)
n∈Z
A ñ∈Z
We have
∞
dt −πtρ 2 −πt (m+ρ)2 −πδt
R1 [ξ, ρ, δ] = −e + e e (A.3)
t m
ξ
∞
dt 1 1 −π m̃2 /t +2iπ m̃ρ −πδt
−e−πtρ + √ + √
2
= e e (A.4)
t t t m
ξ
2e−πξ δ √

= √ + 2π δ Erf πξ δ − Γ 0, πξ ρ 2 + δ
ξ
√ 2
− ln2 sin π ρ + i δ , ξ 1. (A.5)
For the last term in (A.4) we assumed ξ 1 and we used that [30]
∞ ν
ν−1 −bx p −ax −p 2 a 2p √
dx x e = Kν/p 2 ab , Re(b), Re(a) > 0;
p b
0

π −z
K−1/2 (z) = e . (A.6)
2z
In the last integral in (A.4) we set ξ = 0 (the integral has no divergence in ξ or in δ → 0,
e.g., ξ ln δ). Indeed the error @1 induced by doing so vanishes
ξ
dt −π m̃2 /t +2iπ m̃ρ −πδt

|@1 | = e e
t 3/2
0 m̃
∞ dt
∞
dt −π m̃2 t
e 2 e−π m̃t
t 1/2 t 1/2
m̃ 1/ξ m̃>0 1/ξ
∞
2 dt −πt 2ξ 3 3/2
e 1, if ξ 1. (A.7)
1 − e−π/ξ t 1/2 1 − e−π/ξ 2πe
1/ξ
Thus |@1 | vanishes if ξ 1 for δ > 0, ρ ∈ R and the result for R1 is indeed that given by
Eq. (A.5). Further, with the expansions [30],
x
2 2x 2x 3
dt e−t = √ − √ + O x 5 ,
2
Erf[x] ≡ √ if x 1,
π π 3 π
0
(−z)k
−Γ [0, z] = γ + ln z + , for z > 0, (A.8)
k!k
k1
we obtain an approximation for R1 of Eq. (A.5) in the limit ξ → 0:

√
1 √ sin π(ρ + i δ) 2
R1 [ξ, ρ, δ] = − ln 4πe−γ e−2/ ξ − ln √ ,
ξ π(ρ + i δ)

1 1
ξ min 1, , . (A.9)
πδ π(ρ 2 + δ)
If ρ is non-integer (ρ ∈
/ Z) and δ = 0, one finds from (A.5)
∞
dt −πt (m+ρ)2
R1 [ξ, ρ ∈
/ Z, 0] = e
t m
ξ
= √ − Γ 0, πξρ 2 − ln |2 sin πρ|2 , ξ 1, (A.10)

ξ
If ρ is a non-zero integer (ρ ∈ Z∗ )
2
√ 2
R1 [ξ, ρ ∈ Z∗ , δ] = √ − Γ 0, πξρ 2 − ln2 sin πi δ , ξ 1, δ 1. (A.11)
ξ
If ρ → 0 and δ → 0 (the limits commute)
∞
dt −πt m2 1 √
R1 [ξ, 0, 0] = e = − ln 4πe−γ e−2/ ξ , ξ 1. (A.12)
t m ξ
ξ
Eqs. (A.5) and (A.9) were used in the text, Eqs. (19), (47).
For related results on Kaluza–Klein integrals see also Appendices in [11,12].
A.2. Computing R2
We compute the integral:
∞
dt
e−πt (m+ρ)
2
R2 [ξ, ρ] ≡ (A.13)
t 3/2 m
ξ
for ξ → 0 and ρ ∈ / Z. With this notation, integral L2 in Eq. (47) is given by L2 =

R2 [ξ U2 /T2 , ρ1 ]|U2 |.
The constant ρ can be written as
ρ = [ρ] + ∆ρ , with [ρ] ∈ Z, 0 < ∆ρ < 1. (A.14)
∆ρ is the fractional part of ρ, positive definite, irrespective of the sign of ρ. ρ ∈

/ Z thus
∆ρ = 0.
With (A.2) one has
1
dt 1 −π m̃2 /t +2iπ m̃ρ 1 −πtρ 2
R2 [ξ, ρ] = √ e + √ −e
t 3/2 t t
ξ m̃
∞
dt
e−πt (m+ρ)
2
+
t 3/2 m
1
1 1/ξ
dt 1 −πtρ 2 −πt m̃2 +2iπ m̃ρ
= √ −e + dt e
t 3/2 t
ξ m̃ 1
∞
dt
e−πt (m+ρ)
2
+
t 3/2 m
1
∞ ∞
−πt m̃2 +2iπ m̃ρ dt
e−πt (m+ρ)
2
= Iξ + dt e +
t 3/2 m
1 m̃ 1
≡ Iξ + F , ξ 1. (A.15)
We introduced Iξ
1
dt 1
√ − e−πtρ
2
Iξ ≡
t 3/2 t
ξ
1 2 √

− √ e−πρ ξ + 2e−πρ − 1 + 2πρ Erf ρ π − Erf ρ πξ ,
2 2
= (A.16)
ξ ξ
and finite F (since ρ is non-integer)
∞ ∞
−πt m̃2 +2iπ m̃ρ dt
e−πt (m+ρ)
2
F≡ dt e +
t 3/2 m
1 m̃ 1
∞ ∞
dt
e−πt m̃
2 +2iπ m̃ρ
e−πt (m+ρ)
2
= lim t @ dt +
@→0 t 3/2+@ m
1 m̃ 1
1 ∞
dt −π m̃2 /t +2iπ m̃ρ dt
e−πt (m+ρ)
2
= lim e +
@→0 t 2+@ t 3/2+@ m
0 m̃ 1
∞ 1
dt −πt (m+ρ)2 dt √
t e−πtρ
2
= lim e + −1 + ≡ G + H, (A.17)
@→0 t 3/2+@ m
t 2+@
0 0
where
∞
dt
e−πt (m+ρ)
2
G ≡ lim
@→0 t 3/2+@ m
0
∞
dt
e−πt (m+ρ) − e−πtρ
2 2
= lim (A.18)
@→0 t 3/2+@ m
0
∞
dt −πt ∆2ρ −πtρ 2 −πt (m+∆ρ )2 −πt (m−∆ρ )2
= lim e −e + e + e
@→0 t 3/2+@
0 m>0 m>0
(A.19)
1
= lim π Γ 2 +@ −@ − 12
@→0

× ∆ρ 1+2@
− |ρ| 1+2@
+ (m + 1 + ∆ρ ) 1+2@
+ (∆ρ → −∆ρ ) (A.20)
m0

= lim 2π |ρ| − ∆ρ + π 2 +@ Γ − 12 − @
1
@→0

× ζ [−1 − 2@, 1 + ∆ρ ] + ζ [−1 − 2@, 1 − ∆ρ ] (A.21)

= 2π|ρ| + 2πB2 [∆ρ ]. (A.22)
In (A.19), (A.20) we used that ∆ρ = 0 (since 0 < ∆ρ < 1 (ρ ∈
/ Z)) and
1
ζ [z, q] = (q + n)−z , ζ [−1, x] = − B2 [x],
2
n0
1
B2 [x] ≡ + x 2 − x. (A.23)
6
Also
1
dt √ −πtρ 2
√ 1
1

H ≡ lim 2+@
−1 + t e = 1 + |ρ| π Γ − 2 − Γ − 2 , πρ 2
. (A.24)
@→0 t
0
Adding together Eqs. (A.22), (A.24) and using Eq. (A.17) gives
√

F = 2π|ρ| + 2πB2 [∆ρ ] + 1 + |ρ| π Γ − 12 − Γ − 12 , πρ 2 . (A.25)
Eqs. (A.15), (A.16) and (A.25) give

1 2 −πρ 2 ξ
1
R2 [ξ, ρ] = − √ e − 2πρ Erf ρ πξ + 2π |ρ| + + ∆ρ − ∆ρ ,
2
ξ ξ 6
with ξ 1, ρ ∈
/ Z (0 < ∆ρ < 1). (A.26)
We evaluate the error introduced in Eq. (A.15), while computing R2 to ensure it vanishes
for ξ 1:
∞

−πt m2 +2iπmρ
|@2 | ≡ dt e
m

1/ξ
ξ
dt −πm/t −2
2 e = ln 1 − e−π/ξ 1, if ξ 1. (A.27)
t2 π
0 m>0
Eqs. (A.26), (A.27) were used in the text, Eqs. (47) and (48).
A.3. Computing R3
We compute the integral:
∞
dt
e−π m̃2 /t −πt U2 (m1 +ρ1 )
2 2 2 +2iπ m̃ [ρ −U (ρ +m )]
R3 [ξ, ρ1 , ρ2 ] ≡ 2 2 1 1 1 (A.28)
t 3/2 m1
ξ m̃2
for 0 < ξ 1, U1,2 = 0, ρ1 ∈ / Z and real. Therefore, L3 of Eq. (47) is L3 =

R3 [ξ/(T2 U2 ), ρ1 , ρ2 ].
First we introduce ρ1 = [ρ1 ] +∆ρ1 , where [ρ1 ] is an integer and ∆ρ1 is its fractional part
defined as 0 < ∆ρ1 < 1. To evaluate the integral of R3 we use Eq. (A.6) after having set
ξ = 0 in its lower limit (this is allowed for the integral of R3 is finite under the assumptions
for which we evaluate it). Setting ξ = 0 introduces a (vanishing) error in R3 to be evaluated
shortly (Eq. (A.33)). One has
1
R3 [ξ, ρ1 , ρ2 ] = e2iπ m̃2 (ρ2 −U1 (ρ1 +m1 )) e−2π|m̃2 (m1 +ρ1 )U2 |
m
| m̃ 2 |
1 m̃2
1
= e2iπ m̃2 (ρ2 −U1 (ρ1 +m1 )) e−2π|m̃2 (m1 +ρ1 )U2 |
m
| m̃ 2 |
1 m̃2
1
− e2iπ m̃2 (ρ2 −U1 ρ1 )−2π|m̃2 ρ1 U2 |
|m̃2 |
m̃2
1
= e2iπ m̃2 (ρ2 −U1 (m+∆ρ1 ))−2π|m̃2 (m+∆ρ1 )U2 |
m m̃2
| m̃ 2 |

1
2iπ m̃2 (ρ2 −U1 ∆ρ1 )−2π|m̃2 ∆ρ1 U2 |
+ e − (∆ρ1 → ρ1 ) ,
|m̃2 |
m̃2
(A.29)
where we used the notation m = m1 + [ρ1 ] ∈ Z.
With |m + ∆ρ1 | = ±(m + ∆ρ1 ) for m 1 and m −1 respectively, one finds after
some algebra

R3 [ξ, ρ1 , ρ2 ] = − ln 1 − e2iπmU e−2iπ(ρ2 −∆ρ1 U ) 2 1 − e2iπmU e2iπ(ρ2 −∆ρ1 U ) 2
m>0

sin π(ρ2 − Uρ1 ) 2

− 2π|U2 | |ρ1 | − ∆ρ1 + ln . (A.30)
sin π(ρ2 − U ∆ρ1 )
This result may also be written as

ϑ1 (ρ2 − ∆ρ1 U |U ) 2 1

R3 [ξ, ρ1 , ρ2 ] = − ln
η(U ) − 2π|U2 | 6 − ∆ρ1 + |ρ1 |
2
+ ln2 sin π(ρ2 − Uρ1 ) (A.31)
with the special functions η, ϑ1

η(τ ) ≡ eπiτ/12 1 − e2iπτ n ,
n1

ϑ1 (z|τ ) ≡ 2q 1/8 sin(πz) 1 − q n 1 − q n e2iπz 1 − q n e−2iπz ,
n1
q ≡e 2iπτ
. (A.32)
We now evaluate the error introduced by setting ξ = 0 in the integral for R3 . This error
equals
ξ
dt −π m̃2 /t −πt U 2 (m1 +ρ1 )2 +2iπ m̃2 [ρ2 −U1 (ρ1 +m1 )]

|@3 | ≡ e 2 2
t 3/2
m 1 m̃2
0

ξ
dt −π m̃2 /t −πt U 2 (m1 +[ρ1 ]+∆ρ )2
e 2 2 1
m1
t 3/2
m̃2 0

ξ
dt
e−π m̃2 /t −πt U2 (m+∆ρ1 )
2 2 2
=
m
t 3/2
m̃2 0

ξ
dt −π m̃22 /t −πt U22 ∆2ρ
+ e 1 − (∆ρ1 ↔ ρ1 ) . (A.33)
t 3/2
m̃2 0
Each integral in the square bracket vanishes if ξ 1. Indeed with γ standing for ρ1 or ∆ρ1
one has
dt dt
ξ ξ ξ
dt −π m̃22 /t −πt U22 γ 2 −π m̃22 /t
E1 ≡ e 2 e 2 e−π m̃2 /t
t 3/2 t 3/2 t 3/2
m̃2 0 m̃2 >0 0 m̃2 >0 0
ξ
2 dt −π/t 2ξ 3 3/2
e 1 if ξ 1. (A.34)
1 − e−π/ξ t 3/2 1 − e−π/ξ 2πe
0
Similarly, the first integral in (A.33) is vanishing for ξ small enough:
dt
ξ
2
e−π m̃2 /t e−πt U2 (m+∆ρ1 ) + (∆ρ1 ↔ −∆ρ1 )

2 2
E2 ≡ 2
t 3/2
m>0 m̃2 >0 0
dt
ξ
2 2
e−π m̃2 /t e−πt U2 m + e−πt U2 (m−1)

2 2 2
2 3/2
t
m>0 m̃2 >0 0
dt dt
ξ ξ
−π m̃22 /t −πt U22 m2
e−π m̃2 /t .
2
4 3/2
e e +2 3/2
(A.35)
t t
m>0 m̃2 >0 0 m̃2 >0 0
The last integral was already shown to vanish for ξ 1, while the first integral is smaller
than
dt
ξ
e−π m̃2 /t e−πt U2 m

2
4 3/2
t
m>0 m̃2 >0 0
ξ
e−π/t e−πU2 t
2
dt
=4
t 3/2 1 − e−π/t 1 − e−πU22 t
0

4 ξ 5 5/2
1, if ξ min 1, U22 . (A.36)
1 − e−π/ξ πU22 2πe
For the last factor under last integral we used that e−a /(1 − e−a ) < 1/a, a > 0. Eqs. (A.34),
(A.36) set the conditions for which the results for R3 , (A.30), (A.31) hold true:

|@3 | 1, if ξ min 1, U22 . (A.37)
Eqs. (A.31), (A.37) were used in the text, Eqs. (47), (48).
A.4. General Kaluza–Klein integrals
A generic presence in models with compact dimensions is a generalised version of the

integral R2
∞
∗ dt −πt (m+ρ)2 −πδt
R [ξ, ρ, δ, ν] ≡ e , ξ, δ, ν > 0, (A.38)
tν m
ξ
with ξ 1, δ 1, with ν > 0 and ρ real (ξ and δ are UV and IR regulators, respectively).
For future reference we outline the computation of R∗ and of its DR version, following the
approach used for R2 of Appendix A.2. One can have δ = 0 provided that ρ ∈ / Z − {0}. We
write ρ as
ρ = [ρ] + ∆ρ , with [ρ] ∈ Z, 0 ∆ρ < 1, (A.39)
with ∆ρ the fractional part of ρ. Thus ∆2ρ + δ = 0, unless ρ = 0 = δ(= ∆ρ ). With Eq. (A.2)
one has
1
∗ dt 1 −π m̃2 /t +2iπ m̃ρ 1 −πtρ 2
R = √ e + √ −e e−πδt
tν t m̃ t
ξ
∞
dt −πt (m+ρ)2 −πδt
+ e
tν m
1
1 1/ξ
dt 1 −πtρ 2 −πδt dt −πt m̃2 +2iπ m̃ρ−πδ/t
= ν
√ −e e + 3/2−ν
e
t t t
ξ m̃
1
∞
dt −πt (m+ρ)2 −πδt
+ e
tν m
1
∞ ∞
dt dt −πt (m+ρ)2 −πt δ
= Iξ∗ + e −πt m̃2 +2iπ m̃ρ−πδ/t
+ e ,
t 3/2−ν tν m
1 m̃ 1
ξ 1
≡ Iξ∗ + F ∗ , ξ 1, (A.40)
with an obvious notation in the last two steps. Iξ∗ and F ∗ are computed below, Eqs. (A.41)
and (A.42). F ∗ is finite within our assumptions on δ, ρ (∆ρ ). Its integrand is always
∞ on the interval (1, 1/ξ )
exponentially suppressed. In the second line above, the integral
was actually evaluated on (1, ∞). This introduces an error 1/ξ (integrand) which can be
shown to vanish as in Eq. (A.7) if ξ 1 and ν > 0.
One has
1
dt 1
Iξ∗ ≡ √ −e −πtρ 2 −πδt
e
tν t
ξ

ν−1

= π δ + ρ2 Γ 1 − ν, π δ + ρ 2 − Γ 1 − ν, π δ + ρ 2 ξ

− (πδ)ν−1/2 Γ 12 − ν, πδ − Γ 12 − ν, πδξ . (A.41)
Further, with @ → 0 and using (A.2) one can rewrite the finite F ∗ as
∞ ∞
∗ dt −πt m̃2 +2iπ m̃ρ−πδ/t dt −πt (m+ρ)2 −πδt
F ≡ e + e
t 3/2−ν tν m
1 m̃ 1
∞ ∞
dt −πt m̃2 +2iπ m̃ρ−πδ/t dt
e−πt (m+ρ)
2 −πδt
= lim t @
e +
@→0 t 3/2−ν t ν+@ m
1 m̃ 1
1 ∞
dt −π m̃2 /t +2iπ m̃ρ−πδt dt
e−πt (m+ρ)
2 −πδt
= lim e +
@→0 t 1/2+ν+@ t ν+@ m
0 m̃ 1
∞ 1
dt −πt (m+ρ)2 −πδt dt √ −πtρ 2
= lim e + te − 1 e−πδt
@→0 t ν+@ m
t 1/2+ν+@
0 0
∗ ∗
≡G +H . (A.42)
The first integral (denoted G ∗ ) is just a DR version of R∗ and is evaluated below. One has
(@ → 0)
∞ ∞
∗ dt −πt (m+ρ)2 −πδt dt
e−πt (m+ρ) − e−πtρ e−πδt
2 2
G ≡ e =
t ν+@ m
t ν+@ m
0 0
∞
dt −πt ∆2ρ −πtρ 2 −πt (m+∆ρ )2 −πt (m−∆ρ )2 −πδt
= ν+@
e −e + e + e e
t
0 m>0 m>0
ν−1+@ ν−1+@
= Γ [1 − ν − @]π ν−1+@ δ + ∆2ρ − δ + ρ2

ν−1+@
+ Γ [1 − ν − @]π ν−1+@
(m + ∆ρ ) + δ
2
+ (∆ρ → −∆ρ )
m>0
ν−1+@ ν−1+@
= Γ [1 − ν − @]π ν−1+@
δ + ∆2ρ − δ + ρ2

+ ζ [2 − 2ν − 2@, 1 + ∆ρ ] + (∆ρ → −∆ρ )
(−δ)k
+ π ν−1+@ Γ [k + 1 − ν − @]
k!
k1

× ζ [2 + 2k − 2ν − 2@, 1 + ∆ρ ] + (∆ρ → −∆ρ ) , (A.43)

with ∆2ρ + δ = 0. We used the convergent expansion [31] (0 < q/a < 1)

−s Γ [k + s] −q k
a(m + c)2 + q = a −s ζ [2k + 2s, c], (A.44)
k!Γ [s] a
m0 k0
where ζ [x, c] with c = 0, −1, −2, . . . has one singularity (simple pole) at x = 1 and
ζ [x, 1] = ζ [x].
The above result for G ∗ can be simplified for specific cases, if δ 1. If ν is such as
ν = 1 + N ∗ or ν = 1/2 + N ∗ with N ∗ a non-zero natural number, the series in k has
singularities from the Γ and ζ functions respectively (note that @ → 0). One can isolate
such singularities from the rest of the series which can be shown to vanish for δ 1.
For such cases one finds for δ 1 (using ζ [1 − @, q] = −1/@ − ψ(q) + O(@), Γ [−@] =
−1/@ − γ + O(@))
ν−1+@ ν−1+@
G ∗ = Γ [1 − ν − @]π ν−1+@ δ + ∆2ρ − δ + ρ2

+ Γ [1 − ν − @]π ν−1+@ ζ [2 − 2ν − 2@, 1 + ∆ρ ] + (∆ρ → −∆ρ )

(−πδ)ν−1/2 1
− δN ∗ ,ν−1/2
K
+ ψ(∆ρ ) + ψ(−∆ρ ) + ln(4πe )γ
(ν − 1/2)! @
ν
(−δ)m−1
+ δNK
∗ ,ν−1 π ν−1+@
Γ [m − ν − @]
(m − 1)!
m2

× ζ [2m − 2ν − 2@, 1 + ∆ρ ] + (∆ρ → −∆ρ ) ,

(A.45)
K is a notation for the Kronecker delta, equal to 1 for a = b and zero otherwise,
where δa,b
∗
N is a non-zero natural number, ψ(x) ≡ d(ln Γ [x])/dx. The rhs of Eq. (A.45) has a finite
number of terms. The second integral (H∗ ) in Eq. (A.42) is
1
∗ dt √ −πtρ 2 −πδt
H ≡ −1 + te e
t 1/2+ν+@
0

= −(πδ)ν−1/2+@ Γ 12 − ν − @ − Γ 12 − ν − @, πδ

ν+@−1

+ π δ + ρ2 Γ [1 − ν − @] − Γ 1 − ν − @, π δ + ρ 2 . (A.46)
The general result for R∗ of Eq. (A.40) is, with ∆2ρ + δ = 0, ξ 1, δ 1
R∗ = Iξ∗ + F ∗ = Iξ∗ + G ∗ + H∗ , (A.47)

with Iξ given in Eq. (A.41), F∗
in Eq. (A.42), G∗
in Eq. (A.45) (or (A.43)) and H∗ in
Eq. (A.46). Additional assumptions are needed to simplify this result further.
As an example, if ν = 3/2 one has

∞
∗ dt
e−πt (m+ρ)
2 −πδt
R [ξ, ρ, δ, 3/2] ≡
t 3/2 m
ξ
1 2 1/2 1/2
− √ e−π(δ+ρ )ξ − 2π δ + ρ 2
2
= Erf πξ δ + ρ 2
ξ ξ

2 1/2 1 2 1/2
+ 2π δ + ρ + + ∆ρ − ∆ρ + δ
2
6
γ +ψ(∆ρ )+ψ(−∆ρ )

+ πδ ln 4πξ e , (A.48)
with ∆2ρ + δ = 0. One can set δ = 0 to obtain R2 of Eq. (A.2). If also ξ(ρ 2 + δ) 1 then
the term proportional to Erf function is also absent. G ∗ for ν = 3/2 is the DR version of
(A.48) and has a similar form, with the above ξ dependence replaced by πδ/@.
The method presented is particularly useful for cases with ν = N ∗ + 1/2. It also
provides a dimensional regularisation (DR) version of the integral R∗ , Eq. (A.45) and
a relation between the two regularisation schemes.
Appendix B. Case 2(B) for ρ2 ∈

/ Z and ρ1 ∈ Z
We extend the validity of Case 2(B) in the text to situations when ρ2 is non-integer and
ρ1 integer. The method of Case 2(B) is not well-defined for such a case, see integral L2
Eq. (47) which is IR divergent if ρ1 ∈ Z. However, the (formal) limit of the final result
of Case 2(B) for ρ2 non-integer, ρ1 integer is finite and does give the correct result as
we show below by computing separately this case. This finding provides an extension of
Case 2(B) to all cases with ρ1 or ρ2 non-integer with arbitrary, real values for the other.
The correction Ωi can be written as
βi (4) β̄i (4)
Ωi = J + J , (B.1)
4π 0 4π
where we introduced:
∞ ∞
(4) dt −πt M0,0
2 /ν 2 dt −πt ξ |ρ2 −Uρ1 |2 /(T2 U2 )
J0 ≡ e = e
t t
ξ ξ

= Γ 0, πξ |ρ2 − Uρ1 | /(T2 U2 ) , 2
dt −πt M 2 dt − π t |m +ρ −U (m +ρ )|2
∞ ∞
2
m1 ,m2 /µ
J ≡
(4)
e = e T2 U2 2 2 1 1
m ,m
t m ,m
t
1 2 ξ 1 2 ξ
∞
dt − T πUt |m2 +ρ2 −U m|2
= e 2 2
m,m2
t
ξ

− Γ 0, πξ |ρ2 − Uρ1 |2 /(T2 U2 ) + Γ 0, πξρ22 /(T2 U2 ) . (B.2)

In the last step, under the integral, the exponential evaluated for (m1 , m2 ) = (0, 0) was
added and subtracted, we then replaced m1 + ρ1 → m as the new summation index, and
finally isolated the “new” (0, 0) mode from the rest of the series. Further, the integrand in
the final series can be written after Poisson resummation as
− T πUt |m2 +ρ2 −U m|2

e 2 2
m,m2
− T πUt (m2 +ρ2 )2

− T πUt |m2 +ρ2 −U m|2
= e 2 2 + e 2 2
m2 m m2
− T πUt (m2 +ρ2 )2

= e 2 2
m2
1/2
T 2 U2 −πt m2 U2 −π m̃22
T2 U2 U
−πt m2 T 2 +2iπ m̃2 (ρ2 −U1 m)
+ e T 2 + e t 2 .
t m m m̃2
Since ρ2 is assumed non-integer, each of the above series can be integrated separately over
(ξ, ∞) to compute J (4) . Further
∞
dt − T πUt (m2 +ρ2 )2
K1 ≡ e 2 2
t m
ξ 2
1/2
T 2 U2
=2 − Γ 0, πξρ22 /(T2 U2 ) − ln |2 sin πρ2 |2 ,

t
∞
dt −πt
U2 2
T2 m
T2 T2 U2 1/2 π
K2 ≡ (T2 U2 ) 1/2
e = −2 + U2 ,
t 3/2 m
ξ ξ 3
ξ
∞
dt −π m̃22
T2 U2 2 U2
t −πt m T2 +2iπ m̃2 (ρ2 −U1 m)
K3 ≡ (T2 U2 ) 1/2
e
t 3/2 m
ξ m̃2

= − ln 1 − e2iπmU e−2iπρ2 2 1 − e2iπmU e2iπρ2 2 (B.3)
m1
which are valid provided that T2 U2 /ξ max{1/U22 , U22 }. To evaluate K1 we used

Eq. (A.10) while to evaluate K2 we used Eq. (A.12) in Appendix A of Ref. [11] (which
agrees with the limit of ρ → 0 in Eq. (A.26)). K3 can be evaluated in the limit ξ 1 using
Eq. (A.6).
Adding together all the contributions Ki one finds for Ωi

βi Λ2 (R2 sin θ )2 β̄i ∗ 4
Ωi = ln γ − ln 4πe−γ e−T2 T2∗ U2 η(U )

4π πe |ρ2 − Uρ1 |2 4π

β̄i sin πρ2 2
− ln
4π π(ρ2 − Uρ1 )

β̄i
2iπmU −4 2iπmU −2iπρ2 2 2iπmU 2iπρ2 2
− ln 1−e 1−e e 1−e e .
4π
m1
(B.4)
The first term in Ωi is the contribution of the original (0, 0) modes. The second contribution
is due to the tower of Kaluza–Klein modes of non-zero level. The third term (divergent in
the limit ρ2 integer) bears some similarities with the third term in Eq. (24) of the one-
dimensional case. The last term above is suppressed for large U2 and is a two-dimensional
effect. An equivalent form of the above result is
βi Λ2 (R2 sin θ )2
Ωi = ln γ
4π πe |ρ2 − Uρ1 |2

β̄i T ∗ U2 ϑ1 (∆ρ2 |U ) 2

+ T2∗ − ln γ 2 − ln η(U ) , (B.5)
4π πe |ρ2 − Uρ1 |2
where the special function ϑ1 was defined in (A.32) and T2∗ = T2 /ξ . The result (B.5) agrees
with the formal limit ρ1 integer (ρ2 non-integer) of Eq. (52) of Case 2(B) in the text. The
analysis of Case 2(B) is then valid as long as ρ1 or ρ2 is non-integer with arbitrary, real
values for the other.
References
[1] L.J. Dixon, J.A. Harvey, C. Vafa, E. Witten, Nucl. Phys. B 261 (1985) 678;
L.J. Dixon, J.A. Harvey, C. Vafa, E. Witten, Nucl. Phys. B 274 (1986) 285;
L.E. Ibanez, J. Mas, H.P. Nilles, F. Quevedo, Nucl. Phys. B 301 (1988) 157.
[2] Y. Hosotani, Phys. Lett. B 126 (1983) 309;
Y. Hosotani, Phys. Lett. B 129 (1983) 193.
[3] P. Candelas, G.T. Horowitz, A. Strominger, E. Witten, Nucl. Phys. B 258 (1985) 46;
E. Witten, Nucl. Phys. B 258 (1985) 75;
See also: M.B. Green, J.H. Schwarz, E. Witten, Superstring Theory, Cambridge Univ. Press, Cambridge,
1987.
[4] H.P. Nilles, S. Stieberger, Nucl. Phys. B 499 (1997) 3, hep-th/9702110;
B. de Wit, V. Kaplunovsky, J. Louis, D. Lust, Nucl. Phys. B 451 (1995) 53;
E. Kiritsis, C. Kounnas, P.M. Petropoulos, J. Rizos, Nucl. Phys. B 483 (1997) 141;
S. Stieberger, Nucl. Phys. B 541 (1999) 109, hep-th/9807124.
[5] I. Antoniadis, C. Bachas, E. Dudas, Nucl. Phys. B 560 (1999) 93, hep-th/9906039.
[6] T.R. Taylor, G. Veneziano, Phys. Lett. B 212 (1988) 147;
M. Lanzagorta, G.G. Ross, Phys. Lett. B 349 (1995) 319, hep-ph/9501394;
K.R. Dienes, E. Dudas, T. Gherghetta, Nucl. Phys. B 537 (1999) 47, hep-ph/9806292;
For a two-loop calculation in such models see:
D. Ghilencea, G.G. Ross, Nucl. Phys. B 569 (2000) 391, hep-ph/9908369;
Z. Kakushadze, T.R. Taylor, Nucl. Phys. B 562 (1999) 78, hep-th/9905137.
[7] V.S. Kaplunovsky, Nucl. Phys. B 307 (1988) 145;

V.S. Kaplunovsky, Nucl. Phys. B 382 (1988) 436, Erratum;
V.S. Kaplunovsky, hep-th/9205068;
For a completely revised version see: V.S. Kaplunovsky, hep-th/9205070.
[8] D. Bailin, A. Love, Phys. Rep. 315 (1999) 285.
[9] L.J. Dixon, V. Kaplunovsky, J. Louis, Nucl. Phys. B 355 (1991) 649;
See also:
P. Mayr, S. Stieberger, Nucl. Phys. B 407 (1993) 725, hep-th/9303017;
D. Bailin, A. Love, W.A. Sabra, S. Thomas, Mod. Phys. Lett. A 10 (1995) 337;
D. Bailin, A. Love, W.A. Sabra, S. Thomas, Mod. Phys. Lett. A 9 (1994) 2543;
C. Kokorelis, Nucl. Phys. B 579 (2000) 267, hep-th/0001217.
[10] S. Ferrara, C. Kounnas, D. Lust, F. Zwirner, Nucl. Phys. B 365 (1991) 431.
[11] D. Ghilencea, S. Groot Nibbelink, Nucl. Phys. B 641 (2002) 35, hep-th/0204094.
[12] D.M. Ghilencea, Nucl. Phys. B 653 (2003) 27, hep-ph/0212119.
[14] L.E. Ibanez, J.E. Kim, H.P. Nilles, F. Quevedo, Phys. Lett. B 191 (1987) 282.
[16] L.E. Ibanez, J. Mas, H.P. Nilles, F. Quevedo, Nucl. Phys. B 301 (1988) 157.
[17] G. Lopes Cardoso, D. Lust, T. Mohaupt, Nucl. Phys. B 450 (1995) 115, hep-th/9412209;
G. Lopes Cardoso, D. Lust, T. Mohaupt, Nucl. Phys. B 432 (1994) 68, hep-th/9405002.
[18] L.J. Hall, H. Murayama, Y. Nomura, Nucl. Phys. B 645 (2002) 85, hep-th/0107245;
A. Hebecker, J. March-Russell, Nucl. Phys. B 625 (2002) 128, hep-ph/0107039.
[19] H.P. Nilles, S. Stieberger, Phys. Lett. B 367 (1996) 126, hep-th/9510009;
H.P. Nilles, hep-ph/0004064, TASI Lectures, Boulder, CO, 1997, and references therein;
K.R. Dienes, Phys. Rep. 287 (1997) 447, hep-th/9602045.
[20] P. Mayr, S. Stieberger, Phys. Lett. B 355 (1995) 107, hep-th/9504129;
P. Mayr, S. Stieberger, hep-th/9412196;
H.P. Nilles, S. Stieberger, Nucl. Phys. B 499 (1997) 3, hep-th/9702110.
[21] T. Friedmann, E. Witten, hep-th/0211269.
[22] P. Candelas, G.T. Horowitz, A. Strominger, E. Witten, Nucl. Phys. B 258 (1985) 46.
[23] N. Haba, M. Harada, Y. Hosotani, Y. Kawamura, hep-ph/0212035.
[24] J. Polchinski, String Theory, Vol. 1, Cambridge Univ. Press, Cambridge, 1998.
[25] R. Slansky, Group theory for unified model building, Phys. Rep. 79 (1981) 1;
J. Patera, R.T. Sharp, P. Winternitz, J. Math. Phys. 17 (1976) 1972;
J. Patera, R.T. Sharp, P. Winternitz, J. Math. Phys. 18 (1977) 1519, Erratum.
[26] K.R. Dienes, Phys. Rev. Lett. 88 (2002) 011601, hep-ph/0108115.
[27] I. Antoniadis, K. Benakli, M. Quiros, New J. Phys. 3 (2001) 20, hep-th/0108005.
[28] K. Foerger, S. Stieberger, Nucl. Phys. B 559 (1999) 277, hep-th/9901020.
[29] E. Kiritsis, C. Kounnas, Nucl. Phys. B (Proc. Suppl.) 41 (1995) 331, hep-th/9410212;
E. Kiritsis, C. Kounnas, Nucl. Phys. B 442 (1995) 472, hep-th/9501020;
E. Kiritsis, C. Kounnas, hep-th/9507051;
E. Kiritsis, C. Kounnas, Nucl. Phys. B (Proc. Suppl.) 45BC (1996) 207, hep-th/9509017;
P.M. Petropoulos, J. Rizos, Phys. Lett. B 374 (1996) 49, hep-th/9601037;
See also the discussion in: K.R. Dienes, Phys. Rep. 287 (1997) 447, hep-th/9602045.
[30] I.S. Gradshteyn, I.M. Ryzhik, Table of Integrals, Series and Products, Academic Press, New York, 1965.
[31] E. Elizalde, Ten Physical Applications of Spectral Zeta Functions, Springer, Berlin, 1995.
Bound state equation in the Wilson loop approach

with minimal surfaces
F. Jugeau 1 , H. Sazdjian
Groupe de Physique Théorique, Institut de Physique Nucléaire, 2 Université Paris XI,
F-91406 Orsay cedex, France
Received 12 May 2003; received in revised form 24 June 2003; accepted 25 July 2003
Abstract
The large-distance dynamics in quarkonium systems is investigated, in the large-Nc limit, through
the saturation of Wilson loop averages by minimal surfaces. Using a representation for the quark
propagator in the presence of the external gluon field based on the use of path-ordered phase factors,
a covariant three-dimensional bound state equation of the Breit–Salpeter type is derived, in which the
interaction potentials are provided by the energy–momentum vector of the straight segment joining
the quark to the antiquark and carrying a constant linear energy density, equal to the string tension.
The interaction potentials are confining and reduce to the linear vector potential in the static case and
receive, for moving quarks, contributions from the moments of inertia of the straight segment. The
self-energy parts of the quark propagators induce spontaneous breakdown of chiral symmetry with a
mechanism identical to that of the exchange of one Coulomb-gluon. In the nonrelativistic limit, long
range spin–spin potentials are absent; the moments of inertia of the straight segment provide negative
contributions to the spin–orbit potentials going in the opposite direction to those of the pure timelike
vector potential. In the ultrarelativistic limit, the mass spectrum displays linear Regge trajectories
with slopes in agreement with their classical relationship with the string tension.
PACS: 03.65.Pm; 11.10.St; 12.38.Aw; 12.38.Lg; 12.39.Ki
Keywords: QCD; Confinement; Wilson loop; Minimal surfaces; Bound states; Quarkonium
E-mail addresses: jugeau@ipno.in2p3.fr (F. Jugeau), sazdjian@ipno.in2p3.fr (H. Sazdjian).

1 And Université de Cergy-Pontoise.
2 Unité Mixte de Recherche 8608.
doi:10.1016/j.nuclphysb.2003.07.018
222 F. Jugeau, H. Sazdjian / Nuclear Physics B 670 (2003) 221–263
1. Introduction
The Wilson loop [1] appears as one of the most efficient tools for probing the large-
distance properties of QCD. It provides a natural criterion for confinement through the
area law and also participates as a basic ingredient in the formulation of lattice gauge
theories [1,2]. Equations concerning path-ordered phase factors were first obtained by
Mandelstam [3] and analyzed in the QCD case by Nambu [4]. Loop equations were
obtained by Polyakov [5] and Makeenko and Migdal [6,7]. (For reviews, see Refs. [8,9].)
The loop equations actually represent an infinite chain of coupled equations relating
vacuum expectation values of loop operators having as supports different numbers of
closed contours emerging from contours with self-intersections. These equations must
in addition be supplemented with constraint equations [10,11], the so-called Mandelstam
constraints, which are sensitive to the gauge group structure of the theory. Due to their
complexity, the loop equations have not yet allowed a systematic resolution of QCD in
terms of loop variables. In the two-dimensional case, however, explicit expressions of the
Wilson loop averages for various types of contour have been obtained for U (Nc ) gauge
theories [12]. Renormalization properties of the Wilson loop averages were studied in the
framework of perturbation theory in Refs. [13,14], where it was shown that the latter are
multiplicatively renormalizable.
Considerable simplification is obtained in the large-Nc limit [15], corresponding to the
planar diagram approximation of the theory. In that case, apart from the disappearance of
many nonleading terms, it is the factorization property of the Wilson loop average for two
disjoint contours that becomes mostly relevant. Makeenko and Migdal studied the resulting
equations in the above limit pointing out their equivalence with a chain of Schwinger–
Dyson-type equations [6,7]. They showed that for large contours asymptotic solutions exist
corresponding to the minimal surfaces bounded by the loops.
Concerning physical applications, Eichten and Feinberg, using the area law for
the Wilson loop, could obtain the general expression of spin-dependent forces for
quark–antiquark systems to order 1/c2 in terms of color electric and magnetic field
correlators [16]. This problem was also investigated by Gromes [17]. Later on, Prosperi,
Brambilla et al. completed these results by also obtaining the velocity-dependent forces
[18–21]. Using an effective field theory approach, Pineda and Vairo determined the
structure of the full Hamiltonian, including all spin independent potentials [22]. On the
other hand, a wide program of investigations was undertaken by Dosch, Simonov et al. in
the framework of the “stochastic vacuum model” [23].
In recent years, the Wilson loop gained renewal of interest in connection with duality
properties of different field theories in different dimensions manifested through the
AdS/CFT correspondence [24].
The purpose of the present paper is to investigate the properties of the Wilson loop
concerning the bound state problem in QCD in the large-distance regime when the latter
is probed by minimal surfaces. Among possible nonperturbative solutions of the loop
equations at large distances, minimal surfaces appear as the most natural ones [6,7]; they
produce in a simple way the area law needed for confinement and satisfy the factorization
property, valid at large Nc . A complete solution of the loop equations should necessarily
include short-distance effects for which minimal surfaces do not seem sufficient alone to
F. Jugeau, H. Sazdjian / Nuclear Physics B 670 (2003) 221–263 223
provide the appropriate behavior. Therefore, for that regime perturbation theory results
should appropriately be incorporated in the Wilson loop solution; this aspect of the problem
will not, however, be analyzed in the present paper. Other nonperturbative properties of the
theory, related to its possible stringlike behavior, might also arise from contributions of
fluctuations of surfaces around the minimal surface bounded by the loop [25–27].
The main ingredient in our approach is a representation of the quark propagator in
the presence of an external gluon field as a series of terms involving free propagators,
path-ordered phase factors along straight lines and their derivatives. That representation
generalizes to the relativistic case the one used in the nonrelativistic limit [16]. The gauge
invariant quark–antiquark Green function is then represented by a series of terms involving
Wilson loops having as contours skew polygons with an increasing number of sides. Each
Wilson loop average is then replaced by the contribution of the corresponding minimal
surface bounded by the loop. Contrary to the usual two-particle Green functions, the gauge
invariant Green function does not manifestly satisfy a genuine integral equation which
might, as in the Bethe–Salpeter equation case [28], result in a bound state integral equation.
In the present case, the Green function satisfies with the Dirac operators two independent
and compatible equations. Selecting in the large-time limit the total momentum of a bound
state [29] and taking in the center-of-mass frame the equal-time limit for the two particles
allow, with certain mathematical assumptions, the grouping of terms into a form that leads
to a three-dimensional Breit–Salpeter-type wave equation [30,31], where the potentials are
represented by the components of the energy–momentum vector of the straight segment
joining the quark to the antiquark and carrying a constant linear energy density, equal to the
string tension; they involve, apart from the usual confining linear potential, contributions
coming from the moments of inertia of the above segment, which plays the role of the color
flux tube of the quarkonium system.
The self-energy parts of the quark propagators, extracted from the interaction terms,
allow the analysis of the chiral symmetry properties of the system. The situation here is
very similar to that resulting from the exchange of one Coulomb-gluon and the latter has
been studied in the literature. The Schwinger–Dyson equation satisfied by the self-energy
part has a nontrivial solution leading to a spontaneous breakdown of chiral symmetry.
From the nonrelativistic limit of the wave equation one determines the Hamiltonian
of the system to order 1/c2 . Spin–spin potentials are absent from the Hamiltonian. The
contributions of the color flux tube result in new terms for the orbital angular momentum
and for the spin–orbit potentials. In particular, the momentum of the flux tube contributes
with a negative sign to the spin–orbit potential, in opposite direction to the contributions of
the pure timelike vector potential, and may account for the phenomenological observations
made for fine splitting.
At high energies, the mass spectrum displays linear Regge trajectories, the slopes of
which tend to satisfy the classical relationship with the string tension.
The paper is organized as follows. In Section 2, we review the equations satisfied by
the Wilson loop averages and outline the particular status of minimal surfaces within the
set of possible solutions. In Section 3, the representation of the quark propagator in the
presence of an external gluon field is constructed. Section 4 is devoted to the construction
of the representation of the gauge invariant two-particle Green function in terms of a series
of Wilson loops having as boundaries skew polygons. The bound state equation is derived
in Section 5. Section 6 is devoted to the extraction from the interaction terms of the self-
energy parts needed for the quark propagators. Chiral symmetry breaking is studied in
Section 7. In Section 8, the main qualitative properties of the bound state spectrum are
displayed. Summary and concluding remarks follow in Section 9. In Appendix A, the main
properties of minimal surfaces are presented.
2. Loop equations and minimal surfaces
The starting point is the gauge covariant path-ordered phase factor along a line Cyx
joining point x to point y:3
y
U (Cyx , y, x) ≡ U (y, x) = P e−ig x dz Aµ (z) ,
µ
(2.1)
a a a
where Aµ = a Aµ t , Aµ (a = 1, . . . , Nc − 1) being the gluon fields and t the generators
2 a
of the gauge group SU(Nc ) in the fundamental representation, with the normalization
tr t a t b = (1/2)δ ab . Parametrizing the line C with a parameter σ , C = {x(σ )}, 0 σ 1,
such that x(0) = x and x(1) = y, a variation of C induces the following variation of U
(U (x(σ ), x(σ )) ≡ U (σ, σ ), A(x(σ )) ≡ A(σ )):
δU (1, 0) = −igδx α (1)Aα (1)U (1, 0) + igU (1, 0)Aα (0)δx α (0)
1
+ ig dσ U (1, σ )x β (σ )Fβα (σ )δx α (σ )U (σ, 0), (2.2)
0
x
where = ∂x/∂σ and F is the field strength, Fµν = ∂µ Aν − ∂ν Aµ + ig[Aµ , Aν ]. The
functional derivative of U with respect to x(σ ) (0 < σ < 1) is then [5]:
δU (1, 0)
= igU (1, σ )x β (σ )Fβα (σ )U (σ, 0). (2.3)
δx α (σ )
Defining the ordinary derivations with Polyakov’s prescription [5,9], one obtains:
∂ ∂ δU (1, 0)
= igU (1, σ ) ∇µ Fβα (σ ) U (σ, 0), (2.4)
∂x µ (σ ) ∂x β (σ ) δx α (σ )
where ∇ is the covariant derivative, (∇F ) = (∂F ) + ig[A, F ].
The Wilson loop, denoted Φ(C), is defined as the trace in color space of the path-
ordered phase factor (2.1) along a closed contour C and its vacuum expectation value is
denoted W (C):
1
tr P e−ig C dx Aµ (x),
µ
Φ(C) = W (C) = Φ(C) A , (2.5)
Nc
where the factor 1/Nc has been put for normalization and the averaging is defined in the
path integral formalism. The Wilson loop is a gauge invariant quantity. More generally, one
meets insertions of local operators O(x) into the Wilson loop. Their vacuum expectation
3 Loop equations will be written in Minkowski space.

values are:

1
tr P e−ig C dz Aµ (z) O(x)
µ
O(x) W ≡ . (2.6)
Nc A x∈C
The Wilson loop and its average also satisfy equations analogous to Eqs. (2.3) and
(2.4). Considering in Eq. (2.4) three different indices and taking their cyclic permutations
one obtains in the right-hand side the Bianchi identity. For the Wilson loop average the
equation takes the form
∂ ∂ δW (C)
ενµβα = 0. (2.7)
∂x ∂x β δx α
µ
Contraction of the indices µ and β in Eq. (2.4) leads in the right-hand side to the
equation of motion of the gluon field. In the large-Nc limit the quark current can be
neglected and the corresponding term becomes equivalent to the functional derivative
−δ/δAα which acts now on the Wilson loop. (Gauge-fixing and ghost terms mutually
cancel each other in gauge invariant quantities [13,14] and hence can be ignored.) For
contours having a self-intersection point at x, one finds:

∂ ∂ δW (C) g 2 Nc
= i dy α δ 4 (y − x)W (Cyx )W (Cxy ). (2.8)
∂x β ∂xβ δxα 2
C
(The product g 2 Nc is maintained fixed and finite in the above limit [15].) The contours
Cyx and Cxy are the complementary contours separated by the intersection point and
contributing as independent closed contours to the Wilson loop.
Eqs. (2.7) and (2.8) can be considered as basic equations of QCD in the large-Nc limit
in loop space. The factorized structure of the right-hand side of Eq. (2.8) puts severe
restrictions on the class of its possible solutions. For contours not having self-intersections,
Eq. (2.8) becomes:

∂ ∂ δW (C) g 2 Nc
= i dy α δ 4 (y − x)W (C). (2.9)
∂x β ∂xβ δxα 2
C
The points y that contribute to the integral are those that lie in the vicinity of the point x in
space and on the contour.
A first class of solution to Eqs. (2.7) and (2.8) is provided by perturbation theory [13,14].
In the right-hand side of Eq. (2.8), g 2 Nc represents the unrenormalized coupling constant
and the resolution of the equation is accompanied by the renormalization of the theory and
of the Wilson loop itself. However, perturbation theory becomes unstable at large distances
and the search for nonperturbative solutions becomes necessary. Among these, minimal
surfaces appear, for several reasons, as natural candidates. First, they easily reproduce
the area law related to confinement [1]; second, they satisfy, with certain restrictions, the
factorization law in the large-Nc limit; third, they are connected to the classical action
of string theory, which in turn is expected to have an implicit relationship with QCD
[4,25–27]. In the following, we concentrate on the contributions of surfaces to Eqs. (2.7)
and (2.9).
Fig. 1. Prototype of a closed contour with four distinct sides.
Let S be a surface of a given type, that is satisfying a given equation, and having as
contour the closed loop C. Let A be its area. Then a representation of the Wilson loop
average can be given by the following expression:
W (C) = e−iσ A(S,C), (2.10)
where σ is a constant, which will be identified with the string tension.
We shall generally consider as prototypes of contour those having four distinct sides,
with junction points designated by x1 , x1 , x2 , x2 (see Fig. 1).
The surface bounded by this contour will be parametrized with two parameters, σ and
τ , and a point belonging to it will be represented as x(σ, τ ) or y(σ, τ ). We adopt for the
partial derivatives the usual notations
∂x ∂x
x = , ẋ = . (2.11)
∂σ ∂τ
In general, the four sides of the contour will be parametrized as follows: {x1 x1 } with σ = 0,
{x1 x2 } with τ = 1, {x2 x2 } with σ = 1 and {x2 x1 } with τ = 0. The area of the surface is:4
1 1
1/2
A(C, S) = dσ dτ x 2 ẋ 2 − (x .ẋ)2 . (2.12)
0 0
To avoid the occurrence of possible divergences when dealing with functional
derivatives of the area, it is necessary to introduce a short-distance regulator in the
expression of the area. We adopt for the latter the following expression, already suggested
in a more general form in Ref. [7]:

1
A= dσ (x) dσ µν (y) F (x − y),
µν
2
a2 1
dσ µν (x) = dσ dτ x µ ẋ ν − x ν ẋ µ , F = , (2.13)
π ((x − y)2 + a 2 )2
where a is the (positive) regulator and goes to zero at the end of calculations. In the limit
a → 0, F is actually equal to a two-dimensional δ-function. This implies that it is only
4 Formulas related to surfaces will in general be written in Euclidean space.

points x and y lying close to each other that contribute to the integral and therefore a
limited number of terms of the expansion of y about x have to be considered in general.
We are interested in variations of the area of the surface S when local variations are
introduced on its contour C. S being a surface of a given type, that is satisfying a defining
equation, any deformation of the contour introduces corresponding deformations inside
the surface. Let us for definiteness consider deformations on the line τ = 0. The functional
derivative of A with respect to x(σ, 0) is:

δA ν µ

α
= δ µα x (σ, 0) − δ να x (σ, 0) dσ µν (y) F y(σ , τ ) − x(σ, 0)
δx (σ, 0)

δy λ (σ , τ ) ∂
+ dσ (y) µν
dσ µν (z) F (z − y), (2.14)
δx α (σ, 0) ∂y λ
P (λ,µ,ν)
where P (λ, µ, ν) indicates a cyclic permutation of the indices λ, µ, ν inside the sum. The
quantity δy λ (σ , τ )/δx α (σ, 0) depends on the type of surface that is considered.
In order to satisfy the Bianchi identity (2.7) with representation (2.10), the following
equation should hold:
∂ ∂ δA
= 0. (2.15)
∂x γ (σ, 0) ∂x β (σ, 0) δx α (σ, 0)
P (γ ,β,α)

This involves through the right-hand side of Eq. (2.14) the sum P (γ ,β,α)(∂/∂x γ ) ×
(∂/∂x β )(δy λ/δx α ). For arbitrary contours, that quantity does not identically vanish.
Therefore, the only solution to Eq. (2.15) is the constraint

∂
dσ µν (y) F (y − x) = 0, x inside S or on C, (2.16)
∂x λ
P (λ,µ,ν)
which is the defining equation of minimal surfaces. Thus, the Bianchi identity implies that
among surfaces of a given type, only minimal surfaces can be solutions of it. This result
considerably reduces the class of surface type solutions to the loop equations. In order to
incorporate other types of surface in representation (2.10) there remains the possibility of
considering contributions of an infinite sum of all possible surfaces, which might, through
mutual cancellations, satisfy Eq. (2.7). This latter possibility was considered in Ref. [7].
Such a sum can also be considered as representing fluctuations of surfaces around the
minimal surface and corresponding to stringlike contributions [25–27]. Henceforth, we
restrict ourselves to the study of contributions to Eq. (2.10) coming from minimal surfaces.
In order to check Eq. (2.9), we notice that for minimal surfaces the second term of the
right-hand side of Eq. (2.14) is null; one then finds:
∂ ∂ δA 1 x α
= + O(a). (2.17)
∂x β ∂x β δx α a (x 2 )1/2
On the other hand, regularization of the delta function of Eq. (2.9) according to the
prescription
1 2 1
δ 4 (x) → − ∂ , (2.18)
4π 2 x 2 + a 2
and comparison of both sides of Eq. (2.9) (in its Euclidean version) gives for finite σ the
identification
3 g 2 Nc
σ= , (2.19)
4πa 2 2
which necessitates a quadratically vanishing g 2 Nc with a when short-distance interactions
are ignored.
If the unrenormalized coupling constant g 2 Nc vanished as O(a 2) with vanishing a,
then minimal surfaces would define a theory by their own with a finite-dimensional
coupling constant σ . However, we know from the short-distance behavior of QCD, given
by the perturbative solution, that the unrenormalized coupling constant vanishes only
logarithmically. Therefore condition (2.19) cannot be satisfied in general. One must then
interpret the minimal surface contribution to the Wilson loop average as a part of a general
solution in which it represents its large-distance behavior, while the other part includes
the perturbation theory contribution representing its short-distance behavior. We shall
henceforth consider the properties of the large-distance behavior of the theory as deduced
from the minimal surface contribution to the Wilson loop average.
Let us also comment on the factorization property of minimal surfaces. The minimal
surface of two disconnected closed contours lying sufficiently far from each other is the
sum of the minimal surfaces of each contour. This ensures the factorization property of the
Wilson loop average valid at large Nc . However, when the two contours are close to each
other new global solutions may arise that do not lead to factorization. In such cases and
even for single contours having complicated forms, the minimal surface solution should be
chosen locally (therefore not necessarily the absolute one) with global properties being in
agreement with factorization or with other physical requirements.
Extension of that prescription to the case of contours with self-intersections allows min-
imal surfaces to also satisfy the factorization property of the right-hand side of Eq. (2.8).
However, when such contours lie on overlapping surfaces additional contributions might
be necessary to consider, since internal contours then belong to two adjacent surfaces. This
is in particular the case in two-dimensional QCD [12].
Some properties of minimal surfaces are presented in Appendix A. One main property
that we shall use in the present paper concerns the commutativity property of two
successive functional derivatives of the minimal area on its contour C,

δ δ δ δ
− A(C) = 0, (2.20)
δx β (σ ) δx α (σ ) δx α (σ ) δx β (σ )
a feature that is reminiscent of a similar property of the Wilson loop average.
3. The quark propagator in the external gluon field
In dealing with the quarkonium bound state problem within the Wilson loop approach,
one needs, in the path integral formalism, a representation of the quark propagator in the
presence of an arbitrary external gluon field, satisfying the equation

iγ .∂(x) − m − gγ .A(x) S(x, x ) = iδ 4 (x − x ). (3.1)
The usual perturbative representation S(x, x ) = i(iγ .∂ − m − gγ .A)−1 δ 4 (x − x ) is not

very convenient here since at each order of the perturbative expansion in the coupling
constant the gauge covariance of the propagator is lost and the construction of the Wilson
loop including contributions of fermion propagators becomes tricky. A representation that
is well suited to exhibit the Wilson loop structure of the gauge invariant two-particle
Green functions is the Feynman–Schwinger representation [33], which represents the
quark propagator as a quantum mechanical path integral. This representation has, however,
the main drawback that it dissolves the Dirac operator into the path integral and makes it
difficult to obtain from it an equation that easily displays the properties of fermions. In the
present work we shall consider a representation that is based on an explicit use of the free
Dirac propagator accompanied by a path-ordered phase factor. It will have the advantage
of manifestly preserving the main properties of fermions.
The building block of the representation is the gauge covariant composite object,
denoted S̃(x, x ), made of a free fermion propagator S0 (x − x ) (without color group
content) multiplied by the path-ordered phase factor U (x, x ) (Eq. (2.1)) taken along the
straight segment xx :
a a
S̃(x, x ) b ≡ S0 (x − x ) U (x, x ) b . (3.2)
The advantage of the straight segment over other types of line is that in the limit x → x U
tends to unity in an unambiguous way. S̃ satisfies the following equation with respect to x:

iγ .∂(x) − m − gγ .A(x) S̃(x, x )
1
δU (x, x )
= iδ (x − x ) + iγ
4 α
dλ λ S0 (x − x ), (3.3)
δx α (λ)
0
where the segment xx has been parametrized with the parameter λ as x(λ) = λx + (1 −
λ)x . In the above equation, the point x is held fixed; furthermore, the operator δ/δx α (λ)
does not act on the explicit boundary point x of the segment (corresponding to λ = 1, cf.
Eqs. (2.2), (2.3)), this contribution having already been cancelled by the gluon field term
A. A similar equation also holds with respect to x , with x held fixed, with the Dirac and
color group matrices acting from the right.
The quantity −i(iγ .∂(x) −m−gγ .A(x))δ 4(x −x ) is the inverse of the quark propagator
S in the presence of the external gluon field A. Reversing Eq. (3.3) with respect to S −1 ,
one obtains an equation for S in terms of S̃:
1
4 δ
S(x, x ) = S̃(x, x ) − d x S(x, x )γ α
dλ λ S̃(x , x ), (3.4)
δx α (λ)
0
where the operator δ/δx α (λ) acts on the factor U of S̃, along the internal part of the
segment x x , with x held fixed. Using the equation with x , or making in Eq. (3.4) an
integration by parts, one obtains another equivalent equation:
1 ←
4 δ
S(x, x ) = S̃(x, x ) + d x dλ (1 − λ)S̃(x, x ) γ α S(x , x ). (3.5)
δx α (λ)
0
Eqs. (3.4) or (3.5) allow us to obtain the propagator S as an iteration series with respect
to S̃, which contains the free fermion propagator, by maintaining at each order of the
iteration its gauge covariance property. For instance, the expansion of Eq. (3.4) takes the
form:
1
δ
S(x, x ) = S̃(x, x ) − 4
d y1 S̃(x, y1)γ α1
dλ1 λ1 S̃(y1 , x )
δx α1 (λ1 )
0
1
+ d 4 y1 d 4 y2 dλ1 λ1 dλ2 λ2 S̃(x, y1 )γ α1
0

δ δ
× S̃(y ,
1 2y ) γ α2
S̃(y 2 , x ) + ···, (3.6)
δx α1 (λ1 ) δx α2 (λ2 )
the operator δ/δx(λi ) acting on the inner part of the segment yi yi+1 of the phase factor
U (yi , yi+1 ). A verification of Eq. (3.1) can be done through the above iteration series. The
operator ∂/∂x acts on three terms. First, acting on the free fermion propagator contained
in S̃(x, x ), it gives, with the Dirac operator, a δ 4 (x − x ) type term which then removes
the integration over x . Second, it acts on the boundary point x of U (x, x ) and cancels
the gluon field term A. Third, it acts on the inner part of the segment xx of U . This term
is then cancelled by the δ 4 -term coming from the next-order term of the iteration series
under the action of the Dirac operator on the corresponding free propagator, and so forth.
Eqs. (3.4) and (3.5) are generalizations of the representation used for heavy quarks starting
from the static case [16].
The action of the operator δ/δx α (λ) on U can be expressed in terms of an insertion
of the field strength F (Eq. (2.3)). One can check with the first few terms of the series,
using integrations by parts, that one can recover, in perturbation theory with respect to the
coupling constant g, the conventional perturbative expansion of S in terms of g and A.
4. The two-particle gauge invariant Green function
The next step is to consider the quark–antiquark gauge invariant Green function, for
quarks q1 and q2 with different flavors and with masses m1 and m2 :

G(x1 , x2 ; x1 , x2 ) ≡ ψ̄2 (x2 )U (x2 , x1 )ψ1 (x1 )ψ̄1 (x1 )U (x1 , x2 )ψ2 (x2 ) A,q ,q , (4.1)
1 2
the averaging being defined in the path-integral formalism. Here, U (x2 , x1 ) is the phase
factor (2.1) along the straight segment x2 x1 (and similarly for U (x1 x2 )). According to the
conclusion reached in the final part of Appendix A, the dynamics of the system concerning
its energy spectrum not containing stringlike excitations can be probed by considering
between the quark and the antiquark equal-time straight segments (in a given reference
frame, e.g., the rest frame of the bound state); deviations of lines Cx2 x1 from the equal-
time straight segment contribute only to the wave functional of the bound state and not
to its energy. However, for covariance reasons, we first consider general straight segments
x2 x1 (not necessarily equal-time) and it is at a later stage that the equal-time limit will be
taken.
Integrating in the large-Nc limit with respect to the quark fields, one obtains:

G(x1 , x2 ; x1 , x2 ) = − trc U (x2 , x1 )S1 (x1 , x1 )U (x1 , x2 )S2 (x2 , x2 ) A , (4.2)
where S1 and S2 are the two quark propagators in the presence of the external gluon field
and trc designates the trace with respect to the color group.
The Green function G satisfies the following equation with respect to the Dirac operator
of particle 1 acting on x1 :
(iγ .∂(x1) − m1 )G(x1 , x2 ; x1 , x2 )

= −i trc U (x2 , x1 )δ 4 (x1 − x1 )U (x1 , x2 )S2 (x2 , x2 ) A
1
δU (x2 , x1 )
− iγ trc dσ (1 − σ )
α
S1 (x1 , x1 )U (x1 , x2 )S2 (x2 , x2 ) , (4.3)
δx α (σ )
0 A
where the segment x2 x1 has been parametrized with the parameter σ as x(σ ) = (1 −σ )x1 +
σ x2 ; furthermore, the operator δ/δx α does not act on the explicit boundary point x1 of
the segment (cf. Eqs. (2.2), (2.3)), this contribution having already been cancelled by the
contribution of the gluon field A coming from the quark propagator S1 . A similar equation
also holds with the Dirac operator of particle 2:
←
G(x1 , x2 ; x1 , x2 )(−iγ . ∂ (x2 ) − m2 )

= −i trc U (x2 , x1 )S1 (x1 , x1 )U (x1 , x2 )δ 4 (x2 − x2 ) A
1
δU (x2 , x1 )
+ i trc dσ σ S1 (x1 , x1 )U (x1 , x2 )S2 (x2 , x2 ) γ β . (4.4)
δx β (σ )
0 A
Using for S1 and S2 representations (3.4) and (3.5), respectively, one obtains for G
an expansion in a series of terms involving an increasing number of straight segments
between x1 and x1 on the one hand and between x2 and x2 on the other. With each segment
is associated a path-ordered phase factor U ; the union of all such factors, together with
U (x2 , x1 ) and U (x1 , x2 ), forms a Wilson loop along a skew polygon. We can then represent
G in the following form:
∞

G= Gi,j , (4.5)
i,j =1
where Gi,j represents the contribution of the term of the series having (i − 1) points of
integration between x1 and x1 (i segments) and (j − 1) points of integration between x2
and x2 (j segments). We designate by Ci,j the contour associated with the term Gi,j .
A typical configuration for the contour of G4,3 is represented in Fig. 2.
Using for the averages of the Wilson loops appearing in the above series the
representation with minimal surfaces, and designating by Ai,j the minimal area associated
Fig. 2. The contour associated with the term G4,3 .
with the contour Ci,j , one obtains for the latter the following representation:

Gi,j = (−1)i Nc d 4 y1 · · · d 4 yi−1 d 4 z1 · · · d 4 zj −1 S10 (x1 − y1 )
× γ α1 S10 (y1 − y2 ) · · · γ αi−1 S10 (yi−1 − x1 )S20 (x2 − zj −1 )

× γ βj−1 S20 (zj −1 − zj −2 ) · · · γ β1 S20 (z1 − x2 )
1 1
× dτ1 · · · dτi−1 dτ1 · · · dτj −1 (1 − τ1 ) · · · (1 − τi−1 )
0 0
× (1 − τ1 ) · · · (1 − τj −1 )
δ δ δ δ
× ··· ··· e−iσ Ai,j . (4.6)
δx α1 (τ 1) δx αi−1 (τ β
i−1 ) δx 1 (τ1 ) δx βj−1 (τj −1 )
Here, the operators δ/δx(τk ) and δ/δx(τ. ) act on the surface Ai,j through their action
on the segments yk yk+1 and z. z.+1 of the contour, respectively. (The parametrization of
the segments of the line x1 x1 is the opposite of that of Section 3: τ now increases along
the direction x1 x1 to be in accordance with that adopted for the surfaces; cf. comment
after Eq. (2.11).) S10 and S20 are the free quark propagators with masses m1 and m2 ,
respectively.
Using in Eqs. (4.3), (4.4) representations (3.4), (3.5) for the quark propagators, one
obtains:
(iγ .∂(x1) − m1 )G(x1 , x2 ; x1 , x2 )

∞ ∞ 1
δ
= −iδ 4
(x1 − x1 ) G0,j + iγ α
dσ (1 − σ ) G
i,j , (4.7)
α
δx (σ )
j =1 i,j =1 0 x(σ )∈x1 x2
←
G(x1 , x2 ; x1 , x2 )(−iγ . ∂ (x2 ) − m2 )
∞
∞ 1

δ
β
= +iδ 4
(x2 − x2 ) Gi,0 − i dσ σ β Gi,j γ , (4.8)
δx (σ ) x(σ )∈x1 x2
i=1 i,j =1 0
where G0,j and Gi,0 are particular cases of representation (4.6) in which the particle 1 or 2
propagators have been shrunk to a delta-function. These equations can also be obtained by
making the Dirac operators act on Eq. (4.5) and using in the right-hand side representation
(4.6).
Eqs. (4.3) and (4.4) are independent, since the two Dirac operators concern independent
particle variables. They are also compatible; this is evident from the very fact that G exists
and is given by formula (4.2). However, an independent check can also be done by making
the Dirac operator of particle 2 act on Eq. (4.3) and the Dirac operator of particle 1 act on
Eq. (4.4), using in the right-hand sides the properties of the propagators and of the phase
factors U and subtracting from each other the two resulting equations. The result is zero,
due mainly to the facts that the final expressions involve functional derivatives δ/δx(σ ) and
δ/δx(σ ) of the phase factor U (x2 , x1 ) and these commute. A similar verification can also
be done with Eqs. (4.7), (4.8), where now the representation of the Wilson loop averages
by minimal surfaces has been used. One has to use in the right-hand sides properties of the
terms Gi,j , cancellations between contributions of successive Gi,j s along the quark and
antiquark lines and the commutativity property of two successive functional derivatives of
a minimal area localized on its contour (Eq. (2.20) and Appendix A).
Eqs. (4.7), (4.8) are not closed integro-differential equations for G, for once the action
of the functional derivatives δ/δx on the various minimal surfaces has been evaluated one
does not obtain back G on the right-hand sides. This feature makes difficult the search
for a bound state equation in compact form. In this respect, if G has a bound state pole
in momentum space, then the right-hand sides of Eqs. (4.7), (4.8) should also have the
same pole; this is possible only if the actions of the functional derivatives δ/δx on the
partial ingredients Gi,j of G yield among other terms common factors that allow coherent
summations of the Gi,j s to produce again a pole term; otherwise, each Gi,j , containing a
finite number of free quark propagators, cannot produce alone such a pole. In x-space, the
selection of a bound state is made by taking a large time separation between the pairs of
points (x1 , x2 ) on the one hand and (x1 , x2 ) on the other [29].
The independence of Eqs. (4.7), (4.8) means also that they might allow the elimination
of the relative time variable of the two particles prior to any resolution of an eigenvalue
equation and the reduction of the internal dynamics to a three-dimensional space, a feature
which was outlined at the beginning of this section according to the results obtained at
the end of Appendix A. However, such a reduction does not seem easily manageable on
the general forms of Eqs. (4.7), (4.8). Furthermore, arbitrary approximations made in the
right-hand sides of those equations may destroy their compatibility property. This is why
in the following we shall directly study the equal-time limit of the system by considering
the “sum” of the two equations and then, at a later stage, shall indicate how to determine
its relative time evolution law by considering the “difference” of the two equations.
5. Bound state equation
In the large-Nc limit, the mesonic sector of QCD is composed of one-particle states,
which are bound states of a quark–antiquark pair and of gluons [15,34]. Taking in the Green
function (4.1) a large-time separation between the pairs of points (x1 , x2 ) and (x1 , x2 ) and
using the completeness relation one finds:

G(x1 , x2 ; x1 , x2 ) = Φn (x1 , x2 )Φ̄n (x2 , x1 ), (5.1)
n
where Φn is the wave functional of the bound state labelled with the collective quantum
numbers n:
iΦn,α1 ,α2 (x1 , x2 ) = 0|ψ̄2,α2 (x2 )U (x2 , x1 )ψ1,α1 (x1 )|n,
−i Φ̄n,α2 ,α1 (x2 , x1 ) = n|ψ̄1,α1 (x1 )U (x1 , x2 )ψ2,α2 (x2 )|0. (5.2)
Since the lines Cx2 x1 and Cx1 x2 are rigid straight segments completely determined by their
end points, one can consider the wave functionals as functions of the end point coordinates.
Introducing total and relative coordinates and momenta,
1 1
X = (x1 + x2 ), x = x2 − x1 , P = p1 + p2 , p = (p2 − p1 ),
2 2
∂
pa,µ = i µ , a = 1, 2, (5.3)
∂xa
and considering a bound state with total momentum P , one has:
Φ(x1 , x2 ) = e−iP .X φ(P , x). (5.4)
In the large separation time limit, the right-hand side of Eq. (5.1) displays a series
of oscillating functions in the separation time variable. By appropriate projections and
integrations one can select in this series the bound state that will survive in the large time
limit [29]. It should be emphasized that in the above limit only terms that factorize in G into
expressions depending on the line x1 x2 and expressions depending on the line x1 x2 could
survive to the selection operation of the bound state. Terms that still contain expressions
joining line x1 x2 to line x1 x2 would not contribute to the previous operation and could be
discarded.
It is convenient to consider the total momentum P of the selected bound state as
a reference timelike vector and define transverse and longitudinal parts of vectors with
respect to it:
q.P Pµ
qµT = qµ − 2 Pµ , qµL = (q.P̂ )P̂µ , P̂µ = √ , qL = q.P̂ ,
P P2

PL = P 2 , qL |c.m. = q0 , q T 2 c.m. = −q2 , −x T 2 c.m. = r. (5.5)
These decompositions are manifestly covariant. To further simplify the notation we adopt
for the Dirac matrices the following convention: they will be written on the left of the
spinor functions with labels 1 or 2 indicating on which particle indices they act, the particle
2 matrices (the antiquark at x2 ) acting actually from the right; thus:
γ1µ G ≡ (γµ )α1 β1 Gβ1 α2 ,α α , γ2µ G ≡ Gα1 β2 ,α α (γµ )β2 α2 ,
2 1 2 1
γ2µ γ2ν G ≡ Gα1 β2 ,α2 α1 (γν γµ )β2 α2 , γ2µ γ25 G ≡ Gα1 β2 ,α2 α1 (γ5 γµ )β2 α2 . (5.6)
Similar definitions also hold when G is replaced by Φ. Notice that products of γ2 matrices
act from the right in the reverse order.
We next introduce the free Dirac Hamiltonians of the two particles:
h10 = m1 γ1L − γ1Lγ1T .p1T , h20 = −m2 γ2L − γ2L γ2T .p2T . (5.7)
Going back to Eqs. (4.7), (4.8), multiplying the first by γ1L , the second by −γ2L , adding
and subtracting the two equations one finds:

(p1L + p2L ) − (h10 + h20 ) G
∞
∞

= −iδ 4 (x1 − x1 )γ1L G0,j − iδ 4 (x2 − x2 )γ2L Gi,0
j =1 i=1
1 1
δ β δ
+i γ1L γ1α dσ (1 − σ ) α + γ2L γ2 dσ σ β
δx (σ ) δx (σ )
0 0
∞

× Gi,j |x(σ )∈x1x2 , (5.8)
i,j =1

(p1L − p2L ) − (h10 − h20 ) G
∞
∞

= −iδ 4 (x1 − x1 )γ1L G0,j + iδ 4 (x2 − x2 )γ2L Gi,0
j =1 i=1
1 1
δ β δ
+i γ1L γ1α dσ (1 − σ ) α − γ2L γ2 dσ σ β
δx (σ ) δx (σ )
0 0
∞

× Gi,j |x(σ )∈x1x2 . (5.9)
i,j =1
Eq. (5.9) mainly determines the relative time evolution of the Green function, while
Eq. (5.8) mainly determines the dynamical properties of the two-particle system. Since
that equation does not involve the relative energy operator (p2L − p1L ) in its left-hand
side, one is entitled to take in it the equal-time limit xL = 0; the equation becomes in
that case three-dimensional with respect to the transverse relative coordinates x T . After
determining, within a given approximation, the dynamical properties of the system from
Eq. (5.8) in the limit xL = 0, one can go back to Eq. (5.9) and determine, within the same
approximation, its relative time evolution law. We shall henceforth follow this method of
approach.
In the large separation time limit between the pairs of points (x1 , x2 ) and (x1 , x2 )
the delta-functions that are present in the right-hand sides of Eqs. (5.8), (5.9) do not
contribute and can be ignored. In order to extract from the right-hand side of Eq. (5.8)
a bound state wave function Φ, it is necessary that the actions of the functional derivatives
δ/δx(σ ) on the various parts Gi,j yield among other terms common factors that factorize
the Green function G again. The functional derivative δ/δx(σ ) acts on a given Gi,j
(Eq. (4.6)) through the exponential factor containing the minimal area term and yields a
term proportional to δAi,j /δx(σ ). This term may itself be acted on by the other functional
derivatives existing in the definition of Gi,j . One thus ends up with two types of term. The
first contains the first-order derivative δAi,j /δx(σ ), which factorizes the other derivatives
on the left, and the second contains terms involving higher-order derivatives of Ai,j , one
of the derivatives acting along the straight segment x1 x2 . The dominant part in the large-
distance limit comes from the first type of term; higher-order derivatives of the minimal
area tend to weaken the large-distance behavior (cf. Appendix A for the second-order
derivative). In the following we shall mainly concentrate on the contribution of the first
type of term; terms containing second-order derivatives will be considered in Section 6, in
connection with the self-energy parts of the quark propagators.
The term that is retained with Gi,j is an integral over σ of a function proportional to
δAi,j /δx(σ ). According to Eq. (A.1), the latter depends only on the local properties of the
surface at the point x(σ, 0) lying on the straight segment x1 x2 , namely upon the derivatives
x (σ, 0) and ẋ(σ, 0). Because the line Cx2 x1 is a straight segment, one has x (σ, 0) = x,
independent of the form of the surface. In the equal-time limit (xL = 0) the previous
relation becomes x (σ, 0) = x T . The derivative ẋ(σ, 0) depends, however, on the form
of the minimal surface in the vicinity of the straight segment. Using methods of analysis
similar to those used at the end of Appendix A one can show that in the large separation
time limit ẋ(σ, 0) tends to a linear function of sigma. In that case, one can parametrize
it as ẋ(σ, 0) = (1 − σ )ẋ1 + σ ẋ2 , where ẋ1 and ẋ2 are the slopes at the points x1 and
x2 , respectively, ẋ1 = ẋ(0, 0), ẋ2 = ẋ(1, 0). The latter make still reference to the other end
points of the corresponding segments; for the case of the simplest contour C1,1 , these are x1
and x2 . In order to remove any explicit reference to the points of the remote past an operator
representation of the slopes, depending only on points x1 and x2 becomes necessary.
To find such a representation, we consider the simplest contour C1,1 corresponding to
G1,1 . Here, one has ẋ1 = (x1 − x1 ), ẋ2 = (x2 − x2 ), ẋaL < 0 with our parametrization
and because of the facts that xaL → +∞ and xaL → −∞ (a = 1, 2). Considering in

general the cases xL and xL finite in the above limits, one can set, modulo negligible terms,
x1L = x2L . (This is exact in the equal-time cases xL = 0 and xL = 0.) Furthermore, a close
examination of the integrals of the term δA1,1/δx(σ ) shows that the ẋa terms (a = 1, 2)
appear after integration in the forms ẋaα /|ẋaL| and through their orthogonal components to
x1 x2 . An operator representation of the latter can be found by making the quark momentum
operators paµ act on the term G1,1 . They generally yield three different contributions. The
first comes from their action on the corresponding free quark propagator, the second from
the segment x1 x2 and the third from the segment (xa − xa ). Since the terms ẋaα /|ẋaL|
already appear in expressions that are proportional to the string tension σ , one can use for
the latter terms, as a first approximation, free theory expressions; in this case, it is sufficient
to retain the contributions coming from the free quark propagators. The latter, in the large
time limit yield with massive quarks the following dominant behavior:
(xa − xa )µ
paµ Sa0 (xa − xa ) ma Sa0 (xa − xa ),
(xa − xa )2
(xa − xa )L → +∞, a = 1, 2, (5.10)
from which one deduces:

ẋaµ paµ
− ⇐⇒ , a = 1, 2, (5.11)
|ẋaL| |paL |
where the operators paL are the free theory expressions of the individual energies:

paL = ha0 , |paL | = m2a − paT 2 ≡ Ea paT , a = 1, 2, (5.12)
ha0 being defined in Eq. (5.7).
This approximation will retain all terms of order up to O(1/c2) in the nonrelativistic
limit; terms that have been neglected, if they are nonzero, have contributions in the
nonrelativistic limit starting at order O(1/c4).
We next generalize the above evaluation to the higher-order contours Ci,j appearing in
Gi,j . Here, we make the assumption that in the large separation time limit between the pairs
of points (x1 , x2 ) and (x1 , x2 ) the derivatives δAi,j /δx(σ ) can be expanded around the
driving term δA1,1/δx(σ ) of the lowest order surface. Neglecting the higher-order terms
of these expansions, one ends up with the common operators δA1,1 /δx(σ ) to all factors
Gi,j , involving the segment x1 x2 and the momentum operators paµ , a = 1, 2. Those terms
can then be factorized in front of G and interpreted as potentials.
The bound state equation obtained from Eq. (5.8) in the equal-time limit xL = 0 is then:
µ µ
PL − (h10 + h20 ) − γ1L γ1 A1µ − γ2L γ2 A2µ ψ PL , x T = 0, (5.13)
where ψ is the wave function φ in the equal-time limit,

ψ PL , x T ≡ φ PL , xL = 0, x T , (5.14)
and the potentials are defined through the equations (in Minkowski space)
1 1
δA1,1
δA1,1
A1µ = σ dσ (1 − σ ) µ , A2µ = σ dσ σ . (5.15)
δx (σ ) δx µ (σ )
0 0
They can be calculated either by using the Minkowskian version of Eq. (A.1) and then
conditions (5.11), (5.12), or by first writing Eqs. (5.8) and (5.13) in Euclidean space, using
Eq. (A.1) and then passing to Minkowski space. (In Euclidean space the right-hand sides
of Eqs. (5.15) contain an additional (−i) factor.) Since δA1,1/δx(σ ) is orthogonal to x
(Eq. (A.2)), the resulting vectors will satisfy this property. We define transverse vectors
with respect to x with a superscript “t”:
1
qµt = qµ − xµ x.q. (5.16)
x2
However, x itself is orthogonal to the total momentum P in the equal-time limit (xL = 0)
and reduces to x T . The part of the three-dimensional relative momentum pT that is also
orthogonal to x T will be denoted pT t :
1
pµT t = pµT − xµT x T .pT . (5.17)
xT 2
This vector enters in the definition of the relative orbital angular momentum. Defining
the corresponding Pauli–Lubanski vector WL (L referring here to the orbital angular
momentum) as
WLµ = 8µναβ P ν x α pβ , 80123 = +1, (5.18)
one has the relations

WL2 = −P 2 x T 2 pT 2 − x T α x T β pαT pβT − 2ix T .pT = −P 2 x T 2 pT t 2 ,

W 2
− L = L2 . (5.19)
P 2 c.m.
The expression of A1µ is:
E1 E2
A1µ = −σ −x T 2
E1 + E2
T2 2
E1 E2 x P
× gµL 8(p2L ) − 8(p1L ) E1 E2 − 1
(E1 + E2 )2 2WL2

E1 x T 2 P 2 E1 E2 T t
− gµL 8(p1L ) + p µ
E1 + E2 2WL2 E1 + E2

−x PT 2 2 1 −WL 2
1 −WL2
× arcsin + arcsin
−WL2 E2 −x T 2 P 2 E1 −x T 2 P 2

(E2 − E1 ) 1 Tt
− gµL 8(p1L ) + gµL 8(p2L ) − 8(p1L ) − p
(E2 + E1 ) E2 µ
T 2 2
E1 E2 −x P −WL2 −WL2
× 1 − − 1 −
E1 + E2 −WL2 −x T 2 P 2 E22 −x T 2 P 2 E12
T 2 2
1 E1 E2 −x P
− gµL 8(p2L ) − 8(p1L ) + pµT t
2 E1 + E2 −WL2

E1 −WL2 E2 −WL2
× 1− + 1 − .
E1 + E2 −x T 2 P 2 E22 E1 + E2 −x T 2 P 2 E12
(5.20)
Here, 8(p1L ) and 8(p2L ) are the energy sign operators of the free quark and the antiquark,
respectively:
ha0
8(paL ) = , a = 1, 2, (5.21)
Ea
ha0 and Ea being defined in Eqs. (5.7) and (5.12). The expression of A2µ is obtained from
that of A1µ by an interchange in the latter of the indices 1 and 2 and a change of sign of
pT t . In particular, the longitudinal parts of the potentials A1 and A2 add up in Eq. (5.13).
One has for their sum the expression:

E1 E2
A1L + A2L = σ −x T 2
E1 + E2

E1 E2
× 8(p1L ) + 8(p2L )
E1 + E2 E1 + E2

−x T 2 P 2 1 −WL2
× arcsin
−WL2 E2 −x T 2 P 2

1 −WL2
+ arcsin
E1 −x T 2 P 2
T 2 2
E1 E2 −x P
+ 8(p1L ) − 8(p2L )
E1 + E2 −WL2

−WL2 −WL2
× 1− − 1 − . (5.22)
−x T 2 P 2 E22 −x T 2 P 2 E12
For sectors of quantum numbers where WL2 = 0, the expressions of the potentials
become:
1
A1L + A2L = 8(p1L ) + 8(p2L ) σ −x T 2 , (5.23)
2

1 1 1
AT1µt = − (E1 + E2 ) − E1 pµT t σ −x T 2 ,
E1 E2 3 2

1 1 1
AT2µt = + (E1 + E2 ) − E2 pµT t σ −x T 2 . (5.24)
E1 E2 3 2
All expressions of the potentials have been written as classical functions of their arguments,
without taking into account ordering problems. These necessitate a detailed study which
will not be done here. We simply outline some general features that may be useful for the
resolution of the wave equation. (i) Many ordering problems that concern linear momentum
operators do not affect the energy eigenvalues and rather concern the definition of the
kernel of the scalar product of the wave functions; one can pass from one definition to the
other by appropriate changes of function. (ii) The square-root and arcsin functions which
involve the variables x T 2 and 1/Ea2 (a = 1, 2) could be treated in first approximation by
replacing in Ea the radial momentum operator squared by its mean value in the bound state,
or, if the resolution is done in momentum space by replacing x T 2 by its mean value. (iii)
A close study of the chiral properties of the wave equation suggests us to further adopt the
following rules: the doubly transverse momentum operator pµT t , maintaining its definition
of Eq. (5.17), and the energy sign operators 8(paL ) (a = 1, 2) should be placed on the
utmost left.
Eq. (5.13), together with expressions (5.20), is very similar to an equation proposed by
Olsson et al. on the basis of a model where quarks are attached at the ends of a straight
string or a color flux tube [35,36]; the difference mainly concerns the energy sign operators;
in Ref. [35] the equation which was proposed is the Salpeter equation [31], in which
generally the potential is proportional to a global energy projecter; here, the energy sign
operators, though they could be expressed through energy projectors, do not match exactly
the projector of the Salpeter equation; the doubly transverse parts of the vector potentials
do even not have energy sign operators. Apart from this slight difference, however, the
physical significance of Eq. (5.13) is the same as that of Ref. [35]. The vector potentials
Aaµ (a = 1, 2) can be interpreted as representing contributions of the energy–momentum
vector of the straight color flux tube with variable length r; its energy is represented by the
sum A1L + A2L (Eq. (5.22)), while its angular momentum contributes through the doubly
transverse components ATa t .
The norm of the wave function ψ can be obtained (after a few approximations) from
Eq. (5.8) using standard techniques. The result is:

† 1 h1 h2
3 T
d x tr ψ + ψ = 2PL Nc , (5.25)
2 E1 E2
where ha and Ea , a = 1, 2, are the free Dirac Hamiltonian and energy of each particle,
including now the self-energy contributions (cf. Section 6 and Eqs. (6.8)–(6.10)). That
formula is the same as that obtained from the Salpeter equation [31].
Finally, let us mention that Eq. (5.9) can be used to determine the relative time evolution
of the wave function Φ (Eq. (5.2)). Using in the right-hand side of Eq. (5.9) the same
types of approximation as in Eq. (5.8), taking the large separation time limit between the
pairs of points (x1 , x2 ) and (x1 , x2 ) and passing to the bound state wave function, one
can integrate Eq. (5.9) for the latter obtaining the relative time dependence in the form of
an ordered exponential function involving the various operators and potentials appearing in
the equation, the initial value condition being given by the function ψ(PL , x T ) (Eq. (5.14)).
We shall not need, however, that expression in the present work.
6. Quark self-energy
We show in this section that Eq. (5.8), which is at the origin of the bound state equation,
also yields the quark self-energies.
When the interaction kernel of a bound state equation is represented by the mediation of
an effective propagator, one expects to obtain the self-energy contribution by contracting
the lines of the outgoing particles (assumed to be of the same type) through a single particle
propagator and forming a loop with the kernel. In the present case, the interaction part
of the bound state equation (5.13) can be visualized by multiplying it back by the factor
γ1L γ2L . It has three different tensor parts: the first corresponds to vertices with the matrices
γ1L γ2L ; the second and third to vertices with matrices γ1T t .p1T t γ2L and γ1L γ2T t .p2T t ,
respectively. When these terms are incorporated in a two-point loop and integrated in
momentum space, the noninvariant pieces under spatial rotations disappear and what
remains is simply the part of the interaction with the γ1Lγ2L matrices, inside which also
the angular momentum operator has disappeared (cf. Eq. (5.23)). This corresponds to the
situation where the interaction kernel is generated by the mediation of a Coulomb-gluon
Fig. 3. (a) A typical contour C2,1 associated with the term G2,1 of the two-particle propagator. (b) A configuration
where the segments x1 x2 and y1 x1 intersect.
√
propagator, proportional in x-space to δ(xL ) −x T 2 . That result also coincides in form
with that obtained in two-dimensional QCD.
Since the formal expressions of the self-energies are the same in both two and four
dimensions in x-space, we shall directly work in two dimensions, by freezing the transverse
variables with respect to x and considering the simplified case of a plane. A self-energy
contribution for particle 1 will be recognized as depending only on the variables of that
particle, not making reference to the variables of particle 2, except for the directions of the
total momentum P of the bound state, which is chosen as the reference timelike direction,
and of the relative coordinate (x2 − x1 ). The simplest contribution to the self-energy of
particle 1 comes from the term G2,1 in Eq. (5.8), a typical contour of which is shown in
Fig. 3(a).
The term that is relevant here is that in which the operator δ/δx(σ ) (multiplying the γ1
matrices) acts on the term δA2,1/δx(τ ) of G2,1 (Eq. (4.6)). One thus obtains the second-
order functional derivative δ 2 A2,1 /δx µ (σ )δx α (τ ), where τ parametrizes the segment y1 x1
and σ the segment x1 x2 . (We recall that A2,1 is the area of the minimal surface bounded by
the skew polygonal contour C2,1 . More generally, Ai,j is the area of the minimal surface
bounded by the skew polygonal contour Ci,j appearing in the decomposition (4.5), (4.6).)
The expression of the second-order functional derivative of the minimal area has been
given in Eq. (A.3), where now x(σ , 0) and x(σ, 0) have to be replaced by x(σ ) and x(τ ),
respectively. Since x(σ ) and x(τ ) belong to different segments, the terms proportional to
the explicit delta-functions can be dropped; furthermore, since the orthogonal variables to
the surface have been frozen, also the last term of that equation can be ignored; it is only
the second term of the right-hand side of the equation that may contribute:
δ δA2,1
= F x(σ ) − x(τ ) δλα (x2 − x1 )ν − δνα (x2 − x1 )λ
δx µ (σ ) α
δx (τ )

× δλµ (x1 − y1 )ν − δνµ (x1 − y1 )λ , (6.1)
where x(τ ) and x(σ ) are parametrized as x(τ ) = y1 + τ (x1 − y1 )
and x(σ ) = x1 + σ (x2 −
x1 ). In the limit when the regulator a vanishes, the function F , Eq. (2.13), tends to a two-
dimensional delta-function; therefore, the above second-order derivative is nonvanishing
only when the two segments are intersecting (Fig. 3(b)); we then have to integrate with
respect to σ and τ with the weight factors (1 − σ ) and (1 − τ ) (Eqs. (4.6) and (5.8)). The
Fig. 4. Configuration of the contour C2,1 corresponding to the self-energy contribution.
calculation can be done by replacing F by a two-dimensional delta-function involving σ

and τ and the arguments of which can be obtained by first writing the explicit expression
of (x(σ ) − x(τ ))2 with two-dimensional components x 0 and x 3 (in the bound state rest
frame) for instance:

F x(σ ) − x(τ ) = δ x 0 (σ ) − x 0 (τ ) δ x 3 (σ ) − x 3 (τ )

= δ z0 + σ x 0 − τy 0 δ z3 + σ x 3 − τy 3
1 ∂ ∂ 3
= − 3 3 δ z0 + σ x 0 − τy 0 z + σ x 3 − τy 3, (6.2)
2x y ∂σ ∂τ
where we have defined x = x2 − x1 , z = x1 − y1 , y = x1 − y1 . Actually not the full
expression of the above integral is needed, but rather that part which depends only on
the points x1 and y1 , which correspond to the integration endpoints σ = 0 and τ = 0.
Integrating that expression by parts and retaining only the latter endpoints, one obtains:
1
δ δA2,1
dσ dτ (1 − σ )(1 − τ )
δx µ (σ ) δx α (τ )
0

= −δµ0 δα0 δ z0 z3

2
= −δµL δαL δ(x1L − y1L ) − x1T − y1T , (6.3)
where we have dropped additive terms not depending only on z and have done the tensor
calculation in two dimensions and restored at the end the covariant expression.
The presence of the delta-function along the temporal direction means that the segments
x1 x2 and y1 x1 are parallel, or more generally lie in an orthogonal plane to the time direction
(Fig. 4).
In the limit (x1L − x1L ) → ∞, the areas A (x x x y x x ) and A (x x y x x )
2,1 2 2 1 1 1 2 1,1 2 2 1 1 2
become almost equal and one may replace in G2,1 the former by the latter. The expression
(6.3), combined with the multiplicative free quark propagator S10 (x1 − y1 ), then factorizes
G1,1 and plays the role of a self-energy correction. In order to complete the derivation,
one must repeat the same calculations with the higher-order terms Gi,j (i 3, j 1),
where in the second-order derivatives of the type of (6.1), the functional derivative δ/δx(τ )
corresponds to δ/δx(τ1 ) and acts on the second segment y1 y2 of the corresponding contour
(Fig. 2). One thus finds the first-order self-energy correction (in the string tension σ ) that
factorizes the bound state wave function in Eq. (5.13); it corresponds to the effective
Fig. 5. (a) A configuration of the contour C4,1 corresponding to the second-order self-energy correction.
(b) A configuration corresponding to overlapping integrals.
exchange of a Coulomb-gluon, with propagator proportional in momentum space to

δµL δαL 1/((−q T )2 )2 .
The higher-order self-energy corrections are obtained by extending the above procedure
to terms that contain higher numbers of second-order derivatives of the areas Ai,j . The
next correction is provided by the terms Gi,j with i 4 and containing two second-order
derivatives of the area. Let us consider for definiteness the term G4,1 corresponding to the
contour C4,1 (Fig. 5(a)).
The self-energy contribution comes from the term containing the product (δ 2 A4,1 )/
(δx(σ )δx(τ3 ))(δ 2 A4,1 )/(δx(τ1 )δx(τ2 )), where δ/δx(τ1 ) acts on the segment y1 y2 , δ/δx(τ2 )
on y2 y3 and δ/δx(τ3 ) on y3 x1 . The resulting delta-functions imply intersection of the seg-
ments x1 x2 and y3 x1 and parallelism of the adjacent segments y1 y2 and y2 y3 . The two
delta-functions can be integratedas before in an independentway. The result is propor-
tional to the product δ(x1L − y3L ) −(x1T − y3T )2 δ(y1L − y2L) −(y1T − y2T )2 . Notice that
in the case of overlapping integrals, corresponding to the situation where two derivatives
act on the segments x1 x2 and y2 y3 on the one hand and on the segments y1 y2 and y3 x1
on the other, the delta-functions would imply intersections of the segments of each of the
above pairs (Fig. 5(b)); integrating the delta-function of the second pair as before yields
the delta-function δ(y2L − y3L ), the implementation of which prevents the segments x1 x2
and y2 y3 from intersecting and the corresponding integral vanishes. Also notice that in the
two-dimensional limit that we are considering here (freezing of orthogonal deformations
of the areas) higher-order derivatives of the areas vanish, for they would involve derivatives
of the function F outside its support.
The generalization to higher orders is now straightforward. Any high-order contribution
will contain in the term Gi,j a product of second-order derivatives of the area Ai,j , in
which one of them is (δ 2 Ai,j )/(δx(σ )δx(τk )) (k i − 1), where δ/δx(τk ) acts on the
segment yk yk+1 ; the others are nonoverlapping (disjoint or nested) and act on segments
lying between x1 and yk+1 , being thus nested within x1 yk+1 . The remaining part of the
calculation repeats the steps used for the first-order correction.
One thus generates the whole series of self-energy corrections that can be summed into
the Schwinger–Dyson integral equation, corresponding to a Coulomb instantaneous kernel
−σ δµL δνL δ(xL − yL ) −(x T − y T )2 .
√
Defining the momentum space Fourier transform of the function −x T 2 through
analytic continuation of the power parameter [37], one finds:

T T 8π
d 3 x T eip .x −x T 2 = − . (6.4)
(−pT 2 )2
Denoting by S(pL , pT , m) the quark propagator in momentum space with free mass m tied
to the bound state of momentum P and defining the self-energy contribution through the
equation

iS −1 pL , pT , m = γ .p − m − Σ pT , m , (6.5)
one has the Schwinger–Dyson equation:

d 4k 1
Σ pT , m = 8πσ γL S kL , k T , m γL . (6.6)
(2π) 4 (−(p − k T )2 )2
T
Because of the instantaneity of the kernel, the self-energy Σ actually depends only on
the three-dimensional transverse momentum pT . The tensor decomposition of Σ can be
written in the form

Σ pT , m = γ T .pT A −pT 2 , m + B −pT 2 , m . (6.7)
The self-energies of the quark and of the antiquark have now to be incorporated in the
Dirac energy operators in Eq. (5.13) and in the definitions of the energy sign operators
(5.21). Defining

h1 p1T , m1 = h10 + γ1L Σ p1T , m1 = γ1L (m1 + B1 ) − γ1T .p1T (1 − A1 ) ,

h2 p2T , m2 = h20 − γ2L Σ −p2T , m2

= −γ2L (m2 + B2 ) + γ2T .p2T (1 − A2 ) , (6.8)
T2
Ea −pa , ma = h2a = (ma + Ba )2 − paT 2 (1 − Aa )2 , a = 1, 2, (6.9)
ha
8(paL ) = , a = 1, 2, (6.10)
Ea
Eq. (5.13) becomes
µ µ
PL − (h1 + h2 ) − γ1L γ1 A1µ − γ2L γ2 A2µ ψ PL , x T = 0, (6.11)
where the potentials A are given by expressions (5.20) in which Ea and 8a (a =
1, 2) are replaced by expressions (6.9) and (6.10), respectively; furthermore, the self-
energy functions should also be incorporated in the appearances of the doubly transverse
momentum pT t and the orbital angular momentum WL . Since we have replaced p1T t
and p2T t in favor of pT t , the corresponding substitutions may not seem straightforward.
However, the difficulty is circumvented easily. Each energy factor Ea (a = 1, 2) that
appears explicitly in the potentials is reminiscent of a term of the type paT t /Ea ; paT t
undergoes the substitution paT t → (1 − Aa )paT t , where Aa (a = 1, 2) is the self-energy
function, Eq. (6.8). It is then sufficient to replace each Ea in the potentials by Ea /(1 − Aa ),
without modifying the momentum and orbital angular momentum operators.
The quark propagator is not a gauge invariant quantity and therefore could not lead alone
to observable effects. This feature manifests itself through the self-energy equation (6.6),
which, due to the infrared singularity of the integrand, provides an infrared divergent self-
energy. On the other hand, physical observables should be free of infrared singularities.
The treatment of infrared singularities in momentum space depends on the method of
regularization that is adopted. Dimensional regularization, which gives zero for the above-
mentioned singularities, has the advantage of preserving symmetry properties, but the
disadvantage of hiding the distinction between unobservable and observable quantities.
This is why it is necessary to check at every stage of calculations whether physical
quantities are indeed free of infrared singularities by also using an explicit infrared cutoff
method.
In the present approach, we have to check that the wave equation (6.11), which describes
the physical properties of the bound states, is free of infrared singularities. To do this, we
have to isolate the singularities coming from the self-energies. The properties of the latter
will be studied in Section 7; here, we mention the most relevant ones for our purpose.
Designating by V the linear confining potential, Eq. (7.4), it is found that the singularities
of the self-energy parts are contained in the energy factors Ea (Eq. (6.9)), while the
ratios (1 − Aa )/(ma + Ba ), a = 1, 2, are free of singularities; from the integral equations
satisfied by Ea , Eq. (7.6), it is seen that their singularity is represented by the three-
dimensional integral in momentum space of −V /2. (Section 7 deals with the equal-mass
case and mainly with the massless limit; however, the infrared singularity properties are
not affected by the values of the quark masses and could be abstracted from the equations
of that section.) Therefore, the combinations (Ea + V /2) are free of infrared singularities.
Coming back to the potentials Aaµ (Eq. (5.20)) their large-distance behavior in x-space
can be studied by expanding the various functions contained in their expressions in terms
of (1 − Aa )2 L2 /(Ea2 x2 ), a = 1, 2 (in the c.m. frame). Factorizing the function σ r, the
expansions are of the type σ r(1 + O((1 − Aa )2 L2 /(Ea2 x2 ))); therefore the higher-order
terms are nondominant at large distances and could not lead to infrared singularities. For
the present study it is sufficient to keep the leading terms, which are represented by the
expressions (5.23), (5.24).
Considering first the contribution of (A1L + A2L ) and replacing the energy sign
operators by their expressions (6.10), we immediately find that the energy factors and the
potential appear with the combinations (Ea + V /2), which ensure the infrared finiteness of
the result. Next, we consider the contributions of the spacelike potentials ATaµt (Eq. (5.24),
where each Ea should be replaced by Ea /(1 − Aa )). The doubly transverse momentum
operator pT t is equal in the c.m. frame to pt = −(1/x2)x × L. The effective potential that
matters is proportional to σ rpt = −(σ/r)x × L. The orbital angular momentum operator L
acts on spherical harmonics and does not modify the infrared properties of wave functions.
The term x/r has as Fourier transform a function proportional to p/p4 , which does not
lead to infrared singularities, after the angular integrations in convolution integrals are
done. Therefore, the wave equation (6.11) is free of infrared singularities. Explicit cases
of the above cancellations can be found in the equations presented in Section 7. A similar
check can also be found in two-dimensional QCD [38].
7. Chiral symmetry breaking
The presence of the self-energy contributions in the bound state equation (6.11) allows
us to study the possibility of the spontaneous breakdown of chiral symmetry. This is
intimately related to the existence of nonperturbative solutions to the Schwinger–Dyson
equation in the chiral limit when the quark mass m tends to zero [39,40].
Using decomposition (6.7) and considering the bound state rest frame, the integral in
Eq. (6.6) can first be integrated with respect to the energy variable k0 giving rise to two
coupled equations in B and A. It is however more convenient to decompose the functions
B and p(1 − A) along polar combinations (here for m = 0):

B(p) = E(p) sin ϕ(p), p 1 − A(p) = E(p) cos ϕ(p). (7.1)
The self-energy equations then become (in the chiral limit):

1 d 3k
E(p) sin ϕ(p) = − V (p − k) sin ϕ(k), (7.2)
2 (2π)3

1 d 3k
E(p) cos ϕ(p) = p − V (p − k)p̂.k̂ cos ϕ(k),
2 (2π)3
p k
p̂ = , k̂ = , (7.3)
p k
where we have defined the potential V as
1
V (p) = −8πσ , p = |p|, V (x) = σ r, r = |x|. (7.4)
p4
The latter equations in turn can be recombined to decouple the function ϕ from E:

1 d 3k
p sin ϕ(p) = − 3
V (p − k) sin ϕ(k) cos ϕ(p) − p̂.k̂ cos ϕ(k) sin ϕ(p) ,
2 (2π)
(7.5)

1 d 3k
E(p) = p cos ϕ(p) − V (p − k)
2 (2π)3

× sin ϕ(k) sin ϕ(p) + p̂.k̂ cos ϕ(k) cos ϕ(p) . (7.6)
These equations were extensively studied in the literature; they result from the
assumption that confinement is due to the exchange of Coulomb-gluons [41–45]. Using
variational methods, it has been shown that the perturbative vacuum state is unstable under
quark–antiquark pair creation. The existence of a new stable vacuum state is ensured by
the existence of a nontrivial solution to Eqs. (7.2), (7.3).
From Eq. (7.5) one deduces that the function ϕ is an infrared finite quantity.
From Eq. (7.6), after expanding in the integrand k-dependent terms around p, one
deduces that E is infrared singular and its singularity is represented by the integral
(−1/2) d 3 k V (k)/(2π)3 , a property mentioned and utilized at the end of Section 6 for
the checking of the infrared finiteness of the wave equation (6.11). (We use, however,
dimensional regularization for analytic calculations throughout this section.)
The function sin ϕ(p) will be identified later with the Goldstone boson wave function.
Therefore, it should be a normalizable and presumably nodeless function. An analysis of
the above equations shows that the function ϕ behaves at infinity as p−5 and tends at
the origin to π/2. The solution that is found is indeed a monotonically decreasing function
with the above properties. On the other hand, due to the infrared singularity of the potential
V , the energy function E vanishes at some finite value p0 of p and becomes negative for
p < p0 ; it tends to a finite negative value at the origin and behaves as p at infinity. The
functions B and p(1 − A) vanish simultaneously at p0 ; they are negative for p < p0 , with
the limiting values B(0) = E(0) and limp=0 p(1 − A) = 0; asymptotically, B behaves as
p−4 and A as p−2 .
An order parameter for chiral symmetry breaking is the quark condensate ψ̄ψ. For a
given type of quark, say u, it can be defined as minus the trace (in color and Dirac spinor
spaces) of the quark propagator at the origin in x-space:

d 4p
ūu = − tr c,sp S(x)|x=0 = −Nc tr sp S(p). (7.7)
(2π)4
Using definition (6.5) and decomposition (6.7), one finds:

d 3p
ūu = −2Nc sin ϕ(p). (7.8)
(2π)3
We next turn to the bound state equation (6.11). A consistent study of the chiral limit
should be done starting from the situation where the quark masses are different from zero
and then taking the limits m1 = m2 = 0. In the present case, we directly consider the equal
mass case m1 = m2 = m, corresponding to the light quark sectors u and d without isospin
breaking.
In general, the resolution of Eq. (6.11) proceeds by first decomposing the wave function
ψ along a basis of 2 × 2 matrices spanned by the matrices γL and γ5 . In the following, we
consider the rest frame of the bound state (P = 0). One has the decomposition
ψ = ψ1 + γ0 ψ2 + γ5 ψ3 + γ0 γ5 ψ4 , (7.9)
where the functions ψa , a = 1, . . . , 4, are themeselves 2 × 2 matrices on which act
the spin Pauli matrices σ . The quantum numbers of the state are usually defined with
respect to the components that survive in the nonrelativistic limit; these are ψ3 and ψ4 ,
which have the same quantum numbers. For the present problem, we are considering the
sector characterized by the following quantum numbers: total spin s = 0, orbital angular
momentum . = 0, total angular momentum j = 0. We adopt the rule of writing the energy
sign operators present in the potentials (Eqs. (5.20) and (6.9), (6.10)) as well as the factors
((1 − A)/E)pt on the utmost left. The doubly transverse momentum operator pt (Eq.
(5.17)) annihilates S-states; it commutes with x2 and with all scalar operators that act
on S-states in x-space. In the equal-mass case, the transverse potentials At1 and At2 are
opposite to each other; defining
(1 − A) t
At2 = p Ã = −At1 , (7.10)
E
using definition (7.4), introducing the spin operators s1 and s2 of particles 1 and 2, and
removing indices 1 and 2 from various factors which are equal, we obtain the following
four equations:

1
P0 ψ1 + 2(1 − A)(s1 − s2 ).p 1 + V ψ3 = 0, (7.11)
2E
P0 ψ2 = 0, (7.12)

1 1
P0 ψ3 − 2(m + B) 1 + V ψ4 + 2(1 − A)(s1 − s2 ).p 1 + V ψ1
2E 2E
(1 − A)
−2 (s1 − s2 ).pt Ãψ1 = 0, (7.13)
E

1
P0 ψ4 − 2(m + B) 1 + V ψ3 = 0. (7.14)
2E
We have neglected in Eq. (7.13) the internal L2 dependent parts, present in the potentials
(A1L + A2L ) and At acting on ψ1 , which is a P -state, and used the limits (5.23), (5.24);
these neglected parts do not seem to play a crucial role in the following calculations; with
the decomposition (7.10), Ã = σ r/6.
It is found that the ground state solution of the above equations in the limit m = 0
(reaching this limit smoothly) with zero energy is:
ψ1 = ψ2 = ψ4 = 0, ψ3 = 0, (7.15)
with ψ3 satisfying the equation

1 d 3k
E(p)ψ3 (p) = − V (p − k)ψ3 (k). (7.16)
2 (2π)3
This equation is similar in form to Eq. (7.2), indicating that ψ3 is proportional to the self-
energy function sin ϕ:
ψ3 (p) = C sin ϕ(p), (7.17)

with C a constant.
In order to calculate the pion decay constant Fπ , one must again consider the case where
m = 0. The pion decay constant is defined as:
√
0|d̄(0)γµ γ5 u(0)π + (P ) = i 2 Fπ Pµ . (7.18)
This is related to the component ψ4 :
√
d 3k 2
ψ 4 (k) = Fπ P0 . (7.19)
(2π)3 4
The normalization condition (5.25) yields the relation
Nc
CFπ = √ . (7.20)
2
This result is also a consequence of the Ward–Takahashi identity relative to the Green
function
√ T ū(x)γµ γ5 d(x)i d̄(0)γ5 u(0), which implies (in the chiral limit) the equation
2 Fπ 0|i d̄(0)γ5 u(0)|π + = 2ūu.
The normalization condition can also be analyzed by using Eq. (7.14). Integrating that
equation with respect to p and expanding in the second term all quantities with respect
to m to first-order in m and then using the integral equations satisfied by the first-order
self-energy functions, one ends up with the well-known relation of Gell-Mann, Oakes and
Renner [46]: m2π Fπ2 = −2mūu.
In order to be able to calculate Fπ , one needs to know the function ψ4 . This information
comes from Eq. (7.13). Eliminating ψ1 and noticing that ψ4 iself is proportional to P0 , one
can take in all factors multiplying ψ4 the chiral limit. Defining
1
ψ4 ≡ P0 ψ̃4 ≡ P0 gψ3 , (7.21)
2
one then obtains the integral equation satisfied by the function g(p). Eliminating from it
the function E(p), that equation becomes:
pg(p) cos ϕ(p)

1 d 3k
= sin ϕ(p) + V (p − k)
2 (2π)3

× sin ϕ(k) sin ϕ(p) + p̂.k̂ cos ϕ(k) cos ϕ(p) g(p) − g(k)

1 d 3k
+ cos ϕ(p) V (p − k)(p − k).k̂ cos ϕ(k)g(k), (7.22)
3p (2π)3
which is free of infrared singularities; except for the last term, which originates from the
rotational motion part of the flux tube, it is the same as a corresponding one derived in
Refs. [41,43]. The function g has the same asymptotic behavior as sin ϕ and turns out to
be nodeless.
We have made a rough evaluation of the orders of magnitude of the quark condensate
√ of the pion decay constant using analytic approximations. Using
and σ = 0.18 GeV2 ,
σ = 424 MeV, we have found, with Nc = 3, ūu = −(115 MeV) and Fπ = 18 MeV,
3
16% of the latter quantity coming from the rotational motion part of the flux tube. These
results are in agreement with those found in Refs. [43,45], ūu −(100 MeV)3 and
Fπ 11 MeV. The experimental value of Fπ is nearly 93 MeV. The quark condensate
ūu is not a directly measurable quantity; QCD sum rules [47] give the prediction
ūu −(225 MeV)3 at the scale of 1 GeV.
In Ref. [42], numerical calculations have been done with the harmonic oscillator
potential, which makes it difficult to compare predictions. In Refs. [41,44], rather large
values of Fπ and ūu are presented, but these are obtained by fitting parameters in order
to fix the value of the constituent quark mass at 300 MeV, corresponding to big values of
the string tension (or of its equivalents).
The relative smallness (about one order of magnitude) of the theoretical predictions for
ūu and Fπ found here and in Refs. [43,45] seems to indicate that short-distance forces
should have sizable contributions in these quantities; this is corroborated by the facts that,
first, they act in momentum space with a potential having the same sign as the confining
one, and second, they induce much slower asymptotic behavior to the function ϕ. A definite
conclusion about the quantitative aspects of chiral symmetry breaking could be reached
only when both kinds of forces, long-range and short-range, are considered.
8. Properties of the bound state spectrum
We turn in this section to a study of the main qualitative properties of the bound
state spectrum that emerge from the wave equation, leaving to a separate work a more
quantitative study of it, in particular when short-distance effects are incorporated.
8.1. One-particle limit
We begin with the case when one of the particles becomes infinitely massive; this
is achieved by taking one of the masses, m2 , say, to infinity. Defining P0 = m2 + p10 ,
x = x1 − x2 , p = (p1 − p2 )/2 = p1 , the wave equation (6.11) reduces to a Dirac type
equation for the quark in the presence of a vector static potential:
[p10 − h1 − A0 + γ10 γ .A]ψ = 0, (8.1)
where the potentials, in the classical limit, are:

E12 x2 L2 E12 x2 L2
A0 = σ r arcsin − 1 − 8(p 10 ) 1 − 1 − , (8.2)
L2 E12 x2 L2 E12 x2

t 2 2
σ r p E1 x E12 x2 L2 L2
A= arcsin − 1− 2 . (8.3)
2 E1 L2 L2 E12 x2 E1 x 2
(r = |x|, pt is defined in Eq. (5.17), L is the orbital angular momentum operator, h1 , E1
and 8(p10 ) are defined in Eqs. (6.8)–(6.10).) They represent the energy–momentum of a
straight segment of length r with linear energy density σ turning around the point x2 (the
position of the heavy antiquark).
Taking further the limit of a heavy quark, the potentials become to order 1/c2 :

L2 σ r pt
A0 = σ r 1 + , A = . (8.4)
6m21 x2 3 m1
8.2. Nonrelativistic limit
Next, we consider the case of two heavy quarks. The potential A1µ (Eq. (5.20)) becomes
(in the c.m. frame) to order 1/c2 :
2
1 1 m1 1 1 1 L
A10 = σ r + + 2−
2 6 m1 + m2 m1 m2 m1 m2 x 2
2
2
1 m1 m2 1 1 L2
+ − , (8.5)
8 m1 + m2 m41 m42 x2

σr t 2 1
A1 = − p − , (8.6)
2 3m1 3m2
the definitions of x and p being the same as in Eqs. (5.3). A2 is obtained from A1 by
exchange of indices 1 and 2 and replacement of p by −p. The self-energy Σ (Eq. (6.7)),
behaves for large m as:

2σ 1
Σ(p, m) = − +O . (8.7)
πm m2
After having subtracted the masses and made a few changes of function to reach an
explicitly hermitian form, the nonrelativistic Hamiltonian, to order 1/c2 takes the following
form (in the c.m. frame):

p2 2h̄σ 1 1 1 1 1 2 2 h̄2 1 1 σ
H= + σr − + − + p + +
2µ π m1 m2 8 m31 m32 4 m21 m22 r

σ 1 1 1 2 σ L.s1 L.s2
− + − L + 2h̄ +
2
+ 2
6r m21 m22 m1 m2 2r m21 m2

2σ 1 1 2σ 1 1
− 2
− L.s1 − 2
− L.s2 . (8.8)
3r m1 2m1 m2 3r m2 2m1 m2
(µ = m1 m2 /(m1 + m2 ), s1 and s2 are the spin operators of the quark and of the antiquark.)
Several remarks can be made at this stage. First, the Hamiltonian is independent of spin–
spin interactions; this is already evident from the wave equation (6.11) where no direct
interactions involving the spacelike γ -matrices of both quarks exist. The absence of long-
range spin–spin interactions is compatible with experimental data. Second, we notice the
presence of purely orbital angular momentum dependent pieces (proportional to L2 ), the
origin of which is related to the contribution to the rotational motion of the system of the
moment of inertia of the flux tube, represented by the straight segment joining the quark
to the antiquark. The corresponding centrifugal energy produces a global plus sign in front
of those terms; the minus sign results from the additional contributions of the momentum
of the flux tube, which also couples, through the spacelike potentials A1 and A2 to the
quarks (Eq. (8.6)). Those terms were also obtained in Refs. [18,35,36]. Third, we have
two kinds of spin–orbit term. The first, appearing after the term in L2 , comes from the
contribution of a conventional timelike vector interaction represented by the potential σ r,
which is the dominant part of the combination A10 + A20 (Eq. (8.5)). The second type of
contribution, provided by the last two terms of Eq. (8.8), comes from the contributions of
the direct interactions of the momentum of the flux tube with the quarks, represented by
the spacelike potentials A1 and A2 (Eq. (8.6)). The latter terms induce negative signs to
the spin–orbit couplings, in opposite direction to the former one, a feature which is also
observed on phenomenological grounds for the large-distance effects in fine splitting.
Adopting the notations of Ref. [16] for the potentials corresponding to spin–orbit (V1
and V2 ) and to spin–spin interactions (V3 and V4 ), we have:
2 1
V1 = − σ r, V2 = + σ r, V3 = 0, V4 = 0. (8.9)
3 3
Designating by V the nonrelativistic confining potential (Eq. (7.4)), the potentials V , V1

and V2 satisfy the Gromes relation [17]
V + V1 − V2 = 0, (8.10)
which is a consequence of Lorentz covariance.

The expressions of potentials V1 and V2 do not satisfy, however, Buchmüller’s
conjecture about the spin–orbit terms [48], according to which confinement is due to a
pure color electric field in the co-moving frame of the two quarks, thus reducing the spin–
orbit potential to a Thomas precession term and entailing V1 = −σ r and V2 = 0. Leaving
aside the question of a phenomenological determination of these potentials, we notice
that lattice calculations [49,50], which favor the negative sign of V1 , do not, however,
clearly distinguish between the two possibilities above, due to the existing uncertainties. In
order to clarify the structure of the spin–orbit potentials, it is necessary to investigate the
contributions of higher-order terms not taken into account in the potentials (5.20) (of the
type of those considered for the extraction of the self-energy terms). Another independent
check consists in calculating the field correlators appearing in the nonrelativistic expansion
of Ref. [16]. Reviews about bound state problems of quarks can be found in Refs. [51,52].
8.3. Regge trajectories
Finally, we consider the high-energy behavior of the mass spectrum for ultrarelativistic
systems. It has been known for a long time that the linear confining potential of the
static case produces in the ultrarelativistic limit of massless quarks, through the Salpeter
equation, linear Regge trajectories. The inclusion of a flux tube, represented by a straight
string, modifies the relationship of the angular momentum with the total mass of the system
and increases the slope α of the Regge trajectories by an amount of 15–20%, enforcing
the classical relation α = (2πσ )−1 with the string tension σ [36,53].
We have checked these properties on Eq. (6.11), by solving it in an approximate way
that preserves its main qualitative features. The resulting Regge trajectories for the π
and ρ families are presented in Fig. 6, where for comparison we have also presented
the trajectories of the linear confining potential corresponding to the situation where
A10 + A20 = σ r and A1 = A2 = 0 in Eq. (6.11). One verifies, first, the linearity of the
trajectories and, second, the increase of the slopes when the flux tube is present. Similar
trajectories and behaviors are also obtained with the a0 and a1 families.
On phenomenological grounds, one actually determines the value of the string tension
σ from the experimental Regge trajectories using the string theory relation α = (2πσ )−1 .
From the slope of the π -family trajectory one obtains σ = 0.22 GeV2 , while from
the ρ-family trajectory one obtains σ = 0.18 GeV2 [52]. A precise comparison of our
results with those data necessitates a more elaborate resolution of the wave equation and
presumably the inclusion of the short-distance potentials.
Fig. 6. Regge trajectories of the π and ρ families and their first daughter trajectories. Full lines correspond to the
flux tube potential, dotted lines to the linear potential. Inputs: m1 = m2 = 0, σ = 0.18 GeV2 .
9. Summary and concluding remarks
We have investigated, in the large-Nc limit, the large-distance dynamics of QCD

within the quarkonium system through the saturation of Wilson loop averages by minimal
surfaces. The dynamics is described by a wave equation of the Breit–Salpeter type, in
which the interaction potentials are provided by the energy–momentum vector of the
straight segment joining the quark to the antiquark and carrying a constant linear energy
density, equal to the string tension; the latter represents the effective contribution of the
color flux tube of the quarkonium system. In the static case, confinement is realized by the
usual linear potential, while in the general case of moving quarks additional contributions
come from the moments of inertia of the straight segment.
Taking into account the self-energy parts of the quark propagators, it was shown
that chiral symmetry is spontaneously broken with a mechanism identical to that of the
exchange of one Coulomb-gluon.
In the nonrelativistic limit, long range spin–spin potentials are absent, while the
moments of inertia of the flux tube bring contributions to the orbital angular momentum
and spin–orbit potentials with negative signs, in opposite direction, for the latter case, to
the spin–orbit term of the pure timelike vector potential.
In the ultrarelativistic limit, the mass spectrum displays linear Regge trajectories, the
slopes of which tend to satisfy the classical relation with the string tension obtained in
straight string theories.
The potentials that have been isolated in the present work do not represent the
complete interaction kernel of the wave equation, but only the large-distance dominant
contributions to it. The determination of the higher-order terms necessitates further study.
Also, mathematical assumptions made for neglecting certain terms or approximating others
in the large separation time limit need to be analyzed in more detail.
The formalism that has been developed here, centering the Wilson loop representation
on minimal surfaces, ignores the contributions of short-distance effects provided by
perturbation theory. A complete resolution of QCD requires the presence of both large-
and short-distance effects. A corresponding resolution of the loop equations is far from
being trivial, although approximate solutions might still provide a practical framework for
a quantitative investigation of the spectroscopic properties of the theory.
Acknowledgements
We thank J. Stern for stimulating discussions. The numerical calculation of the mass
spectrum used a program based on the inverse iterative method provided to us by
J.-L. Ballot to whom we are grateful. This work was supported in part by the European
Community network EURIDICE under Contract No. HPRN-CT-2002-00311.
Appendix A. Properties of minimal surfaces
The first-order functional derivative of the area (Eq. (2.14)) of a minimal surface (Eq.
(2.16)) at a point on the contour, taken for definiteness on the line τ = 0, is:

δA ν µ

α
= δµα x (σ, 0) − δνα x (σ, 0) dσ µν (y) F y(σ , τ ) − x(σ, 0)
δx (σ, 0)
(x 2 ẋ α − x .ẋx α )
=− , (A.1)
x 2 ẋ 2 − (x .ẋ)2
the second equation resulting in the limit a = 0; it manifestly satisfies the orthogonality
condition
δA
x α (σ, 0) α = 0, (A.2)
δx (σ, 0)
in agreement with Eq. (2.3).
An important quantity in loop dynamics is the second-order functional derivative of
the area at two different points on the contour. For a minimal surface, the variations of the
contour maintain at all stages the minimality property of the area; therefore the second term
of the right-hand side of Eq. (2.14) and all its variations are null and can be ignored. The
second-order functional derivative of the area, at two points on the line τ = 0 as before, is
then:
δ δA
β α
δx (σ , 0) δx (σ, 0)

∂
= (δµα δνβ − δµβ δνα ) δ(σ − σ ) dσ µν (y) F y − x(σ, 0)
∂σ

+ δµα x ν (σ, 0) − δνα x µ (σ, 0) δµβ x ν (σ , 0) − δνβ x µ (σ , 0)

× F x(σ, 0) − x(σ , 0)
∂
+ δµα x ν (σ, 0) − δνα x µ (σ, 0) δ(σ − σ ) β
∂x (σ, 0)

× dσ (y) F y − x(σ, 0)
µν

δy λ (σ , τ )
− δµα x ν (σ, 0) − δνα x µ (σ, 0) dσ dτ
δx β (σ , 0)
γ ν
× y ẏ − y ν ẏ γ δµλ + y µ ẏ γ − y γ ẏ µ δνλ + y ν ẏ µ − y µ ẏ ν δγ λ
∂
× γ F y − x(σ, 0) . (A.3)
∂y
The evaluation of the last integral can be done by expanding δy λ /δx β (σ , 0) about the
point x(σ, 0) that appears in F :
δy λ (σ , τ ) δy λ (σ, 0) ∂ δy λ (σ, 0)

= β + (σ − σ )
β
δx (σ , 0) δx (σ , 0) ∂σ δx β (σ , 0)

∂ δy λ (σ, τ )
+ τ + ···, (A.4)
∂τ δx β (σ , 0) τ =0
where the dots stand for terms that do not contribute in the limit a = 0. One notices that
the first term in the right-hand side of Eq. (A.4) is a derivative that lies on the contour and
hence is equal to δβλ δ(σ − σ ):
δy λ (σ, 0) δx λ (σ, 0)
≡ = δβλ δ(σ − σ ). (A.5)
δx β (σ , 0) δx β (σ , 0)
It is also convenient, to complete the calculations, to stick to the following diagonal
constant metric:

2 x 2
x .ẋ = 0, x = const, ẋ = const, λ ≡
2
. (A.6)
ẋ 2
The result is:
δ δA
δx β (σ , 0) δx α (σ, 0)

∂
= 2 δ(σ − σ ) dσ αβ (y) F y − x(σ, 0)
∂σ

+ 2 δαβ x (σ, 0).x (σ , 0) − x β (σ, 0)x α (σ , 0) F x(σ, 0) − x(σ , 0)

1 2 x α x β
− δ(σ − σ ) x δαβ −
a x 2

x 2 ∂ δy λ (σ, τ ) x λ (σ, 0)x α (σ, 0) ẋ λ (σ, 0)ẋ α (σ, 0)
− δλα − − .
ẋ 2 ∂τ δx β (σ , 0) τ =0 x 2 (σ, 0) ẋ 2 (σ, 0)
(A.7)
The two points x(σ, 0) and x(σ , 0) lying on the boundary, √ the function F (x(σ, 0) −
x(σ , 0)) is equivalent, in the limit a = 0, to 1/(2a x 2 )δ(σ − σ ) and thus cancels
the singular term of the third line. (We assume here that the line under consideration
does not have self-intersections neither displays backtracking.) There remains to evaluate
the quantity (∂/∂τ )(δy λ (σ, τ )/δx β (σ , 0))|τ =0 . This in turn requires a more detailed
knowledge of the derivative δy λ (σ, τ )/δx β (σ , 0). That quantity does not have a compact
expression for general contours. Studies of its main properties can be found in Refs. [27,
32].
Let us first point out the fact that a variation δx on the contour that lies in the tangent
plane to the minimal surface does not modify the internal part of the minimal surface but
only adds to it a new small piece on its boundary within the tangent plane; the variation
thus amounts to a simple extension of the boundary on the same surface. In that case δy
should lie in the tangent plane to the minimal surface at y, being a linear combination of
y and ẏ and corresponding to a reparametrization of the surface. In this respect, it can be
checked that if δy λ contains terms proportional to y λ or to ẏ λ then the contributions of
these in Eq. (A.3) are identically zero. Therefore, the nontrivial contributions to Eq. (A.3)
come only from the transverse variations of δy with respect to the surface, which are
themselves due to transverse variations δx on the contour with respect to that surface.
We shall therefore concentrate on these kinds of variation.
Next, we introduce variations δy on the minimal surface, including the boundaries.
The equation of the new minimal surface is obtained by replacing in the equation of
the background minimal surface y by y + δy. One then introduces the Green function
Gνρ (ξ, η) of the linearized equation in δy, where ξ = (ξ 1 , ξ 2 ) = (σ , τ ), η = (η1 , η2 ) =
(σ, τ ), subject to the boundary condition that it vanishes along the contour C [27]. Using
then the Green identity, the Stokes theorem and the transversality of δy, one obtains for the
transverse variations δy the following expression:

δyµ (η)
2

a ∂

= n (λ ) a Gνµ (ξ, η) , (A.8)
δxν (ξ(λ )) ∂ξ ξ =ξ(λ )
a=1
where the contour C has been parametrized with the parameter λ and where n is an
orthogonal vector to the contour in the tangent plane to the minimal surface.
The derivative of δyµ /δxν along n with respect to η taken on the contour is:

δyµ (η)
2
b ∂
n (λ) b
∂η δxν (ξ(λ )) η=η(λ)
b=1

2
∂ ∂
= n (λ)n (λ ) b a Gνµ (ξ, η)
b a
. (A.9)
∂η ∂ξ ξ =ξ(λ ),η=η(λ)
a,b=1
A consequence of Eq. (A.9) is the symmetry property of its right-hand side with
respect to the variables ξ and η for λ = λ , a feature that is not evident in the left-hand
side. That property is intimately related to the symmetry property of the second-order
variation of the minimal surface with respect to the order of operations. To exhibit this
feature more explicitly we return to the (σ, τ ) parametrization that we were using above.

The derivation na ∂a is equivalent in the limit τ = 0 to the derivation − x 2 /ẋ 2 (∂/∂τ ).
Eq. (A.8) becomes, using the metric (A.6):

δy λ (σ, τ ) x 2 ∂
= − G (σ
, τ ; σ, τ )
δx β (σ , 0) ẋ 2 ∂τ
βλ , (A.10)
τ =0
which in turn yields for Eq. (A.7) the expression
δ δA
δx β (σ , 0) δx α (σ, 0)

∂
=2 δ(σ − σ ) dσ αβ (y) F y − x(σ, 0)
∂σ

x 2 ∂ ∂
+ 2 G (σ
, τ ; σ, τ )
ẋ ∂τ ∂τ βλ
τ =0,τ =0

x λ (σ, 0)x α (σ, 0) ẋ λ (σ, 0)ẋ α (σ, 0)
× δλα − − . (A.11)
x 2 ẋ 2
The symmetry property of the second-order derivative of the minimal area can now be
discussed for the case σ = σ . The Green function can be decomposed along symmetric
Lorentz tensors each multiplied with a scalar function. Because of the boundary condition
of G (G vanishes when ξ or η is on C) the two derivatives ∂/∂τ and ∂/∂τ must
simultaneously act on at least one of the scalar functions in order not to yield zero. In that
case the product of the transverse projecter present in the last term of Eq. (A.11) with the
tensor factors of Gβλ gives a symmetric result, which ensures the commutativity of the two
functional derivatives when acting at two different points of the contour. That symmetry
property is evident when the functional derivatives act on the Wilson loop average; then the
two operators manifestly commute. This study will be completed below by also including
the case of coinciding points.
Concrete calculations can be done when the Green function is specified more explicitly.
To this aim, one can try to construct the Green function by means of an iterative procedure
with respect to powers and derivatives of the curvature of the minimal surface:
∞

Gβλ (ξ, η) = G(i)
βλ (ξ, η), (A.12)
i=0
where the index i is related to the power or the order of the derivative of the curvature.
The lowest-order term of this iteration would correspond to the case of a plane with zero
curvature. The solution to that problem can be explicitly constructed. Using the metric
(A.6), one can use the ansatz
Gβλ (σ , τ ; σ, τ )
(0)

= δβλ (z .y ż.ẏ − z .ẏ ż.y ) − ż.ẏ(zλ yβ − zβ yλ ) + z .ẏ(żλ yβ − żβ yλ )
1
− z .y (żλ ẏβ − żβ ẏλ ) + ż.y (zλ ẏβ − zβ ẏλ ) 2 2 h(0) (σ , τ ; σ, τ ), (A.13)
y ẏ
where we have defined z ≡ x(σ , τ ), y ≡ x(σ, τ ). The function h(0) is assumed to satisfy
the inhomogeneous Laplace equations
C(ξ ) h(0) (ξ ; η) = C(η) h(0) (ξ ; η) = δ 2 (ξ − η), ξ = (σ , τ ), η = (σ, τ ), (A.14)

1 ∂2 ∂2
C(η) = +λ 2, (A.15)
λ ∂σ 2 ∂τ
and the boundary conditions
h(ξ, η)|ξ ∈C = h(ξ, η)|η∈C = 0. (A.16)
The coordinates z and y represent points on the background minimal surface and with
parametrization (A.6) satisfy the Laplace equations
C(ξ ) zµ (ξ ) = 0, C(η) yµ (η) = 0. (A.17)

(0)
The structure of the function Gβλ has been determined in accordance with the following
properties:
(i) when the point z is on the contour, then, according to Eq. (A.8), a tangential
deformation of the contour with respect to the minimal surface implies a tangential
variation δy at the point y;
or ż yields an antisymmetric expression in
(ii) an explicit derivation with respect to zµ µ
the indices β and µ, modulo a tangential term in δy λ , the latter not contributing in
Eq. (A.3);
(iii) the boundary condition (A.5) is satisfied.
Inserting G(0) (1)

βλ in the equations of Gβλ one can reconstruct Gβλ by appropriately
decomposing it along Lorentz tensors and determining its Lorentz scalar parts through
inhomogeneous Laplace equations involving as inhomogeneous terms h(0) and its first-
order derivatives. It should be noticed that the inhomogeneous part of the equations of G
is already saturated by the function G(0)
βλ and therefore the higher-order functions do no
longer have contributions to the delta-functions. The Lorentz tensor parts of the function
G(1)
βλ contain at least one second-order derivative of y or z. The procedure can be continued
to higher orders; the function G(2)βλ will be quadratic in the second-order derivatives of y
and z, or will contain third-order derivatives, and so forth. We shall not develop here the
above procedure, for we shall not need in the present paper the explicit expressions of
(i) (0)
the functions Gβλ (i 1). In many instances, the sole knowledge of the function Gµν is
sufficient to obtain or to check the main properties of the system under study. This is in
particular the case where the background minimal surface from which deformations are
calculated lies in a plane.
Let us now concentrate on the function G(0) µν . We consider a four-sided contour C as in
Fig. 1 with the additional restriction that the four sides correspond to the lines of equations
τ = 0, σ = 0, τ = 1, σ = 1. The explicit expression of the function h(0) , satisfying the
boundary conditions (A.16) is then:

∞
4 sin(nπσ ) sin(nπσ ) sin(mπτ ) sin(mπτ )
h (σ , τ ; σ, τ ) = − 2
(0)
. (A.18)
π (m2 λ + n2 /λ)
n,m=1
Calculating, in view of Eq. (A.10), the derivative of h(0) with respect to τ at τ = 0, one
obtains a series in m that can be summed into a hyperbolic function:

(0) ∂h(0)
f (σ , 0; σ, τ ) ≡ −λ
∂τ τ =0
∞
sin(nπσ ) sin(nπσ ) sinh(nπ(1 − τ )/λ)
=2 . (A.19)
sinh(nπ/λ)
n=1
One checks that δy λ/δx β satisfies, with G(0) βλ and through Eq. (A.10), the boundary
condition (A.5).
We next evaluate the derivative of δy λ /δx β with respect to τ at τ = 0. The derivative
acts on f (0) and on the Lorentz tensor λ β
part of δy /δx . The latter action yields in Eq.
(A.11) the finite part of x ν (∂/∂x β ) dσ αν (y) F (y − x(σ, 0)) multiplied by δ(σ − σ )
coming from the restriction of f (0) on the contour. One finds for the second-order
functional derivative of the area the result:
δ δA
δx β (σ , 0) δx α (σ, 0)

∂
= 2 δ(σ − σ ) dσ αβ (u)F (u − y)
∂σ

∂
+ 2δ(σ − σ )y ν β dσ αν (u) F (u − y)fp
∂y

y 2
− δβα (z .y ż.ẏ − z .ẏ ż.y )
ẏ 2

− (ż.ẏzα yβ + z .y żα ẏβ − z .ẏ żα yβ − ż.y zα ẏβ )

1 ∂f (0) (σ , 0; σ, τ )
× 2 2 , (A.20)
y ẏ ∂τ τ =0
where y ≡ x(σ, 0), z ≡ x(σ , 0) and the subscript “fp” means finite part. The last term of
Eq. (A.20) is invariant under the change of the order of derivations, as can be seen from the
tensor terms and with the use of the defining equations (A.18), (A.19) of f (0) , reflecting
the general property emphasized after Eq. (A.11). The symmetry property of the contact
terms will be established shortly.
The singular part of ∂f (0) (σ , 0; σ, τ )/∂τ |τ =0 when (σ − σ ) approaches zero can be
calculated from Eq. (A.19). It is:

∂f (0) (σ , 0; σ, τ ) 1 ẋ 2
= 2
(σ − σ )−2 + O (σ − σ )0 . (A.21)
∂τ τ =0 (σ −σ )→0 π x
It yields for the contracted second-order derivative δ 2 A/(δx α (σ , 0)δx α (σ, 0)), in which
the contact terms do not contribute, the behavior
δ2 A 2
= − (σ − σ )−2 + O (σ − σ )0 . (A.22)
δx α (σ , 0)δx α (σ, 0) (σ −σ )→0 π
It is in agreement with the leading part of the short-distance behavior obtained in Ref.
[27]. The contracted second-order derivative contains also a logarithmic singularity [27]
which is obtained in the present approach by considering the contributions of the next two
(i)
nonleading terms in the curvature, i.e., the functions Gµν , i = 1, 2 (Eq. (A.12)).
We now check the symmetry property between two successive functional derivatives.
Let us first consider the problem in the Wilson loop case. The commutator of two functional
derivatives on the contour is, using Eqs. (2.3) and (2.6):

δ δ δ δ
− W (C)
∂ ∂
= ig δ(σ − σ ) Fβα x(σ ) W − ig δ(σ − σ ) Fαβ x(σ ) W
∂σ ∂σ

+ igδ(σ − σ )x ν (σ ) ∇β Fνα x(σ ) − ∇α Fνβ x(σ ) W . (A.23)
Expanding in Fαβ (x(σ )) σ about σ and using in the last expression the Bianchi identity
one finds:

δ δ δ δ
− W (C)

∂ ∂
= ig δ(σ − σ ) (σ − σ ) + δ(σ − σ ) Fβα x(σ ) W
∂σ ∂σ
∂ ∂
= ig δ(σ − σ )(σ − σ ) Fβα x(σ ) W = 0. (A.24)
∂σ ∂σ
When the Wilson loop average is saturated by the minimal surface (Eq. (2.10)) the
left-hand side of Eq. (A.23) becomes proportional to the commutator of the functional
derivatives applied on the minimal area. Using Eq. (A.20), one obtains:

1 δ δ δ δ
− A(C)
2 δx β (σ , 0) δx α (σ, 0) δx α (σ, 0) δx β (σ , 0)

∂
= δ(σ − σ ) dσ αβ (y) F y − x(σ, 0)
∂σ

∂

−
δ(σ − σ ) dσ βα (y) F y − x(σ , 0)
∂σ

∂
+ δ(σ − σ )x ν (σ, 0) β
dσ αν (y) F y − x(σ, 0)
∂x (σ, 0)

∂
− α dσ βν (y) F y − x(σ, 0) . (A.25)
∂x (σ, 0) fp
Fig. 7. Contour C made of the three sides x1 x1 x2 x2 of a rectangle and of the arbitrary line C21 lying in an
orthogonal plane to the rectangle.
The right-hand side has the same structure as that of Eq. (A.23) and since the minimal
surface satisfies the Bianchi identity (2.16) the same operations as above lead to the result

δ δ δ δ
− A(C) = 0. (A.26)
δx β (σ , 0) δx α (σ, 0) δx α (σ, 0) δx β (σ , 0)

(Actually, it is the finite part of the quantity (∂/∂x β (σ, 0)) dσ αν (y)F (y − x(σ, 0)) that
is relevant for the nontrivial content of the Bianchi identity, the singular part satisfying the
Bianchi identity as an identity, independent of the type of the surface.)
The diagonal constant metric (A.6) can also be used for the study of the forms of
minimal surfaces in simple cases. In that metric the equation of the minimal surface is given
by the Laplace equation (A.17). Choosing the reference surface in a plane, the orthogonal
deformations of the surface with respect to the plane can be represented with functions
of the type (A.19), while the parallel deformations should be constructed according to
the orthonormality conditions (A.6). However, the sole knowledge of the orthogonal
deformations is sufficient to have a rough idea of the form of the surface. We apply this
study to one particular case of interest.
Let us consider the case where the contour C of Fig. 1 has been restricted in the
following way: the line x1 x1 x2 x2 represents now the three sides of a rectangle (lying in
a plane) while the line x2 x1 , denoted C21 , is arbitrary but lying in an orthogonal plane to
the rectangle and having in it the equation u(σ, 0) = a(σ ) (Fig. 7).
Designating by u(σ, τ ) the orthogonal deformation of the minimal surface with respect
to the plane, its expression can be obtained from Eq. (A.19) by using the Fourier
decomposition of the function a(σ ): a(σ ) = ∞ n=1 an sin(nπσ ). One obtains:
∞
sin(nπσ ) sinh(nπ(1 − τ )/λ)
u(σ, τ ) = an . (A.27)
sinh(nπ/λ)
n=1
In general this minimal surface does not have a simple form; u is different from zero all
over the rectangle, but is mainly peaked in the region τ 0. Let us next take the limit
ẋ 2 → ∞, or λ → 0. Then u goes to zero for any τ = 0. Therefore, in that limit, the
minimal surface shrinks to the surface made of the union of the rectangle x1 x1 x2 x2 x1
and of the minimal surface bounded√ by the straight segment x2 x1 and the line C21 . Taking
the time interval T proportional to ẋ 2 in that limit, we see that T is now contained in
the contribution of the rectangle only. Since the energy spectrum is provided by terms
proportional to T , we conclude that the energy spectrum provided by the previous contour
is the same as that of the rectangle. Therefore the shape of the contour C21 has no influence
on the energy spectrum and contributes only to the wave functional of the corresponding
bound state, made here of a static quark–antiquark pair. The meaning of this result is that
minimal surfaces, irrespective of the shapes of their contours, cannot produce quantum
stringlike excitations in the bound state energy spectrum, since rectangles do not contain
such excitations. This was an expected result, for the minimal surface is the classical action
of the open string and quantum excitations of it should only be searched for in fluctuations
of surfaces about the classical trajectory [25–27]; the present derivation, however, provides
another insight into the same property.
References
[1] K.G. Wilson, Phys. Rev. D 10 (1974) 2445.

[2] J.B. Kogut, Rev. Mod. Phys. 55 (1983) 775.
[3] S. Mandelstam, Phys. Rev. 175 (1968) 1580.
[4] Y. Nambu, Phys. Lett. B 80 (1979) 372.
[5] A.M. Polyakov, Nucl. Phys. B 164 (1979) 171.
[6] Yu.M. Makeenko, A.A. Migdal, Phys. Lett. B 88 (1979) 135;
Yu.M. Makeenko, A.A. Migdal, Phys. Lett. B 97 (1980) 253.
[7] Yu.M. Makeenko, A.A. Migdal, Nucl. Phys. B 188 (1981) 269.
[8] A.A. Migdal, Phys. Rep. 102 (1983) 199.
[9] Yu. Makeenko, Large N gauge theories, hep-th/0001047.
[10] S. Mandelstam, Phys. Rev. D 19 (1979) 2391.
[11] R. Giles, Phys. Rev. D 24 (1981) 2160.
[12] V.A. Kazakov, I.K. Kostov, Nucl. Phys. B 176 (1980) 199;
V.A. Kazakov, Nucl. Phys. B 179 (1981) 283.
[13] V.S. Dotsenko, S.N. Vergeles, Nucl. Phys. B 169 (1980) 527.
[14] R.A. Brandt, F. Neri, M.-A. Sato, Phys. Rev. D 24 (1981) 879;
R.A. Brandt, A. Gocksch, M.-A. Sato, F. Neri, Phys. Rev. D 26 (1982) 3611.
[15] G. ’t Hooft, Nucl. Phys. B 72 (1974) 461.
[16] E. Eichten, F. Feinberg, Phys. Rev. D 23 (1981) 2724.
[17] D. Gromes, Z. Phys. C 22 (1984) 265;
D. Gromes, Z. Phys. C 26 (1984) 401.
[18] A. Barchielli, E. Montaldi, G.M. Prosperi, Nucl. Phys. B 296 (1988) 625;
A. Barchielli, E. Montaldi, G.M. Prosperi, Nucl. Phys. B 303 (1988) 752, Erratum;
A. Barchielli, N. Brambilla, G.M. Prosperi, Nuovo Cimento A 103 (1990) 59.
[19] N. Brambilla, P. Consoli, G.M. Prosperi, Phys. Rev. D 50 (1994) 5878;
N. Brambilla, E. Montaldi, G.M. Prosperi, Phys. Rev. D 54 (1996) 3506.
[20] N. Brambilla, A. Vairo, Quark confinement and the hadron spectrum, hep-ph/9904330.
[21] M. Baldicchi, G.M. Prosperi, Phys. Lett. B 436 (1998) 145;
M. Baldicchi, G.M. Prosperi, Phys. Rev. D 62 (2000) 114024;
M. Baldicchi, G.M. Prosperi, Phys. Rev. D 66 (2002) 074008.
[22] A. Pineda, A. Vairo, Phys. Rev. D 63 (2001) 054007;
A. Pineda, A. Vairo, Phys. Rev. D 64 (2001) 039902, Erratum;
N. Brambilla, A. Pineda, J. Soto, A. Vairo, Phys. Rev. D 63 (2001) 014023.
[23] H.G. Dosch, Phys. Lett. B 190 (1987) 177;
Yu.A. Simonov, Nucl. Phys. B 307 (1988) 512;
H.G. Dosch, Yu.A. Simonov, Phys. Lett. B 205 (1988) 339;
A. Di Giacomo, H.G. Dosch, V.I. Shevchenko, Yu.A. Simonov, Phys. Rep. 372 (2002) 319.
[24] J. Maldacena, Adv. Theor. Math. Phys. 2 (1997) 231;
J. Maldacena, Phys. Rev. Lett. 80 (1998) 4859;
E. Witten, Adv. Theor. Math. Phys. 2 (1998) 505;
S.-J. Rey, S. Theisen, J.-T. Yee, Nucl. Phys. B 527 (1998) 171;
A. Brandhuber, N. Itzhaki, J. Sonnenschein, S. Yankielowicz, Phys. Lett. B 434 (1998) 36;

N. Drukker, D.G. Gross, H. Ooguri, Phys. Rev. D 60 (1999) 125006;
A.M. Polyakov, V.S. Rychkov, Nucl. Phys. B 581 (2000) 116;
A.M. Polyakov, V.S. Rychkov, Nucl. Phys. B 594 (2001) 272;
S.-J. Rey, J.-T. Yee, Eur. Phys. J. C 22 (2001) 379.
[25] T. Egushi, Phys. Rev. Lett. 44 (1980) 126.
[26] M. Lüscher, Phys. Lett. B 90 (1980) 277.
[27] M. Lüscher, K. Symanzik, P. Weisz, Nucl. Phys. B 173 (1980) 365.
[28] E.E. Salpeter, H.A. Bethe, Phys. Rev. 84 (1951) 1232.
[29] M. Gell-Mann, F. Low, Phys. Rev. 84 (1951) 350.
[30] G. Breit, Phys. Rev. 34 (1929) 553;
G. Breit, Phys. Rev. 36 (1930) 383;
G. Breit, Phys. Rev. 39 (1932) 616.
[31] E.E. Salpeter, Phys. Rev. 87 (1952) 328.
[32] R. Courant, Dirichlet’s Principle, Conformal Mapping, and Minimal Surfaces, Springer-Verlag, New York,
1977.
[33] R.P. Feynman, Phys. Rev. 80 (1950) 440;
R.P. Feynman, Phys. Rev. 84 (1951) 108;
J. Schwinger, Phys. Rev. 82 (1951) 664;
T.E. Nieuwenhuis, The Feynman–Schwinger representation of field theory applied to two-body bound states,
Ph.D. Thesis, Utrecht University, 1995, unpublished;
Yu.A. Simonov, J.R. Tjon, Ann. Phys. (N.Y.) 300 (2002) 54.
[34] G. Veneziano, Nucl. Phys. B 117 (1976) 519;
E. Witten, Nucl. Phys. B 160 (1979) 57.
[35] M.G. Olsson, K. Williams, Phys. Rev. D 48 (1993) 417.
[36] D. LaCourse, M.G. Olsson, Phys. Rev. D 39 (1989) 2751;
C. Olson, M.G. Olsson, K. Williams, Phys. Rev. D 45 (1992) 4307;
C. Olson, M.G. Olsson, Phys. Rev. D 49 (1994) 4675.
[37] I.M. Gel’fand, G.E. Shilov, Generalized Functions, Vol. 1, Academic Press, New York, 1964.
[39] Y. Nambu, G. Jona-Lasinio, Phys. Rev. 122 (1961) 345;
Y. Nambu, G. Jona-Lasinio, Phys. Rev. 124 (1961) 246.
[40] M. Baker, K. Johnson, B.W. Lee, Phys. Rev. 133 (1964) B209.
[41] J. Finger, J.E. Mandula, J. Weyers, Phys. Lett. B 96 (1980) 367;
J. Finger, J.E. Mandula, Nucl. Phys. B 199 (1982) 168;
J. Govaerts, J.E. Mandula, J. Weyers, Nucl. Phys. B 237 (1984) 59.
[42] A. Amer, A. Le Yaouanc, L. Oliver, O. Pène, J.-C. Reynal, Phys. Rev. Lett. 50 (1983) 87;
A. Le Yaouanc, L. Oliver, O. Pène, J.-C. Reynal, Phys. Rev. D 29 (1984) 1233;
A. Le Yaouanc, L. Oliver, S. Ono, O. Pène, J.-C. Reynal, Phys. Rev. D 31 (1985) 137.
[43] S.L. Adler, A.C. Davis, Nucl. Phys. B 244 (1984) 469.
[44] R. Alkofer, P.A. Amundsen, Nucl. Phys. B 306 (1988) 305.
[45] J.-F. Lagaë, Phys. Rev. D 45 (1992) 305;
J.-F. Lagaë, Phys. Rev. D 45 (1992) 317.
[46] M. Gell-Mann, R.J. Oakes, B. Renner, Phys. Rev. 175 (1968) 2195.
[47] S. Narison, QCD Spectral Sum Rules, World Scientific, Singapore, 1989, and references therein.
[48] W. Buchmüller, Phys. Lett. B 112 (1982) 479.
[49] A. Huntley, C. Michael, Nucl. Phys. B 286 (1987) 211.
[50] G.S. Bali, K. Schilling, A. Wachter, Phys. Rev. D 56 (1997) 2566.
[51] W. Lucha, F.F. Schöberl, D. Gromes, Phys. Rep. 200 (1991) 127.
[52] G.S. Bali, Phys. Rep. 343 (2001) 1.
[53] M. Ida, Prog. Theor. Phys. 59 (1978) 1661.
Analysis of the axial anomaly on the lattice with

O(a)-improved Wilson action
D. Guadagnoli a , S. Simula b
a Dipartimento di Fisica, Università di Roma “La Sapienza”, Piazzale Aldo Moro 2, I-00185 Roma, Italy
b Istituto Nazionale di Fisica Nucleare, Sezione Roma III, Via della Vasca Navale 84, I-00146 Roma, Italy
Received 10 July 2003; accepted 28 July 2003
Abstract
Flavor singlet and non-singlet axial Ward identities are investigated using the Wilson formulation
of lattice QCD with Clover O(a)-improvement, which breaks explicitly chiral symmetry. The
matching at one-loop order of all the relevant renormalization constants with the continuum MS
scheme is presented. Our calculations include: (1) the contributions arising from the Clover term
of the action; (2) the complete mixing of the gluon operator GG with the divergence of the singlet
axial current; (3) the use of both local and extended definitions of the fermionic bilinear operators.
A definition of the gluon operator GG on the lattice outside the chiral limit is proposed. Our definition
takes into account the possible power-divergent mixing with the pseudoscalar density, generated by
the breaking of chiral symmetry. A non-perturbative procedure for the evaluation of such mixing
constant is worked out. Finally,
the renormalization properties of the composite insertion of the

topological charge operator d 4 x GG(x) relevant for the lattice calculation of the neutron electric
dipole moment, induced by the strong CP -violating term of the QCD Lagrangian, are discussed.
PACS: 11.40.Ha; 12.38.Gc; 13.40.Em; 14.20.Dh
Keywords: Ward identities; Lattice QCD; Topological charge; Electric dipole moment
1. Introduction
The investigation of the axial anomaly is a powerful tool to unravel the structure of the
QCD vacuum. Since the latter is highly non-trivial, the conservation of classical currents
may be spoiled by quantum fluctuations of the vacuum. When this happens an anomaly
E-mail address: simula@roma3.infn.it (S. Simula).
doi:10.1016/j.nuclphysb.2003.07.026
D. Guadagnoli, S. Simula / Nuclear Physics B 670 (2003) 264–288 265
is formed as in case of the flavor-singlet axial vector current. The axial anomaly can have
deep consequences on various physical observables, like the large mass of the η meson
or the smallness of the flavor-singlet nucleon axial coupling constant, which is commonly
known as the proton spin crisis.
The connection between the topological structure of the QCD vacuum and the axial
anomaly has been elucidated by ’t Hooft [1]. A very interesting case where such an
interrelation shows up, is provided by the neutron electric dipole moment (EDM) generated
by the so-called θ -term of the QCD Lagrangian [2], which in the Euclidean space is given
by
g2 2
= iθ g εαβµν Gcαβ Gcµν ,
Lθ = iθ G G (1)
32π 2 64π 2
where Gcµν is the gluon field strength, c the color octet index (c = 1, . . . , 8) and θ a
dimensionless parameter. The θ -term (1) breaks both parity and time reversal symmetries
and therefore it can generate a non-vanishing value of the neutron EDM.1 Available
estimates of the relevant matrix element however are based on phenomenological models,
as the MIT bag model of Ref. [4] or as the effective πN chiral Lagrangian of Ref. [5].
Estimates relying on non-perturbative methods based on the fundamental theory, like lattice
QCD, are still missing to date.
In Ref. [6] a strategy for evaluating the neutron EDM on the lattice induced by
the strong CP -violating term (1) was presented. Such a strategy is based on the
standard definition
of the neutron EDM, involving the insertion of the topological charge

(g 2 /32π 2 ) d 4 x GG(x) in the presence of the charge density operator J0 (see Eq. (4) of
Ref. [6]). In case of three flavors with non-degenerate masses a complete diagrammatic
analysis was performed [6] showing how the axial anomaly governs the replacement of the
topological charge operator with well-defined insertions of the flavor-singlet pseudoscalar
density. The applicability of the method to the case of lattice formulations that break
explicitly chiral symmetry, like the Wilson and Clover actions, was discussed in Ref. [6]
using general arguments.
The aim of this work is twofold: (i) to present a complete one-loop calculation of the
renormalization constants appearing in both singlet and non-singlet axial Ward identities
(WI’s) using Wilson fermions with the Clover O(a)-improvement of the action; (ii) to
provide a definition of the gluon operator GG on the lattice outside the chiral limit, taking
into account its possible power-divergent mixing with the pseudoscalar density. As for
the one-loop matching, our calculations reproduce all the known results and add: (1) the
complete mixing of GG with the divergence of the singlet axial current; and (2) the use
of both local and extended definitions of the fermionic bilinear operators. As for the
calculation of the mixing between a lattice discretization of GG and the pseudoscalar
density in case of lattice formulations breaking chiral symmetry a non-perturbative
procedure is presented. Finally, as a separate issue, the renormalization properties of the
1 The present experimental upper limit on the neutron EDM is d ≡ |d | < 6.3 × 10−26 (e cm) at 90%
N N
confidence level [3], which corresponds to a severe bound on the magnitude of θ . Indeed, using the available
theoretical estimates from Refs. [4,5], one has dN ≈ 3 × 10−16 |θ | (e cm) leading to |θ | 2 × 10−10 . The
smallness of the parameter θ is usually referred to as the strong CP problem.
266 D. Guadagnoli, S. Simula / Nuclear Physics B 670 (2003) 264–288

composite insertion of the topological charge operator d 4 x GG(x) relevant for the lattice
calculation of the neutron EDM are discussed.
The plan of the paper is as follows. In Section 2 the structure of singlet and non-singlet
axial WIs using the Wilson and Clover lattice QCD formulations is briefly recalled to fix
notations and basic definitions. In Sections 3 and 4 the non-singlet and singlet channels are
considered, respectively. All the matching coefficients with the continuum MS scheme are
explicitly calculated at one-loop order both with and without the O(a)-improvement of the
action. The complete one-loop mixing of the gluon operator GG with the divergence of
the singlet axial current is evaluated. The issue of the possible power-divergent mixing
of GG with the pseudoscalar density is addressed and a non-perturbative procedure
for evaluating the mixing constant is proposed. Moreover, in a separate subsection, the
composite insertion of the topological charge operator d 4 x GG(x) relevant for the
lattice calculation of the neutron EDM is considered and its renormalization properties are
discussed. Section 5 is devoted to our conclusions. Finally, all the Feynman rules relevant
for our calculations are collected in Appendix A.
2. Axial Ward identities on the lattice
In the subsequent discussion we will make use of the following definitions and
notations. The QCD action on the lattice is defined (in Euclidean space) as
SLQCD = SF + SU (2)
with SU being the pure Yang–Mills component [7] and SF the Wilson fermion action
1
SF = a 4
− ψ̄(x)(r − γµ )Uµ (x)ψ(x + µ)
x
2a µ

4r
+ ψ̄(x + µ)(r + γµ )Uµ† (x)ψ(x) + ψ̄(x) m0 + ψ(x) , (3)
a
where color and flavor indices are omitted, m0 is the (bare) mass matrix, diagonal in
flavor, and the terms proportional to r are necessary to avoid the fermion doubling. The
improvement of the action is represented by the Clover term
ig0 ar
SC = −a 4 cSW ψ̄(x)σµν Pµν (x)ψ(x), (4)
x µ,ν
4
with Pµν being the usual lattice definition of the field-strength tensor Gµν [8]
1 1
4

Pµν (x) = 2
Ui − Ui† (5)
4a 2ig0
i=1
where the sum is over the four plaquettes in the µ–ν plane stemming from x and taken in
the counterclockwise sense (see also Appendix A).
To obtain axial WIs on the lattice, one starts with the usual definition of the vacuum
expectation value of an operator O(x1 , . . . , xn )

1
O(x1 , . . . , xn ) = d[G] d[ψ] d[ψ̄] O(x1 , . . . , xn )e−S , (6)
Z0
where the fields are defined only on the nodes of the lattice, Z0 is the partition function
and S = SLQCD +SC . Performing local non-singlet axial rotations over the fermionic fields,
namely

λa
ψ(x) → 1 + iα a (x) γ5 ψ(x),
2

λa
ψ̄(x) → ψ̄(x) 1 + iα a (x) γ5 , (7)
2
where λa are the usual SU(3) flavor matrices, and taking into account the invariance of the
measure of integration in Eq. (6), one gets

δS δO
O = (8)
δ(iα a (x)) δ(iα a (x))
with
a
δS λ
O = −∆ x
µ OA a
µ (x) + O ψ̄ (x) , m 0 γ 5 ψ(x)
δ(iα a (x)) 2
a
+ O X (x) + XC (x) .a
(9)
In Eq. (9) we have indicated with ∆xµ the backward derivative in the µ-direction with
respect to x. The non-singlet axial current Aaµ is given by

1 λa
Aaµ (x) = ψ̄(x)Uµ (x)γµ ψ(x + µ) + h.c. , (10)
2 2
and the operators Xa and XC a are the chiral variations of the Wilson and Clover term in the
action, respectively,

r λa λa
Xa (x) = − ψ̄(x) γ5 Uµ (x)ψ(x + µ) + ψ̄(x + µ) γ5 Uµ† (x)ψ(x)
2a µ 2 2

λa
+ (x → x − µ) − 4ψ̄(x) γ5 ψ(x) , (11)
2
ig0 ar λa
XC a
(x) = − cSW ψ̄(x) γ5 σµν Pµν (x)ψ(x). (12)
2 µν
2
Similarly, performing flavor-singlet rotations

ψ(x) → 1 + iα 0 (x)γ5 ψ(x),

ψ̄(x) → ψ̄(x) 1 + iα 0 (x)γ5 , (13)
one obtains

δS δO
O = (14)
δ(iα 0 (x)) δ(iα 0 (x))
where

δS
O 0
= −∆xµ OAµ (x) + 2 Oψ̄ (x)m0 γ5 ψ(x)
δ(iα (x))

+ O X0 (x) + XC
0
(x) . (15)
0
The singlet axial current Aµ and the operators X0 and XC are obtained from the corre-
sponding octet operators (10)–(12), respectively, with the naïve substitution λa /2 → 1.
Eqs. (8), (9) and (14), (15) are expressed in terms of unrenormalized quantities. The
a 0
operators Xa , XC , X0 and XC are dimension-5 operators multiplied by one power of
the lattice spacing; therefore, in any tree-level calculation they vanish identically in the
continuum limit.
With the inclusion of quantum corrections the situation changes drastically. In the
a
non-singlet channel, the operators Xa and XC mix with the axial current and with the
pseudoscalar density: these mixings result in a finite (multiplicative) renormalization
constant for the current Aaµ , and in an additive renormalization constant for the bare mass
m0 , which multiplies the pseudoscalar density. The knowledge of such constants allows
to identify the correct renormalized mass and axial current on the lattice, recalling that
Eqs. (8), (9) should reproduce the corresponding continuum WI in the limit a → 0.
In the singlet channel, the mixings of the operator X0 with the divergence of the (singlet)
axial current and with the (singlet) pseudoscalar density are in general different from those
corresponding to Xa , due to the presence of diagrams involving closed fermion loops
(besides the ones present also in the non-singlet channel). This could cause in principle
the singlet renormalized mass to be different from the non-singlet one: if it were so, we
would be in trouble to identify the continuum limit of the lattice singlet WI, since in the
continuum the renormalized quark masses in both the octet and the singlet channel are the
same. However, using general non-perturbative arguments, it has been shown in Ref. [9]
that renormalized masses are the same also on the lattice, so that, in this respect, a simple
correspondence between the lattice WIs and the continuum ones can be established.
The singlet WI on the lattice must reproduce the anomaly, represented by the term
2NF g02 /32π 2 GG, where GG = 1/2 εµνρσ Gaµν Gaρσ and Gaµν is the usual gluon-field
strength tensor. In a continuum regularization the singlet anomaly is generated by the fact
that the regulator introduced during the renormalization process is not chiral invariant.
Removing the regulator leaves a residual contribution in the action proportional to GG.
On the lattice, which is a regulator, the action (3) with m0 = r = 0 is perfectly chiral
invariant, but it reproduces no anomaly, since it describes 16 quark species, with opposite
chiral charges. A possible way to eliminate the 15 spurious fermions in the continuum
limit, is represented by the Wilson term, which is however a chiral breaking term. The
Wilson term generates the anomaly. Indeed, in the WI (14), (15), the chiral variation of
the Wilson term, namely X0 (x), mixes with GG in such a way to correctly reproduce the
anomalous term present in the continuum [10]. The mixing is independent of r, as long as
r = 0.
Finally, given the anomalous non-conservation of the singlet axial current, the latter
suffers an infinite (logarithmically divergent) renormalization. Analogous divergences
occur for the operator GG, so that one needs a mixing between the two operators to end
up with finite quantities. Furthermore, on the lattice with broken chiral symmetry, one has
to take into account that the discretization adopted for GG may mix with the pseudoscalar
density.
Since the WIs (8), (9) and (14), (15) behave differently under renormalization, they will
be treated separately in the next two sections.
3. Non-singlet axial WIs
In this section we analyze in detail the one-loop structure of the WI (8), (9). For easy
of presentation we will consider the case of degenerate bare masses m0 , which means
{λa /2, m0 } → 2m0 λa /2. First we notice that Xa and XC a
, having the same dimension and
the same quantum numbers, should have an analogous behavior in perturbation theory. An
analysis identical to the one made in Refs. [9,11] for Xa leads to the identification
a

Xa (x) + XC
a
(x) = X (NS) ψ̄(x) λ γ5 ψ(x) + 1 − Z (NS) ∂µ Aaµ (x),
a (x) − 2M (16)
A
2
where the operator X a (x) collectively refers to dimension-5 operators that vanish in the
continuum limit for all matrix elements involving elementary operator insertions. Note
(NS)
that the renormalization constant ZA refers to the extended definition of the non-singlet
axial current, given in Eq. (10).
The mixing coefficients M (NS) and [1−Z (NS) ] appearing in Eq. (16) can be computed at
A
the one-loop level simply by considering the correlator O(x), ψ(y)ψ̄ (z) with amputation
on the external legs, and O given by Xa or XC a . Its Fourier transform can be pictorially
represented by the following diagrams (the corresponding Feynman rules are reported in
Appendix A)
(2)
FT Xa (x)ψ(y)ψ̄(z) amp
= + + + . (17)
One gets

g02 Σ0
(NS) =
M CF + m 0 C , (18)
16π 2 a
where
π
d 4 q 2r 2 4r 2
C = 16π 2
(∆ 1 ∆ 5 − ∆ 3 ) + c SW (−∆3 ∆5 + ∆4 )
(2π)4 ∆22 ∆22
−π
2
2 2r ∆3
+ cSW (∆3 ∆5 − ∆4 ) (19)
∆1 ∆22
and Σ0 is the additive mass renormalization, which appear in the one-loop approximation
of the quark propagator [see Eq. (25)].
The one-loop expression for the (finite) multiplicative renormalization of the axial
current is
(NS)
ZA −1
π 4
g02 d q 2
= CF 16π 2
r
16π 2 (2π)4
−π

1 1

× + 2 5∆1 ∆3 + 4∆1 − 5∆21 + ∆31 − 13∆3 + 4∆4
2∆2 ∆2

r ∆1
2 1
+ 2
∆ 3 − 6∆ 1 + ∆ 2
1 + c SW (4∆3 − ∆4 )
∆2 2∆1 ∆2

1
∆3
+ 2 10∆1 ∆3 − 2∆1 ∆4 − 2∆1∆3 − 8∆3 − 4∆3 + 6∆4 − 4∆7 −
2 2
∆2 2∆2

1
1

+ cSW
2
2
9∆3 ∆4 − 4∆23 − 2∆24 + 4∆3 + ∆23 − 6∆4 + 2∆7
∆1 ∆2 2∆1 ∆2

1 ∆3 2r 2 ∆1
+ (∆1 ∆3 − 6∆3 + 3∆4 ) + 2 (∆3 − 2∆4 ) + cSW (∆4 − ∆3 ∆5 )
4∆2 ∆2 ∆22
r2

+ cSW
2
2
∆1 −8∆3 − 3∆23 + 12∆4 − 4∆7 + 6∆1 ∆3 − 3∆1 ∆4 − ∆21 ∆3
∆2

− ∆3 ∆4 + 4∆23 . (20)
The symbols ∆i stand for the following functions

2
2 qµ 2 qµ
∆1 = sin , ∆2 = sin (qµ ) + 2r
2
sin ,
µ
2 µ µ
2

1 2
2 qµ 4 qµ
∆3 = sin (qµ ), ∆4 = sin cos ,
4 µ µ
2 2

qµ qµ
∆5 = cos2 , ∆6 = cos4 ,
µ
2 µ
2

qµ qµ qµ qµ
∆7 = sin2 cos6 , ∆8 = sin2 cos8 , (21)
µ
2 2 µ
2 2
which for ∆1 , . . . , ∆7 are identical to the corresponding functions used in Ref. [8], while
∆8 is included here for completeness but used only in Section 4. Numerical results for the
(NS)
constants C and ZA at various values of the Wilson parameter r both for cSW = 0 and
for cSW = 1 are reported in Table 1.
Eqs. (19), (20), evaluated at cSW = 0, coincide with the ones obtained in Ref. [11],
where the Clover term was not included. An explicit check of the correctness of the Clover
contributions to Eqs. (19) and (20) can be obtained from the requirement that the WI (8),
(9) with O = ψ(y)ψ̄ (z) should be satisfied. This condition reads as
(NS)
∆xµ ZA Aaµ (x) ψ(y)ψ̄(z)
a

λa λ
= 2 m0 − M (NS)
ψ̄(x) γ5 ψ(x)ψ(y)ψ̄ (z) − δ(x − y) γ5 ψ(y) ψ̄(z)
2 2

λa a
− δ(x − z) ψ(y) ψ̄(z) γ5 + X (x) ψ(y)ψ̄(z) , (22)
2
where we have taken into account Eq. (16). In momentum space, denoting the amputated
Green’s functions with the insertion of the operators ∆µ Aaµ and (ψ̄λa γ5 ψ/2) by Λa∆Aµ
and Λaγ5 , respectively, one gets

a a
(NS) a
ZA (NS) Λaγ (p , p) − γ5 λ S −1 (p) − S −1 (p )γ5 λ ,
Λ∆Aa (p , p) = 2 m0 − M 5
2 2
(23)
where p (p ) denotes momentum of the incoming (outcoming) quark, and S(p) represents
the quark propagator. The diagrams corresponding to Λa∆Aµ are analogous to the ones
considered for Xa and XC a , reported in Eq. (17). For Λa one has to consider only the
γ5
first diagram, being the pseudoscalar current a local operator. After calculating at one
loop all the terms in Eq. (23), including the Clover contributions, one is able to check
Eqs. (18)–(20). The explicit one-loop expression for the lattice inverse quark-propagator
Table 1
(NS)
Numerical results for C and [ZA − 1] for different values of r and cSW .
(NS)
A factor g02 CF /16π 2 is understood for [ZA − 1] [see Eq. (20)]
r cSW = 0 cSW = 1
(NS) (NS)
C ZA − 1 C ZA −1
0.0 0 0 0 0
0.2 8.09 −3.62 5.29 −2.67
0.4 11.41 −5.82 6.20 −4.13
0.6 11.25 −7.00 5.11 −4.87
0.8 10.43 −7.89 3.96 −5.37
1.0 9.64 −8.66 3.07 −5.75
S −1 (p) is

−1 g0 2 1
S (p) = i/
p + m0 − CF Σ0 + i/
pΣ1 (p) + m0 Σ2 (p) (24)
16π 2 a
with Σ0,1,2 given by
π
d 4q ∆3 1 ∆5 ∆1
Σ0 = 16π 2
4
r − − + r2
(2π) ∆1 ∆2 2∆1 2∆2 2∆2
−π

1 r2
+ cSW (∆3 ∆5 − ∆4 ) + cSW
2
(∆3 ∆5 − ∆4 ) , (25)
∆1 ∆2 2∆2
1

Σ1 (p) = γE − F0001 + 2 dx x ln a 2 m20 (1 − x) + a 2 p2 x(1 − x)
0
π
d 4q ∆3 1 ∆3 1
+ 16π 2
+ − + (2∆4 − ∆3 ∆5 )
(2π)4 16∆31 8∆1 8∆1 ∆2 8∆21 ∆2
−π

r2 1 1
+ (2 − ∆1 ) + r cSW
2
(∆4 − 4∆3 ) + −9∆1 + 2∆21
4∆2 2∆1 ∆2 8∆2

1

+ 4∆3 + 4∆5 − 2∆6 + 4 + r cSW 2 2
2
−∆3 ∆4 + 4∆23
8∆1 ∆2
1
+ (−4∆3 ∆5 + 2∆3 ∆6 − 4∆3 + 4∆4 − ∆7 )
8∆1 ∆2

1
+ (−2∆1 ∆3 + 9∆3 − 2∆4 ) , (26)
8∆2
1

2 2
Σ2 (p) = 4 γE − F0000 + dx ln a m0 (1 − x) + a p x(1 − x)
2 2
0
π
4∆1 − ∆2
d 4q 1 r2 ∆2
+ 16π 2 − + + 2∆ 1 ∆ 5 + − 4∆ 3
(2π)4 4∆21 ∆2 4∆2 ∆22 4
−π

∆2 4r 2 r2
+ r 4 −2 12 + cSW 2 (−∆3 ∆5 + ∆4 ) + cSW 2
(∆3 ∆5 − ∆4 )
∆2 ∆2 4∆1 ∆2
4
2 2r ∆1
+ cSW (−∆3 ∆5 + ∆4 ) (27)
∆22
with F0000 = 4.36898, F0001 = 1.31096 and γE = 0.577216. Eqs. (25)–(27) coincide with
the analytical results of Refs. [11,12] (at cSW = 0) and Ref. [8] (at cSW = 1).
The check of Eqs. (18)–(20) via the WI (23) can proceed now through the projection of
p − p
all the quantities onto the structures (/ / )γ5 and γ5 (orthogonal to each other) with the
appropriate normalizations. In other words we calculate the quantities

1

ΠΓ (p , p) = Tr T (p , p)Γ , (28)
NΓ
where the trace is taken over the Dirac indices, T (p , p) stands for any of the terms
appearing in Eq. (23), Γ is equal to (/p − p/ )γ5 or γ5 and NΓ is a normalization constant.
We have calculated all the relevant projections and checked positively the correctness of
Eqs. (19), (20).
Note that the projection of the amputated diagram of the (extended) axial current
over (/ p − p/ )γ5 and of the pseudoscalar density over γ5 can be compared with the
corresponding quantities calculated in the continuum MS-scheme in order to extract the
matching constants between MS and the lattice (see Section 3.1), because the dependence
on the external momenta is cancelled out in the difference between the two schemes.
Finally, from Eq. (22) it can be noticed that the explicit contribution of the operators Xa
and XC a has been traded with a redefinition of the lattice axial current and of the bare mass
m0 , i.e.,
(NS)
∆µ Aaµ −→ ZA ∆µ Aaµ ,
(NS) ≡ mL .
m0 −→ m0 − M (29)
3.1. Matching with the continuum in the non-singlet channel
The Green’s functions calculated with the (bare) fields and operators on the lattice can
be matched to the corresponding ones renormalized in the MS scheme, by rescaling the
fields and the operators on the lattice with appropriate constants. For instance, the bare
1/2
quark field on the lattice ψL (x) can be rescaled through ψL (x) = Zψ ψR (x), where the
subscript R indicates a renormalized (MS) quantity in the continuum. The constant Zψ can
be evaluated via the relation

−1 R
L
S (p) = Zψ S −1 (p) (30)
which simply follows from the definition of the quark propagator. At one-loop order the
relations

−1 L g2 g2
S (p) = i/ p 1 − 0 2 CF Σ1L (p) + mL 1 − 0 2 CF Σ2L (p) ,
16π 16π

−1 R g02 g02
S (p) = i/ p 1− CF Σ1 (p) + mR 1 −
R R
CF Σ2 (p) , (31)
16π 2 16π 2
imply
g02
Zψ = 1 − CF ∆Σ1 , ∆Σ1 = Σ1R (p) − Σ1L (p). (32)
16π 2
Analogously, one can define a matching for the mass on the lattice from the relation
mR = Z m mL , (33)
obtaining, thanks to the use of Eq. (31) and of the definition (32),

g2
Zm = Zψ 1 − 0 2 CF ∆Σ2 . (34)
16π
A similar procedure leads to the definition of the Z’s for the matching of composite
operators. In the non-singlet channel, we are interested in Green’s functions involving
fermion bilinear operators O contracted with two external quark fields, with amputation
on the external legs. The corresponding functions in momentum space will be indicated by
ΛO (p , p). One gets the relation

O (p , p) = Zψ ZO ΛO (p , p)
ΛR L
(35)
which defines ZO as the constant of matching for the operator O. In perturbation theory
one can express the constants in Eq. (35) as
g02
Zψ ZO = 1 + CF ∆O , ∆O = Λ̂R
O − Λ̂O ,
L
(36)
16π 2
where the expressions Λ̂ collect all the one-loop diagrams. The relation (36) implies
g02
ZO = 1 + CF (∆O + ∆Σ1 ). (37)
16π 2
Denoting the subtraction scale in the continuum by the symbol µ, our results for ∆Σ1
and ∆Σ2 are
∆Σ1 = 1 − ln a 2µ2 − γE + F0001

π 4
d q ∆3 1 ∆3
− 16π 2
+ −
(2π)4 16∆31 8∆1 8∆1 ∆2
−π
1 r2
+ (2∆ 4 − ∆ 3 ∆ 5 ) + (2 − ∆1 )
8∆21 ∆2 4∆2

1
+ r 2 cSW (∆4 − 4∆3 )
2∆1 ∆2

1

+ −9∆1 + 2∆21 + 4∆3 + 4∆5 − 2∆6 + 4
8∆2

1
1
+ r 2 cSW
2
−∆ 3 ∆ 4 + 4∆ 2
3 + (−4∆3∆5 + 2∆3 ∆6
8∆21 ∆2 8∆1 ∆2

1
− 4∆3 + 4∆4 − ∆7 ) + (−2∆1 ∆3 + 9∆3 − 2∆4 ) , (38)
8∆2
π 4
1 d q 4∆1 − ∆2 1
∆Σ2 =4 − ln a µ − γE + F0000 − 16π
2 2 2
− +
2 (2π)4 4∆21 ∆2 4∆2
−π
Table 2
Numerical results for Σ0 and ∆Σ1,2 (see Eqs. (25),
(38) and (39)) for different values of r and with
cSW = 0. The choice µ = 1/a is understood
r Σ0 ∆Σ1 ∆Σ2
0.0 0 −6.04 33.17
0.2 −19.79 −7.13 18.63
0.4 −30.70 −8.91 8.35
0.6 −38.29 −10.47 3.86
0.8 −44.96 −11.77 1.55
1.0 −51.43 −12.85 0.10
Table 3
The same as in Table 2 but with cSW = 1
r Σ0 ∆Σ1 ∆Σ2
0.0 0 −6.04 33.17
0.2 −12.04 −6.52 21.81
0.4 −18.31 −7.40 14.69
0.6 −22.98 −8.16 11.95
0.8 −27.44 −8.75 10.74
1.0 −31.99 −9.21 10.10

r2 ∆2 ∆21
+ 2 2∆1 ∆5 + − 4∆ 3 +r
4
−2 2
∆2 4 ∆2
4r 2 r2
+ cSW 2
(−∆3 ∆5 + ∆4 ) + cSW
2
(∆3 ∆5 − ∆4 )
∆2 4∆1 ∆2
4
2 2r ∆1
+ cSW (−∆3 ∆5 + ∆4 ) . (39)
∆22
The numerical results obtained for Eqs. (25), (38) and (39) at various values of r, are
reported in Tables 2 and 3 for the two cases cSW = 0 and cSW = 1, respectively.
As for the quantities ∆I , ∆γ5 and ∆ext
γµ γ5 one obtains
π 4
d q ∆3 ∆3 ∆5
∆I = 4 −1 + ln a µ − F0001 + γE − 16π
2 2 2
− +
(2π)4 16∆1 4∆1 ∆22
3
−π
2
r2 4 ∆1 r2
+ (−∆ 1 ∆ 5 + 3∆ 3 ) + r + c SW (∆3 ∆5 − ∆4 )
4∆22 4∆22 ∆22

r 2 ∆3 4
2 r ∆1
+ cSW
2
(−∆3 ∆5 + ∆4 ) + cSW (∆3 ∆5 − ∆4 ) , (40)
4∆1 ∆22 4∆22
π 4
d q ∆3 ∆3 ∆5
∆γ5 = 4 −1 + ln a µ − F0001 + γE − 16π
2 2 2
− +
(2π)4 16∆31 4∆1 ∆22
−π
2
r2 4 ∆1 r 2 ∆3
+ (∆ 1 ∆ 5 + ∆ 3 ) + r + c 2
SW (∆3 ∆5 − ∆4 )
4∆22 4∆22 4∆1 ∆22
4
2 r ∆1
+ cSW (∆3 ∆5 − ∆4 ) , (41)
4∆22
(2)R (2)
γµ γ5 = ΠJ 5 − Π 5,ext
∆ext
µ Jµ
= −2 + ln a µ − F0001 + γE
2 2
π 4
d q ∆3 ∆3 1
− 16π 2
− + −
(2π)4 3
16∆1 ∆1 ∆2 2 8∆ 1
−π
∆3 1

+ + 2
3∆3 + ∆23 − 5∆4 + 2∆7
4∆1 ∆2 ∆1 ∆2
1 ∆1
+ (∆1 ∆3 − 6∆3 + 2∆4 ) + r 2
2∆22 4∆2
r2
+ −3∆1∆3 − 4∆1 ∆5 + 2∆1 ∆6 + 4∆1 + 3∆21

2∆22
r 4 ∆21 r2
− ∆31 + 7∆3 − 2∆4 + (2 − ∆ 1 ) + c SW −7∆1∆3 + 4∆1 ∆4

2∆22 ∆22

+ 2∆21 ∆3 + 8∆3 ∆5 − 4∆3 ∆6 − 4∆3 − 15∆4 + 6∆7
r2
+ cSW
2
3∆3 ∆4 − 4∆3 ∆7 − 4∆23 ∆5 + 2∆23 ∆6 + 4∆23 + 2∆24
2∆1 ∆22

4
2 r ∆1

+ ∆1 3∆23 − ∆3 ∆4 − ∆1 ∆23 + cSW 2

4∆3 ∆5 − 2∆3∆6
2∆2

− 12∆4 + 4∆7 − 4∆1 ∆3 + 3∆1 ∆4 + ∆21 ∆3 . (42)
We notice that using Eqs. (38) and (42) one has

g02 (NS)
2 γµ γ5 + ∆Σ1 = ZA
CF ∆ext − 1, (43)
16π
(NS)
where ZA is given by Eq. (20).
For completeness, we report also the matching constants for the operators ψ̄(x)γµ ψ(x),
ψ̄(x)γµ γ5 ψ(x) and ψ̄(x)σµν ψ(x) [8]:
π
d 4q ∆3 ∆4
∆γµ = −2 + ln a µ − F0001 + γE − 16π
2 2 2
− +
(2π)4 16∆1 ∆1 ∆22
3
−π
2
r2 4 ∆1 r 2 cSW
+ (∆ 1 ∆ 5 + 3∆ 3 ) + r + (−∆3 ∆5 + ∆4 )
2∆22 ∆22 ∆22

r 2 cSW
2 ∆
3
+ (∆3 ∆5 − ∆4 ) , (44)
2∆1 ∆22
π
d 4q ∆3 ∆4
∆γµ γ5 = −2 + ln a µ − F0001 + γE − 16π
2 2 2
4
− +
(2π) 16∆1 ∆1 ∆22
3
−π
2
r2 4 ∆1 r 2 cSW
+ (−∆ 1 ∆ 5 + 5∆ 3 ) + r + (∆3 ∆5 − ∆4 )
2∆22 ∆22 ∆22
2 ∆
r 2 cSW 3
+ 2
(∆4 − ∆3 ∆5 ) , (45)
2∆1 ∆2
π
d 4q 1
∆3 ∆1
∆σµν = −16π 2
4 2
∆1 ∆3 − 4(∆3 − ∆4 ) + 2r 2 2 + r 4 12
(2π) 3∆1 ∆2 ∆2 ∆2
−π

2r 2 c
SW r 4 cSW
2 ∆
1
+ (∆4 − ∆3 ∆5 ) + (∆4 − ∆3 ∆5 ) . (46)
3∆22 3∆22
The numerical values of the matching constants ∆I , ∆γ5 , ∆γµ , ∆γµ γ5 , ∆σµν and ∆ext
γµ γ5 ,
obtained for various values of r, are collected in Tables 4 and 5. Our results for the local
operators have been explicitly checked against the corresponding ones of Refs. [8,13,14],
γµ γ5 at cSW = 0 coincide with the ones of Ref. [15].
while those for ∆ext
Table 4
Numerical results for various ∆O for different values of r and with cSW = 0
r ∆I ∆γ5 ∆γµ ∆γµ γ5 ∆σµν ∆ext
γµ γ5
0.0 −33.16 −33.16 −8.74 −8.74 0.74 6.04
0.2 −18.63 −26.72 −8.97 −4.92 −0.37 3.52
0.4 −8.34 −19.75 −8.76 −3.05 −1.86 3.09
0.6 −3.85 −15.10 −8.36 −2.73 −2.90 3.47
0.8 −1.55 −11.98 −8.02 −2.80 −3.63 3.88
1.0 −0.10 −9.74 −7.76 −2.94 −4.17 4.19
Table 5
The same as in Table 4 but with cSW = 1
r ∆I ∆γ5 ∆γµ ∆γµ γ5 ∆σµν ∆ext
γµ γ5
0.0 −33.16 −33.16 −8.74 −8.74 0.74 6.04
0.2 −21.80 −27.09 −8.27 −5.62 0.23 3.85
0.4 −14.69 −20.89 −7.45 −4.35 −0.61 3.27
0.6 −11.94 −17.05 −6.82 −4.27 −1.23 3.29
0.8 −10.73 −14.70 −6.40 −4.42 −1.64 3.38
1.0 −10.10 −13.17 −6.12 −4.59 −1.93 3.46
In order to get a further independent check of the expression (19) for the constant C one
can consider the relation
(NS)
g2 Σ0 m0 − M
mR = Zm m0 − 0 2 CF = , (47)
16π a Zγ5
where the first equality follows from the definition (33), while the second is a consequence
of Eq. (22) and of the definition (35). Now, using the one-loop expressions
g02
Zm = 1 − CF (∆Σ1 + ∆Σ2 ),
16π 2

g02 Σ0

M (NS)
= CF + m0 C ,
16π 2 a
g02
Zγ5 = 1 + CF (∆γ5 + ∆Σ1 ), (48)
16π 2
one easily finds
C = −∆Σ2 − ∆γ5 , (49)
which can be checked analytically using Eqs. (19), (39) and (41).
4. Singlet axial WIs
Let us now turn to the singlet WI (14), (15). Because of the anomaly, the operator X0 +
0 , representing the chiral variation of the Wilson and Clover terms, can be written as [11]
XC
X0 (x) + XC
0
(x)

=X 0 (x) − 2M
0 (x) + X (S) ψ̄(x)γ5 ψ(x) + 1 − Z (S) ∂µ Aµ (x)
C A
g02
+ 2NF Z GG sub (x), (50)
32π 2 GG
where, again, X 0 and X
0 indicate dimension-5 operators vanishing in the continuum limit
C
in all matrix elements involving elementary operator insertions, and the symbol GG sub in-
dicates a lattice discretization of the corresponding continuum operator GG. In Eq. (50)
M (S) is the additive mass-renormalization constant, while Z (S) and ZGG
are multiplicative
A
constants in the singlet channel.
Before addressing the calculation of the renormalization constants, we point out that in
Eq. (50) GG sub should be defined in such a way to avoid ambiguities with the pseudoscalar
density. Indeed, let us consider the commonly used discretization GG L obtained through
the use of the symmetric plaquette, viz. [16]

GG L (x) ≡ εµνρσ Tr Pµν (x)Pρσ (x) , (51)
where Pµν is defined in Eq. (5) and the sum over the greek indices is understood.
Expanding the definition (51) in powers of a, one gets
L (x) = GG(x)
GG + a 2 O6 (x) (52)
where GG = 1/2εµνρσ Gaµν Gaρσ and O6 denotes a dimension-6 operator. Since with
Wilson fermions chiral symmetry is broken, there is no guarantee that the operator O6 (x)
does not mix with the pseudoscalar density. In case of mixing one expects by dimensional
reasons a (cubic) divergence, leading to a linear-divergent mixing of GG L with the
pseudoscalar density. Such a divergence has to be subtracted and therefore the operator
GGsub , appearing in Eq. (50), should be related to GG
L by
sub (x) = GG
GG L (x) − :M
L ψ̄(x)γ5 ψ(x) (53)
where :M L contains the 1/a divergence and is in general dependent on the particular
discretization GG L . Up to now, a non-perturbative definition of :M L is not known, and
we anticipate here that at one-loop order we find no mixing, i.e., :M L = O(g 4 ). However
0
a non-perturbative definition of :M L is mandatory, particularly outside the chiral limit. To
this end we require that GG sub should be a chiral singlet.
First of all our requirement on GG sub is needed to guarantee the proportionality between
(m0 − M (NS) ) and (m0 − M (S)), as it is demonstrated in Ref. [9]. This implies that: (i) the
chiral point is the same in the singlet and non-singlet channels; and (ii) the renormalized
mass is the same in the two channels and it can be therefore identified with the continuum
renormalized mass. The latter point follows from a suitable choice of the renormalization
constants for the singlet and octet pseudoscalar densities, which in turn arises from the
transformation properties of both scalar and pseudoscalar densities as members of the same
chiral multiplet.
Let us now consider the non-singlet WI and the case O = S a (y1 , y2 )GG sub (z), where
S (y1 , y2 ) = ψ̄(y1 )λa ψ(y2 )/2 with y1 = y2 ; one has
a

∆xµ Âaµ (x)S a (y1 , y2 )GG sub (z)

= 2mL P5a (x)S a (y1 , y2 )GG sub (z)

− CF δ(x − y1 ) + δ(x − y2 ) P5 (y1 , y2 )GG sub(z)
a
+ X (x) + X a (x) S a (y1 , y2 )GG sub (z) , (54)
C
(NS) a
where P5a (x) = ψ̄(x)λa γ5 ψ(x)/2, P5 (y1 , y2 ) = ψ̄(y1 )γ5 ψ(y2 ), Âaµ (x) = ZA Aµ (x),
mL = m0 − M (NS) and the sum over the flavor index a (a = 1, . . . , 8) is understood. In
Eq. (54) we have considered that δGG sub (z)/δ(iα a (x)) = 0, because GG
sub is assumed to
be a chiral singlet. After integration over the whole space–time in x and z2 one gets

0 = mL sub (z)
d 4 x d 4 z P5a (x)S a (y1 , y2 )GG

− CF P5 (y1 , y2 ) d 4 z GG sub (z)
2 The integration over x is used to cancel out the l.h.s. of Eq. (54), including possible contact terms between
a
Âµ (x) and GGsub (z). The integration over z allows to get rid of the mixing of GG sub (z) with the singlet
R to
∂µ Aµ (x) (see Eq. (59)). Note that such a mixing does not change the property of the renormalized GG
be a chiral singlet.

1 a
+ d 4x d 4z X (x) + X a (x) S a (y1 , y2 )GG
sub (z) . (55)
C
2
We have now to discuss the possible presence of contact terms which may arise in
the first and third terms of the r.h.s. of Eq. (55) when x ≈ z. Note that the operator
S a (y1 , y2 ) is a string of elementary operators taken at different space–time points y1 = y2 ;
therefore it cannot generate contact terms when inserted with composite operators. The
general structure of the possible contact terms in Eq. (55) can be derived using the results
of Ref. [17], where the method of functional integral with the generalized mass term,
σ a ψ̄λa ψ/2 + iπ a ψ̄λa γ5 ψ/2, was developed to derive WIs. The only possible contact
terms should be as follows

d 4 x d 4 z P5a (x)S a (y1 , y2 )O (z),

a
d 4x d 4z X (x) + X C
a
(x) S a (y1 , y2 )O (z),

δO (z)
d 4 x d 4 z S a (y1 , y2 ) b δ5a π b (x), (56)
δπ (x)
where δ5a π b (x) is the chiral variation of the mass field π b (x) and O (z) is an operator
with the same transformation properties and dimension of GG sub (z), i.e., a pseudo-scalar,
chiral-singlet dimension-4 operator, build up with the fields ψ̄(z), ψ(z), σ a (z) and π a (z)
[17]. Since O (z) should be a chiral singlet, the third form of the contact terms (56) is
vanishing. Moreover, the only dimension-4, chiral singlet operator allowed is just the
generalized mass term, i.e.: O = σ a ψ̄λa ψ/2 + iπ a ψ̄λa γ5 ψ/2, which however is not
pseudoscalar. Thus, we conclude that no contact terms are present in Eq. (55). Finally,
according to Refs. [11,17] in the continuum limit the third term in the l.h.s. of Eq. (55)
vanishes and therefore it will be disregarded.
Substituting GG sub (z), defined in Eq. (53), into Eq. (55) one gets

L (z) + mL d 4 x d 4 z P a (x)S a (y1 , y2 )GG
−CF d 4 z P5 (y1 , y2 )GG L (z)

:ML = 5
.
−CF d z P5 (y1 , y2 )P5 (z) + mL d x d z P5 (x)S (y1 , y2 )P5 (z)
4 4 4 a a
(57)
Note that our non-perturbative definition of :M L is gauge-invariant because the loss of
gauge invariance is given by a factor which is the same in the numerator and in the
denominator of the r.h.s. of Eq. (57).
Let us now address the calculation of the renormalization constants M (S) , Z (S) and
A
Z GG
relevant to the singlet WI. Substituting Eq. (50) into the WI (14), (15), one gets
(S)
∆xµ ZA Aµ (x)O

g2
= 2 m0 − M (S) ψ̄(x)γ5ψ(x)O + 2NF 0 Z GG sub (x)O
G G
32π 2
δO
− , (58)
δ(iα a (x))
where we have dropped the term involving X 0 + X 0 . Because of its anomalous non-
C
conservation the singlet current Aµ suffers a logarithmically divergent renormalization
[18]. A mixing between ZA (S)
Aµ and ZGG sub is needed to obtain finite operators [9]
GG
(S)
2 (S)
ARµ (x) = ZA Aµ (x) − g0 ZC ZA Aµ (x),
1 R (x) = 2NF 1 ZGG sub (x) − ZC Z (S) ∂µ Aµ (x),
2NF GG GG A (59)
32π 2 32π 2
where we stress that GG sub (x) = GG L (x) − :M L ψ̄(x)γ5ψ(x). The above redefinition
(S)
can be performed into Eq. (50) by adding and subtracting the counterterm [ZC ZA ·
∂µ Aµ (x)] [6].
We have evaluated at one loop level all the constants appearing in Eq. (59). The constant
(S)
ZA can be obtained by evaluating X0 (x) + XC 0 (x)ψ(y)ψ̄(z): to O(g 2 ) its value is
0
(NS)
identical to that of ZA , Eq. (20), while beyond the one-loop order the two constants will
differ because of the contributions (in the singlet channel) coming from fermion loops. The
same is true for M (S) , which therefore at one-loop is given by Eq. (18), as well as for the
renormalization constants of the pseudoscalar density operators in both the singlet and non-
singlet channels. This implies that at one-loop the renormalized masses (m0 − M (S) )/Zγ(S)
5
and (m0 − M (NS) )/Zγ(NS)
5 coincide and match the corresponding renormalized mass in the
continuum. We stress that the equality between the renormalized masses in the singlet and
non-singlet channels is a more general result, valid at any order of the perturbation theory
and resulting from the property of GG sub to be a chiral singlet, as illustrated in Ref. [9].
The mixing between X0 (x) + XC 0
(x) and GG sub (x) can be computed at one loop by
evaluating the correlator [X (x)+XC (x)]Gβ (y)Gcγ (z). This was carried out in Ref. [10],
0 0 b
where it was shown that the apparent dependence of the correlator on r actually disappears
as far as r = 0, and that the tree level expression 2NF g02 /32π 2 GG is reproduced exactly.
= 1 + O(g0 ), and, since the logarithmic divergence in
Thus, from Eq. (50) one has ZGG 2
the singlet axial current manifests itself at the two-loop level, it follows that g02 ZC = O(g04 )
and consequently ZC = O(g02 ).
The same considerations hold as well for the matching constants between the continuum
and the lattice. For ZGG and ZC appearing in Eq. (59), we can write

4
= 1 + g0 zgg + O g0 ,
2 (2)
Z GG
g02
(2)
2 2 (2)

ZC = 2
CF zgψ ln a µ + z̃gψ + O g04 . (60)
16π
(2)
The coefficient zgg can be computed at one-loop in a pure gauge theory (see Ref. [19] and
references therein quoted), and at this order it is unaffected by the presence of the Clover
term in the action.3 One obtains [19]

1 1 1
zgg = N − 2 + Z0000 + +
(2)
, (61)
4N 8 2π 2
(2)
where N is the dimension of the gauge group and Z0000 = 0.15493. The constants zgψ
(2)
and z̃gψ can be computed by evaluating, in the continuum and on the lattice, the correlator

GG(x)ψ(y) ψ̄(z), with amputated external quark propagators. On the lattice GG should

be GGsub given by Eq. (53). However, we have explicitly checked that the operator GG L
3 Beyond O(g 2 ) the presence of the Clover term in the action will, in general, affect also Z
0 .
GG
does not mix at one-loop with the pseudoscalar density and, therefore, in our leading-order
approximation there is no difference between GG sub and GG L . Thus, on the lattice we
have to compute the following diagrams:

L (x)ψ(y)ψ̄(z) (2) =
FT GG + , (62)
amp
where the ψψg and ψψgg vertices include the contributions of both the Wilson and
Clover action. The second diagram gives a null contribution, since the insertion of GG
contracted with two gluon lines is antisymmetric in the Lorentz indices of the gluons (via
the ε-tensor) and symmetric in their color indices, whereas the ψψgg is symmetric in the
Lorentz indices for the Wilson case and antisymmetric in the color matrices for the Clover
case (see the Feynman rules in Appendix A).4 The first diagram displays a log divergence
as a → 0, in analogy with the continuum calculation. The difference between the two
schemes gives the following result
(2)
zgψ = −6,

(2) d 4 q 3∆3 1 1
z̃gψ = 22 − 6(γE − F0001 ) − 16π 2 −
(2π)4 2∆21 4∆1 ∆2
1

+ 2
10∆3 ∆4 − 30∆3 − 19∆23 + 64∆4 − 46∆7 + 12∆8
4∆1 ∆2
1

+ 45∆3 + 6∆23 − 40∆4 + 12∆7
4∆1 ∆2
1
+ (2∆1 ∆3 − 17∆3 + 6∆4 )
4∆2
1
+ r 2 cSW 8∆3 ∆4 − 3∆3 ∆25 + 3∆3∆6 − 4∆23 ∆5

2∆1 ∆2

+ 10∆4 ∆5 + 2∆4 ∆25 − 2∆4 ∆6 − 8∆5 ∆7 − 14∆7 + 12∆8

1

+ ∆3 ∆25 − ∆3 ∆6 − 2∆4 ∆5 + 2∆7
∆2
1
+ r 2 cSW
2
−10∆3∆4 ∆5 − 2∆3 ∆4 ∆25 + 2∆3 ∆4 ∆6 + 8∆3 ∆5 ∆7
4∆21 ∆2

+ 14∆3 ∆7 − 12∆3 ∆8 − 8∆23 ∆4 + 3∆23 ∆25 − 3∆23 ∆6 + 4∆33 ∆5

1

+ 2∆3 ∆4 ∆5 − 2∆3 ∆7 − ∆3 ∆5 + ∆3 ∆6
2 2 2
. (63)
2∆1 ∆2
4 For sake of completeness we mention that in addition to the diagrams (62) one should consider also the
L − 3 gluon vertex. The latter is trivially vanishing after contraction over
tadpole diagram generated by the GG
color indices.
Table 6
(2)
Numerical results for z̃gψ (see Eq. (63))
for different values of r and cSW
(2) (2)
r z̃gψ , cSW = 0 z̃gψ , cSW = 1
0.0 19.82 19.82
0.2 19.72 19.93
0.4 19.11 19.98
0.6 18.11 19.94
0.8 16.95 19.87
1.0 15.77 19.81
(2)
The numerical values of the constant z̃gψ obtained for various values of r at cSW = 0
and cSW = 1 are reported in Table 6.
4.1. Composite insertions of the topological charge and the neutron EDM
In this subsection we take the opportunity to briefly address a separate issue concerning
the composite insertion of the topological charge relevant for a lattice evaluation of
the neutron EDM induced by the strong θ -term (1). Following Ref. [6] the standard

definition of the neutron EDM, dN , involves the insertion of the topological charge

(g 2 /32π 2 ) d 4 x GG(x) in the presence of the charge density operator J0 . Treating the
θ -term as a perturbation at first order, one has

g2
dN ≡ −iθ d 3
y
y 0 N|J0 (y) d 4
x G
G(x) |N0 , (64)
32π 2
where |N0 is a shorthand for |Nθ=0 . In case of three flavors with non-degenerate masses a
complete diagrammatic analysis was performed in Ref. [6] showing how the axial anomaly
governs the replacement of the topological charge operator with well-defined insertions of
the flavor-singlet pseudoscalar density.5
Thus the question is the possible presence of contact terms in the composite insertion
of the topological charge operator with the electromagnetic (e.m.) current operator, which
would lead to ambiguities in the numerical evaluation of Eq. (64). The contact terms in
Eq. (64) should be of the form Oµ=0 P (x) · δ (4) (x − y), where O P is a local operator of
µ
dimension-3, which transforms as a non-singlet pseudo-vector and is conserved. The last
property derives from the fact that the e.m. current is conserved both with and without the
θ -term in the QCD action at any value of the parameter θ . The non-singlet nature of OµP is
related to the non-singlet nature of the e.m. current operator. The only candidate for OµP is
the non-singlet axial current which however is not conserved (outside the chiral limit).
The absence of contact terms in Eq. (64) can be derived also using the generalized mass
insertion method of Ref. [17] applied to a vector WI with the operator O given by GG.
5 We point out that the final result of Ref. [6] crucially depends on the equality between the renormalized
masses in the singlet and non-singlet WIs. We stress once more that such an equality follows from the property
sub (see Eq. (53)) to be a chiral singlet (cf. Ref. [9]).
of GG
5. Conclusions
A complete one-loop calculation of the renormalization constants appearing in both

singlet and non-singlet axial Ward identities using Wilson fermions with the Clover
O(a)-improvement of the action has been performed. Our calculations include: (1) the
contributions arising from the Clover term of the action; (2) the complete mixing of the
gluon operator GG with the divergence of the singlet axial current; (3) the use of both
local and extended definitions of the fermionic bilinear operators.
In the singlet channel a definition of the gluon operator GG on the lattice outside
the chiral limit has been proposed. Our definition takes into account the possible power-
divergent mixing with the pseudoscalar density, generated by the breaking of chiral
symmetry. No mixing has been found at one-loop order and a non-perturbative definition
of the mixing constant has been developed.
Finally, the renormalization
properties of the composite insertion of the topological

charge operator d 4 x GG(x) relevant for the lattice calculation of the neutron electric
dipole moment have been discussed, showing that no contact terms arise when the
topological charge is inserted with a conserved current.
Acknowledgements
The authors gratefully acknowledge M. Testa for many useful discussions, which have
been essential for clarifying the non-perturbative definition of the mixing between the
gluon operator GG on the lattice and the singlet pseudoscalar density. We warmly thank
V. Lubicz and G. Martinelli for many useful comments and a critical reading of the
manuscript, as well as L. Giusti for valuable discussions.
Appendix A. Feynman rules
The Wilson action is given by

1
SF = a 4
− ψ̄(x)(r − γµ )Uµ (x)ψ(x + µ)
x
2a µ

4r
+ ψ̄(x + µ)(r + γµ )Uµ† (x)ψ(x) + ψ̄(x) m0 + ψ(x) , (A.1)
a
with the following notation {γµ , γν } = 2δµν , σµν = [γµ , γν ]/2, Uµ (x) = exp[ig0 aGµ (x)],
Gµ = Gaµ t a and Tr(t a t b ) = δab /2.
The action is improved to O(a) via the Clover term [20]
ig0 ar
SC = −a 4 cSW ψ̄(x)σµν Pµν (x)ψ(x), (A.2)
x µ,ν
4
with Pµν being the usual lattice definition of the field-strength tensor Gµν [8]
1 1
4

Pµν (x) = 2
Ui − Ui† (A.3)
4a 2ig0
i=1
where the sum is over the four plaquettes in the µ–ν plane stemming from x and taken in
the counterclockwise sense, i.e.,
U1 = Uµ (x)Uν (x + µ)Uµ† (x + ν)Uν† (x),

U2 = Uν (x)Uµ† (x − µ + ν)Uν† (x − µ)Uµ (x − µ),
U3 = Uµ† (x − µ)Uν† (x − µ − ν)Uµ (x − µ − ν)Uν (x − ν),
U4 = Uν† (x − ν)Uµ (x − ν)Uν (x + µ − ν)Uµ† (x). (A.4)
A.1. Propagators
k̂µ k̂ν
δµν − (1 − α)
k̂ 2
≡ Gµν (k) = (A.5)
k̂ 2
with

2 pµ a
p̂µ = sin , p̂2 = p̂µ2 . (A.6)
a 2 µ
≡ S0 (p)

1 2r 2 pλ a −1
= iγλ sin(pλ a) + m0 + sin . (A.7)
a a 2
λ λ
A.2. QCD vertices
The indices W and C refer to the Wilson and Clover action respectively [8]:
≡ Vµa
W
(q, q )

a a
= −g0 t a ij iγµ cos (q + q )µ + r sin (q + q )µ ,
2 2
(A.8)
≡ Vµa
C
(q, q )

a r

a
= −g0 t ij σµν sin (q − q )ν a cos (q − q )µ ,
2 ν 2
(A.9)

ag02 a
= Vρa,σ
W
b (p, p ) = δρσ t a , t b ij iγρ sin (p + p )ρ
2 2

a
− r cos (p + p )ρ .
2
(A.10)
A.3. Extended operators
For simplicity we report only the Feynman rules for flavor-singlet operators.
A.3.1. Operator ∆µ Aµ (x)

a∆µ Aµ (x) = Aµ (x) − Aµ (x − µ),
1
Aµ (x) = ψ̄(x + µ)γµ γ5 Uµ† (x)ψ(x) + ψ̄(x)γµ γ5 Uµ (x)ψ(x + µ) . (A.11)
2
i

=− γµ γ5 sin(pµ a) − sin(pµ a) , (A.12)
a µ
g0
iq a
−ip a
= γρ γ5 t a ij e ρ 2 e µ − eipµ a
2 µ
a

+ e−iqρ 2 eipµ a − e−ipµ a , (A.13)
ag02
=− γρ γ5 δρσ t a , t b ij
4
a

eiqρ 2 e−iqρ 2 e−ipµ a + eipµ a
a
×
µ
a a

− e−iqρ 2 eiqρ 2 eipµ a + e−ipµ a . (A.14)
A.3.2. Operator X0
r
X0 (x) = − ψ̄(x)γ5 Uµ (x)ψ(x + µ) + ψ̄(x + µ)γ5 Uµ† (x)ψ(x)
2a µ

+ (x → x − µ) − 4ψ̄(x)γ5 ψ(x) . (A.15)
r

= γ5 2 − cos(pµ a) − cos(pµ a) , (A.16)
a µ
ig0 r
a iqρ a
ipρ a
=− γ5 t ij e 2 e − e−ipρ a
2
a

+ e−iqρ 2 eipρ a − e−ipρ a , (A.17)
g02 ar iq a −iq a
ip a
= γ5 δρσ t a , t b ij e ρ 2 e ρ 2 e µ + e−ipµ a
4 µ
a a

+ e−iqρ 2 eiqρ 2 eipµ a + e−ipµ a . (A.18)
0
A.3.3. Operator XC
ig0 ar
0
XC (x) = − ψ̄(x)γ5 σµν Pµν ψ(x). (A.19)
2 µν
= ∅, (A.20)

a

= −g0 rγ5 (t a )ij σρν sin(qν a) cos qρ , (A.21)
ν
2

g02 ar a b a

= γ5 t , t ij δρσ σρν sin (q − q )ρ sin(qν a) + sin(qν a)
2 ν
2

1 a a
+ σρσ cos (qρ − qσ ) − cos (qρ − qσ ) − qρ a
2 2 2

a a
− cos (qρ − qσ ) + qσ a − cos (qρ − qσ ) + (qσ − qρ )a
2 2

a a a
+ cos (qρ + qσ ) − cos (qρ + qσ ) − qρ a − cos (qρ + qσ ) − qσ a
2 2 2

a
− cos (qρ + qσ ) − (qσ + qρ )a . (A.22)
2
L
A.3.4. Operator GG
L = εµνρσ Tr(Pµν Pρσ ).
GG (A.23)

4 a a
= − 2 δab εναµβ cos pν cos rµ sin(pα a) sin(rβ a). (A.24)
a 2 2
αβ
References
[1] G. ’t Hooft, Phys. Rev. D 14 (1976) 3432.

[2] G. ’t Hooft, Phys. Rev. Lett. 37 (1983) 8.
[3] P.G. Harris, et al., Phys. Rev. Lett. 82 (1999) 904.
[4] V. Baluni, Phys. Rev. D 19 (1979) 2227.
[5] R.J. Crewther, et al., Phys. Lett. B 88 (1979) 123;
R.J. Crewther, et al., Phys. Lett. B 91 (1980) 487, Erratum.
[6] D. Guadagnoli, V. Lubicz, G. Martinelli, S. Simula, JHEP 04 (2003) 019, hep-lat/0210044.
[7] K.G. Wilson, Phys. Rev. D 10 (1974) 2445;
K.G. Wilson, in: A. Zichichi (Ed.), New Phenomena in Subnuclear Physics, Plenum, New York, 1977.
[8] E. Gabrielli, G. Heatlie, G. Martinelli, C. Pittori, C.T. Sachrajda, Nucl. Phys. B 362 (1991) 475.
[9] M. Testa, JHEP 9804 (1998) 2, hep-th/9803147.
[10] L.H. Karsten, J. Smit, Nucl. Phys. B 183 (1981) 103.
[11] M. Bochicchio, L. Maiani, G. Martinelli, G.C. Rossi, M. Testa, Nucl. Phys. B 262 (1985) 331.
[12] A. Gonzalez Arroyo, F.J. Yndurain, G. Martinelli, Phys. Lett. B 117 (1982) 437.
[13] G. Martinelli, Y.-C. Zhang, Phys. Lett. B 123 (1983) 433.
[14] S. Capitani, et al., Nucl. Phys. B 593 (2001) 183.
[15] G. Martinelli, Y.-C. Zhang, Phys. Lett. B 125 (1983) 77.
[16] J. Mandula, G. Zweig, J. Govaerts, Nucl. Phys. B 228 (1983) 91.
[17] L. Maiani, G. Martinelli, G.C. Rossi, M. Testa, Nucl. Phys. B 289 (1987) 505.
[18] W.A. Bardeen, Nucl. Phys. B 75 (1974) 246.
[19] B. Allés, A. Di Giacomo, H. Panagopoulos, E. Vicari, Phys. Lett. B 350 (1995) 70.
[20] G. Heatlie, G. Martinelli, C. Pittori, G.C. Rossi, C.T. Sachrajda, Nucl. Phys. B 352 (1991) 266.
MSW-resonant fermion mixing during reheating

Tsuneto Kanai a , Shinji Tsujikawa b
a Department of Physics, Graduate School of Science, The University of Tokyo, 7-3-1, Hongo, Bunkyo-ku,
Tokyo 113-0033, Japan
b Institute of Cosmology and Gravitation, University of Portsmouth, Mercantile House,
Portsmouth PO1 2EG, United Kingdom
Received 27 June 2003; accepted 5 August 2003
Abstract
We study the dynamics of reheating in which an inflaton field couples two flavor fermions through
Yukawa-couplings. When two fermions have a mixing term with a constant coupling, we show
that the Mikheyev–Smirnov–Wolfenstein (MSW)-type resonance emerges due to a time-dependent
background in addition to the standard fermion creation via parametric resonance. This MSW
resonance not only alters the number densities of fermions generated by a preheating process but
also can lead to the larger energy transfer from the inflaton to fermions. Our mechanism can provide
additional source terms for the creation of superheavy fermions which may be relevant for the
leptogenesis scenario.
PACS: 98.80.Cq
1. Introduction
Inflationary scenario [1] is widely accepted as a most reasonable extension to the

standard big bang cosmology. This paradigm not only explains why the present universe is
approximately homogeneous, isotropic and spatially flat, but also provides the quantitative
predictions about the origin of large-scale structure [2]. Fortunately, within observational
errors, all existing and constantly accumulating data have continuously confirmed this
paradigm for these past 10 years [3–6]. We are now entering the golden age where the
inflationary physics can be probed by upcoming high-precision observational data.
E-mail address: kanai@light.phys.s.u-tokyo.ac.jp (T. Kanai).
doi:10.1016/j.nuclphysb.2003.08.008
290 T. Kanai, S. Tsujikawa / Nuclear Physics B 670 (2003) 289–306
In addition to the generation of large-scale density perturbations, the inflationary

paradigm provides us the origin of elementary particles through the reheating process
after inflation. It is now well known that the reheating can begin from the nonperturbative,
violent particle creation, which is called preheating [7–11]. The bosonic particles coupled
to the inflaton are efficiently created by parametric resonance during preheating. This
nonequilibrium phase is followed by a perturbative decay of the inflaton characterized
by the Born process. The reheating stage completes after the thermalization of produced
particles.
In the context of the production of spin 1/2 (and higher) quantum fields, this has
traditionally been considered as the less interesting domain compared to the bosonic
preheating due to the Pauli-blocking [12]. However in the last five years there has been
a growing interest for the application of the fermionic particle creation to cosmology in
a number of profound ways [13–23]. For example, superheavy particles with mass up
to mχ = 1017 –1018 GeV can be resonantly produced during preheating [17], which may
have mediated leptogenesis [24]. The gravitational creation of massive and noninteracting
particles [25,26] potentially leads to an interesting possibility for the origin of the long
lived dark matter and for the cosmic rays observed beyond the Greisen–Zatsepin–Kuzmin
(GZK) cut-off [27].
When we consider particle physics theories such as supergravity, gravitino production
can provide us a useful tool to constrain particle physics models in the early universe [28].
In a perturbative theory of reheating [29], the thermal production of gravitinos places a
constraint Tr 108 GeV [28,30], on the reheat temperature. Early studies of gravitino
production [31] argued that thermal gravitino abundances may be enhanced due to the
nonperturbative dynamics of preheating. Within a few years many works have been devoted
to the nonthermal production of gravitinos during preheating [32–38]. In particular it was
found in Ref. [36] that the gravitino creation is suppressed relative to the superpartner of
the inflaton (inflatino) for a model of two scalar fields including a supersymmetry breaking
field. Similar conclusion has been reached in Ref. [38] for a more realistic supergravity
inflation model.
Nilles et al. derived general fomalisms for the quantization of coupled bosonic and
fermionic fields in a time-dependent external background [37] along with a formalism for
nonthermal gravitino production. In the context of supergravity, it is natural to consider a
system of multiple fields, since more than one scalar field generally contribute to the Kähler
potential [39]. Unfortunately the full lagrangian derived from supergravity generically has
very complicated forms [35]. Even for the simplest model with a massive inflaton and a
Polonyi field [37], it is not so easy to understand how the mixing occurs between fermionic
fields without complex and detailed numerical simulations.
In this work we shall focus on the simpler system where two fermionic fields coupled
to the inflaton are mixing each other during reheating. Since we are interested in general
results, we will work with the spin 1/2 Dirac equation rather with the gravitino (Rarita–
Schwinger) equations. This is particularly important to understand clearly what happens
for the coupled system of spin 1/2 fermions. In the case of multiple scalar fields, it was
shown in Ref. [40] that the mixing between the inflaton and another scalar field gives rise
to the oscillations between the perturbation of the inflaton and the perturbations of other
T. Kanai, S. Tsujikawa / Nuclear Physics B 670 (2003) 289–306 291
degrees of freedom during inflation. In particular the probability of oscillations are found
to be resonantly amplified around the horizon crossing.
In the context of coupled fermions we shall show that the Mikheyev–Smirnov–
Wolfenstein (MSW)-type resonance [41] emerges due to the oscillation of the inflaton
during reheating. This is completely analogous to the MSW resonance of neutrino
oscillations which occurs when neutrinos propagate in a background of slowly varying
electron densities. In fact our scenario can also be regarded as a system of neutrino
oscillations between two different flavor eigenstates in a time-dependent background. In
addition to the MSW resonance between two fermions, inclusion of the mixing term
shifts the resonance points where fermions are created by preheating, thereby changing
the occupation numbers of fermions derived in the standard analysis [15–17]. In fact we
will show that the energy density transferred from the inflaton to fermions can get larger
by taking into account the mixing between fermions.
This paper is organized as follows. In the next section, we review the formalism for the
quantization of coupled fermionic fields developed in Ref. [37]. In Section 3 the structure
of the MSW resonance in reheating will be discussed in addition to the particle creation
due to parametric resonance. In Section 4 we shall show the numerical results about the
number densities of fermions and the energy transfer from the inflaton to fermions. We
give our conclusions and discussions in the final section.
2. Formalism for the quantization of fermions
Let us consider a model where the inflaton, φ, couples to fermions of two flavor
eigenstates, ψα and ψβ , through the Yukawa couplings hα and hβ ,

√ 1 1
S = d 4 x −g R − ∂µ φ∂ µ φ − V (φ) + i ψ̄α (γ̄ µ Dµ − mα )ψα − hα φ ψ̄α ψα
2κ 2 2

+ i ψ̄β (γ̄ µ Dµ − mβ )ψβ − hβ φ ψ̄β ψβ − δ(ψ̄α ψβ + ψ̄β ψα ) ,
(2.1)
where R is a scalar curvature and κ 2 /(8π) ≡ G = m−2
pl is a Newton’s gravitational constant
with mpl 1.22 × 1019 GeV a Planck mass. In this work we adopt the quadratic potential
1
V = m2φ φ 2 , (2.2)
2
where the inflaton mass, mφ , is constrained to be mφ 1013 GeV from the COBE
normalization [2].
In the action (2.1) mα and mβ are bare masses of fermions, γ̄ µ is the curved-space
αβ
Dirac matrices, and Dµ ≡ ∂µ + (1/4)γαβ ωµ is the spin-1/2 covariant derivative, where
αβ
ωµ is the spin connection. γα denotes the Dirac matrices in Minkowski spacetime with
γαβ ≡ γ[α γβ] [12]. The last term in Eq. (2.1) corresponds to the mixing between two flavor
eigenstates. Hereafter we shall assume that the strength of the mixing, δ, is constant in
time.
When the inflaton oscillates around the potential minimum, fermions are resonantly
created through the Yukawa interactions. This fermionic preheating was investigated by a
number of authors [13–21], but all of these analysis are restricted to the case of a single
fermionic field coupled to the inflaton. In this work we shall extend this to the more generic
case where the inflaton can decay to multiple fermionic fields. In fact we will show that an
interesting effect emerges in addition to the standard resonant production of fermions.
For the coupled system of fermions with a time-dependent background, a formalism
was developed in Ref. [37] to compute the occupation number of fermions. We first review
the general formalism of coupled fields and then apply it to our 2-fermionic system.
Let us consider N -fermionic fields, ψi , whose action is given by

√
S = d 4 x −g ψ̄i i δij γ̄ µ Dµ − Mij ψj . (2.3)
In our case the mass matrix Mij corresponds to

mα + hα φ δ
Mij = . (2.4)
δ mβ + hβ φ
We shall adopt a flat Friedmann–Robertson–Walker (FRW) metric with the line element
ds 2 = dt 2 − a 2 (t)δij dx i dx j , (2.5)
where a(t) is the scale factor. In this work we do not take into account the effect of metric
perturbations that can provide additional source terms for the production of fermions
[23] (see also Refs. [42,43]). Then two matrices are connected through the relation,
γ̄ 0 = γ 0 , γ̄ i = γ i /a, γ̄ µ Dµ = γ̄ µ ∂µ + (3/2)H γ 0 , with H ≡ ȧ/a being the Hubble rate
(a dot denotes the derivative with respect to the cosmic time, t). Redefining a new field,
Ψi ≡ a 3/2ψi , the action (2.3) reduces to

S = d 4x Ψ i iδij γ̄ µ ∂µ − Mij Ψj , (2.6)
which leads to Dirac equation of the curved spacetime,

iδij γ̄ µ ∂µ − Mij Ψj = 0, (2.7)
together with the canonical Hamiltonian,

H = d 3x Ψ −i γ̄ i ∂i + M Ψ. (2.8)
We diagonalize the mass matrix Mij with an orthogonal matrix Uij , whose new
eigenstate, Φ, is given by Φ ≡ U −1 Ψ . Then the equation of motion (2.7) becomes
µ
i γ̄ ∂µ + iγ 0 U T U̇ − µ Φ = 0, (2.9)
where µ ≡ U T MU is a diagonal matrix. The time-dependence of the unitary transfor-
mation matrix Uij generates the nonzero off-diagonal term iγ 0 U T U̇ in Eq. (2.9), which
leads to the mixing between the Φi states. Therefore, the system cannot be regarded as the
decoupled N state when the fermion has a time-dependent mass matrix as given by (2.4).
We decompose the field Φ into Fourier modes

d 3 x ik·x j k jk
Φ= e us (k, t)bsk (k) + vs (k, t)ds†k (−k) , (2.10)
(2π)3/2
where the quantities,
ij 1 ij ij T ij 1 ij ij T
us ≡ √ u+ χs , su− χs and vs ≡ √ v+ χs , sv− χs
2 2
are spinor matrices with χ+ and χ− being the eigenvectors of the helicity operator.
T are the solutions of Eq. (2.7) as well
When Ψi satisfies Eq. (2.7), the fields ΨiC ≡ C Ψ i
(here C is a charge conjugation matrix). Therefore we can impose the relation us (k) =
C v̄sT (−k), which allows for dealing with the u± matrices only. Taking the momentum k
along the third axis, the u± satisfy the equations of motion
k µ
u̇± = −i u∓ ∓ i u± − Qu± , (2.11)
a a
where Q ≡ U T U̇ .
From the requirement of the quantization of the field Φ, we have the following
anticommutation relation

†
bri (k), bsj (p) = δ (3)(k − p)δrs δij ,

†
dri (k), dsj (p) = δ (3)(k − p)δrs δij . (2.12)
The solutions for Eq. (2.11) can be expanded as

µ 1/2 −i η ω dη µ 1/2 i η ω dη
u± = I ± e A∓ I ∓ e B
ω ω

µ 1/2 µ 1/2
≡ I± α∓ I ∓ β, (2.13)
ω ω

where ω = k 2 I 2 + µ2 and η ≡ a −1 dt is a conformal time. Here I is a unit matrix.
The “Bogolyubov matrices” α and β satisfy the following relation from Eqs. (2.12) and
(2.13)
αα † + β ∗ β T = I,
αβ † − β ∗ α T = 0, (2.14)
together with the differential equation

µ k
αij = −iωil − Kil+ αlj + − il2 + Kil− βlj ,
2ωil

µil k
βij = 2
− Kil− αlj + iωil − Kil+ βlj , (2.15)
2ωil
where a prime denotes the derivative with respect to η and

1 µik 1/2 µlj 1/2
Kij± ≡ Iik + Qkl Ilj ±
2 ωik ωlj
1/2
1 µik µlj 1/2
± Iik − Qkl Ilj ∓ . (2.16)
2 ωik ωlj

Here ωij = k 2 I + µ2ij .
In order to derive the occupation number for the fermionic eigenstates, we introduce
new operators, b̂i (k, t) and d̂i (k, t), which are given by
b̂i (k, t) = αij (k, t)bj (k) − βij∗ (k, t)dj† (−k),
d̂i† (k, t) = βij (k, t)bj (k) + αij∗ (k, t)dj† (−k). (2.17)
Then we obtain the following diagonalized Hamiltonian after normal ordering

1
H= d 3 k ωi b̂i† b̂i + d̂i† d̂i . (2.18)
a
The initial vacuum state, |0η0 , corresponds to the no particle state with
α(η0 ) = I, β(η0 ) = 0, (2.19)
which satisfies the conditions (2.14). The occupation number density for the ith fermionic
eigenstate is given by

ni (η) = 0η0 |b̂i† b̂i |0η0 = 0η0 |d̂i† d̂i |0η0 = β ∗ β T ii , (2.20)
which means that particle and antiparticles are produced with the same amount.
In the case of two fermionic fields, we have
n1 = |β11 |2 + |β12 |2 , (2.21)
n2 = |β21 | + |β22 | .
2 2
(2.22)
In what follows we shall apply this formula to the flavor oscillation between two fermions
during reheating.
3. Mechanism of the MSW-resonance during reheating
In this section we shall analyze the mechanism of the MSW-resonance during reheating.
To do so let us write the orthogonal matrix, U , in the form

cos θ (t) sin θ (t)
U= . (3.1)
− sin θ (t) cos θ (t)
Here θ (t) is the time-varying effective mixing angle, given by
2δ
tan θ (t) = , (3.2)
7m + 7hφ(t) + (7m + 7hφ(t))2 + 4δ 2
where 7m ≡ mα − mβ , and 7h ≡ hα − hβ . The dynamics of the system is different

depending on the model parameters we consider. In what follows we shall discuss two
simple cases first and next proceed to more generic cases.
3.1. Flavor mixing without Yukawa couplings (hα = hβ = 0)
In the absence of Yukawa couplings (hα = hβ = 0), preheating production does not take
place. Fermions can be created only gravitationally, which was discussed in Refs. [18,26]
in the case of the single fermion. When δ = 0 two fermions are generated independently
through the gravitational process, in which case the amount of the each fermion is described
by the analysis in Refs. [18,26].
When δ = 0 the mixing angle is given by
2δ
tan θ = . (3.3)
7m + (7m)2 + 4δ 2
The strength of the mixing is dependent on the relative ratio between 7m and δ. When
|7m| δ the mixing angle is small (θ 0). In the case of |7m| δ, we have a strong
mixing with θ π/4. Since tan θ is constant in time in the absence of Yukawa couplings,
two fermions oscillate each other as in neutrino oscillations in the vacuum [44]. In this case
the U T U̇ term vanishes in Eq. (2.9), which means that the state Φ corresponds to the true
mass eigenstate. Note that when Yukawa couplings are nonzero but the same (hα = hβ = 0)
we also have a constant mixing angle given by Eq. (3.3).
3.2. Preheating without flavor mixing (δ = 0 and hα = 0, hβ = 0)
If we take into account the Yukawa couplings, fermions are created nongravitationally
through the preheating process. In the absence of the mixing (δ = 0), the equations of
motion for two fermions take the decoupled forms
µ
i γ̄ ∂µ − (mα + hα φ) Φα = 0, (3.4)
µ
i γ̄ ∂µ − (mβ + hβ φ) Φβ = 0. (3.5)
The two fermionic states evolve independently for δ = 0. Fermions are created resonantly
i ≡ mi +hi φ, change rapidly (here i = α, β).
when the effective masses of the fermions, meff
This takes place around mi = 0 [17], corresponding to the inflaton value
eff
mi
φ∗,i = − . (3.6)
hi
When the condition, mi > |hi φ|, is satisfied at the beginning of reheating, the inflaton does
not pass through the resonance point (3.6). Therefore we require the condition, mi < |hi φ|,
to lead to parametric excitation of fermions.1 Since inflation ends around φ = 0.3mpl , it is
possible to generate heavy fermions whose masses are of order 1017 –1018 GeV unless the
1 As long as this condition is satisfied, fermions can be created even during inflation. This was analyzed in
Ref. [19].
couplings hi are not small [17]. In order to have an efficient particle creation, it is required
that the dimensionless parameter,
h2i φ02
qi ≡ , (3.7)
4m2φ
is much larger than unity (here φ0 0.28mpl is the value of the inflaton at the beginning
of reheating). The maximum momenta kimax have a dependence, kimax ∝ (qi /mi )γ , with γ
slightly larger than 1/3 [17].
The fermion creation takes place at each time when the inflaton crosses the resonance
point (3.6). This phase comes to an end when the amplitude of the inflaton drops down
below the value, |φ∗,i |. For the decoupled system given by Eqs. (3.4) and (3.5), two
fermions are resonantly generated around two different resonance points (3.6) without
mixing each other.
3.3. Structure of MSW resonance during reheating (δ = 0 and hα = 0, hβ = 0)
Let us now consider the general case where the couplings hα , hβ and δ are nonzero.
In this case, we can expect two kinds of nonadiabatic phenomena: one is the preheating
resonance explained in Section 3.2 and another is the MSW resonance induced by the time-
dependent mixing angle, θ (t). The latter emerges from the combined effect of the mixing
of two fermions and the Yukawa interactions with the time-varying inflaton field. This is
analogous to the MSW resonance of solar neutrinos where the slowly varying electron
density in the Sun leads to the conversion from one flavor neutrino to another [41]. The
difference is that our case corresponds to the time-dependent background, whereas the
MSW oscillation of solar neutrinos corresponds to the space-dependent background.
From Eq. (3.2) we find that the mixing angle is maximum (θ = π/4) when the inflaton
crosses the value
7m
φMSW = − . (3.8)
7h
Therefore the system undergoes the MSW resonance provided that the amplitude of the
inflaton is larger than |φMSW | at the beginning of reheating. The sharpness of the MSW
resonance is dependent on the strength, δ. As shown in Fig. 1, the resonance is sharper
with decreasing δ. When δ is small, Eq. (3.2) clearly indicates that tan θ (t) suddenly jumps
up from a small value to unity when φ crosses φMSW = −7m/7h.
The eigenvalues of the mass matrix Mij given in Eq. (2.4) are written as

1 2
µi ≡ µii = (mα + mβ ) + (hα + hβ )φ(t) ± 7m + 7hφ(t) + 4δ 2 , (3.9)
2
which are the hyperbolic functions in terms of φ. At the MSW resonance point (3.8), the
difference between the two eigenvalues, 7µ ≡ µ1 − µ2 , takes its minimum value 2δ.
In Fig. 2 we plot the two mass eigenvalues as a function of φ(t). The MSW resonance
occurs when two curves approach close to each other. Note that in the case of δ = 0 we
have two straight lines, µ1 = mα + hα φ and µ2 = mβ + hβ φ, which cross at φ = φMSW . In
the presence of the mixing (δ = 0), these two straight lines are replaced by two parabolic
Fig. 1. Plots of sin2 2θ as a function of φ for mα = 4mφ , mβ = 5mφ , hα = 5.7 × 10−4 , and hβ = 1.6 × 10−4 .
In this case the MSW resonance occurs around φMSW = 2.0 × 10−3 mpl . The resonance is sharper for smaller δ.
Fig. 2. Two eigenvalues of the mass matrix Mij , as a function of φ(t). The MSW resonance occurs around
φMSW = −7m/7h. The fermion production occurs when two hyperbolic curves cross the φ axis. The production
points are shifted by taking into account the mixing effect (δ = 0).
curves. This means that the values of φ when µ1 and µ2 cross zero change for δ = 0.
Since the fermions are created around µ1 0 and µ2 0 due to the effect of preheating,
these production points are shifted by taking into account the δ coupling. This nontrivially
affects the qualitative details of the occupation number of fermions, as we will show later.
In the conventional MSW resonance of solar neutrinos, the matter-induced electron
neutrino mass A(ne ) corresponds to the inflaton in our case. As a flavor neutrino propagates
toward the outgoing direction in the Sun, it undergoes the MSW resonance due to the
decreasing electron density and changes to another flavor neutrino. This property is quite
similar to our scenario. The difference is that in our case the inflaton can cross the
MSW resonance point (3.8) for more than once due to the oscillating inflaton with a
decreasing amplitude. In addition particle production due to preheating makes the system
a bit complicated. In the next section we will numerically investigate the dynamics of the
system in order to make the accurate qualitative analysis.
In order to understand when the MSW resonance becomes important, let us return back
to the basic equation (2.9). In our case the matrix, Q = U T U̇ , is given by

0 1 2δ ξ̇ 0 1
Q = θ̇ = 2 , (3.10)
−1 0 4δ + ξ 2 −1 0

where ξ ≡ 7m + 7hφ(t) − (7m + 7hφ(t))2 + 4δ 2 . Then Eq. (2.9) can be written as

µ1 0 0 1 Φ1
i γ̄ µ ∂µ − − iγ0 θ̇ = 0. (3.11)
0 µ2 −1 0 Φ2
This indicates that the mass eigenstate is no longer the true eigenstate in the presence
of the off-diagonal terms. Defining new states as Φ 1 = exp(−iγ0 7µ t)Φ1 and Φ 2 =
2
7µ
exp(−iγ0 2 t)Φ2 , we find that in an adiabatic regime (φ̇ ≈ 0) Eq. (3.11) reduces to

7µ 1 0 0 1 1
Φ
i γ̄ µ ∂µ − − iγ0 θ̇ = 0. (3.12)
2 0 −1 −1 0 Φ2
In order to give a criterion to the adiabaticity of the time-evolution of the system, we define
[45]

7µ/2
Γ ≡ . (3.13)
θ̇
When Γ is large (Γ 1), the off-diagonal terms can be neglected and the mass eigenstates
become the true eigenstates. The system effectively reduces to the two independent states
discussed in Section 3.2. For smaller Γ we have to take into account the correction to the
adiabatic approximation. Since Γ takes its minimum value at the MSW resonance point
(3.8), the minimum adiabaticity parameter during the half period of the inflaton oscillation
is characterized by the value

7µ/2 8δ 2
Γn =
min = , (3.14)
θ̇ φ=φMSW 7hφ̇(tn ) φ=φMSW
where φ̇(tn ) is the value of φ̇ in the nth MSW resonance. For smaller δ, we have a
stronger nonadiabaticity as shown in Fig. 1. When the inflaton evolves sufficiently slowly
(φ̇(tn ) ≈ 0) around φ = φMSW , the MSW resonance occurs adiabatically. The MSW
resonance ends when the amplitude of the inflaton drops down under |φMSW |.
4. Numerical analysis
In this section we shall numerically study the dynamical system of two coupled fermions
for the potential (2.2). During reheating the background evolution is characterized as [8]
a ∝ t 2/3 , (4.1)
φ = φ̃(t) sin mφ t, with φ̃(t) ∝ 1/(mφ t). (4.2)
Fig. 3. The evolution of the occupation numbers of fermions for mα = 6mφ , mβ = 7mφ , hα = 7.0 × 10−4 ,
hβ = 3.1 × 10−4 , k = 8mφ and δ = 10−3 mφ . n1 and n2 correspond to the occupation numbers corresponding to
the heavy mass and the light mass, respectively. At the initial stage both the preheating and the MSW resonances
occur in turn, which is followed by the period of the MSW resonance only. After the amplitude of the inflaton
drops down under |φMSW |, the MSW resonance comes to an end.
The initial amplitude of the inflaton corresponds to φ̃(0) 0.07mpl and φ̃(t) gradually
decreases due to the cosmic expansion. If we consider another scalar field χ coupled to
φ with an interaction (1/2)g 2 φ 2 χ 2 , χ particles can be efficiently produced by parametric
resonance [8]. This can alter the background evolution in the later stage of preheating due
to the backreaction effect of created particles. In this paper we do not take into account the
bosonic coupling and its resulting backreaction for simplicity.
The fermion creation due to preheating is significant around µ1 = 0 and µ2 = 0, which
is characterized by the value of φ as

1
φpre,i = −(mα hβ + mβ hα ) ± (mβ hα − mα hβ )2 + 4hα hβ δ 2 , (4.3)
2hα hβ
where i = 1, 2. Note that this reduces to φ1,α = −mα / hα and φ2,β = −mβ / hβ for δ = 0.
These values are shifted by including the δ coupling.
As long as |φMSW | and |φpre,i | are smaller than φ̃(0), both the MSW and the preheating
resonances occur when the inflaton crosses two resonance points of Eqs. (3.8) and (4.3).
In Fig. 3 we plot one evolution of the occupation numbers of fermions given in Eqs. (2.21)
and (2.22). This case corresponds to φpre,1 = −8.6 × 10−3 mpl , φpre,2 = −2.3 × 10−2 mpl
and φMSW = 2.6 × 10−3 mpl , whose absolute values are smaller than φ̃(0). Therefore both
the MSW and the preheating resonances take place in turn as shown in Fig. 3 (see also
Fig. 4). Since we are considering the case of |φpre,i | > |φMSW |, the system enters a phase
where only the MSW resonance occurs with the decrease of φ̃(t) (see Fig. 3). The MSW
resonance comes to an end when φ̃(t) drops down under |φMSW |. After that the system
enters a stable period during which both resonances do not occur.
Of course the dynamics of the system differs significantly depending on the model
parameters we consider. When φ̃(0) > |φMSW | > |φpre,i |, both the MSW and the preheating
Fig. 4. The period during which both the preheating and the MSW resonances occur in Fig. 3 (0 < mφ t < 50).
While the preheating resonance can change the total occupation numbers (nT = n1 + n2 ), the MSW resonance
keeps nT unchanged.
resonances occur at the initial stage but the former terminates after φ̃(t) drops down
|φMSW |. In the case of |φpre,i | > φ̃(0) > |φMSW | (or |φMSW | > φ̃(0) > |φpre,i |), we can
expect only the MSW resonance (or the preheating resonance). When {|φpre,i |, |φMSW |} >
φ̃(0), both resonances do not take place.
The number densities of fermions are altered by taking into account the coupling δ.
This effect works in following two ways. One is the MSW resonance between the states
Φ1 and Φ2 , and another is the shift of the preheating resonance points. First, let us
consider the case where the former effect can be important. One example is plotted in
Fig. 5. In this case we have φpre,1 = −mα / hα −7.5 × 10−2 mpl , φpre,2 = −mβ / hβ
−8.3 × 10−3 mpl and φMSW 8.3 × 10−3 mpl for δ = 0, which corresponds to the case
|φpre,1 | φ̃(0) > |φpre,2 | φMSW (here we used mφ 1013 GeV). Since |φpre,1 | is a bit
larger than the value φ̃(0) 0.07mpl, the occupation number n1 is much smaller than
n2 as seen in Fig. 5. If we include the coupling δ, the situation is different due to the
MSW resonance. In the case shown in the lower panel of Fig. 5, the value of δ is chosen
so that two preheating resonance points are hardly shifted relative to the case of δ = 0.
Then the particle creation due to preheating is mostly the same even taking into account
the coupling δ. In contrast the MSW resonance around φ = φMSW leads to the mixing
of fermions through the third term in the l.h.s. of Eq. (3.12). In fact from Eq. (3.14) the
adiabaticity parameter is estimated as Γ1min = 4.2 × 10−3 for the first MSW resonance,
implying a strong nonadiabaticity. Although φ̇(tn ) becomes smaller for larger n, the system
satisfies the nonadiabatic condition, Γnmin 1, until the amplitude of φ decreases under
|φMSW |. Since two states Φ1 and Φ2 mix sufficiently through the MSW resonance, the
occupation numbers n1 and n2 become the same order as seen in Fig. 5.
We shall also evaluate the ratio between the energy density of fermions and that of the
inflaton

ρtr 2µ̃i 1 2 2 −1
= 2
dk k ni (k) 2 m φ , (4.4)
ρφ π 2 φ 0
i=1,2
Fig. 5. The spectra of the occupation numbers of fermions for mα = 10mφ , mβ = 6mφ , hα = 1.3 × 10−4 ,
hβ = 7.2 × 10−4 . The upper and lower panels corresponds to the case of δ = 0 and δ = 0.1mφ , respectively. Two
number densities become the same order by taking into account the coupling δ.
Fig. 6. The energy density of fermions relative to that of the inflaton for various values of δ and mα . The other
model parameters are mβ = 6mφ , hα = 1.3 × 10−4 , and hβ = 7.2 × 10−4 . For larger mα , |φMSW | tends to
exceed the amplitude of the inflaton, thereby yielding the smaller energy transfer to the fermions.

where µ̃i ≡ 12 [mα + mβ ± (mα − mβ )2 + 4δ 2 ]. We are interested in the energy density
transferred from the inflaton to two fermions after the system enters the stable period. In
Fig. 6 we plot the ratio ρtr /ρφ as a function of mα /mφ with the same values of mβ , hα and
hβ as in Fig. 5. When δ = 0 the total number densities of fermions are dominated by n2
(equivalently the state β), which means that the ratio ρtr /ρφ hardly changes for mα > 10mφ
even if we vary the value of mα (see Fig. 6). We completely reproduced the numerical value
ρtr /ρφ = 5.9 × 10−8 for mβ = 6mφ and hβ = 7.2 × 10−4 that was derived in Ref. [20] in
the single field case. We find from Fig. 6 that the energy density of fermions gets larger by
taking into account the coupling δ. This is particulary seen for smaller values of mα . As
Fig. 7. The energy density of fermions relative to that of the inflaton for the parameters mβ = 60mφ ,
hα = 1.3 × 10−3 , and hβ = 7.2 × 10−3 . The energy transfer is larger than that in Fig. 6.
mα being increased, |φMSW | tends to approach the initial amplitude of φ, thereby resulting
the smaller ratio ρtr /ρφ that is close to the case of δ = 0.
The energy transfer from the inflaton to fermions increases for larger masses of
fermions. This is because the ratio ρtr /ρφ is proportional to the mass µ̃i . When mβ =
60mφ , ρtr /ρφ becomes of order 10−5 as shown in Fig. 7. Again the energy transfer gets
larger in the presence of the δ coupling. In Figs. 6 and 7 we find that ρtr /ρφ increases up to
O(2) times by involving the mixing term. This can be understood that two fermions mix
sufficiently due to the MSW resonance, thereby yielding the twice contribution to the total
occupation number of fermions.
Let us next consider the shift of the preheating resonance points through the nonzero δ.
In Fig. 8 we show one example of the plot of the ratio ρtr /ρφ as a function of δ. When
δ = 0 one has φpre,1 = −0.20mpl and φpre,2 = −0.073mpl in this case, which means that
both resonance points are outside of the amplitude of φ, i.e., {|φpre,1 |, |φpre,2 |} > φ̃(0)
0.07mpl. Therefore preheating production is inefficient for δ = 0. With the increase of δ,
however, one of the resonance points (φpre,2 ) enters the inside of the inflaton oscillation
with the increase of δ. Note that another resonance point (φpre,1 ) always satisfies |φpre,1 | >
φ̃(0), since it is shifted toward large negative values. The growth of ρtr /ρφ saturates around
δ 80mφ . This means that one of the resonance points (φpre,2 ) reaches deep inside the
inflaton amplitude (|φpre,2 | φ̃(0)), in which case the energy transfer does not change
significantly even increasing the value of δ. In the limit of δ → ∞ we have φpre,i → ±∞
from Eq. (4.3), implying that too large values of δ tend to suppress the fermion creation.
We shall also mention the case where two resonance points (4.3) are inside the
amplitude of φ for δ = 0 but one or two of the resonance points are shifted outside the
inflaton oscillation by including the coupling δ. In this case the total number density and
the energy density of fermions should be smaller than in the case of δ = 0. In fact we have
numerically found that this is the case.
As we already mentioned, important quantities which characterize our system are
φMSW , Γnmin , φpre,1 , and φpre,2 , which are determined by mα , mβ , hα , hβ , and δ. Another
Fig. 8. The energy density of fermions relative to that of the inflaton for various values of δ. The other model
parameters are mα = 2.0 × 102 mφ , mβ = 1.0 × 102 mφ , hα = 1.0 × 10−3 , and hβ = 1.3 × 10−3 . The energy
transfer increases by including the δ coupling, provided that it works so that resonance points given by Eq. (4.3)
are shifted inside the amplitude of the inflaton.
important quantity is an initial amplitude of the inflaton, which is fixed as φ̃(0) 0.07mpl
in our model. Depending on the strength of the masses and the couplings, we can have a
variety of the results which are the combination of the MSW and the preheating resonances.
5. Conclusions and discussions
We have investigated the influence of the flavor mixing on fermionic preheating.

Fermions are resonantly created by the effect of preheating when fermions are coupled to
the inflaton through Yukawa couplings. We have considered a two-field system of fermions
which mixes through a coupling δ. We found that in addition to the standard preheating
resonance the MSW-type resonance occurs due to a time-dependent background of the
oscillating inflaton with a decreasing amplitude. This is analogous to the MSW resonance
of neutrino oscillations which occurs when neutrinos propagate in a space-dependent
background of the varying electron density.
Our system can be characterized by the MSW resonance point φMSW given by Eq. (3.8)
and by the preheating resonance points φpre,i (i = 1, 2) given by Eq. (4.3). In order
for these resonances to occur, it is crucial that |φMSW | and |φpre,i | are smaller than
the initial amplitude of the inflaton, φ̃(0) 0.07mpl . Depending on the values of the
masses (mα , mβ ) and the couplings (hα , hβ , δ), we have a variety of results that are the
combination of the MSW and the preheating resonances. For example, when the condition
φ̃(0) > |φpre,i | > |φMSW | is satisfied, both resonances occur in turn at the initial stage,
which is followed by the stage of the MSW resonance only. The system enters a stable
period after the amplitude of the inflaton drops down under |φMSW | (see Figs. 3 and 4).
When the system evolves nonadiabatically around φ = φMSW [corresponding to
Γnmin 1 in Eq. (3.14)], two fermionic states mix sufficiently through the MSW resonance.
Therefore even if one of the fermions are not created enough by preheating, the number
density of each fermion can be the same order after exhibiting a number of MSW
resonances (see Fig. 5). This means that the total energy density of fermions can increase up
to O(2) times by including the coupling δ. This was actually confirmed by our numerical
simulations in Figs. 6 and 7.
Inclusion of the coupling δ also leads to the shift of preheating resonance points, φpre,i .
Therefore even if |φpre,i | is larger than φ̃(0) for δ = 0, this effect can make the values of
|φpre,i | shift inside the amplitude of the inflaton. As shown in Fig. 8, the number density of
fermions and the energy transfer from the inflaton to fermions can increase by taking into
account the coupling δ.
The above mechanism works even when one of the fermions does not couple to
the inflaton through a Yukawa coupling (hα = 0 or hβ = 0). As long as the condition
|φMSW | < φ̃(0) is satisfied (this automatically means that 7h is nonzero), the MSW
resonance is crucially important when we discuss the fermion production in a coupled
system of a time-varying background.
It is straightforward to generalize our work to the case of the N flavor mixing. The
system possesses N(N − 1)/2 MSW resonance points and N preheating resonance points.
In this case it is expected that the energy transfer from the inflaton to fermions can increase
up to about O(N) times by involving the mixing term. The increased fermion production
can affect the leptogenesis scenario via the decay of heavy Majorana neutrinos.
In our analysis, we did not take into account the resonant production of scalar particles
coupled to the inflaton. This can lead to the change of the background evolution around the
final stage of preheating due to the backreaction effect of created particles [8]. In addition,
the mode–mode coupling between fermions and scalar particles can be important in the
nonlinear stage of preheating. It is certainly of interest to extend our analysis to such a
nonlinear regime.
Acknowledgements
We are grateful to Shin’ichiro Ando, Bruce Bassett, Naofumi Fujishiro, Masahiro

Kawasaki, Marco Peloso and Masahide Yamaguchi for useful discussions. S.T. is thankful
for financial support from the JSPS (No. 04942).
References
[1] A.H. Guth, Phys. Rev. D 23 (1981) 347;

K. Sato, Mon. Not. R. Astron. Soc. 195 (1981) 467;
K. Sato, Phys. Lett. B 99 (1981) 66.
[2] A.D. Linde, Particle Physics and Inflationary Cosmology, Harwood, Chur, 1990;
E.W. Kolb, M.S. Turner, The Early Universe, Addison–Wesley, New York, 1990;
A.R. Liddle, D.H. Lyth, Cosmological Inflation and Large-Scale Structure, Cambridge Univ. Press,
Cambridge, 2000.
[3] P. de Bernardis, et al., Nature (London) 404 (2000) 955;
A.E. Lange, et al., Phys. Rev. D 63 (2001) 042001.
[4] S. Hanany, et al., Astrophys. J. Lett. 545 (2000) L5;
A. Balbi, et al., Astrophys. J. Lett. 545 (2000) 1.
[5] N.W. Halverson, et al., astro-ph/0104489;

C. Pryke, et al., astro-ph/0104490.
[6] D.N. Spergel, et al., astro-ph/0302209;
H.V. Peiris, et al., astro-ph/0302225.
[7] J.H. Traschen, R.H. Brandenberger, Phys. Rev. D 42 (1990) 2491;
Y. Shtanov, J. Traschen, R.H. Brandenberger, Phys. Rev. D 51 (1995) 5438.
[8] L. Kofman, A. Linde, A.A. Starobinsky, Phys. Rev. Lett. 73 (1994) 3195;
L. Kofman, A. Linde, A.A. Starobinsky, Phys. Rev. D 56 (1997) 3258.
[9] D. Boyanovsky, H.J. de Vega, R. Holman, J.F.J. Salgado, Phys. Rev. D 54 (1996) 7570.
[10] S. Khlebnikov, I.I. Tkachev, Phys. Rev. Lett. 77 (1996) 219;
S. Khlebnikov, I.I. Tkachev, Phys. Lett. B 390 (1997) 80;
S. Khlebnikov, I.I. Tkachev, Phys. Rev. Lett. 79 (1997) 1607.
[11] T. Prokopec, T.G. Roos, Phys. Rev. D 55 (1997) 3768.
[12] N.D. Birrell, P.C.W. Davies, Quantum Fields in Curved Space, Cambridge Univ. Press, Cambridge, 1980.
[13] D. Boyanovsky, M. D’attanasio, H.J. de Vega, R. Holman, D.-S. Lee, Phys. Rev. D 52 (1995) 6805;
D. Boyanovsky, H.J. de Vega, Y.J. Ng, D.-S. Lee, S.-Y. Wang, Phys. Rev. D 59 (1999) 105001.
[14] S.A. Ramsey, B.L. Hu, A.M. Stylianopoulos, Phys. Rev. D 57 (1998) 6003.
[15] J. Baacke, K. Heitman, C. Patzold, Phys. Rev. D 58 (1998) 125013;
J. Baacke, C. Patzold, Phys. Rev. D 62 (2000) 084008.
[16] P.B. Greene, L. Kofman, Phys. Lett. B 448 (1999) 6;
P.B. Greene, L. Kofman, Phys. Rev. D 62 (2000) 123516.
[17] G.F. Giudice, M. Peloso, A. Riotto, I.I. Tkachev, JHEP 9908 (1999) 014.
[18] V. Kuzmin, I. Tkachev, Phys. Rev. D 59 (1999) 123006;
V. Kuzmin, I. Tkachev, Phys. Rep. 320 (1999) 199.
[19] D.J. Chung, E.W. Kolb, A. Riotto, I.I. Tkachev, Phys. Rev. D 62 (2000) 043508.
[20] M. Peloso, L. Sorbo, JHEP 0005 (2000) 016.
[21] S. Tsujikawa, B.A. Bassett, F. Viniegra, JHEP 08 (2000) 019.
[22] S. Tsujikawa, H. Yajima, Phys. Rev. D 64 (2001) 023519.
[23] B.A. Bassett, M. Peloso, L. Sorbo, S. Tsujikawa, Nucl. Phys. B 622 (2002) 393.
[24] T. Asaka, K. Hamaguchi, M. Kawasaki, T. Yanagida, Phys. Lett. B 464 (1999) 12;
T. Asaka, K. Hamaguchi, M. Kawasaki, T. Yanagida, Phys. Rev. D 61 (2000) 083512;
H. Goldberg, Phys. Lett. B 474 (2000) 389;
R. Barbieri, P. Creminelli, A. Strumia, N. Tetradis, Nucl. Phys. B 575 (2000) 61;
G. Mangano, G. Miele, Phys. Rev. D 62 (2000) 063514.
[25] V. Kuzmin, I.I. Tkachev, JETP Lett. 68 (1998) 271.
[26] D. Chung, E.W. Kolb, A. Riotto, Phys. Rev. D 59 (1999) 023501;
D. Chung, E.W. Kolb, A. Riotto, Phys. Rev. D 60 (1999) 063504.
[27] K. Greisen, Phys. Rev. Lett. 16 (1966) 748;
G.T. Zatsepin, V.A. Kuzmin, JETP Lett. 4 (1966) 78.
[28] T. Moroi, Effects of the gravitino on the inflationary universe, PhD thesis, hep-ph/9503210;
S. Sarkar, Rep. Prog. Phys. 59 (1996) 1493.
[29] A.D. Dolgov, A.D. Linde, Phys. Lett. B 116 (1982) 329;
L.F. Abbott, E. Fahri, M. Wise, Phys. Lett. B 117 (1982) 29.
[30] J. Ellis, A. Linde, D. Nanopoulos, Phys. Lett. B 118 (1982) 59;
M. Kawasaki, T. Moroi, Prog. Theor. Phys. 93 (1995) 879.
[31] H. Fujisaki, K. Kumekawa, M. Yamaguchi, M. Yoshimura, Phys. Rev. D 54 (1996) 2494.
[32] A.L. Maroto, J.R. Pelaez, Phys. Rev. D 62 (2000) 023518.
[33] A.L. Maroto, A. Mazumdar, Phys. Rev. Lett. 84 (2000) 1655.
[34] G.F. Giudice, I.I. Tkachev, A. Riotto, JHEP 9908 (1999) 009;
G.F. Giudice, A. Riotto, I.I. Tkachev, JHEP 9911 (1999) 036.
[35] R. Kallosh, L. Kofman, A. Linde, A. Van Proeyen, Phys. Rev. D 61 (2000) 103503;
R. Kallosh, L. Kofman, A. Linde, A. Van Proeyen, Class. Quantum Grav. 17 (2000) 4269.
[36] H.P. Nilles, M. Peloso, L. Sorbo, Phys. Rev. Lett. 87 (2001) 051302.
[37] H.P. Nilles, M. Peloso, L. Sorbo, JHEP 0104 (2001) 004.
[38] P.B. Greene, K. Kadota, H. Murayama, hep-ph/0208276.

[39] D.H. Lyth, A. Riotto, Phys. Rep. 314 (1999) 1.
[40] N. Bartolo, S. Matarrese, A. Riotto, Phys. Rev. D 64 (2001) 083514.
[41] L. Wolfenstein, Phys. Rev. D 17 (1978) 2369;
L. Wolfenstein, Phys. Rev. D 20 (1979) 2634;
S.P. Mikheyev, A.Y. Smirnov, Sov. J. Nucl. Phys. 42 (1985) 913, Yad. Fiz. 42 (1985) 1441.
[42] A.L. Maroto, Phys. Rev. D 64 (2001) 083006;
A.L. Maroto, Phys. Rev. D 65 (2002) 083508.
[43] B.A. Bassett, G. Pollifrone, S. Tsujikawa, F. Viniegra, Phys. Rev. D 63 (2001) 103515.
[44] Z. Maki, M. Nakagawa, S. Sakata, Prog. Theor. Phys. 28 (1962) 870.
[45] T.K. Kuo, J. Pantaleone, Rev. Mod. Phys. 61 (1989) 937.
3D two-color QCD at finite temperature

and baryon density
Gerald V. Dunne a , Shinsuke M. Nishigaki a,b
a Department of Physics, University of Connecticut, Storrs, CT 06269-3046, USA
b Department of Material Science, Faculty of Science and Engineering, Shimane University,
Matsue 690-8504, Japan
Abstract
We study the phase diagram for two-color QCD in three-dimensional spacetime, as a function of
temperature and baryon chemical potential, using the low-energy effective Lagrangian approach. We
show one-loop renormalizability at zero temperature, and then use the one-loop effective Lagrangian
at finite temperature and chemical potential to show that at low temperature there is a critical line
separating the normal and diquark phase, with this critical line ending at a tricritical point. This phase
structure is qualitatively similar to that found recently by Splittorff et al. for two-color QCD in four-
dimensional spacetime, although the details are quite different, due to the different symmetries and
the different loop and infrared properties of three-dimensional spacetime.
PACS: 11.10.Kk; 11.30.Hv; 11.30.Qc; 12.39.Fe
Keywords: Chiral Lagrangian; Chemical potential; Finite temperature; Baryon superconductivity
1. Introduction
A systematic approach to the investigation of the symmetry breaking structure of

fermion-gauge theories is to use low-energy effective Lagrangians [1,2] to study the Gold-
stone modes corresponding to the spontaneously broken global symmetries [3]. Ideally, one
would like to understand the phase diagram of fermion-gauge theories in the (T , µ) plane,
where T is the temperature, and µ is the (baryon number) chemical potential. However, it
is not known how to analyze general fermion-gauge systems at finite baryon density, since
the baryon number chemical potential makes the Euclidean Dirac operator non-Hermitian
E-mail address: mochizuki@riko.shimane-u.ac.jp (S.M. Nishigaki).
doi:10.1016/j.nuclphysb.2003.07.024
308 G.V. Dunne, S.M. Nishigaki / Nuclear Physics B 670 (2003) 307–328
and the Boltzmann weight complex.1 This problem can be avoided, as advocated in [4–8],
by considering “two-color QCD”, for which the fundamental representation of SU(2) is
pseudoreal. From studies of 4D theories, it has long been appreciated that the SU(2) the-
ory exhibits exotic types of spontaneous breakdown of global symmetry [9]. Quarks and
charge-conjugated antiquarks are combined into an extended flavor multiplet, which is ex-
pected to break into its extended vector subgroup. This has the consequence that in lattice
simulations the Boltzmann weight is real even at finite (baryon number) chemical potential
µ. For an even number of flavors, the Boltzmann weight is also positive definite. Thus,
analytic predictions can be quantitatively compared with Monte Carlo simulations in lat-
tice gauge theory [10,11], provided the lattice regularization respects the relevant flavor
symmetry group. Such studies have recently been carried out [12] in 4D QCD with quarks
in pseudoreal (and real) representations at finite µ.
In [13] we studied the T = 0 case of 3D two-color QCD, and found that the phase
structure predicted by the tree-level effective potential was qualitatively the same as the
4D case studied in [5,6,14], even though the details of the symmetry breaking in 3D are
completely different from in 4D. This universality between 3D and 4D corresponds to
the existence of two different breakings of the original global symmetry by a mass and
a chemical potential term. In this paper we extend the 3D analysis of [13] to include
nonzero temperature effects via the one-loop effective potential, as was done recently for
the 4D system in [15]. The loop effects introduce another difference between 3D and 4D,
as infrared physics is more significant in 3D. As a result, we find some similarities with the
4D case, but also some important differences.
In 3D, with an even number NF of flavors of massless complex fermions, denoted by
the NF /2 pairs ψf , χf , one can predict spontaneous flavor symmetry breaking patterns
along similar lines to 4D QCD. In 3D the generic situation is that the continuous part of
the global symmetry group is broken according to [16,17]
U(NF ) → U(NF /2) × U(NF /2) (1)
by the quark–antiquark condensate
N
F /2

ψ̄f ψf − χ̄f χf . (2)
f =1
Evidence for such a symmetry breaking pattern has been observed in 3D lattice simulations
[18] with gauge group SU(3). This pattern of flavor symmetry breaking can also be
predicted for 3D QCD at large NC using the Coleman–Witten argument [19]. The 3D
symmetry breaking pattern in (1) is for quarks in a complex representation of the gauge
group, and is expected to apply to a U(1) theory and to SU(NC ) theories with number
of colors NC 3. For SU(2), with fundamental quarks, the symmetry breaking pattern is
expected to be different again, due to the pseudoreality of the fundamental representation.
The pseudoreality of the fundamental representation of SU(2) means that the naive U(NF )
flavor symmetry is extended to USp(2NF ), and the continuous part of this global symmetry
group is predicted to break down in 3D as [20]
1 Note that there has been some interesting recent progress in overcoming this problem [31,32].
G.V. Dunne, S.M. Nishigaki / Nuclear Physics B 670 (2003) 307–328 309
USp(2NF ) → USp(NF ) × USp(NF ). (3)

This is different from the generic 3D symmetry breaking pattern in (1), and also is different
from 4D theories where the standard flavor symmetry breaking patterns are
U(NF )L × U(NF )R → SU(NF )V , for SU(NC 3)
U(2NF ) → USp(2NF ), for SU(2). (4)
In 4D the U(NF ) flavor symmetries are first broken to SU(NF ) symmetries by the axial
anomaly, and then broken by the chiral condensate, with the net breakings as shown in (4).
Physically, the differences between the 3D and 4D cases reflect the differences between
the anomalous discrete symmetries of parity and chirality in 3D and 4D, respectively.
In Section 2 we review the global symmetries and their cosets that are relevant for the
low energy effective Lagrangian description of the 3D two-color QCD system, with quarks
in the fundamental representation. In Section 3 we compute the one-loop effective action
at finite temperature and finite baryon chemical potential. This result is used in Section 4
for a Landau–Ginzburg description of the phase diagram, using the rotation angle α which
characterizes the tree-level vacuum state as an order parameter. Section 5 contains our
conclusions and some discussion.
2. Symmetries and effective Lagrangian
2.1. Enlarged flavor symmetry
The fermionic part of the microscopic Lagrangian for two-color QCD with NF = 2n
flavors of quarks in Euclidean 3D space is

Lf = ψ̄/Dψ + χ̄/Dχ + m ψ̄ψ − χ̄χ − µ ψ † ψ + χ † χ . (5)
Here ψ = ψfi , ψ † = ψf∗ i , χ = χfi , χ † = χf∗ i are independent two-component spinor fields,
with the color index i = 1, 2 and flavor index f = 1, . . . , n being suppressed. Pauli matrices
denoted as σν , with ν = 1, 2, 3, are employed to represent the Euclidean Dirac matrix
algebra, and those denoted as τα , with α = 1, 2, 3, are employed to represent the gauge
group algebra. We choose x3 to be the Euclidean time direction and define ψ̄ = ψ † σ3 ,
/ = σν Dν , and the covariant derivative is Dν = ∂ν + iAν ,
χ̄ = χ † σ3 . The Dirac operator is D
and the gauge field Aν = Aαν τα is Hermitian and su(2)-valued.
Due to the pseudoreality of the SU(2) Dirac operator, the kinetic part of the
Lagrangian (5) is invariant under a symmetry group larger than the apparent U(2n)
symmetry. One can put ψ, χ , ψ̃ , and χ̃ into a single flavor 4n-plet,
 
ψ
 χ 
Ψ = 
 σ2 τ2 ψ̄ T  , (6)
σ2 τ2 χ̄ T
so that
1 m µ
/ IˆΨ + Ψ T σ2 τ2 M̂Ψ − Ψ T iσ1 τ2 ĈΨ,
Lf = Ψ T σ2 τ2D (7)
2 2 2
where the 4n × 4n matrices Iˆ, M̂ and Ĉ are

 
−1n

−12n  1n  12n
Iˆ = , M̂ = 
,
 Ĉ = . (8)
12n 1n 12n
−1n
The kinetic term in the microscopic Lagrangian (7) has extended flavor symmetry group
being the unitary symplectic group USp(4n),
Ψ → SΨ, S T IˆS = Iˆ, S † S = 14n . (9)
The above extension of the flavor symmetry group is analogous to what happens in two-
color QCD in 4D, where the conjugated right-handed spinor σ2 τ2 ψR∗ transforms as the
left-handed spinor ψL does under gauge and Lorentz transformations, so that the chiral
SU(NF )L × SU(NF )R symmetry is extended to SU(2NF ).
In the 3D case, the flavor group USp(4n) is broken down to USp(2n) × USp(2n)
explicitly by the mass term in (7), or spontaneously by the quark–antiquark condensate (2),
if formed. In the latter case, the Goldstone manifold is thus given by the coset space
USp(4n)/(USp(2n) × USp(2n)), which has 4n2 independent degrees of freedom, and can
be parametrized as
Σ = SΣc S T , Σc = M̂ † , (10)
where

iΠ(x)
S(x) = exp , Π(x) = πa (x)Xa . (11)
2F
The fields πa are the Goldstone modes, and the 4n2 generators Xa span the subspace
usp(4n) − (usp(2n) ⊕ usp(2n)). The choice of Σc = M̂ † leads to the block representation
of the generators Xa as
 
P Q
1  P† QT 
Π=  , (12)
2 Q ∗ −P 
∗
Q† −P T
where P and Q are n × n complex matrices, each having 2n2 degrees of freedom. The
chemical potential term in (7) explicitly breaks the extended flavor group USp(4n) down to
its unextended U(2n) subgroup, while in the presence of both mass and chemical potential
terms, the surviving global symmetry becomes U(n) × U(n), which is the intersection of
USp(2n) × USp(2n) and U(2n).
2.2. Low-energy effective Lagrangian
Coupling the fermionic Lagrangian (7) to the SU(2) gauge field gives the microscopic
Lagrangian
1
L=− tr Fµν Fµν + Lf . (13)
2g 2
In the low energy ( ΛQCD3 ) regime, where fundamental particles are confined and
Goldstone bosons dominate, we can define a low energy effective theory. The kinetic term
of the effective Lagrangian describing the Goldstone modes Σ, parametrized as in (10)
and (11), should be invariant under the action of the global USp(4n) group
Σ(x) → sΣ(x)s T , s ∈ USp(4n) (14)
with s in the antisymmetric tensor representation. Invoking the standard symmetry
principles of chiral perturbation theory, and the principle of local flavor symmetry, the
effective Lagrangian is
F2
L= tr ∇ν Σ∇ν Σ † − F 2 M 2 tr(M̂Σ), (15)
2
where the flavor-covariant derivatives are [5]

∇ν Σ = ∂ν Σ − µ ΣBν† + Bν† Σ ,

∇ν Σ † = ∂ν Σ † + µ Σ † Bν + Bν Σ † (16)
with

12n
Bν = B̂δν,3 , B̂ = ĈI = . (17)
−12n
2.3. Phases at tree level
At zero temperature, the static part of the effective Lagrangian (15), i.e., the effective
potential, determines the vacuum alignment of the field Σ. Introducing a dimensionless
parameter ξ = 2µ/M to represent the chemical potential, we find

ξ2
Lst (Σ) = F M − tr(M̂Σ) −
2 2
tr B̂Σ B̂Σ − nξ .
† 2
(18)
4
The above two terms in Σ compete for the direction of the condensate which we denote
by Σ̄ . The condensate that gives the global minimum of the static tree level effective
Lagrangian is
Σ̄α = M̂ † cos α + Jˆ sin α, (19)
which is parametrized by α where

cos α = min 1, ξ −2 . (20)
2.3.1. Normal phase

When ξ < 1 (i.e., µ < M/2), the vacuum orientation of the condensate does not depend
on ξ and is given by Σ̄ = M̂ † . Expanding Σ around M̂ † using the Goldstone field defined
in (11) and (12) we find
1 M2 1
L = Lst (Σ̄α ) + tr ∂ν P † ∂ν P + tr P † P + tr ∂ν Q† ∂ν Q
2 2 2
2
M
− 2µ tr Q† ∂3 Q + − 2µ2 tr Q† Q + · · · . (21)
2
From this we can identify the excitations. The baryon charge b of the excitations are: b = 0
for P and P † ; b = 2 for Q; and b = −2 for Q† . The pole masses for these excitations are
M for P and P † ,
M − 2µ for Q, M + 2µ for Q† . (22)
Each mode belongs to a representation of dimension n2 .
2.3.2. Diquark condensation phase

In the regime ξ > 1 (i.e., µ > M/2), the vacuum condensate has U(n) degeneracy
corresponding to n2 true Goldstone modes. This change of massless modes indicates a
second-order phase transition at ξ = 1. When ξ > 1 the configuration (19) begins to rotate
away from M̂ † according to cos α = ξ −2 . This rotation can be written as
α
Σ̄α = sα M̂ † sαT = sα2 M̂ † , sα = ei 2 X2 , (23)
where X2 is the generator that rotates M̂ into Jˆ. We parametrize the fluctuation around the
vacuum Σ̄α as
Σ = sα S M̂ † S T sαT , (24)
where S are generated by the unrotated generators Xa .
In this phase we expand around the rotated value of the condensate Σ̄α , as in (24). To
second order
1 1 1 1
L = Lst (Σ̄0 ) + tr ∂ν PS† ∂ν PS + tr ∂ν PA† ∂ν PA + tr(∂ν QR )2 + tr(∂ν QI )2
2 2 2 2

+ iMξ −1 tr Q†R ∂3 QI − Q†I ∂3 QR
M2 2
+ ξ tr PS PS† + ξ −2 tr PA PA† + ξ 2 − ξ −2 tr Q2I + · · · . (25)
2
Here PS,A = (P ± P T )/2, and QR,I are the real and imaginary parts of the complex
matrix Q. The PS,A fields have conventional dispersion relations with masses
M2
2µ for PS , for PA , (26)
2µ
and lie in representations of dimension n(n + 1) and n(n − 1), respectively.
On the other hand, the dispersion laws for the Q fields are unconventional due to the
linear time derivative terms in (25), and are determined by the secular equation

[E 2 − p2 ] 2iEMξ −1
det = 0. (27)
−2iEMξ −1 [E 2 − p2 − M 2 (ξ 2 − ξ −2 )]
Diagonalizing, we find one true Goldstone field, which we denote Q̃, and its orthogonal
complement denoted Q̃† . These excitations have masses

4
M
0 for Q̃, 2µ 1 + 3 for Q̃† , (28)
2µ
and each lies in a representation of dimension n2 . Note that fields with this unconventional
form of dispersion also appear in the context of kaon condensation [21,22], and gauge
symmetry breaking via Bose condensation [23].
3. One-loop free energy at finite density and temperature
The bosonic low energy effective Lagrangian for 3D two-color QCD reads
F2
L= tr ∇ν Σ(x)∇ν Σ(x)† − F 2 M 2 tr M̂Σ(x) + L4 , (29)
2
where the phenomenological constant F 2 has dimensions of mass in 3D, and L4 is a
collection of terms of order O(∇, M)4 . We take M 0 without loss of generality.
3.1. Normal phase
Within the normal phase (the boundary of which will be determined in the next section),
the Lagrangian reads
1
L = −4nF 2 M 2 + P (x)† −∂ 2 + M 2 P (x)
2
1
+ Q(x)† −∂ 2 − 4µ∂3 + M 2 − 4µ2 Q(x) + L4 . (30)
2
Thus the one-loop free energy is given by

Ω = −4nF 2 M 2 + n2 tr log −∂ 2 + M 2

+ n2 tr log −∂ 2 − 4µ∂3 + M 2 − 4µ2 + L4 Σ = M̂ †

d 3p
= −4nF M + n
2 2 2
3
log p2 + M 2
(2π)

d 3p
+ n2 3
log p2 + 4iµp3 + M 2 − 4µ2 + Ω4 . (31)
(2π)
Here Ω4 is a constant that is a combination of the contact term couplings. The sum-integral
in the above is defined as
∞
d p f (p3 , p) = 2πT
d
d d−1 p f (2nπT , p). (32)
n=−∞
The first sum-integral in (31) is standard

∞ ∞
2 e−t M − n22
2
dd p 1
− log p + M 2
= dt e 4tT
(2π)d (4π)d/2 t d/2+1
0 N=−∞
d/2
∞
8(−d/2) d MT NM
= d/2
M +4 N −d/2 Kd/2
(4π) 2π T
N=1
d=3 M3 T 2 M M
−→ + M Li2 e− T + T Li3 e− T . (33)
6π π
Note that the zero temperature part is finite in odd-dimensional spacetime, and the finite
temperature correction is expressed in terms of the polylogarithm function [24]
∞ k
x
Lin (x) = . (34)
kn
k=1
The second sum-integral in (31) is the conventional effective action for a bosonic field at
finite chemical potential,

dd p
− d
log (p3 + 2iµ)2 + p2 + M 2
(2π)
∞ ∞
e−t M − n22
2
1 2µN
= dt e 4tT exp
(4π)d/2 t d/2+1 T
0 N=−∞
d/2
∞
8(−d/2) d MT −d/2 MN 2µN
= M +4 N Kd/2 cosh (35)
(4π)d/2 2π T T
N=1
d=3 M 3 T 2 − M−2µ M+2µ
−→ + M Li2 e T + Li2 e− T
6π 2π M−2µ M+2µ
+ T Li3 e− T + Li3 e− T . (36)
Once again, the zero T part is finite due to the odd dimensionality, and the finite T
corrections are expressed in terms of polylog functions.
Bringing together the contributions (33) from the P modes, and the contribution (36)
from the Q modes, the one-loop effective action in the normal phase, at finite T and µ, is

3
2 M T2 −M − M
Ω = −4nF M − n
2 2
+ M Li2 e T + T Li3 e T
6π π
3
M T 2 − M−2µ M+2µ
− n2 + M Li2 e T + Li2 e− T
6π 2π
M−2µ
− T
− M+2µ
+ T Li3 e + Li3 e T + Ω4 . (37)
3.2. Diquark condensation phase
We now compute the free energy, to 1-loop order, by expanding about the rotated
vacuum

α
Σα = sα M̂ sα , sα = exp i X2 ,
† T
(38)
2
which minimizes the tree level effective potential in the diquark condensation phase. We
parameterize the fluctuations around Σα as
Σ(x) = sα S(x)M̂ † S(x)T sαT ,
 
P Q

iΠ(x) 1P
† QT 
.
S(x) = exp , Π(x) =  (39)
2F 2 Q∗ −P 
∗
Q† −P T
The quadratic Lagrangian reads

L = −F 2 µ2 tr B̂Σα B̂ΣαT + 14n − F 2 M 2 tr M̂Σα

1 1 2
+ PS (x) −∂ + M1 + M3 PS (x)
† 2 2
2 4

1 1 2
+ PA (x) −∂ + M2 + M3 PA (x)
† 2 2
2 4

1 −∂ 2 + M2 M 3 ∂3
+ Q(x)† 1
Q(x) + L4 . (40)
2 M3 ∂3 −∂ 2 + M22
Here we have used the notations of [14,15], expressing M1 , M2 and M3 in terms of the
overall mass scale M, the baryon chemical potential µ, and the rotation angle α which
characterizes the rotated vacuum (38):
M12 = M 2 cos α − 4µ2 cos 2α,

M22 = M 2 cos α − 4µ2 cos2 α,
M32 = 16µ2 cos2 α. (41)
There are n(n + 1) PS modes, n(n − 1) PA modes, and (n2 + n2 ) Q modes. Integrating
out these modes, one obtains the one-loop free energy

Ω = −F 2 µ2 tr B̂Σα B̂ΣαT + 14n − F 2 M 2 tr M̂Σα

1 1
+ n(n + 1) tr log −∂ 2 + M12 + M32
2 4

1 1
+ n(n − 1) tr log −∂ 2 + M22 + M32
2 4

1 2 −∂ + M1
2 2 M3 ∂3
+ n tr log + Ω4
2 M3 ∂3 −∂ 2 + M22
2
2 M1 + M2 M32
2
= −4nF +
2 4

n(n + 1) d p 3 1 2
+ log p + M1 + M3
2 2
2 (2π)3 4

n(n − 1) 3
d p 1 2
+ log p 2
+ M 2
2 + M
2 (2π)3 4 3

n2 d 3 p
+ 3
log p2 + M12 p2 + M22 + p32 M32 + Ω4 . (42)
2 (2π)
Here Ω4 (α) = L4 [Σα ] is given by substituting Σ(x) = Σα into the O(∇, M)4 terms of
the effective Lagrangian.
The PS and PA mode contributions in (42) are precisely as for the P modes in the
normal phase in (33), with M replaced by M12 + M32 /4 for the PS modes, and by

M22 + M32 /4 for the PA modes. On the other hand, the final term in (42) is nontrivial
because of the quartic nature of the operator. One approach is to factor this quartic as
2
p + M12 p2 + M22 + p32 M32
2
M1 − M22 2
= (p3 + iMy) + p + M z (p3 − iMy) + p + M z −
2 2 2 2 2 2 2 2
,
2
(43)
where we have introduced the convenient dimensionless parameters y and z, as in [14,15],

M3 1 M12 + M22 M32
y= , z= + . (44)
2M M 2 4
One can now expand the final term in (42) in terms of the difference
M12 − M22 = 4µ2 sin2 α. (45)

Such an expansion proves useful in the Landau–Ginzburg analysis for the region where the
order parameter α is small, as studied in the next section.
Then the contribution of these Q-type modes is

dd p
GQ ≡ − d
log p2 + M12 p2 + M22 + p32 M32
(2π)

dd p
=− d
log (p3 + iMy)2 + p2 + M 2 z2 (p3 − iMy)2 + p2 + M 2 z2
(2π)
∞
1 M12 − M22 2k
+
k 2
k=1

d p d 1
× .
(2π) ((p3 + iMy) + p + M z ) ((p3 − iMy)2 + p2 + M 2 z2 )k
d 2 2 2 2 k
(46)
The logarithmic term in the second line of (46) factors in terms of conventional bosonic
effective actions at finite chemical potential:

dd p
G(0)
Q ≡ − d
log (p3 + iMy)2 + p2 + M 2 z2
(2π)
× (p3 − iMy)2 + p2 + M 2 z2
∞ ∞
e−t (Mz) − n22
2
2 MyN
= dt e 4tT cosh
(4π)d/2 t d/2+1 T
0 N=−∞
8(−d/2)
= 2 (Mz)d
(4π)d/2
∞
MzT d/2 −d/2 Mz N My N
+8 N Kd/2 cosh
2π T T
N=1
d=3(Mz)3 T 2 − M(z−y) M(z+y)
−→ + Mz Li2 e T + Li2 e− T
3π π M(z−y) M(z+y)
+ T Li3 e− T + Li3 e− T . (47)
The finite-temperature parts of (47) correspond to the corrections for masses M(z ± y).
The k 1 terms in the expansion (46) are more difficult to evaluate in a simple closed
form. A natural approach is to use Feynman parameters:

dd p 1
(2π) ((p3 + iMy) + p + M z ) ((p3 − iMy)2 + p2 + M 2 z2 )k
d 2 2 2 2 k
1
8(2k) d d p (s(1 − s))k−1
= ds
8(k)2 (2π)d (p2 + M 2 (z2 − y 2 ) + (2s − 1)2 iMyp3 )2k
0
1 ∞
1 k−1
e−t M
d−1 2 (z2 −y 2 )
= d−1
ds s(1 − s) dt t 2k−1− 2
8(k)2 (4π) 2
0 0
∞

×T e−t (2πnT )2 +(2s−1)2 iMy(2πnT )
n=−∞
1 ∞
1
dt t 2k−1−d/2e−t M
2 (z2 −y 2 )
e−t M
2 y 2 (1−2s)2
= ds (s(1 − s))k−1
8(k)2 (4π)d/2
0 0
∞

n2 MN
× exp − + y(1 − 2s)
4tT 2 T
N=−∞
MN
∞ d 1 z2 − y 2 u
23− 2 −4k π − 2 N 2k− 2
d d
du uk−1 K d2 −2k T
= √
8(k)2 TM 1−u (z2 − y 2 u)k− 4
d
N=1 0
MN √
× cosh y 1 − u + (T = 0 part)
T
MN
3
−4k − 32 ∞ 3 1 z2 − y 2 u
d=3 2 2 π N 2k− 2 du uk−1 K 32 −2k T
−→ √
8(k)2 TM 1−u (z2 − y 2 u)k− 4
3
N=1 0
MN √
× cosh y 1 − u + (T = 0 part).
T
(48)
Motivated by the Landau–Ginzburg analysis of the next section, we consider just the
k = 1 term, since higher terms are at least of order α 8 . In d = 3 the k = 1 term in (48)
reduces to a simple integral:

(z+y)M/T
d=3 1 1 z+y dx
G(1)
Q −→ log + . (49)
8πMy 2 z−y x(ex − 1)
(z−y)M/T
Here the first term corresponds to the T = 0 contribution, while the integral represents the
finite T correction.
Collecting together the contributions of the PS , PA and Q modes, we obtain the
following expression for the one-loop effective action in the diquark phase, at finite
temperature and chemical potential
Ω = −4nF 2 M 2 z2

n(n + 1) (M12 + M32 /4)3/2
−
2 6π

2 M 2 M32
T 1
+ M1 +
2 3
Li2 exp − M1 +
2
π 4 T 4

1 M2
+ T Li3 exp − M12 + 3
T 4

n(n − 1) (M22 + M32 /4)3/2
−
2 6π

2 2
T2 M 1 M
+ M22 + 3
Li2 exp − M22 + 3
π 4 T 4

1 M32
+ T Li3 exp − M2 +
2
T 4

n2 M 3 z3 T 2 −(z−y)M/T (z+y)M
− + Mz Li2 e + Li2 e− T
2 3π π
(z−y)M (z+y)M
+ T Li3 e− T + Li3 e− T

(z+y)M/T
(M12 − M22 )2 1 z+y dx
+ log +
32πyM 2 z−y x(ex − 1)
(z−y)M/T
4
+ O M12 − M22 + Ω4 . (50)
Note that the higher order corrections are at least of order (M12 − M22 )4 ∼ µ8 sin8 α.
4. Landau–Ginzburg model
In the Landau–Ginzburg approach [25], we regard the rotation angle α as an order

parameter. For small α, the shape of the effective potential can be used to deduce
information about the phase structure of the system [26,27]. The vanishing of the
coefficient of the term quadratic in α determines a critical line in the T and µ plane.
4.1. Landau–Ginzburg model at zero temperature
It is important first to analyze the Landau–Ginzburg model at T = 0 in order to

understand the effect of one-loop renormalization. Since we expect the critical behavior
to be near the tree-level T = 0 critical point where µ = M/2, we expand Ω in terms of the
rotation angle α and also the deviation of chemical potential from M/2:
µ 1
µ̄ ≡ − . (51)
M 2
Substituting the definitions of M1 , M2 , M3 , z and y into the T = 0 parts of the one-loop
effective potential in (50) and expanding, one finds

ΩT =0 1 M 2 2n + 1 M
= const + α + 2 + µ̄α 2
−4nF 2 M 2 32π F 2 8π F 2

+ O α 4 log α 2 − 4µ̄ . (52)
Note that the higher-order terms in (50) are not relevant as they contribute terms at least of
order α 8 . Also, we neglect the quartic contact terms Ω4 . This is motivated by the 4D case
[15] where these terms are negligibly small. The dimensionless combination M/F 2 in (52)
is the loop expansion parameter. The vanishing of the coefficient of α 2 leads to the critical
point (at one-loop)
1 M
µ̄ = − . (53)
64π F 2
This can be rewritten as

1 1 M
µ= M 1− , (54)
2 32π F 2
which shows that the critical point is shifted by a finite one-loop correction from the tree-
level T = 0 critical point µ = M/2. Actually, this shift is just a mass renormalization since
the one-loop renormalized pion mass mπ is

8(1 − d/2) M d−2 d=3 1 M
mπ = M 1 + −→ M 1 − (55)
8(4π)d/2 F 2 32π F 2
as computed in Appendix A. Thus, this result means that one-loop renormalization has the
effect of shifting the T = 0 critical point from µ = M/2 to µ = mπ /2, where mπ is the
one-loop renormalized value of M.
4.2. Landau–Ginzburg model at finite temperature
At finite temperature, the shape of the free energy as a function of the order parameter
α changes as we move through the T and µ plane. Given the explicit expression (50) it
is straightforward to plot (using, e.g., M ATHEMATICA) the free energy for chosen values
of T and µ. This gives some guidance as to the parameter region in which to search for a
critical line. We find that the appropriate critical region of parameters is
mπ
0<µ− mπ g T mπ , (56)
2
where we have defined the natural dimensionless self coupling constant of pions by
M
g= . (57)
32πF 2
In this region the temperature is far smaller than the mass of massive modes (the P ’s and
half of the Q’s), but is far larger than the mass of the almost Goldstone modes (the other
half of the Q’s). We also have the intrinsic applicability limit of the low energy effective
Lagrangian that other hadronic modes are not excited: i.e., mπ ΛQCD3 .
We expand the Landau–Ginzburg free energy (normalized by that at µ = T = 0) up to
the quartic order in α,
Ω
= c0 + c2 α 2 + c4 α 4 + O α 6 . (58)
−4nF M
2 2
Using the expression (50), the coefficient of the quadratic term is given explicitly by
1 2µ2 1
c2 = − + 2 + −2nM 3 + 4(1 + 2n)µ2 M
2 M 32πF 2 M 2
M
+ 2 −nM 2 + 4(1 + n)µ2 T log 1 − e− T
M−2µ
+ −M 2 + 4µM + 4µ2 nT log 1 − e− T
M+2µ
+ −M 2 − 4µM + 4µ2 nT log 1 − e− T
M−2µ M+2µ
− 4nµT 2 Li2 e− T − Li2 e− T . (59)
Similarly, from (50) the coefficient of the quartic term is given explicitly by

1 µ2 1 µ4
c4 = − + 10nM 3
− 56(1 + 2n)Mµ 2
+ 48(2 + 3n)
24 3M 2 768πF 2M 2 M
3n(M 2 − 4Mµ − 4µ2 )2 3n(M 2 + 4Mµ − 4µ2 )2
+ M−2µ + M+2µ
e T −1 M e T −1 M
6(nM 4 − 8(1 + n)M 2 µ2 + 32(1 + n)µ4 )
+ M
eT −1 M
M−2µ
+ 2n M 2 − 4Mµ − 40µ2 T log 1 − e− T
M+2µ
+ 2n M 2 + 4Mµ − 40µ2 T log 1 − e− T
M
+ 4 nM 2 − 16(1 + n)µ2 T log 1 − e− T

(M+2µ)/T
M−2µ M+2µ dx x
+ 8nµT 2 Li2 e− T − Li2 e− T + 12nµ3 coth .
x 2
(M−2µ)/T
(60)
The separation between the tree level part and the one-loop part is apparent in these
expressions for c2 and c4 . We stress that no approximation has been used so far, except
that (58) is an expansion for small values of the order parameter α.
In the parameter region (56) we may obtain more compact approximate expressions
for c2 and c4 by the use of the expansion formula of polylogarithms with positive integer
indices,
ψ(n) + γ − log ? ζ(n − k)
Lin e−? = (−?)n−1 + (−?)k , (61)
(n − 1)! k!
k0
k=n−1
and the formula

Λ
dx x 2
coth = + log Λ + O ? 0 + O e−Λ /Λ , (62)
x 2 ?
?
valid for 0 < ? 1 Λ. We define the dimensionless deviation parameter
2µ
δ =1− ≡ −2µ̄, (63)
M
which is of the same infinitesimal order as the coupling g. Then we find the following
simple approximate formulas for c2 and c4 :

T T
c2 −δ + g 1 − 2n log , (64)
M Mδ
1 5n T
c4 − + g . (65)
24 24 Mδ
In the Landau–Ginzburg approach, the critical line is characterized by the condition c2 = 0,
and the tricritical point is characterized by the conditions c2 = c4 = 0 [26]. Expressing
dimensionful quantities in the unit of the renormalized pion mass mπ , rather than the bare
pion mass M, the critical conditions given by the approximated coefficients (65) read
mπ T
critical line: µ − = ngT log , (66)
2 mπ − 2µ + gmπ

mπ g
tricritical point: µtri = 1+ ,
2 1 − 2 log(5ng)
5
mπ
Ttri = . (67)
n(5 − 2 log(5ng))
Fig. 1. Plot of the phase diagram for two flavors (n = 1) and coupling g = 0.002 which corresponds to
M/F 2 = 0.2. The solid line is the critical line in (66), and the x marks the tricritical point (67). The dotted
line is an approximation based on expanding the effective potential to sixth order in α and demanding it to have
the form Ω(α) = c6 α 2 (α 2 − const)2 + const. Note that the scale has been magnified to show just the vicinity of
the critical line.
We have confirmed numerically that these expressions approximate very well the critical
line and tricritical point derived numerically from the vanishing of the full expressions
for c2 and c4 in (59) and (60), for the parameter region of concern. Accordingly our
approximation is indeed self-consistent in that the critical region is within the region (56).
An example of the phase diagram as given by (66) and (67) is exhibited in Fig. 1.
4.3. The critical line and a weakly interacting Bose gas
The critical line (66) has an interpretation in terms of a weakly interacting Bose
gas, as in the (3 + 1)-dimensional case studied in [14,15], where the diquark phase
was argued to be a Bose condensate. The fact that the same structure appears in our
(2 + 1)-dimensional model is quite interesting, as Bose condensation is rather different
in two spatial dimensions. So we show here explicitly how the usual Bose gas results are
consistent with the Landau–Ginzburg effective Lagrangian approach of this paper.
Recall that for a free relativistic Bose gas in d spatial dimensions, the charge density is
given by [28,29]

8 d+1 µ µ
ρ = (d+1)/2 T d gd M
2
T ,M − gd MT ,−M . (68)
π
The first term corresponds to the particles and the second to the antiparticles. For d = 2
[29],
µ (µ−M)/T M
g2 MT , M = Li2 e −
T
log 1 − e(µ−M)/T . (69)
In the (2 + 1)-dimensional two-color QCD system discussed in this paper, the Q

excitations have baryon charge 2, so we should replace the chemical potential in (68), (69)
by µ → 2µB , and the charge density in (68) by ρ → nB /2. At low temperature the lightest
modes, the Q modes, dominate. There are n2 such modes, where we recall that the number
of flavors is Nf = 2n. Furthermore, at low temperature only the particles contribute and so
we expect
T 2 M 2µB
nB = 2n2 g2 T , M . (70)
2π
In the Landau–Ginzburg approach discussed in the previous section, the critical line
separating the normal and diquark phases is determined by the vanishing of the coefficient
c2 of the term quadratic in the order parameter α in the expansion of the free energy (58).
At one-loop, the vanishing of c2 in (59) near the critical point µB ≈ mπ /2 can be written
as [here we use slightly less restrictive approximations than were used in obtaining (66)]

mπ nT 2 (2µ −M)/T M (2µB −M)/T

µB − ≈ Li 2 e B
− log 1 − e
2 32πF 2 T
nT 2 2µB
= g2 M T , M . (71)
32πF 2
Combining these two results (70) and (71) we see that

mπ
nB = 16Nf F µB −
2
. (72)
2
In the effective Lagrangian approach the renormalized vacuum energy is2
mπ n2B
Evac = nB + + ···, (73)
2 32F 2 Nf
where the coefficients are fixed by the symmetries. Since the baryon number chemical
potential is
∂Evac
µB = (74)
∂nB
we see that the relation (72) follows directly. This demonstrates the consistency of the
Landau–Ginzburg effective Lagrangian approach with the standard Bose gas results, and
shows that the critical line separating the normal and diquark phases describes a free Bose
gas in two spatial dimensions.
2 Note that in (73) we differ from [14,15] by a factor of 2 in the interaction term. This is because only half of
the Q modes contribute to the low T Bose gas, and so only they should be selected out of the general Π field in
this limit.
5. Conclusion and discussion
In this paper we have used the low energy effective field theory method at finite
temperature and finite baryon density to investigate the phase structure of three-dimen-
sional parity-invariant SU(2) QCD with fundamental quarks. We computed the one-loop
effective potential at both T = 0 and finite T , at nonvanishing baryon number chemical
potential. At T = 0, the tree-level critical point µ = M/2 between the normal and diquark
phase receives a finite shift at one-loop, but this shift corresponds exactly to the mass
renormalization, so that at one-loop the T = 0 critical point is µ = mπ /2, where mπ
is the one-loop renormalized pion mass. At nonzero temperature we found a simple
expression (50) for the effective potential, suitable for a Landau–Ginzburg analysis in
terms of the order parameter α which describes the rotation of the tree-level vacuum away
from the normal phase vacuum alignment. The subsequent Landau–Ginzburg analysis of
this free energy revealed a critical line in the (µ, T ) plane, given approximately by (66).
This line separates the normal and diquark phases, and ends in a tricritical point whose
location is given approximately by (67). This general structure, with a critical line ending
in a tricritical point, is qualitatively similar to that found in the 4D case [15], except that the
T log T temperature dependence for the critical line in (66) becomes a power law ∼ T 3/2
in the 4D case.
However, the details of the 3D phase diagram are much less clear than in the 4D case,
due to the presence of infrared divergences, which become more pronounced at higher
loops, where they will appear as powers. In a restricted region of the relevant coupling
and parameters (such as g, T /M and µ̄), the one-loop effective potential does describe
the correct physics. The one-loop calculation indicates that some transition happens in
the vicinity of the critical line (66), and the one-loop Landau–Ginzburg analysis suggests
that this is a line of second-order phase transitions, as is the case in 4D [15]. However,
in 3D there are infrared divergences as µ̄ tends to zero (µ → M/2), which become more
significant at higher loops. This suggests that a nonperturbative analysis may be necessary
to identify the true nature of the phase transition. There is another reason to think that
in the 3D case the critical line may not be one of second-order phase transitions: in
two spatial dimensions, only truly massless modes Bose condense—massive modes, no
matter how light, do not Bose condense [29]. If the Q̃ modes are truly massless even
at nonzero temperature, we would expect to be able to dimensionally reduce the system
to two dimensions, in which case the Mermin–Wagner–Coleman theorem tells us that
a continuous symmetry cannot be broken (in the absence of long-range interactions).
What then is the fate of the diquark phase at nonzero T ? Perhaps nonperturbative effects
generate an exponentially small, ∼ e−F /T (rather than actually zero), mass for the
2
modes (which at tree level are exactly massless), and the critical line is actually a cross-
over between a normal and diquark phase, until the tricritical point is reached, after
which it presumably becomes a line of first-order transitions. We mention that such a
nonperturbative exponentially small mass has been observed in the thermodynamics of
the large N O(N) sigma model in 3D [30]. This would mean that the standard perturbative
effective potential approach to two-color QCD in 3D is good for describing the physics on
some portion of the critical line where it is a noninteracting Bose gas, but away from this
region the perturbative analysis is insufficient. It would be very interesting to investigate
the relevant nonperturbative effects, perhaps from a large NF analysis, or from a lattice
study of 3D two-color QCD.
Acknowledgements
We thank A. Kovner, V. Miransky, I. Shovkovy, K. Splittorff and J. Verbaarschot for

helpful discussions. We also thank the US DOE for support through the grant DE-FG02-
92ER40716.
Appendix A. One-loop renormalization
In this appendix we perform the one-loop renormalization of the pion mass and
wave function. As the theory is renormalizable at each order of the chiral perturbation,
renormalization constants are common for both normal and diquark phase. Thus we shall
concentrate on the simpler case of the normal phase, and set Ω4 = 0 as mentioned in
Section 4.1. We substitute (39) at α = 0 into the tree-level term of (29) and expand in P
and Q fields up to the quartic order. We write
(1) (2) (3) (4)
Pab = φab + iφab , Qab = φab + iφab , a, b = 1, . . . , n. (A.1)
(1) (1)
Without loss of generality we choose to compute φ11 (p)φ11 (−p). The terms that
contribute to this Green’s function are:
4 n
1 (α) 2 (α)
Ltree = φab −∂ + M 2 φab
2
α=1 a,b=1
4 n

M 2 (1) 4 (1) 2 (α) 2 4
(α) 2 (α) 2
− φ11 + 2 φ11 φ11 + φa1 + φ1a
96F 2
α=2 α=1 a=1
4
1 (1) 2 (α) 2 (α) 2 (α)2
4 n
− φ 4 ∂φ11 + ∂φa1 + ∂φ1a
96F 2 11
α=2 α=1 a=1
4
1 (1)2 (α) 2 (α) 2 (α) 2
4 n
− ∂φ11 4 φ11 + φa1 + φ1a . (A.2)
96F 2
α=2 α=1 a=1
(1) (1) (α) (α)
There exist vertices of the form φ11 ∂ν φ11 φ1a ∂ν φ1a , a = 1, but they do not contribute to
(1) (1) !
φ11 φ11 .
Using these vertices, we have
(1) (1) !
φ11 (p)φ11 (−p)

1 M2 dd q 1
= 2 p 2
+ M 2
+ 4 · 3
(p + M ) 2 2 96F 2 (2π) q + M 2
d 2

M2 dd q 1
+ 2 · 1 3 + 8(n − 1)
48F 2 (2π)d q 2 + M 2

1 dd q −q 2
− 2 · 1 12 + 8(n − 1)
96F 2 (2π) q + M 2
d 2

1 dd q 1
− 2 · 1 12 + 8(n − 1) (−p ) 2
96F 2 (2π)d q 2 + M 2

1 2n + 1 dd q 1
= 2 p 1+
2
(p + M 2 )2 12F 2 (2π)d q 2 + M 2

n−1 dd q 1
+ M2 1 +
6F 2 (2π)d q 2 + M 2

1 2n + 1 8(1 − d/2) d−2
= 2 p 2
1 + M
(p + M 2 )2 12F 2 (4π)d/2

n − 1 8(1 − d/2) d−2
+M 1+
2
M . (A.3)
6F 2 (4π)d/2
" dd q
We have discarded a term proportional to the divergent integral (2π) d in the dimensional
regularization.
The wave function renormalization should absorb the factor multiplying p2 ,

2n + 1 8(1 − d/2) d−2
φ(p) = φ(p)ren 1 + M . (A.4)
24F 2 (4π)d/2
Accordingly,

(1) (1) ! 1 1 8(1 − d/2) d−2
φ11 (p)ren φ11 (−p)ren = 2 p +M 1−
2 2
M .
(p + M 2 )2 4F 2 (4π)d/2
(A.5)
At p = 0,

(1) (1) ! 1 8(1 − d/2) M d−2 1
φ11 (0)ren φ11 (0)ren = 2 1 − ≡ 2. (A.6)
M 4(4π)d/2 F 2 mπ
That is,

8(1 − d/2) M d−2
mπ = M 1 + . (A.7)
8(4π)d/2 F 2
Note that the n-dependent part of the coefficients of p2 and M 2 in (A.3) are identical,
leading to the n-independent mass renormalization.
References
[1] J. Gasser, H. Leutwyler, Chiral perturbation theory to one loop, Ann. Phys. 158 (1984) 142;
J. Gasser, H. Leutwyler, Chiral perturbation theory: Expansions in the mass of the strange quark, Nucl. Phys.
B 250 (1985) 465.
[2] S. Weinberg, The Quantum Theory of Fields, Vol. 2: Modern Applications, Cambridge Univ. Press,
Cambridge, UK, 1996, Section 19.
[3] J.J.M. Verbaarschot, T. Wettig, Random matrix theory and chiral symmetry in QCD, Annu. Rev. Nucl. Part.
Sci. 50 (2000) 343, hep-ph/0003017.
[4] S. Hands, J.B. Kogut, M.P. Lombardo, S.E. Morrison, Symmetries and spectrum of SU(2) lattice gauge
theory at finite chemical potential, Nucl. Phys. B 558 (1999) 327, hep-lat/9902034.
[5] J.B. Kogut, M.A. Stephanov, D. Toublan, On two-color QCD with baryon chemical potential, Phys. Lett.
B 464 (1999) 183, hep-ph/9906346.
[6] J.B. Kogut, M.A. Stephanov, D. Toublan, J.J.M. Verbaarschot, A. Zhitnitsky, QCD-like theories at finite
baryon density, Nucl. Phys. B 582 (2000) 477, hep-ph/0001171.
[7] D. Toublan, J.J.M. Verbaarschot, Dirac spectra and real QCD at nonzero chemical potential, in: K.A. Olive,
M.A. Shifman, M.B. Voloshin (Eds.), Minneapolis 2002, Continuous Advances in QCD, World Scientific,
Singapore, 2002, pp. 265–290, hep-th/0208021.
[8] J.J.M. Verbaarschot, What really matters, Talk at the Channel Meeting, Plymouth, August 2002.
[9] M.E. Peskin, The alignment of the vacuum in theories of technicolor, Nucl. Phys. B 175 (1980) 197.
[10] E. Dagotto, F. Karsch, A. Moreo, The strong coupling limit of SU(2) QCD at finite baryon density, Phys.
Lett. B 169 (1986) 421.
[11] C. Baillie, K.C. Bowler, P.E. Gibbs, I.M. Barbour, M. Rafique, The chiral condensate in SU(2) QCD at finite
density, Phys. Lett. B 197 (1987) 195.
[12] J.B. Kogut, D. Toublan, D.K. Sinclair, The phase diagram of four flavor SU(2) lattice gauge theory at nonzero
chemical potential and temperature, Nucl. Phys. B 642 (2002) 181, hep-lat/0205019;
J.B. Kogut, D. Toublan, D.K. Sinclair, SU(2) lattice gauge theory at nonzero chemical potential and
temperature, Talk at Lattice 2002, hep-lat/0208076.
[13] G.V. Dunne, S.M. Nishigaki, Two-color QCD in 3D at finite baryon density, Nucl. Phys. B 654 (2003) 445,
hep-ph/0210219.
[14] K. Splittorff, D. Toublan, J.J.M. Verbaarschot, QCD with two colors at finite baryon density at next-to-
leading order, Nucl. Phys. B 620 (2002) 290, hep-ph/0108040.
[15] K. Splittorff, D. Toublan, J.J.M. Verbaarschot, Thermodynamics of chiral symmetry at low densities, Nucl.
Phys. B 639 (2002) 524, hep-ph/0204076.
[16] R.D. Pisarski, Chiral symmetry breaking in three-dimensional electrodynamics, Phys. Rev. D 29 (1984)
2423.
[17] A.P. Polychronakos, Symmetry breaking patterns in (2 + 1)-dimensional gauge theories, Phys. Rev. Lett. 60
(1988) 1920.
[18] P.H. Damgaard, U.M. Heller, A. Krasnitz, T. Madsen, A quark–antiquark condensate in three-dimensional
QCD, Phys. Lett. B 440 (1998) 129, hep-lat/9803012.
[19] S.R. Coleman, E. Witten, Chiral symmetry breakdown in large-N chromodynamics, Phys. Rev. Lett. 45
(1980) 100.
[20] U. Magnea, The orthogonal ensemble of random matrices and QCD in three dimensions, Phys. Rev. D 61
(2000) 056005, hep-th/9907096.
[21] T. Schafer, D.T. Son, M.A. Stephanov, D. Toublan, J.J. Verbaarschot, Kaon condensation and Goldstone’s
theorem, Phys. Lett. B 522 (2001) 67, hep-ph/0108210.
[22] V.A. Miransky, I.A. Shovkovy, Spontaneous symmetry breaking with abnormal number of Nambu–
Goldstone bosons and kaon condensate, Phys. Rev. Lett. 88 (2002) 111601, hep-ph/0108178.
[23] F. Sannino, K. Tuominen, Spontaneous symmetry breaking in gauge theories via Bose–Einstein condensa-
tion, hep-ph/0303167.
[24] M. Abramowitz, I. Stegun, Handbook of Mathematical Functions, Dover, 1998.
[25] E.M. Lifshitz, L.P. Pitaevskii, Statistical Physics: Part 2, Theory of the Condensed State, Pergamon Press,
Oxford, 1980.
[26] K. Huang, Statistical Mechanics, 2nd Edition, Wiley, New York, 1987.
[27] N. Goldenfeld, Lectures on Phase Transitions and the Renormalization Group, Addison–Wesley, Reading,
MA, 1992.
[28] J.I. Kapusta, Bose–Einstein condensation, spontaneous symmetry breaking, and gauge theories, Phys. Rev.
D 24 (1981) 426.
[29] H.E. Haber, H.A. Weldon, Thermodynamics of an ultrarelativistic Bose gas, Phys. Rev. Lett. 46 (1981) 1497;
H.E. Haber, H.A. Weldon, Finite temperature symmetry breaking as Bose–Einstein condensation, Phys. Rev.
D 25 (1982) 502.
[30] B. Rosenstein, B.J. Warr, S.H. Park, Thermodynamics of the O(N )-invariant sigma model in (2 + 1)
dimensions, Nucl. Phys. B 336 (1990) 435.
[31] Z. Fodor, S.D. Katz, A new method to study lattice QCD at finite temperature and chemical potential, Phys.
Lett. B 534 (2002) 87, hep-lat/0104001.
[32] Z. Fodor, S.D. Katz, Lattice determination of the critical point of QCD at finite T and µ, JHEP 0203 (2002)
014, hep-lat/0106002.
Erratum
Erratum to: “Asymptotics of the perturbative

series for fB ∗ /fB ”
[Nucl. Phys. B 663 (2003) 280] ✩
F. Campanario, A.G. Grozin, T. Mannel
Institut für Theoretische Teilchenphysik, Universität Karlsruhe, Germany
Received 28 July 2003
PACS: 12.38.Cy; 12.39.Hg
There is a sign error in the second equation in Appendix B. The correct formula is

1 αs (m) 2
αs (m) = αs (m) 1 + (32CA − 39CF )TF + ··· .
9 4π
(m/m̂)
Due to this fact, several coefficients in the multiline formula for c3 in Appendix B
should be corrected:
Line Before Replace

13 CF TF − 2263 2029
18 → − 18
15 CA TF − 3541 3733
18 → − 18
20 CF2 TF + 115 275
6 →− 6
21 CF CA TF + 3856 3712
9 →+ 9
22 2T
CA F + 322 1858
27 → + 27
28 CF2 CA TF − 2417 1559
3 →− 3
28 2T
CF CA F − 3301 3775
9 →− 9
29 3T
CA F − 2051 4739
18 → − 18
✩
doi of original article: 10.1016/S0550-3213(03)00354-7.
E-mail address: a.g.grozin@inp.nsk.su (A.G. Grozin).
doi:10.1016/j.nuclphysb.2003.08.005
332 F. Campanario et al. / Nuclear Physics B 670 (2003) 331–332
The last term of the coefficient of αs3 in (5.8) should be corrected accordingly:
17538918028079 17341442069927
+ →+ .
767418048576 767418048576
Numerical changes are very small. In Table 2, the exact coefficient at L = 3 should be
corrected: 56.63 → 56.37. Eq. (3.34) becomes
N0 = 0.288 · (1 + 0.075 + 0.630 + · · ·) ≈ 0.491.
Our conclusions are not affected.
Nuclear Physics B 670 [FS] (2003) 335–358
Comments on nonunitary conformal field theories

Terry Gannon
Department of Mathematical Sciences, University of Alberta, Edmonton, T6G 2G1 Canada
Received 19 May 2003; accepted 30 July 2003
Abstract
As is well known, nonunitary RCFTs are distinguished from unitary ones in a number of ways,
two of which are that the vacuum 0 does not have minimal conformal weight, and that the vacuum
column of the modular S matrix is not positive. However, there is another primary field, call it o,
which has minimal weight and has positive S column. We find that often there is a precise and useful
relationship, which we call the Galois shuffle, between primary o and the vacuum; among other things
this can explain why (like the vacuum) its multiplicity in the full RCFT should be 1. As examples
we consider the minimal WSU(N) models. We conclude with some comments on fractional level
admissible representations of affine algebras. As an immediate consequence of our analysis, we get
the classification of an infinite family of nonunitary WSU(3) minimal models in the bulk.
PACS: 11.25.Hf
1. Introduction
In any RCFT, each chiral sector Ha (a module of the chiral algebra, i.e., of the
vacuum sector H0 ) has a Hermitian inner product , a , obeying vx, ya = x, ω(v)ya for
x, y ∈ Ha and all v in the chiral algebra V = H0 , where ω is an anti-linear anti-involution.
In a unitary RCFT, this inner product is in addition positive-definite (i.e., x, x > 0 when
x = 0).
From the point of view of string theory (being a quantum field theory), the requirement
of unitarity seems natural,1 and indeed much of the work on fusion rings and on what we
E-mail address: tgannon@math.ualberta.ca (T. Gannon).

1 Perhaps though this belief is somewhat naive. As pointed out to the author by Christoph Schweigert, in string
theory the matter CFT is coupled to the (super-)ghost CFT, and the latter is, of course, nonunitary. Physically,
what is required is the unitarity of the corresponding BRST cohomology, since the physical states live there.
doi:10.1016/j.nuclphysb.2003.07.030
336 T. Gannon / Nuclear Physics B 670 [FS] (2003) 335–358
will call modular data has assumed it. But from the perspective of, e.g., statistical models, at
least on a torus, that requirement seems unnecessarily restrictive. Indeed, some of the better
known RCFTs are nonunitary, such as the Yang–Lee model (c = −22/5). Presumably
(but see [2]!), most RCFTs will be nonunitary—for example, the Virasoro minimal model
(p, p ), with central charge c = 1 − 6 (p − p )2 /pp , is unitary iff |p − p | = 1.
The Wess–Zumino–Witten models [3], describing strings living on Lie group manifolds
and corresponding to affine algebras at positive integer levels (see, e.g., [4]), will always
be unitary, as will their GKO cosets [5]. On the other hand, ‘admissible representations’
[6] of affine algebras at fractional levels do not (yet) have a direct interpretation as an
RCFT, although their cosets have one as nonunitary RCFTs—in fact, this is a powerful
way to construct nonunitary theories. For example, the Virasoro minimal (p, p ) model
p
corresponds to the coset ( su(2)m ⊕ su(2)1 )/
su(2)m+1 , where m = p−p − 2 ∈ Q.
In this paper we compare nonunitary and unitary RCFTs. We isolate in Section 3 what
seems to be a key uncertainty in nonunitary theories: the multiplicity in the RCFT (what
we will denote below multn (o) = Moo ) of the primary o with minimal conformal weight.
We learn that in any RCFT it will be greater than 0, but the question of its precise value
turns out to be important. We conjecture in Section 4 that this multiplicity should equal
1; this would imply, for nonunitary RCFT, irreducibility of NIM-reps and finiteness results
for modular invariants and NIM-reps, among other things. This multiplicity issue did not
arise in the Virasoro minimal model classification [7], because as we will see in Section 4,
modular invariance there forces this multiplicity to be 1.
We find that in many theories, the primary o is related to the vacuum 0 in a certain simple
way—we will say such theories possess the Galois Shuffle (or GS) property. In particular,
this holds for the nonunitary WN = WSU(N) and WSO(2N) minimal models, at arbitrary
rank N and arbitrary values p, p (these include, of course, the Virasoro minimal models).
Related comments are made for the affine algebras at fractional level. Because of their role
in the quantum Drinfeld–Sokolov reduction [8], this suggests that the GS property may be
fairly typical among RCFTs.
Nevertheless, we know that not all RCFTs possess this property. When the GS property
holds for a given RCFT, there are many nontrivial consequences, as we will see in
Sections 4 and 6. One of these will give us for free a nonunitary RCFT classification.
Let us briefly describe another consequence.
First, recall that there are several unitary matrices
S which diagonalise the fusion rules,
c
i.e., for which the fusion coefficients Nab can be formally recovered from Verlinde’s
formula when the modular matrix S is replaced with S. In particular, take the modular
matrix S and permute the columns, and multiply each column by an arbitrary phase. But
typically the symmetry condition S = S t will be lost. However, there are two generic ways
to find symmetric unitary matrices S which diagonalise the fusions: simple-currents and
the Galois symmetry. We will discuss these generic constructions next section.
Now, the GS property implies that the fusion ring of the nonunitary theory can be
diagonalised by a symmetric S-matrix S which obeys all the usual properties of a unitary
The relationship between the (non)unitarity of the matter CFT and the unitarity of that cohomology space isn’t
obvious. For a careful review of the relation of CFT to string theory, see, e.g., [1], especially Section 6.
T. Gannon / Nuclear Physics B 670 [FS] (2003) 335–358 337
theory (e.g., the 0th column is strictly positive), and that

S will be obtained from the ‘sick’
modular S matrix by a generic construction (a ‘Galois shuffle’). (By ‘sick’ here we only
mean that the vacuum column of S is not positive.) We call S the unitarisation of the
nonunitary theory. In other words, the fusion rings of nonunitary and unitary theories are
indistinguishable, at least when the nonunitary theory obeys the GS property.
To give a simple explicit example, the fusions of the nonunitary Yang–Lee model are
diagonalised, of course, by its ‘sick’ modular S matrix

2 − sin(2π/5) sin(π/5)
S=√ .
5 sin(π/5) sin(2π/5)
But the Yang–Lee fusions are also diagonalised by the unitarisation

2 sin(π/5) sin(2π/5)
S=√ ,
5 sin(2π/5) − sin(π/5)
which is the modular S matrix of the (unitary) WZW model for the affine algebras G 2

and F4 at level 1. However, as we will discuss below, the unitarisation will not in general
correspond to a WZW model. Incidentally, the tensor product of the G 2,1 WZW model
with the Yang–Lee model is an example of a theory which does not obey GS.
What does this unitarisation mean at the level of the RCFT? Indirectly, we can see
the corresponding unitary theory in, e.g., the fusion ring and the list of cylindrical and
toroidal partition functions. Typically these partition functions will be in one-to-one
correspondence with those of the nonunitary theory, and we will use this correspondence
below to obtain the classification of a new infinite family of nonunitary theories. But can
we see the unitarisation directly inside the nonunitary theory, perhaps at the level of the
chiral algebras? This is still unclear.
2. Background material
The characters χa (τ ) = TrHa q L0−c/24 of an RCFT carry a representation of SL2 (Z):

χa (−1/τ ) = Sab χb (τ ), (1a)
b

χa (τ + 1) = Tab χb (τ ). (1b)
b
The subscripts a, b here label the (finitely many) chiral primary fields φ. The ‘modular
matrices’ S and T are unitary and symmetric, and T is diagonal with entries exp[2πi (ha −
c
24 )], where ha are the conformal weights and c is the central charge. The matrix T has
finite order, i.e., all numbers ha and c are rational. The matrix C = S 2 is the order-
2 (or order-1) permutation matrix called charge-conjugation. Then (ST )3 = (T S)3 = C.
Moreover,
∗
Sab = SCa,b = Sa,Cb . (1c)
These properties should hold for any RCFT [9].
c
The fusion coefficients Nab can be expressed using Verlinde’s formula [10]:
Sad Sbd S ∗
c
Nab = cd
, (2)
S0d
d
where ‘0’ denotes the vacuum. We have h0 = 0 and C0 = 0. It is convenient to define the
fusion matrices Na by (Na )bc = Nab c
. Then Verlinde’s formula says that S simultaneously
diagonalises all matrices Na , with eigenvalues Sab /S0b .
In the case of nonunitary RCFTs (as discussed next section), this so-called ‘vacuum’ 0
does not possess all the usual properties expected of a physical vacuum, and so perhaps a
better name for it would be ‘identity’.
Definition 1. By modular data we mean here unitary matrices S, T such that T is diagonal
and of finite order, S is symmetric, (ST )3 = S 2 =: C is an order-2 (or order-1) permutation
matrix, C0 = 0, and the fusion coefficients Nab c given by (2) are all nonnegative integers
Z+ := {0, 1, 2, . . .}.
See, e.g., [11,12] for some of the basic properties of modular data, although those papers
primarily specialise to what we call below unitary modular data (see Definition 2 in the next
section). There are many additional properties which the modular data of a healthy RCFT
should obey. For one important example [13], for any primaries a, b ∈ Φ, the numbers
defined by
1
∗1
a ∗ 2 −2
Z(a, b) := T002 Tbb2 Ncd Sbc S0d Tdd Tcc
c,d∈Φ
must always be integers satisfying
|Z(a, b)| Naa

b
, (3a)
Z(a, b) ≡ Naa
b
(mod 2). (3b)
In particular, Z(a, 0) ∈ {±1, 0} is the Frobenius–Schur indicator. Also, [14] gives formulas
for the traces of S, ST and (ST )2 , among other things. These identities and conditions
would not play any special role in this paper, for reasons we will give later, and so are
ignored in Definition 1. We turn next to an additional condition which will be used below.
Form the vector χ(i) := (χa (i)) (each character converges throughout the upper half-
plane, so these character values are defined). Since each character

χa (τ ) = q ha −c/24 Aa,n q n
n=0
has nonnegative coefficients Aa,n , each component χa (i) of χ (i) will be strictly positive.
Because τ = i is a fixed-point of the transformation τ → −1/τ , χ (i) will be an eigenvector
of S with eigenvalue 1:

Sab χb (i) = χa (i).
b
Fact 1. S has a strictly positive eigenvector, corresponding to eigenvalue 1.
The one-loop torus partition function [15]

Z(τ ) = Mab χa (τ ) χb (τ )∗
a,b
is modular invariant: Z(−1/τ ) = Z(τ ) = Z(τ + 1), etc., so the coefficient matrix M lies
in the commutant of the SL2 (Z) representation: SM = MS, T M = MT . We also know
that the coefficients Mab are multiplicities, and so lie in Z+ . Also, the multiplicity of the
vacuum must be 1: M00 = 1.
A vacuum-to-vacuum amplitude of fundamental importance in boundary CFT, is the
cylindrical partition function (see, for instance, [16–19] and references therein)

Zαβ (t) = nβaα χa (it).
a
β β
The coefficients naα will be nonnegative integers. Define matrices na by (na )αβ = naα .
c are required to form a representation, called a NIM-rep, of the fusion ring:
These matrices
na nb = c Nab nc . We can simultaneously diagonalise the na : Eq. (2) is replaced with
∗
ψαd Sad ψβd
nβaα = , (4)
S0d
d
where each d ∈ Φ appears in (4) with a certain multiplicity multn (d) 0 independent of
a, α, β.
There is a compatibility condition between the toroidal and cylindrical partition
functions: the multiplicity multn (d) ∈ Z+ of each primary d in (4) must equal the diagonal
entry Mdd of the corresponding modular invariant M. For instance, the diagonal modular
invariant M = I corresponds to the fusion matrix choice na = Na .
Two subtleties which we will avoid this paper are the gluing automorphism of [20],
and the fact that the characters χa (τ ) are usually not linearly independent. These play no
special role in our discussion, and the interested reader can consult the relevant
literature.
To fix notation, note that we have chosen S to correspond to the matrix 01 −1 0
rather than
its inverse, which also appears in the literature.
Next, let us review the lesser known Galois action of RCFTs [21]. The entries of S will
lie in some cyclotomic extension of Q—that just means each entry Sab can be written as
a polynomial pab (x), with rational coefficients, evaluated at some root of unity x = ξL :=
exp[2πi/L]. In fact, L can be taken to be the order of the matrix T [22,23]. By the ‘Galois
group’ of this cyclotomic extension, we mean the multiplicative group of integers mod L
which are coprime to L—we write this Z× ×
L . For example, Z12 = {1, 5, 7, 11}. Any such
×
, ∈ ZL acts on the number pab (ξL ) by sending it to pab (ξL, )—we denote this ‘Galois

automorphism’ by σ, . For instance: σ, fixes all rationals; σ, cos 2π La = cos 2π ,a L ;
√ √
σ, i = ±i for , ≡ ±1 (mod 4), respectively; and σ, 2 = ± 2 depending on whether or
not , ≡ ±1 (mod 8). What is so special about the σ, is that they are symmetries of the
cyclotomic numbers: σ, (u+v) = σ, (u)+σ, (v) and σ, (uv) = σ, (u)σ, (v), for any numbers
u, v in the Lth cyclotomic extension Q[ξL ] of Q.
The point [21] is that to any , ∈ Z×L , there is a permutation a → σ, a of the primary
fields, and a choice of signs /, (a) = ±1, such that
σ, (Sab ) = /, (a)Sσ, a,b = /, (b)Sa,σ, b . (5a)

This Galois action should be regarded as a generalisation of charge-conjugation. In
particular, , = −1 corresponds to the familiar Galois automorphism σ−1 = ∗ (complex
conjugation), with signs /−1 (a) = +1 and permutation given by the charge-conjugation
σ−1 a = Ca. It was proved in [22] that

c c
hσ, a − ≡ , ha −
2
(mod 1), (5b)
24 24
where we can take , to be coprime to the order L of the matrix T (in fact, it suffices here to
take , in (5b) coprime to the denominator of the rational number ha − c/24). The partition
functions M and n respect this Galois action [21,24]:
Mab = Mσ a,σ b , (5c)

multn (a) = multn (σ a). (5d)
As mentioned in the Introduction, there are infinitely many different unitary matrices

S which diagonalise the fusion coefficients. We conclude this section by mentioning a
generic way to construct symmetric unitary matrices S. Choose any simple-current J and
any Galois automorphism σ, , where the integer , is coprime to the order L of T . Define
the following matrices:

Sab := σ, (SJ a,J b ) = exp 2πi, QJ (b) + QJ (J ) /, (b)Sa,σ, J b , (6a)
Tab := δab (TJ a,J b ), . (6b)
This implies S = σ, (P SP t ) and T = σ, (P T P t ), where P is the (orthogonal) permutation
matrix Pab = δb,J a , and these define a representation of the modular group SL2 (Z):
(
S T)3 = S 2 = C. This new representation will typically be inequivalent to that of the
original S, T . Note that
S is manifestly symmetric. That S diagonalises the fusions can be
verified either by direct calculation (using the fact that the fusion coefficients are rational
numbers), or by noting from the right-side of (6a) that the bth column of S has been
permuted by σ, ◦ J , and multiplied by the phase exp[2πi,(QJ (b) + QJ (J ))]/, (b). Any
such S, T will automatically obey the constraints (3) and identities in [14], provided S, T
do, and provided we choose the squareroot T1/2 := P T ,/2P t (this is why nothing will be
gained in this paper by manifestly imposing those conditions on modular data). Of course
an additional (trivial and well-known) construction is to tensor by any 1-dimensional
S, T
representation of SL2 (Z), i.e., to choose any (not necessarily primitive) 6th root t of unity
and to replace S → t −3
S, T → t T (t should be a 6th root of unity, in order to preserve the
positivity of C).
At least sometimes, the matrices S, T in (6) can be realised by ‘characters’. Let L be
the order of T , and suppose , is a perfect square m2 (mod L). For instance, if L or L/2 is
a power of any odd prime, then half of the ,’s in Z× L would be perfect squares mod L. Let
m be a multiplicative inverse (mod L) of m, and choose any matrix A ∈ SL2 (Z) obeying

a b m 0
A= ≡ (mod L).
c d 0 m
Such an A will always exist and, with respect to the SL2 (Z) representation defined by (1),
will correspond to the matrix Gm defined by [22,23]

(Gm )ab = ST m ST m ST m C ab = /m (a)δb,σ, (a) .
Define the functions

aτ + b
a (τ ) := χJ m a
χ = /m (a)χJ σma (τ ). (6c)
cτ + d
Then these ‘characters’ realise our matrices
S, T in the sense of (1):

a (−1/τ ) =
χ b (τ ),
Sab χ
b

a (τ + 1) =
χ Tab χ
b (τ ).
b
are integers but may not be positive—in
Note that the coefficients of these ‘characters’ χ
fact,
S may not satisfy Fact 1 even if this ‘character’ interpretation exists.
More generally however (e.g., for the Yang–Lee model of Section 1 or the c = −25/7
minimal model studied in Section 6), the modified modular matrices S, T would not be
realised by merely permuting the characters (i.e., reordering the primaries), and their direct
meaning in terms of the nonunitary theory is still unclear.
To our knowledge, Eqs. (6a), (6b) and the related remarks are new (but for some
thoughts in this direction see Section 5 of [25]).
3. Unitary vrs nonunitary RCFTs
In this section we describe the distinctions between the more familiar unitary RCFTs,
and the more complicated and common nonunitary RCFTs.
The assumption of unitarity certainly does yield some simplifications. For instance,2 as
is well known, the conformal weight h = 1 fields in the vacuum sector will always form
a (complex) Lie algebra g, and the inner product , 0 will define an invariant symmetric
bilinear form for g. When the RCFT is unitary, the positive-definiteness of that invariant
form implies that g will be reductive [26,27]—i.e., a direct sum of an abelian Lie algebra
Cm with a number of simple Lie algebras. Otherwise, when the theory is nonunitary, that
Lie algebra belongs to a much wider class called self-dual. Any such Lie algebra can be
constructed from a reductive Lie algebra by a sequence of ‘double extensions’ [28] (see
also [29]). Self-dual Lie algebras have appeared in the physics literature: for instance, the
(2 + 1)-dimensional Poincaré group ISO(2, 1) used by Witten [30] to relate Chern–Simons
2 I thank Matthias Gaberdiel for helpful discussions on this point.

with (2 + 1)-dimensional gravity has a self-dual Lie algebra; in [31–33] and references
therein, families of self-dual Lie algebras have been explicitly related to CFT. Self-dual
Lie algebras are precisely those for which the Sugawara construction works.
We are more interested here in the differences at the level of modular data, and at the
level of the toroidal M and cylindrical na partition functions. From this perspective, the key
observation is that, in a unitary theory, the vacuum 0 is the unique primary with minimal
conformal weight h. Hence, for τ = /i (/ ≈ 0)

0 < χa (τ ) = Sab χb (−1/τ ) ≈ Sa0 χ0 (−1/τ ) (7)
b
and thus Sa0 = S0a > 0.
Definition 2. By unitary modular data, we mean modular data obeying the additional
requirement that Sa0 > 0 for all a ∈ Φ.
A word of warning should be made: the modular data of any unitary RCFT will be
unitary modular data. However, the converse may not hold: given some unitary modular
data, there may or may not be a unitary RCFT which realises it—e.g., Fact 1 or constraint
(3) may be violated.
In any modular data, Sa0 is a nonzero real number (reality follows because Sa0 ∗ =
Sa,C0 = Sa0 ; it must be nonzero in order for (2) to be defined). What replaces the inequality
Sa0 > 0 in a nonunitary theory?
First note that there must be one and only one chiral primary, call it o, with
Sao = Soa > 0 ∀a. (8)

To see this, number all the primaries a1 = 0, a2, . . . , an , and take m ∈ R to be
large enough
that the sum N := ni=1 mi Nai of fusion matrices has distinct eigenvalues i mi Sai b /S0b
(in fact, all these eigenvalues will be distinct, for all but finitely many values of m). Then
for some phase ϕ and some primary o ∈ Φ, the column ϕSo will be the unique (strictly
positive) Perron–Frobenius eigenvector [34] of the strictly positive matrix N . But Fact 1
implies that the phase ϕ equals +1 (if we do not use Fact 1, i.e., for arbitrary modular data
not necessarily corresponding to an RCFT, we only get that ϕ = ±1). By unitarity of S,
there can be only one chiral primary obeying (8).
Definition 3. By the minimal primary of an RCFT, we mean the unique chiral primary
o ∈ Φ obeying (8) for all primaries a ∈ Φ.
Fact 2. In any RCFT, no primary can have smaller conformal weight than that of the
minimal primary o.
To see this, let 0 , 0 , . . . be the primaries with minimal conformal weight. The argument
of (7) shows that for small / > 0, the leading terms of χa (/i) will be
Sa0 χ0 (i//) + Sa0 χ0 (i//) + · · · ,

that is to say
0 Sa0 A0 ,0 + Sa0 A0 ,0 + · · ·

for all a ∈ Φ (recall that Ab,n are the coefficients of χb ). By unitarity of S, and inequality
(8), this forces o to be one of the primaries 0 , 0 , . . . .
Every RCFT has a unique minimal primary. In a unitary RCFT, the minimal primary
will be the vacuum 0. In all cases known to this author, there is a unique primary with
minimal conformal weight (it does not seem to be known that this should always be the
case). Recall that the conformal weight is (essentially) the energy, so it is tempting to
identify the minimal primary with the true vacuum. On the other hand, the true vacuum
should be invariant under the Poincaré group and in particular translations, yet typically
the descendent L−1 o will not vanish (see the explicit example in Section 6). We will find
that in a nonunitary theory, all of the familiar properties usually ascribed to the vacuum are
now distributed between 0 and o.
The quantum-dimension is the positive number Sao /S0o . It is the maximal (Perron–
Frobenius [34]) eigenvalue of the fusion matrix Na , and so obeys
Sao S0o , (9a)

|Sao S0b | |S0o Sab |, (9b)

Sao Sbo Sco
c
Nab min , , . (9c)
S0o S0o S0o
See, e.g., [12] for proofs. When there is a unique primary with minimal conformal weight,
then that primary, of course, must be o, and the quantum-dimension Sao /S0o equals the
limit limq→1 χa (q)/χ0(q).
A simple-current j is by definition any primary with quantum-dimension 1: i.e.,
Sj,o = S0o . Any simple-current corresponds to a permutation J and a phase (‘monodromy
charge’) a → QJ (a) ∈ Q such that Nj,ab =δ
b,J a , J 0 = j , and [35]
SJ a,b = e2πiQJ (b) Sab . (10a)

We will usually identify the simple-current j with the permutation J . From (10a) we
get QJ J (a) ≡ (QJ (a) + QJ (a)) (mod 1). Hence if J n = id, then QJ (a) ∈ n1 Z. Also,
QJ (σ, a) ≡ ,QJ (a) (mod 1) and σ, (J a) = J , σ, a, for all Galois automorphisms σ, .
Although (8) requires QJ (o) ∈ Z for all simple-currents J , one distinction with the
more familiar unitary theories is that the monodromy charge QJ (0) of the vacuum can
be a half-integer. This introduces the following modifications in the formulas applying
to the unitary theories: Nab c = 0 implies Q (a) + Q (b) ≡ (Q (c) + Q (0)) (mod 1),
J J J J
so QJ (J a) ≡ (QJ (J ) + QJ (a) + QJ (0)) (mod 1). Expanding SJ,J in two ways gives
QJ (J ) + QJ (0) ≡ (QJ (J ) + QJ (0)) (mod 1). Finally (see [12] for the arguments),

hJ a − ha ≡ hJ + QJ (0) − QJ (a) (mod 1), (10b)

2hJ ≡ QJ (0) − QJ (J ) (mod 1). (10c)
Note that for nonunitary theories, a simple-current is defined by Sj o = S0o , rather than the
simpler but potentially weaker Sj 0 = ±S00 (but see Consequence 2(vi) next section).
Both the vacuum 0 and the minimal primary o obey C0 = 0 and Co = o, because
their S-columns are real. Also, neither 0 nor o can be fixed by any simple-current J = id
(because otherwise Sa0 = 0 or Sao = 0 for any a with QJ (a) ∈ / Z).
There are arguments valid for unitary RCFTs which break down for nonunitary ones,
because the vacuum 0 and the minimal primary o are distinct. For instance, consider the
proof that for a given choice of unitary modular data, there are only a finite number of
possible modular invariants M. In particular, we get the bounds [36]

Sa0 Sb0 Sa0 Sb0
Mab = ∗
Sac Mcd Sbd S2 S0c Mcd S0d = ,
c,d∈Φ 00 c,d∈Φ
S2 00
using the triangle inequality, (9b), and the uniquess of the vacuum M00 = 1. For nonunitary
modular data, this argument breaks down, as it yields the bound Mab Moo Sao Sbo /S0o 2 .
All other finiteness proofs for modular invariants known to this author, similarly break
down for nonunitary modular data, because we do not have a bound on the multiplicity
Moo . Note, however, that this argument does give us the interesting fact that in any RCFT
Moo 1.
For example, consider the nonunitary modular data
 
1 1 2 2 2 2 −3 −3
 1 1 2 2 2 2 3 3 
 
 2 2 4 −2 −2 −2 0 0 

1  2 2 −2 4 −2 −2 0 
0 
S=  , (11a)
6  2 2 −2 −2 −2 4 0 0 
 
 2 2 −2 −2 4 −2 0 0 
 
−3 3 0 0 0 0 3 −3
−3 3 0 0 0 0 −3 3

T = diag 1, 1, 1, 1, ξ3, ξ3∗ , −1, 1 .
Number the primaries from 0 to 7, and let the 0th one be the vacuum. Then the minimal
primary o is ‘1’. This data has the same fusions as the quantum-double of the finite group
S3 (which is its ‘uniformisation’); although S obeys Fact 1, we know of no RCFT which
realises it. It has infinitely many modular invariants: for example, choose any integer
m 0, then
7

Z= |χa | + m|χ1 + χ3 + χ7 |2
2
(11b)
a=0
defines a modular invariant. Here, Moo = m + 1 can be arbitrarily large.
The same thing was noticed before, in a related context: there are infinitely many
modular invariants for the admissible representations of affine algebra su(2) at fractional
(nonintegral) level k [37]. Why this did not happen for the nonunitary Virasoro minimal
models, is explained by Fact 3 below. It is somewhat surprising that, although there are
infinite numbers of su(2) modular invariants at each fractional level, there are only a finite
number of modular invariants for their cosets (namely, the nonunitary minimal models)
— after all we tend to think that a source of coset modular invariants are products of the
modular invariants of the component affine algebras. Of course, the explanation, ultimately,
is field identification.
Another useful fact about modular invariants in unitary modular data, is that for any
simple-currents j, j ∈ Φ, Mj,j = 0 or 1, and Mj,j = 1 iff the following selection rule
holds for all a, b ∈ Φ:
Mab = 0 ⇒ QJ (a) ≡ QJ (b) (mod 1). (12)
The proof of this also collapses for nonunitary RCFTs, for the same reason; the argument
instead tells us that MJ o,J o Moo with equality only if the above selection rule holds.
Similar remarks are valid for the cylindrical partition functions (i.e., the NIM-reps).
Any NIM-rep is uniquely decomposable into a direct sum of irreducible NIM-reps. For
any NIM-rep na of unitary modular data, the multiplicity multn (0) of the vacuum in
(4) precisely equals the number of irreducible NIM-reps which compose na . Each of
these irreducible summands would correspond to a family of boundary conditions which
completely decouples from the other families. The uniqueness of the vacuum then tells us
that the NIM-rep of any unitary RCFT must be indecomposable [24]. The finiteness result
for NIM-reps of unitary modular data follows from this (see [24] for details). For nonunitary
modular data, the number of irreducible NIM-reps composing a given NIM-rep equals the
multiplicity multn (o) of the minimal primary o. A priori, a NIM-rep corresponding to a
nonunitary RCFT may be decomposable; hence, the finiteness result [24] for NIM-reps
breaks down unless we know that the multiplicity of o is bounded.
4. The Galois Shuffle and consequences
For the reasons given last section, it is important to answer the following:
Question 1. What is the multiplicity multn (o) = Moo of the minimal primary?
The Virasoro minimal models were classified in [7]; checking the answer there, we get
multiplicity Moo = 1 for all of them. We can explain and generalise this result:
Fact 3. In any WN (p, p ) minimal model, for any positive integers N, p, p , modular
invariance of the 1-loop partition function forces the multiplicity multn (o) = Moo of the
minimal primary to be 1.
We prove this in Section 6. Fact 3 is new.
Conjecture 1. The multiplicity multn (o) = Moo of the minimal primary in any RCFT is 1.
This makes some sense, as there are reasons for thinking of the minimal primary as the
true vacuum.3 It can also (see Consequence 2(ii)) follow from the Galois shuffle property
given shortly. Collecting results from last section, we get
Consequence 1. Suppose the multiplicity multn (o) = Moo of the minimal primary in a
given RCFT is 1. Then:
(i) There will be only finitely many modular invariants M with the given (chiral) modular
data—in fact, we get the inequality Mab Sao Sbo /S0o
2 ;
(ii) Any NIM-rep corresponding to a modular invariant M will be indecomposable. There

are only finitely many indecomposable NIM-reps.
In a superficially independent direction, further consequences occur when we assume

more.
Definition 4. We say an RCFT has the GS (or Galois shuffle) property, if there is a simple-
current Jo (possibly the identity), and a Galois automorphism σo (possibly the identity),
such that the minimal primary o in the theory is related to the vacuum by o = Jo σo 0.
Of course, this is trivially obeyed by any unitary RCFT. We verify in Section 6 that:
Fact 4. All WN minimal models possess the GS property.
The same is true for the minimal WSO(2N) models, but we will not prove it here.
The modular data for the minimal Wḡ models, for ḡ nonsimply laced, is a little more
complicated and we have not looked at them yet.
Modular data may not obey this property. For example, the 1-dimensional modular data
S = (−1), T = (−1) does not obey it. On the other hand, that 1-dimensional modular data
violates Fact 1 and so can not be realised by an RCFT. A more serious example is the
2,1 with the Yang–Lee model. That this does not have the GS property
tensor product of G
can be seen from Consequence 2(iv) below: this model has central charge c = −8/5,
and four primaries, with conformal weights 0, ±1/5, 2/5, so ho = −1/5, hJ = 0 and the
congruence in Consequence 2(iv) reduces to ,2 ≡ −2 (mod 15), which has no solutions.
More generally, the tensor product of a nonunitary model with its unitarisation (defined
below) typically will not satisfy the GS property.
Even though the GS property does not hold for all RCFTs, when it holds for a
given RCFT (which is often, as we will see) it has several consequences. We begin
by listing some of its immediate consequences, without assuming the Conjecture 1.
The most important of these are Consequences 2(ii), (iv), (v) and (viii)—in particular,
(iv) reduces to a statement about quadratic residues and becomes a stronger constraint
3 Nevertheless, as pointed out to the author by Philippe Ruelle, curious things can happen in even the nicest
unitary RCFTs. For example, the Ising model is unitary, but it has nonunitary observables, according to [38]. That
paper also suggests that the conformal weight 0 primary there need not be the identity field. In [39] we see that
the 3-state Potts model on the cylinder can have a degenerate vacuum, for some choices of boundary conditions.
the more distinct primes divide the order of T . Consequences 2(i), (iii), (iv) can be
interpreted as constraining the possible simple-currents Jo and Galois automorphism σo .
Consequences 2(vi), (vii) are of technical interest.
Consequence 2. Let 0 be the (chiral) vacuum and o the minimal primary. Suppose
o = Jo σo 0 for some Galois automorphism σo and some simple-current Jo . Then:
(i) The simple-current Jo has order 1 or 2;

(ii) If Jo = id, then the multiplicity multn (o) = Moo of the minimal primary must be 1
(i.e., the Conjecture must hold for this RCFT).
(iii) For any simple-current J , QJ (Jo ) ∈ Z and QJo (J ) ≡ (QJo (0) + QJ (0)) (mod 1).
(iv) The conformal weight ho of the minimal primary satisfies
c
ho ≡ (1 − ,2 ) + hJ (mod 1)
24
for some integer , coprime to the denominator of c/24, and for some simple-current
J of order 1 or 2 (so 4hJ ∈ Z).
(v) The fusion rules are exactly reproduced by modular data , which is unitary in the
S, T
sense of Definition 2 (this unitarisation is explicitly given in Eq. (13) below).
(vi) A primary j ∈ Φ will be a simple-current, iff Sj 0 = ±S00 .
(vii) |Sa0 | So0 .
(viii) The list of modular invariants M of S, T with Moo = 1 is identical to that of
the modular invariants M of the unitarisation which obey M
S, T Jo ,Jo = 1; the
(indecomposable) NIM-reps of S, T can be identified with those of the unitarisation

S, T.
First note that Eq. (1c), Cσ = σ C, and C0 = 0, tell us that each column S,σ 0 is
real, for any Galois σ . Hence Consequence 2(i) follows from Eq. (10a), because both
columns S,σo 0 = S,J −1 o and So are real. Consequence 2(ii) follows from Eq. (5c).
o
To see Consequence 2(iii), note that the inequality SJ,o > 0 requires QJ (Jo σo 0) ∈ Z.
In Consequence 2(iv), take J = Jo and σ, = σo , and use Eqs. (5b) and (10b). Since
Sj 0 = ±S00 iff Sj,Jo = ±S0,Jo iff σo (Sj,Jo ) = ±σo (S0,Jo ), we get Consequence 2(vi).
Applying σo ◦ σo−1 = id. to Sa0 and using Eq. (9a), gives Consequence 2(vii). Since the
fusions of S and S are identical (as was discussed after Eqs. (6)), so are their NIM-reps.
The statement in Consequence 2(viii) about modular invariants follows from Eq. (5c): an
integral matrix M commutes with S and T iff the integral matrix M := P MP t commutes
with S and T.
The unitarisation
S, T mentioned in Consequence 2(v) is given explicitly as follows.
Recall Eqs. (6a), (6b). Put J = Jo and σ = σo . Define the following matrices:

Sab := /σ (0)σ (SJ a,J b ) = /σ (0)(−1)QJ (a)+QJ (b)+QJ (0) σ (Sab ), (13a)
Tab := δab /σ (0)(TJ a,J b), , (13b)
where , ∈ Z corresponds to σ , i.e., σ acts on roots of unity by σ (ξ ) = ξ , . Requiring the
vacuum 0 in the unitarised modular data to have conformal weight 0, we find that the
unitarisation
S, T has central charge
ĉ ≡ (,c − 24,hJo + 12/) (mod 24),

where (−1)/ := /σ (0).
As explained after (6b), the unitarisation obeys all of the conditions stated in
Definitions 1 and 2, and an easy Galois argument shows that it automatically obeys the
additional conditions (3) and identities given in [14] (provided S, T do). It is conceivable,
however, that S, T may not always be realisable by an RCFT—for instance, S could
conceivably violate Fact 1. Nevertheless, in the few cases this author has been able to
check, the uniformisation always seems to be realised by a completely healthy unitary
RCFT. We return to this in the concluding section.
Although the minimal primary o is uniquely determined from the modular data, the
simple-current Jo and Galois automorphism σo may not be unique. There can be more
than one unitarisation of nonunitary modular data. In particular, let G0 be the set of all
Galois automorphisms σ such that σ (S00 ) = ±S00 . Then σo will be unique only up to G0 :
σo can be replaced by any other σ in its coset G0 σo ; but once the Galois automorphism σo
has been chosen, the simple-current Jo will be uniquely determined. This duplication G0
has in fact been previously studied in [40], under the name ‘Galois currents’.
Further results hold if we assume both the Conjecture and the GS property.
Consequence 3. Let 0 be the (chiral) vacuum and o the minimal primary. Suppose
o = Jo σo 0 for some Galois automorphism σo and some simple-current Jo . Assume that
the multiplicity multn (o) = Moo in the full RCFT is 1. Then:
(i) The list of possible toroidal and cylindrical partition functions are identical for S, T as
for its unitarisation
S, T.
(ii) The multiplicity multn (Jo ) = MJo ,Jo = 1.
(iii) We have the symmetry MJo a,Jo b = Mab .
(iv) We have the selection rule Mab = 0 ⇒ QJo (a) ≡ QJo (b) (mod 1).
(v) If Jo = id, the NIM-rep n is 2-colourable.
First let us define 2-colourability of the NIM-rep. Let B be the set labelling the rows and
columns of the nim-rep (i.e., the vertices of the fusion graphs, equivalently the labels of the
boundary states). We can assign each x ∈ B to a number qx = 0, 12 such that the selection
rule
y
nax = 0 ⇒ QJo (a) + qx − qy ∈ Z
holds for all a ∈ Φ, x, y ∈ B. For example, the NIM-reps of the Virasoro minimal models, at
least those compatible with the modular invariants, can be found in [41]. It can be verified
explicitly that all of them are 2-colourable. The NIM-reps built out of tadpoles are not
2-colourable, but they also are not compatible with any modular invariants.
Consequence 3 is now easy to prove. Consequence 3(i) is just 2(viii). The Galois
symmetry (5c) with σ = σo tells us that 1 = M00 = MJo o,Jo o . Because MJo o,Jo o = Moo ,
we get 3(ii), and the rest automatically follows.
Consequence 3(i) reduces the classification of nonunitary models to that of the more
familiar unitary ones. In Section 6 we give the unitarisation of the WN minimal models,
and we find for instance that the affine algebra su(3) gives the uniformisation of certain W3
models. We thus obtain for free the classification of those W3 models, in the bulk. With a
little additional work, the classification of all nonunitary W3 minimal models (in the bulk)
can be obtained [42].
Consequence 4. The complete list of W3 minimal models at (p, 4) is in exact one-to-one

correspondence with the list of su(3) level p − 3 modular invariants. In particular, there
are precisely four modular invariants for each value of p, except for p = 3 (where there
is only one) and p = 5 (where there are only two). Those four modular invariants are all
constructed using simple-currents and/or charge-conjugation in the standard way. Each of
those modular invariants has a compatible NIM-rep.
Of course, p there must be odd (since it must be coprime to 4). It is important to note
that Consequence 4 is a theorem independent of the general validity of the conjecture,
because Facts 3 and 4 tell us that the minimal W3 models obey all the hypotheses of
Consequence 3. The proof of Consequence 4, and the explicit correspondence between
the W3 modular invariants and those of su(3), will be given at the end of Section 6.
5. Comments on affine algebras at fractional level
Let g = Xr(1) be any affine algebra. For the relevant facts and notation about integrable
representations of g, and corresponding SL2 (Z) representations, see, e.g., [4].
The so-called admissible representations of g at fractional level k = t/u [6] share many
properties with the better known integrable representations. Their importance for RCFTs,
and this paper, lies in the quantum Drinfeld–Sokolov reduction method for constructing
W-algebras (see [8] for a review) and their parallel use in GKO cosets for nonunitary
models. But a natural question, which has received much attention (see, e.g., [43] and
references therein), is: is there an RCFT which corresponds more directly to the admissible
representations, roughly in the way that the integrable data is directly realised by the Wess–
Zumino–Witten models [3]? As is well known, naively placing the admissible S matrix
S adm into Verlinde’s formula fails to produce nonnegative integer fusions (indeed, a general
though conjectural expression for these ‘fusions’, manifestly demonstrating that they can
be integers of either sign, is given in [44]). This means that the desired correspondence is
more subtle, if it exists at all. Indeed, for su(2) the generally accepted wisdom of [45,46]
has now been challenged by [47] and independently by [48]. For example, [48] suggests
that a CFT corresponding to admissible fusions will be both nonunitary and quasi-rational.
In this section we make a couple of preliminary remarks, relating Galois to the
admissible S matrix S adm . This underlies our calculations in Section 6, and (we believe)
lends support to our expectation that the GS property will be quite common among RCFT.
Let g be the dual Coxeter number of g = Xr . Write the level k = t/u, where gcd(t, u) =
1, t ∈ Z, u ∈ {1, 2, 3, . . .}. Put k I = u(k + g) − g and k F = u − 1, for reasons to be clear
soon. The integrable case is recovered when the denominator u = 1. Admissibility requires
k ( u1 − 1)g, and we will assume for most of this section an additional coprimeness
condition: for g =su(N ), the denominator u must be coprime to N ; for g =so(N), sp(N),
F4 , or E7 , u must be odd; and for g = E6 or G2 , u must not be a multiple of 3. This
coprimeness condition is necessary for the direct Galois interpretation to be given shortly;
when it fails the Galois interpretation is slightly more subtle (see the discussion on [49]
when u is even, given at the end of this section).
I
The admissible highest weights Padmk consist of pairs λ = (λI , λF ), where λI ∈ P+k is
an integrable highest weight of g of level k I ∈ Z+ , and where λF belongs to a certain finite
F k = P kI × P kF .
set of level k F ∈ Z+ weights of g. For instance, for su(2), λF ∈ P+k , so Padm + +
F F
For su(3), λF belongs to this disjoint union of P+k with the subset of P+k with first Dynkin
+ I
label λF1 1. In general, the cardinality of Padm will be !P+k !ur where r is the rank of the
Lie algebra ḡ. The modular S matrix S adm is given by [6]

adm
Sλµ = ±Fm exp 2πi λI + ρ µF + λF µI + ρ − (k + g) λF µF

−2πi I I

(k)
× det(w) exp w λ + ρ µ + ρ = ϕλµ SλI µI , (14)
k+g
w∈W
where Fm is some constant (independent of λ, µ), and the sign ±1 depends on λ, µ (it is
given explicitly in, e.g., Eq. (2.46) of [50]). W here is the (finite) Weyl group of ĝ. The
modulus |ϕλµ | = u−r/2 is constant, and the matrix S (k) is the sum over the Weyl group W ,
normalised by ur/2 Fm .
S adm is unitary and symmetric, but S does not have a column of constant phase, and
therefore Verlinde’s formula (2) will yield negative ‘fusions’ Nλµ ν . So the admissible
modular matrices S adm , T adm do not constitute modular data (although they define a
representation of SL2 (Z) as before). In modular data, the charge-conjugation matrix
C = S 2 is a permutation matrix; for the admissible data, C = S 2 is a signed permutation
matrix. Typically, there will not be a unique admissible weight λ ∈ Padm k
with minimal
conformal weight.
Returning to formula (14), our observation in this section is simply that, up to a sign
(k I )
independent of λ, µ, Sλ(k)I µI = ±σ (SλI µI ), where σ is a Galois automorphism depending
I
on u but not λ, µ, and where S (k ) is the integrable ‘WZW’ S matrix for g at integer
level k I .
What this means is that the λI part of admissible representations is understandable. If
for instance we adjusted the phases ϕ → ϕ in (14) appropriately, the admissible S matrix
S adm would become a factorised matrix S fac = ϕ ⊗ S (k) which, when placed in Verlinde’s
I
formula (2), would yield the factorised nonnegative integer fusions Nλ F µF ν Nλ(kI µ)I ν
F I
I
where N (k ) are the usual integrable WZW fusions for g at level k I , and where the
fractional part N has simple-current fusions.
For example, for su(2) at level k = t/u, write λ = (λI , λF ) =: (l, l ), µ =: (m, m ),
where λI ∈ P+t +2u−2 (su(2)) = {0, 1, . . ., t + 2u − 2} and λF ∈ {0, 1, . . . , 2u − 1}. Then

2 l (m+1)+m (l+1) −πit l m /u π(l + 1)(m + 1)
Sλµ =
adm
(−1) e sin .
u2 (k + 2) k+2

If we were to drop the (−1)l (m+1)+m (l+1) factor, then the resulting factorised S matrix S fac
(t +2u−2)n
would yield the factorised fusions Nl ,m n Nlm where Nl ,m n = δn ,(l +m ) (mod 2u)
and N (t +2u−2) are su(2)t +2u−2 fusions. The Galois automorphism σ involved here
corresponds to any , coprime to u satisfying , ≡ u (mod 8(t + 2u)).
A more subtle idea along these lines was suggested in [49], who in effect changed the

sign (−1)l (m+1)+m (l+1) to (−1), m+m , (for u odd) and to (−1)(, −1)(m+1)+(m −1)(,+1)
(for u even). The resulting fusions are again in the factorised form: su(2)t +2u−2 fusions
times Z2u fusions. Although these fusions differ from those in, e.g., [45,46], they are
consistent with the in-depth analysis of [47] for su(2) at level k = −4/3. Incidentally,
this modular data of [49] is nonunitary; its unitarisation corresponds to the Galois
automorphism with , odd and congruent to u (mod t + 2u).
What value if any the observation in this section has, is unclear at this point. However,
it is implicit in the following section.
6. Examples
The classic examples of nonunitary RCFTs come from the Virasoro minimal models.
Let p > p 2 be coprime integers. The primary fields ϕrs here are parametrised by all
pairs (r, s) where 1 r p − 1, 1 s p − 1, and p s < pr. So the number of primaries
is (p − 1)(p − 1)/2. The conformal weight hrs of ϕrs is
(pr − p s)2 − (p − p )2
hrs = . (15a)
4pp
The S matrix has entries

2 +rs p p
Srs;r s = 2
(−1) 1+sr
sin π rr sin π ss . (15b)
pp p p
The central charge of the (p, p ) minimal model is c = 1 − 6(p − p )2 /pp . It is unitary
iff p − p = 1. The vacuum 0 corresponds to primary (r, s) = (1, 1). The simple-current
is j = (p − 1, 1), corresponding to permutation J (r, s) = (p − r, s) (if pr + p s < pp )
or (r, p − s) (otherwise); the monodromy charge in (10a) is Q(r, s) = (1 + pr + p s)/2.
There is a unique primary o = (ro , so ) with minimal conformal weight, corresponding to
the unique (r, s) ∈ Φ obeying pr − p s = 1.
The Galois group here can be taken to be Z× 8pp . When ro and so are both odd, choose

any , obeying , ≡ ro (mod 2p ) and , ≡ so (mod 2p), and take Jo = (1, 1). When instead ro
is even, so both so and p will be odd, choose , ≡ (p − ro ) (mod 2p ) and , ≡ so (mod 2p),
and Jo = (p − 1, 1). Finally, if so is even, then both ro and p will be odd, and choose
, ≡ ro (mod 2p ) and , ≡ (p − so ) (mod 2p), and again Jo = (p − 1, 1). In all cases, by
the Chinese remainder theorem, there’s a unique solution in the range 1 , 2pp ; it
will lie in Z×
8pp (i.e., be coprime to 2, p and p ), so let σo be the corresponding Galois
automorphism. Then o = Jo σo (1, 1), showing that the GS property holds here.
For a concrete example, consider the (7,3) minimal model (c = −25/7), with primaries
{(1, 1), (1, 2), (2, 1), (2, 2), (2, 3), (2, 4)}
taken in that order. Then

 
−d a d −a −b b
 a b a b d d 
2 d a d a −b −b 

Srs;r s =  ,
7  −a b a −b d −d 
 
−b d −b d −a −a
b d −b −d −a a

25 −5 235
Trs;r s = diag exp πi , exp πi , exp πi ,
84 84 84

121 43 1
exp πi , exp πi , exp πi ,
84 84 84
where a = sin(π/7) ≈ 0.434, b = sin(2π/7) ≈ 0.782, and d = sin(3π/7) ≈ 0.975. The
positive column (namely, the second) corresponds to the minimal primary o = (1, 2) (note
that 7 · 1 − 3 · 2 = 1). The Galois automorphism σo corresponds to , = 19, and the simple-
current Jo = (2, 1) is also nontrivial. The unitarisation is
 
a b a b d d
 b d −b −d −a a 
 
2  a −b −a b −d d 
S=  ,
7  b −d b −d a a 
 
d −a −d a b −b
d a d a −b −b

97 −53 −29
T = diag exp πi , exp πi , exp πi ,
84 84 84

73 19 −23
exp πi , exp πi , exp πi .
84 84 84

Incidentally, this unitarisation is the modular data for affine algebra su(2) ⊕ E8 at level
(5,1); more generally, the (p, 3) Virasoro minimal model has unitarisation corresponding
8,1 ’s, whose only
to su(2) at level p − 2, whenever p is odd (ignoring a certain number of E
purpose is to adjust c by an appropriate multiple of 8).
The characters χrs of the minimal models have been explicitly calculated [51]: for
example, for the (7,3) model the characters of the vacuum and minimal primaries are

χ0 = χ11 = q 25/168 1 + q 2 + q 3 + 2q 4 + · · · ,

χo = χ12 = q −5/168 1 + q + q 2 + 2q 3 + · · · .
In particular we see that, although the translation operator L−1 kills the vacuum 0, it does
not kill the minimal primary o. This illustrates one of the ways the minimal primary does
not behave like a vacuum.
More generally, we can consider the minimal WN models (cf. [44,50,52,53]). These
theories are also parametrised by a pair p > p N of coprime integers. As before,
they are unitary iff p = p + 1. The primaries consist of all (J, J )-orbits (J i λ, J i µ),
p p
where λ ∈ P++ , µ ∈ P++ —by ‘P++ m ’ we mean all N -tuples ν of positive integers,
i

with N−1 i=0 νi = m. Algebraically, P++ are the level m − N highest weights of s
m
u(N),
shifted by the Weyl vector ρ = (1, 1, . . . , 1). The simple-current J of su(N) takes weight
ν = (ν0 , ν1 , . . . , νN−1 ) ∈ P++
m to (ν
N−1 , ν0 , ν1 , . . . , νN−2 ) ∈ P++ . So there are a total of
m
p p
( N−1 )( N−1 )/N primaries. The matrix T is given by the conformal weights
|pλ − p µ|2 − (p − p )2 (N − 1)N(N + 1)/12

hλµ ≡ (mod 1) (16a)
2pp
and central charge

(p − p )2
c = (N − 1) 1 − N(N + 1) .
pp
The S matrix is
(p/p ) (p /p)
S(λ,µ)(λ ,µ ) = α exp −2πi[t (λ)t (µ ) + t (µ)t (λ )]/N Sλ,λ Sµ,µ , (16b)

where α is some irrelevant constant, where t (λ) = i iλi , and where S (m) here is the
usual (‘integrable’) affine S matrix, expressed as usual as an alternating sum over the
(finite) Weyl group (see, e.g., Eq. (14.57) of [4]), formally evaluated at (fractional) level
m − N . The vacuum 0 corresponds to the (J, J )-orbit containing (ρ, ρ), or more precisely
((p − N + 1, 1, . . . , 1), (p − N + 1, 1, . . . , 1)) (recall that (k + 1, 1, . . . , 1) ∈ P++
k+N
corresponds to the vacuum in the su(N)k WZW theory, and projects to the Weyl vector
ρ of su(N)). An order-N simple-current of this minimal model is (J, id) (or rather its
(J, J )-orbit); it has monodromy charge
pt (λ) − p t (µ) N − 1
Q(λ, µ) ≡ + (mod 1).
N 2
As we will see below, it generates all of the simple-currents of the model.
The Galois group here can be taken to be the multiplicative group Z× 4Npp . Choose
any odd integer ,, coprime to N , obeying the congruences ,p ≡ 1 (mod p ) and ,p ≡
1 (mod p) (this is always possible, by the Chinese remainder theorem and the fact that
gcd(p, p ) = 1). Let σ, be the Galois automorphism corresponding to ,, and write
p
(o , o ) ∈ P++ × P++ for the image σ, 0 of the vacuum under σ, . We claim that (J i o , o )
p
is the minimal primary o, for some i.

To see this, first note that for any primaries (λ, µ), (λ , µ ), we have
(p/p ) (p /p) (p ) (p )
S(λ,µ)(λ ,µ ) Sλλ S S
(p) S S (p) S
= µµ = σ −1 λλ σ −1 µµ = λλ µµ ,
S (p/p ) (p /p) ,
S S S S
(p) , (p ) (p) (p )
0(λ ,µ )
ρλ S ρµ S ρλ ρµ ρλ ρµ
where σ,−1 (λ , µ ) = (λ , µ ). Implicit in this calculation is that we can write Sλλ =
(p)
(p )
ψλλ Sλλ (similarly for Sµµ ), where ψλλ = exp[2πit (λ + ρ)t (λ + ρ)/pN] is a root of
is in the number field Q[exp[2πi/p]]. This means that |σ S (p)
unity, and Sλλ m λλ | = |σm Sλλ |
depends only on the value of m (mod p). The maximum value of the right-side is the
product of WZW quantum dimensions, achieved by the vacuum (λ , µ ) = 0 (recall (9b)).
In particular, the choice (λ , µ ) = (o , o ) achieves this maximum, for all primaries (λ, µ),
as must the minimal primary (λ , µ ) = o (again by (9b)). Thus, for every primary (λ, µ)
there is a phase ϕλµ such that S(λ,µ)(o ,o ) /S0(o ,o ) = ϕλµ S(λ,µ),o /S0,o . Taking the absolute
value of
S(λ,µ)(o ,o ) S(λ ,µ )(o,o ) (λ ,µ ) S(λ ,µ )(o ,o )
= N(λ,µ)(λ ,µ )
S0(o ,o ) S0(o ,o )
S0(o ,o )
(λ ,µ )
and using the triangle inequality, we find that the assignment (λ, µ) → ϕλµ defines a
grading on the WN fusion ring at (p, p ). Hence (λ, µ) → ϕλµ S(λ,µ)0 /S00 defines a 1-
dimensional representation of the fusion ring, and so
S(λ,µ)0 S(λ,µ),A
ϕλµ = (17)
S00 S0,A
for some primary A. We want to show A is a simple-current, i.e., that SA,o = S0,o . By
unitarity of the matrix S, the norm of the 0-column must equal that of the A-th column,
and so (17) implies |S00 | = |S0,A |. Substituting o for (λ, µ) in (17) and using positivity (8)
now concludes the proof of Fact 4: A is a simple-current and Ao = (o , o ).
This argument also showed that the quantum-dimension S(λ,µ)o /S0,o equals the product
of the WZW su(N) quantum-dimensions of λ and µ at levels p − N and p − N ,
respectively. Since the only simple-currents of su(N)k are the ones J i corresponding to the
order-N cyclic symmetry of the su(N) Dynkin diagram [54], the only simple-currents of
the WN minimal models are the (J, J )-orbit of those primaries (J i , id). Hence A = (J i , id)
for some i.
In fact, we know the order of A = Jo must be 1 or 2 (see, e.g., Consequence 2(i),
although this can be easily seen directly). Thus when N is odd, Jo = id, and when N is
even, the only possibilities are Jo = id or Jo = (J N/2 , id).
Next, let’s give a proof of Fact 3. By Consequence 2(ii) it is now automatic for N
odd, since then Jo = id. So let N be even and Jo = (J N/2 , id). Let M be any modular
invariant of WN . From MT = T M we see that if M(λ,µ)(λ ,µ ) = 0, then (16a) requires
|pλ − p µ|2 ≡ |pλ − p µ |2 (mod 2pp ). But an easy calculation shows (for su(N)) that
N|λ|2 ≡ −t (λ)2 (mod N). Thus if M(λ,µ)(λ ,µ ) = 0, then the monodromies QJo = NQ/2
of (λ, µ) and (λ , µ ) are equal mod 1. Note then from M = SMS ∗ that

∗
MJo ,Jo = SJo ,(λ,µ) M(λ,µ)(λ ,µ ) S(λ ,λ ),J
o
(λ,µ),(λ ,µ )

∗
= S0,(λ,µ) M(λ,µ)(λ,µ ) S(λ ,λ ),0 = M00 = 1.
(λ,µ),(λ ,µ )
Thus by (6c) and the fact that o = σ, Jo , we get that Moo = 1, which is Fact 3.
We conclude this section, and this paper, with the proof of Consequence 4. It is helpful
to remember that the simple-current Jo for W3 must be trivial, so all we have to account for
here is the Galois automorphism. The minimal (p, 4) W3 models have unitarisation S equal
to the modular S matrix of the affine algebra su(3) at level p − 3. Explicitly, the primary
p−3
λ = (λ0 , λ1 , λ2 ) ∈ P+ (su(3)) corresponds to the W3 primary given by the (J, J )-orbit
[λ + ρ] := (J (2, 1, 1), J i (λ0 + 1, λ1 + 1, λ2 + 1)). The T matrix enters into the modular
i
invariant classification only via the selection rule
Mab = 0 iff ha ≡ hb (mod 1), ∀a, b ∈ Φ,

and in both our cases T -invariance reduces to (in the W3 notation) the selection rule
p
M[λ][µ] = 0 iff λ21 + λ1 λ2 + λ22 ≡ µ21 + µ1 µ2 + µ22 (mod 3p), ∀λ, µ ∈ P++ .
Thus, using that primary field correspondence, the modular invariants and NIM-reps of
these two different pairs of modular data can be identified. The modular invariants are
classified in [55]. Of course, the exceptional su(3) modular invariants all occur for odd
level and so do not arise here. Many NIM-reps for su(3) were first given in [56], and other
NIM-reps were obtained in other papers, culminating with the announcement in [57] that
there is a unique NIM-rep for each su(3) modular invariant. Ocneanu’s argument requires
the full structure of his von Neumann subfactor theory (see, e.g., [58] for a review) and it
is not yet known if all this has a necessary counterpart in RCFT. In any case a NIM-rep for
each su(3) modular invariant (and hence each W3 (p, 4) model) can be found in, e.g., [41].
More generally, the WN minimal models at (p, p ) = (p, N + 1) will for most p
correspond to the WZW su(N) model at level p − N .
7. Concluding remarks
This paper compares nonunitary and unitary RCFT. In a nonunitary theory, some of the
properties we would like to ascribe to the vacuum 0 instead belong to what we call the
minimal primary o, which corresponds to the positive column of the modular S matrix.
Question 2. Must there be exactly one primary with minimal conformal weight?
In all examples of (healthy) nonunitary RCFTs known to this author, the answer
to Question 1 is ‘yes’ and it is tempting to conjecture that it always must be. The
affine algebras at fractional level can have several different admissible highest weight
representations with minimal conformal weight (see, e.g., [50]), but they do not have a
direct interpretation as an RCFT and so do not constitute a counterexample. When there
is only one primary with minimal conformal weight, that primary must be the minimal
primary o (hence the name).
A fundamental question here is the multiplicity of the minimal primary o in the full
RCFT. We showed that this multiplicity must be 1 for any minimal WN model, and we
conjectured in Section 4 that it always equals 1.
The relation between 0 and o lies at the heart of this paper. We identify a property which
many (but not all) RCFTs obey, namely that the minimal primary o and the vacuum 0 in
the chiral theory are related to each other by what we call the Galois shuffle. Once again,
all minimal WN models obey this GS property.
Question 3. Is there a characterisation of the nonunitary RCFTs which obey the GS

property? How typical is it?
Our reasons for suspecting that it’s quite typical are that we know many RCFTs
which obey it, and also that many W-algebras can be constructed from the affine
algebras at fractional level, and the latter respects it (see Section 5). On the other hand
Consequence 2(iv) in Section 4 indicates that it is quite nontrivial.
We learned in Section 4 that the GS property has many nice consequences. Roughly
speaking, it means that the nonunitary theory behaves almost like a unitary theory. That
corresponding theory (or rather its modular data) is called the ‘unitarisation’ of the unitary
theory. The unitarisation obeys the same fusion rules as the nonunitary theory, and their
toroidal and cylindrical partition functions will be in one-to-one correspondence.
Question 4. Can we see the unitarisation directly inside the nonunitary RCFT? In
particular, it makes sense at the chiral level, so can we see it at the level of the chiral
algebras?
Question 5. Will the unitarisation always equal the modular data of a completely healthy
unitary RCFT?
As mentioned in Section 4, the unitarisation will always satisfy conditions (3) and the
identities in [14], and all the properties of unitary modular data, but nevertheless it still
might not be realised as the modular data of an actual theory. A simple critical case is
given by the Virasoro minimal models at (p, 2), where p 1 is odd: we find in Section 6
that its unitarisation can be expressed as

2 ab
Sab = √ sin π ,
p p
2
±a 3±1
Tab = δab exp −π + ,
2p 12
where the primaries a, b consist of all odd numbers 1 a, b p − 2, and where we take
the upper signs (i.e., ‘+’) in the formula for T if p ≡ +1 (mod 4), and otherwise take the
lower signs (i.e., ‘−’). The question is, can this modular data be realised by the characters
of an RCFT. It obeys all properties (e.g., Fact 1) this author has been able to check. For
p = 3, 5, 9, 11, this data coincides with that of the WZW models with affine algebras
E8 ⊕ E8 ⊕ E8 at level (1,1,1), F 4 at level 1, G2 at level 2, and F
4 at level 2. For the
other values of p, this author has been unable to find a healthy RCFT which realises it (but
that certainly doesn’t mean that none exists). Curiously, for arbitrary (odd) p, that matrix
S coincides with the matrix ψ diagonalising (in the sense of Eq. (4) above) the spurious
su(2) level p − 2 NIM-rep called the tadpole in [56]. In any case, if an RCFT realisation
can be found for this modular data for all odd p, this would lend support to the thought that
the uniformisation of a nonunitary RCFT is the modular S and T matrices for a healthy
unitary RCFT.
Another reason the existence of a unitarisation might be interesting is the following.
A deep relationship between von Neumann subfactors and RCFT has been developed by
Ocneanu, Evans, and others (see, e.g., [36,58] for reviews). Although subfactors cannot
recover the entire RCFT (e.g., they don’t see the chiral algebra or even its character), they
can realise for example the fusions, the modular S- and T -matrices, the 1-loop modular
invariant partition functions, and the NIM-reps of RCFTs. The subfactor picture (at present)
can only realise unitary RCFTs. It is tempting to speculate that perhaps subfactors realise
instead the unitarisation and corresponding modular invariants and NIM-reps, even if
S, T
the unitarisation were not to correspond to a completely healthy RCFT.
Question 6. Not all RCFTs have a unitarisation. However, given any nonunitary RCFT,
can we always find a unitary RCFT with identical fusion rules?
Again, this is true of all examples known to this author. Of course, it holds for any RCFT
obeying the GS property.
Acknowledgements
I thank Peter Bántay, David Evans, Matthias Gaberdiel, Christoph Schweigert, Mark
Walton and Jean-Bernard Zuber for helpful comments. This research was supported by
NSERC.
References
[1] C. Schweigert, J. Fuchs, J. Walcher, Conformal field theory, boundary conditions, and applications to string
theory, hep-th/0011109.
[2] W. Eholzer, M.R. Gaberdiel, Commun. Math. Phys. 186 (1997) 61.
[3] D. Gepner, E. Witten, Nucl. Phys. B 278 (1986) 493.
[4] J. Fuchs, C. Schweigert, Symmetries, Lie Algebras and Representations, Cambridge Univ. Press, Cambridge,
UK, 1997.
[5] P. Goddard, A. Kent, D. Olive, Commun. Math. Phys. 103 (1986) 105.
[6] V.G. Kac, M. Wakimoto, Proc. Natl. Acad. Sci. USA 85 (1988) 4956.
[7] A. Cappelli, C. Itzykson, J.-B. Zuber, Commun. Math. Phys. 13 (1987) 1.
[8] P. Bouwknegt, K. Schoutens, Phys. Rep. 223 (1993) 183.
[9] G. Moore, N. Seiberg, Commun. Math. Phys. 123 (1989) 177.
[10] E. Verlinde, Nucl. Phys. B 300 (1988) 360.
[11] J. Fuchs, Fortschr. Phys. 42 (1994) 1.
[12] T. Gannon, Modular data: the algebraic combinatorics of conformal field theory, math.QA/0103044.
[13] P. Bántay, Phys. Lett. B 394 (1997) 87.
[14] P. Bántay, On generalizations of Verlinde’s formula, hep-th/0007164.
[15] J. Cardy, Nucl. Phys. B 270 (1986) 186.
[16] J. Cardy, Nucl. Phys. B 324 (1989) 581.
[17] G. Pradisi, A. Sagnotti, Y.S. Stanev, Phys. Lett. B 381 (1996) 97.
[18] R.E. Behrend, P.A. Pearce, V.B. Petkova, J.-B. Zuber, Phys. Lett. B 444 (1998) 163.
[19] J. Fuchs, C. Schweigert, Nucl. Phys. B 530 (1998) 99.
[20] A. Recknagel, V. Schomerus, Nucl. Phys. B 531 (1998) 185.
[21] A. Coste, T. Gannon, Phys. Lett. B 323 (1994) 316.
[22] P. Bántay, Commun. Math. Phys. 233 (2003) 423.
[23] A. Coste, T. Gannon, Congruence subgroups and RCFT, math.QA/9909080.
[24] T. Gannon, Nucl. Phys. B 627 (2002) 506.
[25] D. Gepner, Foundations of rational quantum field theory, I, hep-th/9211100.
[26] A.B. Zamolodchikov, Theor. Math. Phys. 65 (1986) 1205.
[27] P. Goddard, in: V.G. Kac (Ed.), Infinite-Dimensional Lie Algebras and Groups, World Scientific, 1989,
p. 237.
[28] A. Medina, P. Revoy, Ann. Scient. Éc. Norm. Sup. 18 (1985) 553.
[29] J.M. Figueroa-O’Farrill, S. Stanciu, J. Math. Phys. 37 (1996) 4121.
[31] C.R. Nappi, E. Witten, Phys. Rev. Lett. 71 (1993) 3751.
[32] B.H. Lian, Commun. Math. Phys. 163 (1994) 307.
[33] A. Giveon, O. Pelc, E. Rabinovici, Nucl. Phys. B 462 (1996) 53.
[34] F.R. Gantmacher, The Theory of Matrices, Chelsea, New York, 1990.
[35] A.N. Schellekens, S. Yankielowicz, Phys. Lett. B 227 (1989) 387.
[36] J. Böckenhauer, D.E. Evans, Modular invariants from subfactors, math.OA/0006114.
[37] S. Lu, Phys. Lett. B 218 (1989) 46.
[38] E. Lapalme, Y. Saint-Aubin, J. Phys. A 34 (2001) 1825.
[39] P. Ruelle, J. Phys. A 32 (1999) 8831.
[40] P. Bántay, JHEP 0303 (2003) 025.
[41] R.E. Behrend, P.A. Pearce, V.B. Petkova, J.-B. Zuber, Nucl. Phys. B 579 (2000) 707.
[42] E. Beltaos, T. Gannon, in preparation.
[43] A.Ch. Ganchev, V.B. Petkova, G. Watts, Nucl. Phys. B 571 (2000) 457.
[44] P. Mathieu, M.A. Walton, Prog. Theor. Phys. Suppl. 102 (1990) 229.
[45] H. Awata, Y. Yamada, Mod. Phys. Lett. A 7 (1992) 1185.
[46] B. Feigin, F. Malikov, Lett. Math. Phys. 31 (1994) 315.
[47] M.R. Gaberdiel, Nucl. Phys. B 618 (2001) 407.
[48] F. Lesage, P. Mathieu, J. Rasmussen, H. Saleur, Nucl. Phys. B 647 (2002) 363.
hep-th/9301121.
[49] S. Ramgoolam, New modular Hopf algebras related to rational k sl(2),
[50] P. Mathieu, D. Sénéchal, M.A. Walton, Int. J. Mod. Phys. A 7 (Suppl. 1B) (1992) 731.
[51] A. Rocha-Caridi, in: J. Lepowsky, S. Mandelstam, I.M. Singer (Eds.), Vertex Operators in Mathematics and
Physics, Springer, 1985, p. 451.
[52] V.A. Fateev, S. Lykyanov, Int. J. Mod. Phys. A 3 (1988) 507.
[53] E. Frenkel, V. Kac, M. Wakimoto, Commun. Math. Phys. 147 (1992) 295.
[54] J. Fuchs, Commun. Math. Phys. 136 (1991) 345.
[55] T. Gannon, Commun. Math. Phys. 161 (1994) 233.
[56] P. Di Francesco, J.-B. Zuber, Nucl. Phys. B 338 (1990) 602.
[57] A. Ocneanu, talk, Kyoto, 2000.
[58] D.E. Evans, Y. Kawahigashi, Quantum Symmetries on Operator Algebras, Oxford Univ. Press, London,
1998.
Energy crisis or a new soliton in the

noncommutative CP(1) model?
Subir Ghosh
Physics and Applied Mathematics Unit, Indian Statistical Institute, 203 B.T. Road, Calcutta 700108, India
Abstract
The non-commutative (NC) CP(1) model is studied from field theory perspective. Our formalism
and definition of the NC CP(1) model differs crucially from the existing one [Phys. Lett. B 498
(2001) 277, hep-th/0203125, hep-th/0303090].
Due to the U (1) gauge invariance, the Seiberg–Witten map is used to convert the NC action to
an action in terms of ordinary spacetime degrees of freedom and the subsequent theory is studied.
The NC effects appear as (NC parameter) θ-dependent interaction terms. The expressions for static
energy, obtained from both the symmetric and canonical forms of the energy momentum tensor, are
identical, when only spatial noncommutativity is present. Bogomolny analysis reveals a lower bound
in the energy in an unambiguous way, suggesting the presence of a new soliton. However, the BPS
equations saturating the bound are not compatible to the full variational equation of motion. This
indicates that the definitions of the energy momentum tensor for this particular NC theory (the NC
theory is otherwise consistent and well defined), are inadequate, thus leading to the “energy crisis”.
A collective coordinate analysis corroborates the above observations. It also shows that the above
mentioned mismatch between the BPS equations and the variational equation of motion is small.
PACS: 11.15.-q; 11.15.Tk; 11.90.+t
Keywords: CP(1) model; Noncommutative field theory; Seiberg–Witten map
1. Introduction
Non-commutative (NC) field theories have turned into a hotbed of research activity after
its connection to low energy string physics was elucidated by Seiberg and Witten [1,2].
E-mail address: subir_ghosh2@rediffmail.com (S. Ghosh).
doi:10.1016/j.nuclphysb.2003.08.016
360 S. Ghosh / Nuclear Physics B 670 [FS] (2003) 359–372
Specifically, the open string boundaries, attached to D-branes [3], in the presence of a two-
form background field, turn into NC spacetime coordinates [1]. (This phenomenon has
been recovered from various alternative viewpoints [4].) The noncommutativity induces an
NC D-brane world volume and hence field theories on the brane become NC field theories.
Studies in NC field theories have revealed unexpected features, such as UV–IR mixing
[5], soliton solutions in higher-dimensional scalar theories [6], to name a few. The inherent
nonlocality (or equivalently the introduction of a length scale by θ —the noncommutativity
parameter), of the NC field theory is manifested through these peculiar properties, which
are absent in the corresponding ordinary spacetime theories. Also, solitons in NC CP(1)
model have been found [7], very much in analogy to their counterpart in ordinary
spacetime. The present work also deals with the search for the solitons in the NC CP(1)
model. The difference between our field theoretic analysis and the existing framework [7]
is explained below. In fact, we closely follow the conventional field theoretic approach
in ordinary spacetime [8]. The Seiberg–Witten map [1] plays a pivotal role in our
scheme. The Bogolmolny analysis of the static energy reveals a lower bound, protected
by topological considerations. However, we encounter a small discrepancy between the
BPS equations and the variational equation of motion. Although “small” in an absolute
sense, the mismatch is conceptually significant for the reasons elaborated below. The
above conclusions, drawn from field theoretic analysis, will be corroborated and quantified
explicitly in a collective coordinate framework.
It appears natural to attribute the above mentioned problem to the definition of the
energy functional of the NC CP(1) model in particular, and of the NC field theories in
general (since the BPS equations are derived directly from the energy of the system).
As it is well known, there are complications in the definition of the energy–momentum
(EM) tensor in NC field theory [9,10]. In general, it is not possible to obtain a symmetric,
gauge invariant and conserved EM tensor. There are two forms of EM tensor in vogue:
a manifestly symmetric form [10], obtained from the variation of the action with respect
to the metric, and the canonical form [9], following the Noether prescription. The former
is covariantly conserved whereas the latter is conserved. Interestingly, we find that in the
particular case that we are considering, that is NC CP(1) model in 2 + 1 dimensions,
with only spatial noncommutativity, expressions for the static energy, obtained from both
the derivations, are identical. Moreover the expression for energy is gauge invariant. We
show that there is a Bogomolny-like lower bound in the energy. However, the subsequently
derived BPS equations (that saturate the lower bound), does not fully satisfy the equation
of motion.
Let us put our work in its proper perspective. As such, our result in no way questions the
consistency of the existing literature [7] on CP(1) solitons since our model differs crucially
from the one considered in [7]. In particular, we have adopted a different NC generalization
of the CP(1) constraint. (We have provided conceptual and technical reasons for allowing
such a difference.) Thus it is not expected that the results of [7] will be reproduced. In fact,
the energy profile of the localized structure that we uncover, is more sharply peaked and
has an O(θ ) correction, with respect to the ordinary spacetime CP(1) soliton. (This will
be made explicit in the collective coordinate analysis at the end.) Surprisingly, the BPS
equations remain unchanged, although the variational equation of motion is altered.
S. Ghosh / Nuclear Physics B 670 [FS] (2003) 359–372 361
All the same, we emphasize that, we have provided a well defined NC gauge theory,
which conforms to the expected features of such a system. Hence, the clash between the
BPS equations and the equation of motion that is revealed here, can have a deeper bearing
on the structure of a general NC gauge theory, indicating that the traditional lore of field
theory in ordinary spacetime should be applied with greater care in the context of NC field
theory.
Our methodology and its difference from the existing one [7] is explained below. There
are two basic approaches in studying an NC field theory:
(i) The appropriate NC field theory is constructed in terms of NC analogue fields (ψ̂)
of the fields (ψ) with the replacement of ordinary products of fields (ψϕ), by the
Moyal–Weyl ∗-product (ψ̂ ∗ ϕ̂),
i
ψ̂(x) ∗ ϕ̂(x) = e 2 θµν ∂σµ ∂ξν ψ̂(x + σ )ϕ̂(x + ξ )|σ =ξ =0
i
= ψ̂(x)ϕ̂(x) + θ ρσ ∂ρ ψ̂(x)∂σ ϕ̂(x) + O θ 2 . (1)
2
The hatted variables are NC degrees of freedom. We take θ ρσ to be a real constant
antisymmetric tensor, as is customary [1] (but this need not always be the case [11]).
The NC spacetime follows from the above definition,
ρ σ
x , x ∗ = iθ ρσ . (2)
Note that the effects of spacetime noncommutativity has been accounted for by the
introduction of the ∗-product. For gauge theories the Seiberg–Witten map [1] plays a
crucial role in connecting φ̂(x) to φ(x). This formalism allows us to study the effects
of noncommutativity as θ ρσ dependent interaction terms in an ordinary spacetime field
theory format. This is the prescription we will follow.
(ii) An alternative framework is to treat the NC theories as systems of operator valued
fields and to directly work with operators on the quantum phase space, characterized
by the noncommutativity condition (2). On the NC plane, the coordinates satisfy
a Heisenberg algebra [x 1 , x 2 ]∗ = x 1 ∗ x 2 − x 2 ∗ x 1 = iθ 12 = i 12θ = iθ which in
the complex coordinates reduces to the creation annihilation operator algebra for the
simple harmonic oscillator. Thus to a function in the NC spacetime, through Weyl
transform, one associates an operator acting on the Hilbert space, in a basis of a simple
harmonic oscillator eigenstates.
The investigations on the NC CP1 solitons carried out so far [7,12] exploit the latter
method. As it turns out, a major advantage is that structurally, the NC system with its
dynamical equations, energy functionals, etc., are similar to their ordinary spacetime
counterpart. This happens because the ∗-products of (i) are replaced by operator products
in (ii) and the spacetime integrals are replaced by trace over the basis states in the Hilbert
space.
In the present work, our aim is to study the NC CP1 solitons in the former field theoretic
approach. From past experiences [13] we know this to be a perfectly viable formalism.
Indeed, since the NC spacetime physics is not that much familiar or well understood, it
is imperative that one explores different avenues to reach the same goal, to gain further
insights. Also we would like to point out that since solitons are already present in the CP1
model at θ = 0 (i.e., ordinary spacetime), unlike the noncommutative solitons of the scalar
theory [6] it is natural to analyze the fate of the solitons under a small perturbation (which
is a small value of θ in the present case). The small θ -results of [7,12] are perfectly well
defined.
The paper is organized as follows: Section 2 contains a short recapitulation of the CP(1)
solitons. This will help us fix the notations and in fact, identical procedure will be pursued
in the NC theory as well. The detailed construction of our version of the NC CP(1) model
is provided in Section 3. Section 4 discusses the energy–momentum tensor of the model.
Section 5 consists of the Bogomolny analysis in the NC theory. Section 6 is devoted to the
collective coordinate analysis. The paper ends with a conclusion in Section 7.
2. CP(1) soliton: a brief digression
Let us digress briefly on the BPS solitons of CP(1) model in ordinary spacetime. Later
we will proceed with the NC theory in an identical fashion. The gauge invariant action,

†
S = d 3 x D µ φ Dµ φ + λ φ † φ − 1 , (3)
where Dµ φ = (∂µ − iAµ )φ defines the covariant derivative and the multiplier λ enforces
the constraint, the equation of motion for Aµ leads to the identification,
Aµ = −iφ † ∂µ φ. (4)
Since the “gauge field” Aµ does not have any independent dynamics one is allowed to
make the above replacement directly in the action. Obviously the infinitesimal gauge
transformation of the variables are,
δφ † = −iλφ † , δφ = iλφ, δAµ = ∂µ λ. (5)

From the EM tensor
†
Tµν = (Dµ φ)† Dν φ + (Dν φ)† Dµ φ − gµν D σ φ Dσ φ, (6)
the total energy can be expressed in the form,

2
E = d 2 x |D0 φ|2 + (D1 ± iD2 )φ ± 2πN, (7)
where the last term denotes the topological charge

1
N ≡ d 2 x n(x) = d 2 x ij (Di φ)† Dj φ
2πi

1 ij 1
= d x 2
Fij = d 2 x F12 , (8)
4π 2π
corresponding to the conserved topological current. The Bogomolny bound follows from
(7),
E 2π|N|, (9)
with the following saturation conditions (BPS equations) obeyed by the soliton,
2
|D0 φ|2 = (D1 ± iD2 )φ = 0. (10)
It can be checked that the solutions of the BPS equations belong to a subset of the full set
of solutions, that satisfy the variational equation of motion.
3. Construction of the NC CP(1) model
Let us now enter the noncommutative spacetime. The first task is to generalize the
scalar gauge theory (3) to its NC version, keeping in mind that the latter must be ∗-gauge
invariant. The NC action is,

† †
Ŝ = d 3 x D̂ µ φ̂ ∗ D̂µ φ̂ = d 3 x D̂ µ φ̂ D̂µ φ̂, (11)
where the NC covariant derivative is defined as

D̂µ φ̂ = ∂µ φ̂ − i Âµ ∗ φ̂.
Depending on the positioning of Âµ and φ̂, the covariant derivative can act in three ways,
D̂µ φ̂ = ∂µ φ̂ − i Âµ ∗ φ̂
= ∂µ φ̂ + i φ̂ ∗ Âµ

= ∂µ φ̂ − i φ̂ ∗ Âµ − Âµ ∗ φ̂ , (12)
which are termed, respectively, as fundamental, anti-fundamental and adjoint representa-
tions. We have chosen the fundamental one.1 Notice that for the time being we have not
considered the target space (CP(1)) constraint. We will return to this important point later.
The NC action (11) is invariant under the ∗-gauge transformations,

δ̂ φ̂ † = −i λ̂ ∗ φ̂ † , δ̂ φ̂ = i λ̂ ∗ φ̂, δ̂ Âµ = ∂µ λ̂ + i λ̂, Âµ ∗ . (13)
We now exploit the Seiberg–Witten map [1,16] to revert back to the ordinary spacetime
degrees of freedom. The explicit identifications between NC and ordinary spacetime
counterparts of the fields, to the lowest nontrivial order in θ are,

1
Âµ = Aµ + θ σρ Aρ ∂σ Aµ − ∂µ Aσ ,
2
1 ρσ 1
φ̂ = φ − θ Aρ ∂σ φ, λ̂ = λ − θ ρσ Aρ ∂σ λ. (14)
2 2
1 This is the first difference between our model and [7] who use the anti-fundamental representation. In fact,
in [7], it is difficult to proceed with the fundamental definition [14]. On the other hand, in the present work, the
choice between the first and second definition is not very important as it affects the overall sign of θ only. Similar
type of situation prevails in [13,15].
As stated before, the “hatted” variables on the left are NC degrees of freedom and gauge
transformation parameter. The higher order terms in θ are kept out of contention as there
are certain nonuniqueness involved in the O(θ 2 ) mapping. The significance of the Seiberg–
Witten map is that under an NC or ∗-gauge transformation of Âµ by,

δ̂ Âµ = ∂µ λ̂ + i λ̂, Âµ ∗ ,
Aµ will undergo the transformation
δAµ = ∂µ λ.
Subsequently, under this mapping, a gauge invariant object in conventional spacetime will
be mapped to its NC counterpart, which will be ∗-gauge invariant. This is crucial as it
ensures that the ordinary spacetime action that we recover from the NC action (11) by
applying the Seiberg–Witten map will be gauge invariant. Thus the NC action (11) in
ordinary spacetime variables reads,

†
Ŝ = d 3 x D µ φ Dµ φ

1 † 1 †
+ θ αβ Fαµ (Dβ φ)† D µ φ + D µ φ Dβ φ − Fαβ D µ φ Dµ φ .
2 2
(15)
The above action is manifestly gauge invariant. The equation of motion now satisfied by
Aµ is,

1 αβ
i −2iAµ φ φ + φ ∂µ φ − ∂µ φ φ 1 − θ Fαβ
† † †
2
1 α β † α † β †
+ θαµ ∂ D φ Dβ φ − ∂β D φ D φ + D β φ D α φ
2
1 † †
− θαβ ∂ α D β φ Dµ φ + (Dµ φ)† D β φ + iFµα φ † D β φ − D β φ φ = 0.
2
(16)
Remember that so far we have not introduced the CP1 target space constraint in the NC
spacetime setup. Let us assume the constraint to be identical to the ordinary spacetime
one, i.e.,2
φ † φ = 1. (17)
2 This is the second difference between our model and that of [7], where θ -correction terms are present in
the CP(1) constraint. This is a serious difference as it drastically alters the structures of the model in [7] from
ours. Apart from the conceptual reasoning given above, there also appears a technical compulsion. We would
like to obtain perturbative θ -corrections to the ordinary spacetime CP(1) model. Incorporating θ -corrections in
the CP(1) constraint as in [7] in our system will lead to a differential equation for the multiplier λ, instead of an
algebraic one as in the ordinary spacetime case. This will change the φ-equation of motion in a qualitative way.
We stress that our model is a perfectly well defined NC theory which, incidentally, is distinct from the existing
NC CP(1) model [7].
The reasoning is as follows. Primarily, after utilizing the Seiberg–Witten map, we have
returned to the ordinary spacetime and its associated dynamical variables and the effects
of noncommutativity appears only as additional interaction terms in the action. Hence, it
is natural to keep the CP1 constraint unchanged. Alternatively, the above assumption can
also be motivated in a roundabout way. Remember that the CP(1) constraint has to be
introduced in a ∗-gauge invariant way. In order to introduce the CP1 constraint directly in
the NC action (11) or (15), the constraint term λ(φ † φ − 1) has to be generalized to a
∗-gauge invariant one by the application of the (inverse) Seiberg–Witten map. This is quite
straightforward but needless because as soon as we apply the Seiberg–Witten map to the
∗-gauge invariant constraint term, we recover the earlier ordinary spacetime constraint.
This allows us to write,
Aµ = −iφ † ∂µ φ + aµ (θ ), (18)
with aµ denoting the O(θ ) correction, obtained from (16), (17). For θ = 0, Aµ reduces
to its original form. Note that aµ is gauge invariant. Thus the U (1) gauge transformation
of Aµ remains intact, at least to O(θ ). Keeping in mind the constraint φ † φ = 1, let us
now substitute (18) in the NC action (15). Since we are concerned only with the O(θ )
correction, in the θ -term of the action, we can use Aµ = −iφ † ∂µ φ. However, in the
first term in the action, we must incorporate the full expression for Aµ given in (18).
Remarkably, the constraint condition conspire to cancel the effect of the O(θ ) correction
term aµ . Finally it boils down to the following: the action for the NC CP1 model to O(θ )
is given by (15) with the identification Aµ = −iφ † ∂µ φ and φ † φ = 1.
4. Energy–momentum tensor for the NC CP(1) model
Our aim is to study the possibility of soliton solutions for the action (15). Let us try to
derive the Bogomolny bound and BPS equations in the present case. The first task is to
compute the EM tensor.
We follow [10] in computing the symmetric form of the EM tensor by coupling the
model with a weak gravitational field and get,

1 †
S
Tµν = 1 − θ αβ Fαβ (Dµ φ)† Dν φ + (Dν φ)† Dµ φ − gµν D σ φ Dσ φ
4
1 αβ
+ θ Fαµ (Dβ φ)† Dν φ + (Dν φ)† Dβ φ
2

+ Fαν (Dβ φ)† Dµ φ + (Dµ φ)† Dβ φ
†
− gµν Fασ (Dβ φ)† D σ φ + D σ φ Dβ φ . (19)
S stands for the symmetric form of the EM tensor. In the static situation,
The Tµν
φ̇ = 0 → A0 = F0i = D0 φ = 0
and the static energy density simplifies to,

1 †
S
T00 = 1 + θ 12 F12 D i φ D i φ, (20)
2
where Fµν is also expressible in the form

Fµν = −i (Dµ φ)† Dν φ − (Dν φ)† Dµ φ = −i (∂µ φ)† ∂ν φ − (∂ν φ)† ∂µ φ .
Now we discuss the canonical form of the EM tensor. Remembering that the indices of
adjacent φ’s are summed, the expanded form of the Lagrangian is,
†
L̂ = D µ φ Dµ φ + λ φ † φ − 1
i
− θ αβ 2∂µ φ † ∂β φ∂α φ † ∂ µ φ + 2∂µ φ † ∂β φφ † ∂α φφ † ∂ µ φ
2
− 2∂β φ † ∂µ φφ † ∂α φφ † ∂ µ φ − ∂α φ † ∂β φ∂ µ φ † ∂µ φ

− ∂α φ † ∂β φφ † ∂ µ φφ † ∂µ φ . (21)
The canonical energy–momentum tensor is,
δ L̂ δ L̂
Tµν = ∂ν φ † + ∂ν φ − gµν L̂
δ(∂ µ φ † ) δ(∂ µ φ)

1
= 1 − θ αβ Fαβ Dµ φ † Dν φ + (µ ↔ ν)
4

− iθ ∂ν φ † ∂β φ ∂α φ † ∂µ φ + φ † ∂α φφ † ∂µ φ + (µ ↔ ν)
αβ

− ∂β φ † ∂ν φφ † ∂α φφ † ∂µ φ − (µ ↔ ν)

1 1 α † †
− iθµα − ∂ν φ ∂ φ + ∂ φ ∂ν φ D µ φ Dµ φ
† α
2 2

− ∂σ φ † ∂ν φ ∂ α φ † ∂ σ φ + φ † ∂ α φφ † ∂ σ φ

+ ∂σ φ † ∂ α φ ∂ν φ † ∂ σ φ + φ † ∂ν φφ † ∂ σ φ + ∂ν φ † ∂σ φφ † ∂ α φφ † ∂ σ φ

− ∂ φ ∂σ φφ ∂ν φφ ∂ φ − gµν L̂.
α † † † σ
(22)
Note that θ αβ part is symmetric. In the energy density T00 the contribution coming from the
nonsymmetric parts in the θ -contribution drop out if only space–space noncommutativity
is assumed, i.e., θ 0i = 0. For this special case, in the static limit, the above θ -contribution
completely drops out and the energy density reduces to

1 †
T00 = −L̂ = 1 + θ 12 F12 D i φ D i φ.
N
(23)
2
Clearly this is identical to the static energy (20) obtained from the symmetric form Tµν S .
Indeed, it is satisfying that in this particular case, both the canonical and symmetric forms
of the EM tensor lead to the same expression of the static energy, which is manifestly gauge
invariant and conserved (as it comes from the canonical form).
Interestingly to O(θ ), the noncommutativity effect factors out from the ordinary
spacetime result. Also notice that in the two spatial dimensions that we are considering,
the θ -term in the energy density is proportional to the topological charge density n(x) in
(8). This is because the expression for the topological current remains unchanged since
the dynamical variables as well as the CP(1) constraint is unaltered in our model. We
specialize to only space–space noncommutativity, θ ij = θ ij , θ 0i = 0, and find,
(θ) †
T00 = π θ n(x) D i φ D i φ, (24)
where the superscript S or N is dropped.
5. Analysis of the Bogomolny bound
In order to obtain the Bogomolny bound, we follow the same procedure as that of the
CP1 model in ordinary spacetime and rewrite the static energy functional in the following
form,

2
E= d 2 x (1 + π θ n) (D1 ± iD2 )φ ± 2πn

2
= d 2 x (1 + π θ n)(D1 ± iD2 )φ ± 2πn ± 2π 2 θ n2
2
1 2
= d 2 x 1 + π θ n(x) (D1 ± iD2 )φ
2

± 2πN ± 2π θ d 2 x n2 (x) + O θ 2 .
2
(25)
Now individually all the terms in the energy expression are positive definite. Hence we
obtain the Bogomolny bound to be

E N + 2π θ 2
d 2 x n2 (x) (26)
and the saturation condition is

2
1
1+ π θ n(x) (D1 ± iD2 )φ 2 = 0. (27)
2
The BPS equation turns out to be,
D 1 φ = ±iD 2 φ. (28)
Thus we find that the BPS equation remains unchanged and there is a O(θ ) correction in
the static energy of the soliton. Note that both of the above results do not agree with [7,12].
But this is not unexpected since as we have mentioned before, the defining conditions
of the NC CP(1) models are different. However, we repeat that a priori there is nothing
inconsistent in our NC model.
Finally, we are ready to discuss the curiosity. It appears that solutions of the BPS
equations (28) do not satisfy the equation of motion for φ. A straightforward computation
yields the dynamical equation for φ,

1
D µ 1 − θ αβ F αβ Dµ φ
4
1 †
+ θ αβ iDα D σ φ Dσ φDβ φ + Dβ Fαµ D µ φ + D µ Fαµ Dβ φ
2
†
− iDα (Dβ φ)† D µ φ + D µ φ Dβ φ Dµ φ
†
+ iDµ (Dβ φ)† D µ φ + D µ φ Dβ φ Dα φ − λφ = 0. (29)
(Details of the derivation are provided in Appendix A.) To get λ, contract by φ † and use
φ † φ = 1. For the time being, instead of writing the full equation of motion, we want to
check the consistency of the programme only, that is whether the solution of the BPS
equation satisfies the equation of motion, which they should. Since the BPS equation
remains unchanged here, the θ -term in the equation of motion should vanish for those
solutions that satisfy the BPS equation as well. So we consider the equation of motion in
a simplified setting where the BPS equation is satisfied and only θ 12 is nonzero and obtain
for λ
1
λ = φ † D i Di φ + θ 12 φ † D i (F12 Di φ). (30)
2
Putting λ back in the equation of motion, we get
1
D i Di φ − φ † D i Di φφ + θ 12 D i (F12 Di φ) − φ † D i (F12 Di φ)φ = 0. (31)
2
This equation can be rewritten as
D i Di φ − φ † D i Di φφ
1
+ θ 12 ∂ i F12 Di φ − φ † ∂ i F12 Di φφ + F12 D i Di φ − φ † D i Di φφ = 0. (32)
2
Clearly the last term, that is 12 θ 12 F12 (D i Di φ − φ † D i Di φφ) ≈ O(θ 2 ) and can be dropped.
The term θ 12 φ † ∂ i F12 Di φφ = θ 12 ∂ i F12 φ † Di φφ = 0. However the remaining O(θ )-term,
1 12 i
2 θ ∂ F12 Di φ does not vanish. This is the purported mismatch between the BPS equations
and the full equation of motion. This brings us to the last part—the collective coordinate
analysis, where we can check explicitly the above conclusions in a simplified setup.
6. Collective coordinate analysis
We consider the topological charge N = 1 sector. As we have discussed before,

expression for the topological current and subsequently the charge remains same (in our
NC CP(1) model) as that of the ordinary spacetime CP(1) model. This means that we can
use the same parameterizations as before [8] to introduce the collective coordinates. As a
first approximation, only the zero mode arising from the global U (1) invariance is being
quantized. In the O(3) nonlinear sigma model, the N = 1 sector is characterized by [8]

na = r̂ sin g(r) , cos g(r) , a = 1, 2, 3,
Fig. 1.
with the constraint na na = 1 and the boundary conditions g(0) = 0; g(∞) = π . Keeping
in mind the O(3)–CP(1) duality and the Hopf map na = φ † σ a φ, the soliton profile in the
CP(1) variables is of the form,

φ1 cos g2
φ≡ = , (33)
φ2 sin g2 ei(ϕ+α(t ))
where the gauge (φ 1 )∗ = φ 1 has been used and r, ϕ refer to the plane polar coordinates.
α(t) is the collective coordinate. Substituting the above choice (33) in the static energy
expression in (20) leads to,

sin(g)g 2 sin2 (g)
E(r) = 1 + θ g + , (34)
2r r2
where g = dg
dr . In Fig. 1 the effect of the θ -correction is shown where the following simple
form of g(r) is considered,

g(r) ≈ π 1 − e−µr . (35)
One can clearly see that with typical values of the parameters (θ = 1, µ = 1) the energy
density for the NC case is more sharply peaked. (In reality, θ should be smaller.) This
assures us of the rationale of our previous Bogomolny analysis. Next we look in to the
equation of motion.
It is straightforward check that the profile (33) satisfies the BPS equation as well as
the equation of motion for the ordinary spacetime situation, θ = 0. For θ = 0, the BPS
equations are once again satisfied since they remain unaltered. So we concentrate only on
the problem term 12 θ 12 ∂ i F12 Di φ in the equation of motion (32). With the particular form
Fig. 2.
of g(r) in (35) we obtain

1 12 i θg g g g g (g )2 g
θ ∂ F12 Di φ = sin − 2 sin + cos
2 4 2r 2 2r 2 4r 2

sin g2 X(r)
× ≡ . (36)
− cos g2 eiϕ Y (r)
In Fig. 2 we again use the g(r) given in (35) and compare the magnitude of the above
expression with a typical term,

S(r)
∂ ∂i φ ≡
i
,
T (r)
occurring in the equation of motion (32). For consistency, the expressions in (36) should
have vanished. However, even for θ = 1 (which is quite a large value), the mismatch term
is small in an absolute sense.
7. Conclusions
In this paper, we have attempted to recover the soliton solutions in the noncommutative
CP(1) model, discovered earlier [7]. The U (1) and NC U (1) gauge invariances in the
CP(1) and NC CP(1) models, respectively, requires the use of the Seiberg–Witten map, to
convert the NC action to an action comprising of ordinary spacetime dynamical variables.
The effects of noncommutativity are manifested as interaction terms. For theoretical as
well as technical reasons, we found it convenient to keep the CP(1) constraint unchanged,
i.e., without any θ correction.
From the above action, we construct both the symmetric and canonical forms of the
energy–momentum tensor. For only spatial noncommutativity, both the above forms reduce
to an identical (gauge invariant) expression for the static energy. The Bogomolny analysis
yields a lower bound in the energy, hinting at the presence of a new type of soliton.
However, the resulting BPS equations do not match completely with the full variational
equation of motion. The present model is otherwise a perfectly well defined NC field
theory with the expected features. Hence we conclude that inadequacy in the definitions
of the energy–momentum tensor in an NC field theory is responsible for this failure. The
above phenomena are nicely visualized in a collective coordinate framework. The above
awkward situation clearly demands further study.
Finally, as a future work we mention that inclusion of the Hopf term (in the form of
Chern–Simons term in CP(1) variables), in the NC theory would indeed be interesting.
The Hopf term was introduced [8,17] to impart anyonic behavior to the CP(1) solitons.
The exact form of the NC version of the Chern–Simons term is known [15]—it is a “non-
Abelian” generalization of the Chern–Simons term. In our formalism, application of the
Seiberg–Witten map will reduce it to the ordinary Chern–Simons term [15] but there will
appear O(θ ) correction terms since the gauge field of the Chern–Simons term is actually a
nonlinear combination of the CP(1) variables.
Another interesting problem is the reconstruction of the NC CP(1) model of [7] in our
framework.
Acknowledgements
It is a pleasure to thank Professor Hyun Seok Yang for fruitful correspondence. Also I
thank Professor Avinash Khare for a helpful discussion. Lastly I am indebted to Bishwajit
Chakraborty for a free access to his notes on CP(1) model.
Appendix A
To get the φ-equation of motion, we consider variation of φ † and exploit the relations,
µ †
D φ φ = φ † D µ φ = 0, (A.1)
µ † µ † † µ †
δ D φ = δ ∂ φ + φ ∂ φ φ = ∂ µ δφ † + δφ † ∂ µ φ φ † + φ † ∂ µ φ δφ † ,

δ D µ φ = δ ∂ µ φ − φ † ∂ µ φ φ = − δφ † ∂ µ φ φ. (A.2)
µ † ν
In the action the terms are products of the generic form (D φ) (D φ)X(x). The
variation of (D µ φ) will reproduce

µ † ν µ † † µ
D φ δ D φ X=− D φ φ δφ ∂ φ X = 0,
by using (A.1). Similarly, the variation of (D µ φ)† will yield

† µ † † µ † † µ † ν
δ Dµ φ Dν φ X = ∂ δφ + δφ ∂ φ φ + φ ∂ φ δφ D φX

= − δφ † D µ D ν φX
by partial integration and using (A.1). The above identities simplifies the computations
considerably and leads to Eq. (29).
References
[1] N. Seiberg, E. Witten, JHEP 9909 (1999) 032.

[2] For reviews see, for example:
M.R. Douglas, N.A. Nekrasov, Rev. Mod. Phys. 73 (2001) 977;
R.J. Szabo, Quantum field theory on noncommutative spaces, hep-th/0109162.
[3] J. Polchinski, Phys. Rev. Lett. 75 (1995) 4724.
[4] C.S. Chu, P.-M. Ho, Nucl. Phys. B 550 (1999) 151;
V. Schomerus, JHEP 06 (1999) 030;
R. Banerjee, B. Chakraborty, S. Ghosh, Phys. Lett. B 537 (2002) 340.
[5] S. Minwalla, M. Van Raamsdonk, N. Seiberg, Noncommutative perturbative dynamics, hep-th/9912072.
[6] R. Gopakumar, S. Minwalla, A. Strominger, JHEP 0005 (2000) 020.
[7] B.-H. Lee, K. Lee, H.S. Yang, Phys. Lett. B 498 (2001) 277;
K. Furuta, et al., Low energy dynamics of noncommutative CP1 solitons in 2 + 1 dimensions, hep-
th/0203125;
H. Otsu, et al., New BPS solitons in (2 + 1)-dimensional noncommutative CP1 model, hep-th/0303090.
[8] M. Bowick, D. Karabali, L.C.R. Vijewardhana, Nucl. Phys. B 271 (1986) 417.
[9] J.M. Grimstrup, et al., hep-th/0210288.
[10] A. Das, J. Frenkel, Phys. Rev. D 67 (2003) 067701, hep-th/0212122.
[11] H.S. Snyder, Phys. Rev. 71 (1947) 38;
For some later works, see, for example:
S. Doplicher, K. Fredenhagen, J.E. Roberts, Phys. Lett. B 331 (1994) 39;
K. Morita, Lorentz-invariant noncommutative QED, hep-th/0209234;
S. Ghosh, Phys. Rev. D 66 (2002) 045031.
[12] J. Murugan, R. Adams, Comments on noncommutative sigma models, hep-th/0211171.
[13] S. Ghosh, Phys. Lett. B 558 (2003) 245;
S. Ghosh, Phys. Lett. B 563 (2003) 112.
[14] H.S. Yang, private communications.
[15] N. Grandi, G.A. Silva, Phys. Lett. B 507 (2001) 345.
[16] B. Jurco, et al., hep-th/0104153.
[17] See also, B. Chakraborty, S. Ghosh, R.P. Malik, Nucl. Phys. B 600 (2001) 351;
B. Chakraborty, Mod. Phys. Lett. A 17 (2002) 115.
Globally conformal invariant gauge field theory

with rational correlation functions ✩
N.M. Nikolov a , Ya.S. Stanev a,b , I.T. Todorov a,c
a Institute for Nuclear Research and Nuclear Energy, Tsarigradsko Chaussee 72, BG-1784 Sofia, Bulgaria
b INFN, Sezione di Roma II, Via della Ricerca Scientifica 1, I-00133 Roma, Italy
c Section de Mathématiques, Université de Genève, 2-4 rue du Lièvre, cp 240, CH-1211 Genève, Switzerland
Abstract
Operator product expansions (OPE) for the product of a scalar field with its conjugate are presented
as infinite sums of bilocal fields Vκ (x1 , x2 ) of dimension (κ, κ). For a globally conformal invariant
(GCI) theory we write down the OPE of Vκ into a series of twist (dimension minus rank) 2κ
symmetric traceless tensor fields with coefficients computed from the (rational) 4-point function
of the scalar field.
We argue that the theory of a GCI hermitian scalar field L(x) of dimension 4 in D = 4 Minkowski
space such that the 3-point functions of a pair of L’s and a scalar field of dimension 2 or 4 vanish
can be interpreted as the theory of local observables of a conformally invariant fixed point in a gauge
theory with Lagrangian density L(x).
PACS: 11.25.Hf; 11.10.-z; 11.15.-q; 11.15.Tk; 11.30.-j
1. Introduction
The present paper offers a new step in the realization of the program set up in [19,20]
and [21] of constructing a 4-dimensional conformal field theory (CFT) model with rational
correlation functions of observable fields.
✩
A brief preview of this paper is contained in Proc. 3rd Int. Sakharov Conf. on Physics, Vol. 2, 2002, hep-
th/0211106.
E-mail addresses: mitov@inrne.bas.bg (N.M. Nikolov), stanev@roma2.infn.it (Ya.S. Stanev),
todorov@inrne.bas.bg (I.T. Todorov).
doi:10.1016/j.nuclphysb.2003.08.006
374 N.M. Nikolov et al. / Nuclear Physics B 670 [FS] (2003) 373–400
Global conformal invariance (GCI) allows to write down close form expressions
for correlation functions involving (each) a finite number of free parameters. Thus,
the truncated 4-point function w4t of a neutral scalar field of (integer) dimension d
depends on [d 2 /3] real parameters ([a] standing for the integer part of the positive
number a). In addition to GCI it satisfies the constraints of locality and energy positivity.
Operator product expansion (OPE) provides a method of taking the remaining condition
of Wightman (i.e., Hilbert space) positivity into account. We organize systematically
OPE into mutually orthogonal bilocal fields Vκ (x1 , x2 ) of dimension (κ, κ) defined by
the condition that Vκ can be expanded into an (infinite) sum (of integrals) of twist
2κ symmetric traceless tensor local fields. An algorithm is given for computing the
full contribution of each Vκ to the 4-point function. The effectiveness of this approach
is enhanced by the fact that V1 (x1 , x2 ), which involves an infinite sum of conserved
tensor fields, satisfies the d’Alembert equation in each argument and the 4-point function
0|V1 (x1 , x2 )V1 (x3 , x4 )|0 is rational by itself. For the model of a d = 2 neutral scalar field
φ(x) the OPE of two φ’s can be summed up simply as [20]
φ(x1 )φ(x2 ) = 0|φ(x1 )φ(x2 )|0 + (12)V1(x1 , x2 ) + :φ(x1 )φ(x2 ): (1.1)
where (12) is the free 0-mass 2-point function,
1
(12) = , ρ12 = x12
2
+ i0x12
0
,
4π 2 ρ12
x12 = x1 − x2 , x 2 = x2 − x02 , (1.2)
and the normal product :φ(x1 )φ(x2 ): defined by (1.1) is non-singular for x1 = x2 . The
simplicity of V1 in this model has been exploited in [20] to prove that it can be written as
a sum of normal products of (mutually commuting) free 0-mass fields.
We focus in the present paper on the study of a model generated by a (neutral)
scalar field which can be interpreted as a (gauge invariant) Lagrangian density L(x).
Taking into account the above cited triviality result for d = 2 we require that the OPE
of L(x1 )L(x2 ) − 0|L(x1 )L(x2 )|0 does not involve any field of dimension lower than 4
(which just amounts to excluding a possible scalar field of dimension 2). This requirement
decreases the number of free parameters in w4t from five to four. It already excludes most
of the standard renormalizable interaction Lagrangians (like Yukawa and ϕ 4 ) but allows
for the Lagrangian L and for the pseudoscalar topological term L of a pure gauge theory
1
L(x) = − tr Fµν (x)F µν (x) , (1.3)
4

= 1 εκλµν tr Fκλ (x)Fµν (x) .
L(x) (1.4)
4
The invariance of L under conformal rescaling of the metric is made manifest by writing
the action density in terms of the Yang–Mills curvature form
1
F (x) := Fµν (x) dx µ ∧ dx ν (1.5)
2
and its Hodge dual, ∗ F :

L(x) |g| dx 0 ∧ dx 1 ∧ dx 2 ∧ dx 3 = tr ∗ F (x) ∧ F (x) (1.6)
N.M. Nikolov et al. / Nuclear Physics B 670 [FS] (2003) 373–400 375
(where |g| is the absolute value of the determinant of the metric tensor, |g| = 1 for the
Minkowski metric). (It does not seem superfluous to reiterate that the conformal invariance
of the gauge field Lagrangian singles out D = 4 as the dimension of spacetime [6].)
It seems appropriate to demand further invariance of the theory under “electric–
magnetic duality”,1 i.e., under the change F → ∗ F . Taking into account the fact that the
Hodge star defines a complex structure in Minkowski space, we deduce that the Lagrangian
(1.6) changes sign under ∗:
∗ ∗
( F ) = −F ⇒ L( ∗ F ) = −L(F ). (1.7)
It follows that odd point correlation functions of L(x) should vanish. Requiring space
reflection invariance we also find that the vacuum expectation values of an odd number
(and any number of L’s) should vanish. This leads us to the assumption that no
of L’s
(pseudo)scalar field of dimension 4 appears in the OPE of two L’s thus eliminating one
more parameter in the expression for w4t .
The case of a non-abelian gauge field is distinguished by the existence of a non-trivial
3-point function of F which agrees with local commutativity of Bose fields. Consistency of
the equations of motion satisfied by the 2- and 3-point functions of F with OPE, however,
requires the use of indecomposable representations of the conformal group C which makes
unpractical exploiting the compositeness of L (Section 5.2).
The paper is organized as follows.
We outline basic ideas and review earlier work and notation in Section 2. In particular,
we make precise the notation of an elementary positive energy representation of C, and
indicate an extension of the present approach to any even spacetime dimension D.
Section 3.1 provides a general treatment of the OPE of a pair of conjugate scalar fields
of dimension d ∈ N in terms of bilocal fields Vκ (x1 , x2 ). It also reproduces the formula for
the expansion of Vκ into an infinite series of symmetric traceless tensor fields of twist 2κ.
Section 3.2 displays the (crossing symmetrized) contribution of twist 2 (conserved) tensors
and their light cone expansions for d = 2 and d = 4.
In Section 4.1 we study systematically the truncated 4-point function of the d = 4 field
L(x) and discuss the special case of the Lagrangian L0 (4.17) of the free Maxwell field.
In Section 4.2 we analyse the conformally invariant (rational) factor f1 of the harmonic
4-point function
0|V1 (x1 , x2 )V1 (x3 , x4 )|0 = (13)(24)f1(s, t), (1.8)
displaying a basis of solutions jν (s, t) of the “conformal Laplace equation” (4.7) which
admits a natural crossing symmetrization. Section 4.3 displays the operator content and
the light cone expansion in twist 4 and twist 6 tensor fields in the case of general L.
The restrictions on the parameters of the truncated 4-point function coming from
Wightman positivity of the singular part of the s-channel OPE are analysed in Section 5.1.
Section 5.3 is devoted to a summary of results and concluding remarks.
Appendix A summarizes the derivation of the main result of [20] referred to in the text
discussing on the way the possibility for a similar treatment of the d = 4 case of interest
(and the difficulties lying ahead).
1 The authors thank Dirk Kreimer for this suggestion.

2. Consequences of GCI (a synopsis)
2.1. Elementary positive energy representations of the conformal group
The idea that quantum fields should be subdivided into local observables and “charged
fields” (that are relatively local to the observables and act as intertwiners among different
superselection sectors) emerged from several decades of work of Haag and collaborators
reviewed in [16]. It has been implemented in 2-dimensional conformal field theory (CFT)
models in which chiral currents (including the stress–energy tensor) play the part of
local observables while primary (with respect to a given chiral algebra) “chiral vertex
operators” correspond to “charged fields” intertwining between different superselection
sectors. (Among the numerous reviews on 2D CFT we cite the textbook [7] and the
earlier article [15] which refers to the axiomatic quantum field theory framework.) Local
observable fields are assumed to satisfy Wightman axioms [24]. More specifically, we
demand that their set contains both the stress energy tensor Tµν (x) (cf. [18]) and the
Lagrangian density L(x).
To set the stage we fix the class of local field representations of the quantum mechanical
conformal group C(D) = Spin(D, 2) under consideration (cf. [9,17,25]). The forward tube,
1/2
T+ = x + iy; y 0 > |y| := y12 + · · · + yD−1 2
, (2.1)
the primitive analyticity domain of the vector valued function ψ(x + iy)|0 for any
Wightman field ψ, appears as a homogeneous space of C(D),
T+ = C(D)/K(D) for K(D) = Spin(D) × U (1), (2.2)

K(D) being the maximal compact subgroup of C(D). An elementary positive energy
representation of C(D) is one induced by a (finite-dimensional) irreducible representation
of K(D). Any such representation gives rise to a transformation law for local fields defined
(as operator valued distributions) on the (conformally) compactified Minkowski space M D
which appears as a (part of the)
of T+ . MD also appears as a homogeneous
boundary

space of C(D) with respect to a D+1 2 + 1 parameter “parabolic subgroup” H(D) ⊂ C(D)
defined as a semidirect product H(D) = ND (Spin(D − 1, 1) × SO(1, 1)) of a D-
dimensional abelian subgroup ND of “special conformal transformations” with the direct
product of the (quantum mechanical) Lorentz group and the 1-parameter subgroup of C(D)
of uniform dilations:

MD C(D)/H(D) S1 × SD−1 /Z2 . (2.3)
The resulting elementary (local field) representation [25] can be defined as a positive
energy representation of C(D) induced by a finite-dimensional irreducible representation
of H(D) (energy positivity fixing the representation of the discrete centre—Z4 for even
D—of C(D), see [17]). In the case D = 4 of interest (studied in [17]),

C ≡ C(4) = SU(2, 2) ∼ = Spin(4, 2) ,

M
M(≡ 4 ) = U (2) = S1 × S3 /Z2 , (2.4)
the elementary representations of C are labeled by triples (d, j1 , j2 ) of non-negative half

integers, which admit two equivalent to each other interpretations. Viewed as labels of a K-
induced representation d stands for the minimal eigenvalue of the conformal Hamiltonian
(defined below) while (j1 , j2 ) label the irreducible representations of SU(2) × SU(2).
Alternatively, referring to a H-induced representation, d is the conformal dimension
(which may take all positive values if we substitute C by its universal covering) and
(j1 , j2 ) stands for a (2j1 + 1)(2j2 + 1)-dimensional representation of the Lorentz group
(a spin-tensor of 2j1 undotted and 2j2 dotted indices). The elementary positive energy
representation of C are, in general, neither unitary nor irreducible. They are proven [17] to
be unitary (or to admit an irreducible unitary subrepresentation) iff
(a) d 1 + j1 + j2 for j1 j2 = 0,
(b) d 2 + j1 + j2 for j1 j2 > 0. (2.5)
For a symmetric traceless tensor with # indices (i.e., for 2j1 = 2j2 = #) the counterpart of
(2.5) can be readily written for any spacetime dimension D:
D−2
d d0 := for # = 0,
2
d D − 2 + # for # = 1, 2, . . . , (2.6)
d0 being the dimension of a free 0-mass scalar field. It is quite remarkable that conventional
relativistic (quantum) fields fit precisely the above definition (rather than transforming
under Wigner’s unitary irreducible representations of the Poincaré group, cf. [5] or [24]).
Furthermore, a free scalar, Weyl spinor and Maxwell tensor correspond to the lower limit
of the above series (a), for weights (1, 0, 0), ( 32 , 12 , 0) and (2, 1, 0) + (2, 0, 1), respectively,
while conserved symmetric traceless tensors (starting with the vector current (3, 12 , 12 )) are
twist two fields (fitting the lower limit of the series (b) in (2.5) and of (2.6)).
In all the latter cases it is the subspace of solutions of the corresponding field equations
or conservation law that carries a unitary representation of C.
Note further that the conformal Hamiltonian, the generator of the centre, U (1), of the
maximal compact subgroup, Spin(D) × U (1)/Z2 of C(D) (S(U (2) × U (2)) for D = 4)
is positive definite whenever the Minkowski space energy is (cf. [23]) and has a discrete
spectrum belonging to the set {d + n; n = 0, 1, 2, . . .} for the positive energy representation
of C of weight (d, j1 , j2 ).
2.2. Huygens principle and rationality of conformally invariant correlation functions
Global conformal invariance (GCI) and local commutativity in four-dimensional (4D)

Minkowski space M imply the Huygens principle which can be stated in the following
strong form [19]. Let ψ(x) be a GCI local (Bose or Fermi) field of weight (d, j1 , j2 ); then
d + j1 + j2 should be an integer and
2 d+j1 +j2
x12 ψ(x1 )ψ ∗ (x2 ) − (−1)2j1 +2j2 ψ ∗ (x2 )ψ(x1 ) = 0 (2.7)
for all x1 , x2 in M. The Huygens principle and energy positivity (more precisely, the rela-
tivistic spectral conditions) imply rationality of correlation functions ([19, Theorem 3.1]).
Wightman positivity in 4D restricts the degree of singularities of truncated (≡ connect-

ed) n-point functions (which can only occur for coinciding arguments): they must be
strictly lower than the degree of the pole of the corresponding 2-point propagator. This
allows to determine any correlation function up to a finite number of (constant) parameters.
In particular, the truncated 4-point function of a hermitian scalar field φ of (integer)
dimension d ∈ N can be written in the form ([20, Section 1])
w4t ≡ wt (x1 , x2 , x3 , x4 ) := 1234 − 1234 − 1324 − 1423

d−2
ρ13 ρ24
= (12)(23)(34)(14) Pd (s, t) (2.8)
ρ12 ρ23 ρ34 ρ14
where
1 . . . n := 0|φ(x1 ) . . . φ(xn )|0, (2.9)
(12) is the free massless propagator (1.2), s and t are the conformally invariant cross-ratios
(so that the prefactor in (2.8) is a rational function of ρij ):
ρ12 ρ34 ρ14 ρ23

s= , t= , (2.10)
ρ13 ρ24 ρ13 ρ24
ρij are defined in (1.2) and Pd (s, t) is a polynomial in s and t of total degree 2d − 3
(P1 (s, t) ≡ 0):

Pd (s, t) = cij s i t j . (2.11)
i,j 0
i+j 2d−3
Furthermore, local commutativity implies the crossing symmetry conditions
s12 Pd (s, t) = Pd (s, t) = s23 Pd (s, t) (2.12)
where sij is the substitution exchanging the arguments xi and xj :
s12 Pd (s, t) := t 2d−3 Pd (s/t, 1/t),

s23 Pd (s, t) := s 2d−3 Pd (1/s, t/s). (2.13)
Invariance under (2.13) implies the symmetry property s13 Pd (s, t) := Pd (t, s) = Pd (s, t)
where s13 = s12 s23 s12 = s23 s12 s23 . (The Wightman function w4 of a neutral scalar field
is in fact symmetric—as a rational function—under the group S4 of permutations of the
four position variables xa , a = 1, 2, 3, 4. The normal subgroup Z2 × Z2 of S4 , however,
generated by s12 s34 and s14 s23 leaves the conformal cross ratios s and t invariant, so
that only the factor group S4 /Z2 × Z2 S3 acts effectively on Pd (s, t).) The number
of independent crossing symmetric polynomials is [d 2 /3] (the integer part of d 2 /3: 1 for
d = 2, 3 for d = 3, 5 for d = 4).
3. OPE in terms of bilocal fields
3.1. Bilocal fields as infinite series of symmetric tensor fields of a given twist
It appears convenient, at least in analysing a (truncated) 4-point function, to organize

the infinite series of integrals of local tensor fields in an OPE into a finite sum of bilocal
fields. In order to avoid purely, technical complications we shall exhibit the basic idea in
the simplest case of the product of a d-dimensional scalar field ψ(x) (in M = M4 ) with its
conjugate. We look for an expansion of the form

d−1
ψ ∗ (x1 )ψ(x2 ) = 12 + (12)d−κ Vκ (x1 , x2 ) + :ψ ∗ (x1 )ψ(x2 ):, (3.1)
κ=1
where (12) is the free massless propagator defined in (1.2), 12 is the 2-point function
12 = 0|ψ ∗ (x1 )ψ(x2 )|0 = Nψ (12)d , and Vκ (x1 , x2 ) is a bilocal conformal field of
dimension (κ, κ) which is assumed to have an expansion in a series of local (hermitian)
symmetric traceless tensor fields
O2κ,# (x; ζ ) = O2κ,µ1 ...µ# (x)ζ µ1 . . . ζ µ# , ✷ζ O2κ# (x; ζ ) = 0 (3.2)

(for ✷ζ := ∂ 2 /∂ζ 2 − ∂ 2 /∂ζ02 ) of dimension 2κ + # (i.e., of fixed twist 2κ). We shall
make use of the fact that the harmonic polynomial O2κ,#(x, ζ ) in the auxiliary variable
ζ is uniquely determined by its values on the light cone ζ 2 = 0 [2]. Whenever such an
expansion is valid, it can be written quite explicitly:2
∞
1
Vκ (x1 , x2 ) = Cκ# Kκ# (α, ρ12 ✷2 )O2κ,# (x2 + αx12 ; x12) dα, (3.3)
#=0 0
where
∞
[α(1 − α)]#+κ+n−1 (− 4z )n
Kκ# (α, z) =
B(# + κ, # + κ) n!(# + 2κ − d0 )n
n=0

0(λ + n)
(λ)n = , (3.4)
0(λ)
B is the Euler beta function, 0(µ + ν)B(µ, ν) = 0(µ)0(ν), and the d’Alembert operator
✷2 acts on x2 for fixed x12 . The normal product in (3.1) has a similar expansion
∞

:ψ ∗ (x1 )ψ(x2 ): = κ−d
ρ12 Vκ (x1 , x2 ) (3.5)
κ=d
thus involving all higher (even) twists.
2 Covariant conformal OPE were proposed in [13], soon after the pioneer work of Wilson [29]. Further
developments are reviewed in [9,22,27], and [14]; we found useful the recent presentation [10].
Remark 3.1. A field V (x1 , x2 ) is called bilocal if it commutes with any local field φ(x3 )
for spacelike x13 and x23 . We note that GCI does not imply the Huygens principle for
bilocal fields; hence, their correlation functions need not be rational. It is noteworthy that
for any even spacetime dimension D, Vd0 does have rational correlation functions as will
be demonstrated in Section 3.2 below using conservation of twist 2d0 tensors. We shall
also see (in Section 4.3) that this is not the case for Vκ if κ > d0 .
In order to verify (3.3), (3.4) we need the general conformally invariant expression for
the 2- and 3-point functions [27]
#
0|O2κ,#(x1 , ζ1 )O2κ,# (x2 , ζ2 )|0 = δ## Nκ# (12)2κ ζ1 R(x12 )ζ2
2
for ζ1,2 = 0, (3.6)
3 #
0|Vκ (x1 , x2 )O2κ,# (x3 , ζ )|0 = Aκ# (13)κ (23)κ X12 ζ , (3.7)
where we have introduced the symmetric tensor
(ζ1 x12 )(ζ2 x12 )
ρ12 ζ1 R(x12 )ζ2 = ζ1 r(x12 )ζ2 := ζ1 ζ2 − 2 r(x)2 = 1 (3.8)
ρ12
and the vector
x13 x23 3 2 ρ12
3
X12 = − of square X12 = . (3.9)
ρ13 ρ23 ρ13 ρ23
The integrodifferential operator in the right-hand side of (3.3) is characterized by the
property to transform a 2-point function of type (3.6) into a 3-point one, (3.7) (see [10]):
1 # −2κ 3 ζ )#
(X12
dα Kκ# (α, ρ12 ✷2 ) x12 R y(α) ζ ρy(α) = for ζ 2 = 0, (3.10)
(ρ13 ρ23 )κ
0
where y(α) = x23 + αx12 = αx13 + (1 − α)x23 , ρy = y 2 + i0y 0 . The above relations are
particularly easy to verify for ρ12 = 0 as the second argument of Kκ# then vanishes and we
obtain the light-cone expansion of Vκ .
Eqs. (3.3), (3.6) and (3.10) imply that the constants Aκ# in (3.7) are proportional to the
expansion coefficients Cκ# in (3.3). Furthermore, a simple analysis shows that only their
product is invariant under rescaling of O2κ,# (for fixed Vκ ).
We emphasize that the expansion (3.3) is universal: only the coefficients Cκ# depend on
the field ψ in (3.1).
It follows from (3.1) that each Vκ satisfies the symmetry condition
[Vκ (x1 , x2 )]∗ = Vκ (x2 , x1 ); (3.11)
taking into account the reality of O2κ,# we deduce
∗
Cκ# = (−1)#Cκ# . (3.12)
If ψ is hermitian then so is Vκ , and Cκ2#+1 vanish:
Cκ# = 0 for odd # if Vκ (x1 , x2 ) = Vκ (x2 , x1 ) (ψ ∗ = ψ). (3.13)
The tensor fields O2κ,# transform under elementary representations of C. It is convenient

to work with such conformal fields, since they are mutually orthogonal under vacuum
expectation values (see, e.g., [25] and references therein). It follows that each term in the
expansion (3.1) is orthogonal to all others:
0|Vκ |0 = 0 = 0|Vκ (x1 , x2 )Vλ∗ (x3 , x4 )|0 for κ = λ,

0|Vκ (x1 , x2 ):ψ(x3 )ψ ∗ (x4 ):|0 = 0, (3.14)
as the normal product (3.5) is expanded in higher twist fields.
From now on we restrict attention to the study of the OPE algebra of a hermitian GCI
scalar field of dimension d.
For a given (rational) truncated 4-point function we can compute the invariant under
rescaling products
Bκ# := Aκ2# Cκ2# (3.15)

(for O2κ2# → λO2κ2# , Aκ2# → λAκ2# , Cκ2# → λ−1 Cκ2# , Bκ# → Bκ# with A and C
introduced in (3.7), (3.3)) using the light cone expansion of Vκ (x3 , x4 ) in the 4-point
function
0|Vκ (x1 , x2 )Vκ (x3 , x4 )|0 = (13)κ (24)κ fκ (s, t). (3.16)
As a consequence of (3.13) the conformally invariant amplitude fκ is s12 symmetric:
s12 fκ (s, t) := t −κ fκ (s/t, 1/t) = fκ (s, t). (3.17)

Combining (3.3), (3.7), (3.15) and using the standard integral representation for the
hypergeometric function we find
∞

fκ (0, t) = Bκ# (1 − t)2# F (2# + κ, 2# + κ; 4# + 2κ; 1 − t). (3.18)
#=0
The symmetry condition (3.17) is reflected in a known transformation formula for the
hypergeometric function:

z
F a, a; 2a; z−1 = (1 − z)a F (a, a; 2a; z) for a = 2# + κ, z = 1 − t. (3.19)
3.2. Crossing symmetrized contribution of conserved tensors
The case κ = d0 (2.6) is of particular interest since the expansion then comprises all
conserved tensors O2d0 ,# (x; ζ ) =: T# (x, ζ ) (T2 being the stress–energy tensor). The 3-point
functions (3.7) are all harmonic in x1 and x2 ; hence, so is Vd0 :
✷1 Vd0 (x1 , x2 ) = 0 = ✷2 Vd0 (x1 , x2 ). (3.20)

Applying ✷1 to both sides of (3.16) for κ = d0 we find
4st
✷1 (13)d0 (24)d0 fd0 (s, t) = ρ24 (13)d0 (24)d0 ∆fd0 (s, t) = 0 (3.21)
ρ12 ρ14
where

∂2 ∂2 ∂2 D ∂ ∂
∆=s + t + (s + t − 1) + + . (3.22)
∂s 2 ∂t 2 ∂s∂t 2 ∂s ∂t
For D = 4 the operator ∆ has been introduced in [11] (in the context of N = 4
supersymmetric Yang–Mills theory3). The general solution of (3.21) can be written in
terms of the chiral variables u and v exploited in [10] (a similar procedure is used in
Section 7 of [11]; see also Appendix A to [20]). It is given by
g(u) − g(v)
fd0 (s, t) = for s = uv, t = (1 − u)(1 − v). (3.23)
(u − v)d0
(For euclidean xj the variables u and v are complex conjugate to each other.) For a d-
dimensional field (d > d0 ) we are looking for a solution f = fd0 of ∆f = 0 (3.21) such
that the product of t d−1 f is a polynomial in s and t of overall degree not exceeding 2d − 3
(as a consequence of (2.11)). It can be obtained by viewing Eq. (3.21) as a Cauchy problem
with initial condition (satisfying the symmetry property (3.17))
0 −1
d−d (1 − t)2ν −d0
f (0, t) = 1 + t −d0 aν = t f (0, 1/t) . (3.24)
tν
ν=0
It thus depends on d − d0 parameters a0 , a1 , . . . , ad−d0 −1 . Noting that for s = 0 = v we
have t = 1 − u, we can write the solution in the form (3.23) with
0 −1
d−d u2ν
u−d0 g(u) = f (0, 1 − u) = 1 + (1 − u)−d0 aν . (3.25)
(1 − u)ν
ν=0
Proposition 3.1. The equation

1
✷1 Kd0 ,# (α, ρ12 ✷y )O2d0,# (y; x12) dα = 0 for y = x2 + αx12 (3.26)
0
where Kκ# is given by (3.4) is necessary and sufficient for the conservation of the
(traceless) twist 2d0 tensor O2d0 # :
∂2
O2d0 ,# (y; x12) = 0. (3.27)
∂y∂x12
Sketch of proof. We shall verify that (3.4) and (3.27) imply (3.26). (The proof of the
sufficiency of (3.26) for the validity of (3.27) uses the same computation.) The statement
follows from the differentiation formula
n
✷1 ρ12 O2d0 ,# (y; x12)

∂ ∂2
= ρ12 4n n + # + d0 + α + αρ12 + α✷y O2d0 ,# (y; x12) (3.28)
∂α ∂y∂x12
3 The authors thank Emery Sokatchev for this remark.

by integrating by parts the term involving ∂/∂α.
In the case D = d = 4 (d0 = 1) of interest we find a 3-parameter family of solutions:

f1 (s, t) = a0 i0 (s, t) + a1 i1 (s, t) + a2 i2 (s, t),

−1 1−t 2 s
i0 (s, t) := 1 + t , i1 (s, t) := (1 + t − s) − 2 ,
t t

(1 − t)4 (1 − t)2 s 2 (1 − t)2

i2 (s, t) := (1 + t − 2s) − 6s + (1 + t) 1 + . (3.29)
t3 t2 t2 t
In both cases, d = 2 and d = 4 (D = 4), we can compute B1# (d) from (3.18). The result is
2a0
B1# (2) = 4# (3.30)
2#
(we have used c for a0 in [20])
1
B1# (4) = 4# 2a0 + 2#(2# + 1) 2a1 + (2# + 3)(# − 1)a2 . (3.31)
2#
The condition for the absence of a d = 2 scalar field in the OPE is given by a0 = 0
implying the vanishing of V1 for x12 → 0:
µ
a0 = 0 ⇒ V1 (x1 , x2 ) = Cx12 x12
ν
Tµν (x1 , x2 ). (3.32)
The proportionality coefficient C can be chosen in such a way that Tµν (x) := Tµν (x, x) is
the stress energy tensor of the theory, normalized by the standard Ward–Takahashi identity
(cf. [18]).
We note that the system (3.18) is overdetermined: each B1# has to satisfy two conditions
to fit the coefficients to (1 − t)2# and (1 − t)2#+1 . Thus the existence of a solution provides
a non-trivial consistency check. The result (3.30) was proven analytically using the integral
representation for the hypergeometric functions (see Appendix A of [20]); Eq. (3.31) was
derived analytically for small values of # and verified numerically for 2# 300.
Wightman (i.e., Hilbert space) positivity implies
Bκ# = Aκ2# Cκ2# = Nκ2# Cκ2#
2
0 (3.33)
(since Cκ2# is real, according to (3.12), while Nκ2# is the normalization of the positive
definite 2-point function (3.6) so it should be positive). The condition B1# > 0 is indeed
verified for non-negative aν with a positive sum. This is only a necessary condition for
Wightman positivity. A necessary and sufficient positivity condition was established in the
d = 2 case; it says that c (= a0 ) should be a positive integer (see [20, Theorem 5.1]). As the
ingredients in the derivation of this result have a more general significance we review them
in Appendix A with an eye to a possible generalization to the theory of the Lagrangian
field L(x) (of dimension d = 4).
Knowing the singular part of the OPE (3.1) allows, after crossing symmetrization to
reconstruct the complete 4-point function. The result for the d = 2 case is:

1234 = 0|φ(x1 )φ(x2 )φ(x3 )φ(x4 )|0

= (1 + s23 + s13 ) 1234 + 12 (12)(34)0|V1(x1 , x2 )V1 (x3 , x4 )|0
= 1234 + 1324 + 2314 + wt (x1 , . . . , x4 ) (3.34)
where, in the conventions of [20] (for d = 2),

c 1
j # = (j #) 2
(j #) = , j <# , (3.35)
2 4π 2 ρj #

wt (x1 , x2 , x3 , x4 ) = c (12)(23)(34)(14) + (13)(23)(24)(14)

+ (12)(13)(24)(34) (3.36)
for

0|V1 (x1 , x2 )V1 (x3 , x4 )|0 = c (13)(24) + (14)(23) . (3.37)
Remark 3.2. The factor 12 in the second term in the braces in Eq. (3.34) reflects a special
case of the following phenomenon. In general, we wish that the symmetrized contribution
of Vκ (x1 , x2 ) to the truncated 4-point function,

Fκ (x1 , x2 , x3 , x4 ) = S (12)d−κ (34)d−κ 0|Vκ (x1 , x2 )Vκ (x3 , x4 )|0 , (3.38)
where S stands for an yet unspecified symmetrization, κ = 1, . . . , d − 1, is not just
symmetric under any permutation of its arguments but that the difference
Fκ (x1 , x2 , x3 , x4 ) − (12)d−κ (34)d−κ 0|Vκ (x1 , x2 )Vκ (x3 , x4 )|0
is of order (12)d−κ (34)d−κ s (or smaller) for s → 0. This second condition forces us
to use the standard symmetrization 1234 → (1 + s23 + s13 )1234 = 1234 +
1324 + 1423 for the disconnected term in the middle part of (3.34) while applying
2 (1 + s23 + s13 ) to the second term. We shall see in Section 4.2 that there exist a
1
basis of rational harmonic functions whose symmetrized contribution satisfying the above
condition is proportional to the standard symmetrization. We further note that the function
F1 , which is rational by itself, can be viewed as providing a minimal model for the truncated
4-point function of φ.
4. General form of the 4-point function of L(x). Operator content
4.1. s-channel contributions to the 4-point function for arbitrary twists
We shall now exploit a result of Dolan and Osborn [10] which permits to write
down closed form expressions for fκ (s, t) (3.16) given the rational 4-point function for
d = 4(= D). More precisely, keeping in mind the analysis of s-channel positivity we shall
study the difference
w(x1 , x2 , x3 , x4 ) − 1234
= 1324 + 1423 + wt (x1 , x2 , x3 , x4 )

1
= (2π)−8 (ρ12 ρ23 ρ34 ρ14 )−2 P (s, t) + N 2 s 2 t 2 + t −2 (4.1)
st
with P (s, t)(= P4 (s, t)), a polynomial in s and t of overall degree 5 satisfying the crossing
symmetry conditions (2.12), (2.13) (see (4.28)). The rational function in the brackets in the
right-hand side of (4.1) can be expressed as an infinite sum of even twist contributions
∞

N 2 s 3 t t 2 + t −2 + P (s, t) = t 3 s κ−1 fκ (s, t), (4.2)
κ=1
thus extrapolating (3.16) beyond the range κ = 1, 2, 3. Albeit fκ are, in general, not rational
functions of s and t they are determined by the polynomial P and the constant N 2 . Indeed
Eq. (3.10) of [10] allows to write down the following extension of (3.23) (d0 = 1) to any
κ > 1:
1
fκ (s, t) = F (κ − 1, κ − 1; 2κ − 2; v)gκ (u)
u−v

− F (κ − 1, κ − 1; 2κ − 2; u)gκ (v) . (4.3)
We note that the normalization N contributes to κ d (= 4) only. Then it has to be taken
into account when verifying Wightman positivity condition (3.33). The hypergeometric
functions F (a, a; 2a; v), a = 1, 2, . . . , are expressed as linear combinations of log(1 − v)
with rational in v coefficients:
∞ n−1
v 1 1
F (1, 1; 2; v) = = ln ,
n v 1−v
n=1
∞

6nv n−1 6 2−v 1

F (2, 2; 4; v) = = 2 ln −2 , etc.
(n + 1)(n + 2) v v 1−v
n=1
The OPE of the bilocal field Vκ (x1 , x2 ) in terms of twist 2κ rank 2# symmetric
traceless tensors corresponds, according to (3.18), to an expansion of gκ (u) in terms of
hypergeometric functions of the same type:
∞

gκ (u) = ufκ (0, 1 − u) = Bκ# u2#+1 F (2# + κ, 2# + κ; 4# + κ; u). (4.4)
#=0
(Eq. (4.4) is obtained from (4.3) just using gκ (0) = 0, F (a, a; 2a; 0) = 1.) The functions
fκ (0, t) can be determined recursively from the relations

κ−1
1−κ −3
fκ (0, t) = lim s t P (s, t) − ν−1
s fν (s, t) ,
s→0
ν=1
f1 (0, t) = t −3 P (0, t) (4.5)
and (4.3) which imply that the operator ĥ,

−3 1 u v
ĥt P (s, t) := P (0, 1 − u) − P (0, 1 − v) = f1 (s, t)
u − v (1 − u)3 (1 − v)3
(4.6)
defines a (rational) harmonic projection of the product t −3 P (s, t). In fact, the function
(13)(24)f1(s, t) is harmonic in both x12 and x34 for fixed x23 while its difference with
(13)(24)t −3P (s, t) vanishes on either light cone (ρ12 = 0 and ρ34 = 0) since t −3 P (0, t) =
f1 (0, t) according to (4.5). Thus Eq. (4.3) can be viewed as an extension of the notion of
harmonic projection to arbitrary twists.
4.2. A basis of crossing symmetrized conformal harmonic functions
We shall now display a basis jν (s, t) of rational solutions of the conformal Laplace
equation (3.21),
∆jν (s, t) = 0, ν = 0, 1, 2

∂2 ∂2 ∂2 ∂ ∂
∆ = s 2 + t 2 + (s + t − 1) +2 +2 (4.7)
∂s ∂t ∂s∂t ∂s ∂t
that are symmetric under s12
t −1 jν (s/t, 1/t) = jν (s, t) (4.8)

and are eigenfunctions of the operator
ĥ(1 + s23 + s13 ). (4.9)

We shall verify that such a basis is given by
jα (s, t) = iα (s, t), α = 0, 1,

j2 (s, t) = i1 (s, t) + i2 (s, t),

t 3 j2 (s, t) = 1 + t 3 (1 + s − t)2 − s − 3s(1 − t)2 , (4.10)
where iν are defined in (3.29). It will become clear on the way that i2 is not an eigenvector
of the operator (4.9) as j1 and j2 correspond to different eigenvalues. Set indeed

Iν = (1 + s23 + s13 ) t 3 iν (s, t) , ν = 0, 1, 2 (4.11)
we find
I0 (s, t) = s 2 (1 + s) + t 2 (1 + t) + s 2 t 2 (s + t),
I1 (s, t) = s(1 + s)(1 − s)2 + t (1 + t)(1 − t)2

+ st (s − t) s 2 − t 2 − 2Q1 , Q1 := 1 + s 2 + t 2 ,

I2 (s, t) = (1 + s)(1 − s)2 2 − 3s + 2s 2 + (1 + t)(1 − t)2 3 − 3t + 2t 2 − 2

+ st 4Q1 − (s + t) 5s 2 − 8st + 5t 2 . (4.12)
The simplest way to compute the harmonic projection ĥ(t −3 Iν (s, t)) consists in looking at
the “initial conditions” for s = 0. We see that while Iα (0, t) = t 3 iα (0, t) for α = 0, 1, we
have

I2 (0, t) = 2(1 − t)4 (1 + t) + t (1 − t)2 (1 + t) = t 3 2i2 (0, t) + i1 (0, t) . (4.13)
It follows that i2 is not an eigenfunction of the operator (4.9) but j2 (4.10) is one (with
eigenvalue 2). This suggests introducing a new basis of symmetrized twist two polynomial
factors in the truncated 4-point function:

Jα (s, t) = Iα (s, t) = (1 + s23 + s13 ) t 3 jα (s, t) , α = 0, 1,
1 1
J2 (s, t) = (1 + s23 + s13 ) t 3 j2 (s, t) = I1 (s, t) + I2 (s, t)
2 2

= 1 + t 3 (1 + s − t)2 − s − 3s(1 − t) + s 3 (1 + t − s)2 − t , (4.14)
the Jν are distinguished by the properties

ĥ t −3 Jν (s, t) = jν (s, t), ν = 0, 1, 2,

J0 (s, t) − t 3 j0 (s, t) = s 2 1 + t 3 + s 3 1 + t 2 ,

J1 (s, t) − t 3 j1 (s, t) = s(1 − t) 1 − t 3 − s 2 1 + t 3 − s 3 (1 + t)2 + s 4 (1 + t),

J2 (s, t) − t 3 j2 (s, t) = s 3 1 + t + t 2 − 2s 4 (1 + t) + s 5 . (4.15)
Comparing (4.15) with (4.5) we deduce that the differences J1 − t 3 j1 , J0 − t 3 j0 , and

J2 − t 3 j2 involve twist 4 and higher, twist 6 and higher, and twist 8 and higher, respectively.
The expressions (3.29) for f1 and the corresponding crossing symmetric amplitude F1 can
be rewritten in the basis (4.10), (4.14) with the result
f1 (s, t) = a0 j0 (s, t) + a12j1 (s, t) + a2 j2 (s, t), a12 = a1 − a2 ,

1
F1 (x1 , x2 , x3 , x4 ) = (12)2 (23)2 (34)2(14)2
st
× a0 J0 (s, t) + a12 J1 (s, t) + a2 J2 (s, t) . (4.16)
We note that the truncated 4-point function of the (vacuum) Maxwell Lagrangian,
1
L0 (x) = − :Fµν (x)F µν (x): (4.17)
4
(or a sum of mutually commuting expressions of this type) is a special case of (4.16). Here
Fµν is a free electromagnetic field with 2-point function
0|Fµ1 ν1 (x1 )Fµ2 ν2 (x2 )|0

= 4Dµ1 ν1 µ2 ν2 (x12)

= ∂µ1 (∂µ2 ην1 ν2 − ∂ν2 ην1 µ2 ) − ∂ν1 (∂µ2 ηµ1 ν2 − ∂ν2 ηµ1 µ2 ) (12); (4.18)
we have
4π 2 Dµ1 ν1 µ2 ν2 (x) = Rµ1 µ2 (x)Rν1 ν2 (x) − Rµ1 ν2 (x)Rν1 µ2 (x), (4.19)

the symmetric tensor R(x) being defined in (3.8). The OPE of the product of two L0 ’s has
the form
8
L0 (x1 )L0 (x2 ) = 120 + 3
T (x1 , x2 ; x12) + :L0 (x1 )L0 (x2 ):, (4.20)
π 2 ρ12
where 120 = 3(πρ12)−4 , and T (x1 , x2 ; x12) (a multiple of V1 ) is the harmonic bilocal
field
1 µ
T (x1 , x2 ; x12) = :F σ τ (x1 )Fσ τ (x2 ):x122
− x12 :F σ µ (x1 )Fσ ν (x2 ):x12
ν
4
1 2
= x12 (r12 )µν (r12 )σ τ :Fµσ (x1 )Fντ (x2 ):. (4.21)
4
In verifying ✷1 T (x1 , x2 ; x12) = 0 one should use both the tracelessness of T , i.e.,
✷y T (x1 , x2 ; y) = 0 and the two Maxwell equations,
dF (x) = 0 = d ∗ F (x), (4.22)
where F (x) is the 2-form (1.5). The second expression for T (x1 , x2 ; x12) (4.21) makes
obvious its conformal invariance. The 4-point function of L0 is computed from (3.17)–
(3.19) using repeatedly the triple-product formula (of [22])
1 x13 x12
r(x12 )r(x23)r(x13 ) = r X23 , X231
= − (4.23)
ρ13 ρ12
(see Appendix B of [20]). The contribution of T to the 4-point function is given by the
solution (3.29) of (3.21):
π 4 0|T (x1 , x2 ; x12)T (x3 , x4 ; x34)|0 = (13)(24)f1(s, t)
for a0 = 0, a1 = a2 , (4.24)
it follows that w4t from (4.1) assumes the form

t
wL 0
(x1 , x2 , x3 , x4 ) = (1 + s12 + s23 )c0 tr D12 D 23 D34 D 14
c0
= (ρ12 ρ23 ρ34 ρ14 )−2 (st)−1 J2 (s, t), (4.25)
8π 8
D12 := Dµ1 ν1 µ2 ν2 (x12 ), D 23 := D µ2 ν2 µ3 ν3 (x23 ), etc., (4.26)
corresponding to a1 = a2 = 25 c0 , a0 = a12 = 0. In fact, a finite sum of expressions of type
(4.17) will give rise to a discrete subset of such values with a positive integer c0 .
4.3. General form of w4t . Higher twist contributions
The most general GCI truncated 4-point function of a d = 4 scalar field contains, in
addition to F1 (4.16), two more terms corresponding to crossing symmetrized twist 4
contributions. They can be written as stQi (s, t), i = 1, 2, where Qi are crossing symmetric
polynomials of degree 2:
Q1 = 1 + s 2 + t 2 , Q2 = s + t + st,
t Qj (s/t, 1/t) = Qj (s, t) = s 2 Qj (1/s, t/s),
2
j = 1, 2. (4.27)
Thus, the polynomial P (s, t) of Eq. (4.1) can be presented in the form:
P (s, t) = a0 J0 (s, t) + a12J1 (s, t) + a2 J2 (s, t)
+ st[b1 Q1 (s, t) + b2 Q2 (s, t)]. (4.28)
Remark 4.1. The above basis of crossing symmetric polynomials Jν , Qj and the Pν , Qj
basis of [20], where

P0 (s, t) = 1 + s 5 + t 5 , P1 (s, t) = s 1 + s 3 + t 1 + t 3 + st s 3 + t 3 ,

P2 (s, t) = s 2 (1 + s) + t 2 (1 + t) + s 2 t 2 (s + t) = I0 (s, t) , (4.29)
are related by
J0 = P2 , J1 = P1 − P2 − 2stQ1 ,
J2 = P0 − 2P1 + P2 + stQ1 . (4.30)
We shall now apply the procedure outlined in Section 4.1 to compute the twist 4 and 6
contributions in (4.28). It follows from (4.4), (4.15) and (4.28) that

1 1 1
f2 (0, 1 − u) = a12 1 − u + + (b 1 − a 12 ) 1 + + b2
(1 − u)3 (1 − u)2 1−u
∞
g2 (u)
≡ = B2# u2# F (2# + 2, 2# + 2; 4# + 4; u) (4.31)
u
#=0
yielding

4# + 1 −1
B2# = #(# + 1)(2# + 1)(2# + 3)a12 + 2(# + 1)(2# + 1)b1 + b2 .
2#
(4.32)
We see, in particular, that the absence of a scalar field of dimension 4 in the OPE of two
L’s implies
B20 = 2b1 + b2 = 0 (⇒ 123 = 0). (4.33)
The difference between P (s, t) (4.28) and the twist 2 (Eq. (4.16)) and twist 4 part,
computed from (4.3), defines t 3 f3 (0, t) as the coefficient to the s 2 term:

P (s, t) − t 3 [f1 (s, t) + sf2 (s, t)] = s 2 t 3 f3 (0, t) + O s 3 . (4.34)
A computer aided calculation (using M APLE) gives
∞
g3 (u)
= B3# u2# F (2# + 3, 2# + 3; 4# + 6; u)
u
#=0
1 1 b2 − b1 1 b1 + 2b2 1 a12 + 2a0
= a12 + a0 + + +
2 2 1−u 2 (1 − u)2 2 (1 − u)3

1 1 2 ln(1 − u)
− (2b1 + b2 ) + − (2b1 + b2 ) , (4.35)
2 u u2 u3
with

1 4# + 3 −1
B3# = (# + 1)(# + 2)(2# + 1)(2# + 3)(2a0 + a12 )
2 2# + 1

+ 2(# + 1)(2# + 3)(b1 + 2b2 ) − 2b1 − b2 . (4.36)
Remarkably, g3 (u) is rational precisely when the condition (4.33), reflecting the electric–
magnetic duality, is satisfied.
5. Implications for the gauge field Lagrangian
5.1. Restrictions from OPE, Hodge duality, and Wightman positivity
As discussed in the introduction the gauge field Lagrangian (1.3) is characterized by the
absence of scalar fields of dimensions 2 and 4 in the OPE of two L’s. According to (3.32)
and (4.33) this implies
a0 = 0 = 2b1 + b2 . (5.1)
This allows to write the polynomial P (s, t) in (4.1) as
P (s, t) = a J1 (s, t) + aJ2 (s, t) + b[Q1 (s, t) − 2Q2 (s, t)] (5.2)
where we have set a2 = a, a12 = a, b1 = b(= −b2 /2); the difference
Q1 (s, t) − 2Q2 (s, t) = (1 − s − t)2 − 4st (5.3)
is characterized by having a (second order) zero at t = 1 for s = 0 (and being negative for
euclidean xij ).
The restrictions on the three remaining constants, a, a , and b, coming from the
positivity condition (3.33) for Bkl given by (3.31), (4.32) and (4.36) are
a + a > 0, a, a 0; (5.4)
(# + 1)(2# + 1)a + b 03 for # > 0, (# + 2)(2# + 1)a 6b for # 0. The last two
inequalities imply
1
−3a b a . (5.5)
3
In particular, if a = 0 it would follow that also b = 0 and we would end up with
the truncated 4-point function that is a multiple of the one of the free electromagnetic
Lagrangian (4.17).
Remark 5.1. If we allow for a positive a0 in the 4-point function (including, say, a scalar
field contribution in the Lagrangian) then, according to (4.36), the restriction (5.5) would
be replaced by a more general one
1
−3a b (2a0 + a ) (5.6)
3
which leaves room for a (non-zero) positive b even for a = 0. A numerical analysis
indicates that the most general GCI 4-point function involves a non-trivial open domain
in the 5-dimensional projective space of the parameters aν , bj and the 2-point function
normalization, N 2 , defined up to a common positive factor, in which Wightman positivity
is verified for all twists.
5.2. Difficulties in exploiting the compositeness of L
It is instructive to understand the restrictions, coming from (5.1), within an axiomatic

treatment of non-abelian gauge field theory. At the same time we shall try to answer the
question: can we extract more information from the expression (1.3) for L in terms of
Fµν than just saying that the OPE of L(x1 )L(x2 ) contains neither L nor a scalar field of
dimension 2?
In (perturbative) quantum electrodynamics amplitudes with an odd number of photon
external legs (and no external charged particles) vanish because of charge conjugation
invariance (Furry’s theorem). The general conformally invariant 3-point function of a
Maxwell field Fµν (x), on the other hand, violates local commutativity of Bose fields and
should hence be also set equal to zero (without having to assume any discrete symmetry).
This last argument fails for a non-abelian gauge field
1
F (x, ω) := ωµν Fµν a
(x)ta , [ta , tb ] = ifabc tc ωµν = −ωνµ (5.7)
2
where ta are orthonormal hermitian matrices generating the defining representation of
a (compact, semi-simple) non-abelian gauge group G, fabc is the totally antisymmetric
tensor of (real) structure constants of the Lie algebra G of G (fabc = εabc , the Levi-Civita
tensor for G = su(2), ta = 12 σa , a, b, c = 1, 2, 3), ω is a constant skew-symmetric tensor
(or differential form) introduced for notational convenience. While the general (gauge and)
conformally invariant 2-point function of F a (x, ω) = 12 ωµν Fµνa (x) coincides with the free
Maxwell one, (4.18),
0|F a (x1 , ω1 )F b (x2 , ω2 )|0 = NF δ ab D(x12 ; ω1 , ω2 ) (5.8)

where we have introduced a contracted form of (4.19),
µ ν µ ν
4π 2 D(x; ω1 , ω2 ) = Rµ1 µ2 (x)Rν1 ν2 (x)ω1 1 1 ω2 2 2 , (5.9)
there is a 2-parameter family of local gauge and conformally invariant 3-point functions:

W abc (x1 , ω1 , x2 , ω2 , x3 , ω3 ) = f abc N1 W (1) + N2 W (2) , (5.10)
W (1) (x1 , ω1 , x2 , ω2 , x3 , ω3 )
1
= X23µ R
1 ν1 µ3
(x13 )X31µ2
R (x23 ) − X23µ
2 ν2 ν3
1
R
1 ν1 µ2
3
(x12 )X12µ R (x23)
3 ν2 ν3
µ1 ν1 µ2 ν2 µ3 ν3
+ X31µ2 Rµ1 ν2 (x12 )X12µ3 Rν1 ν3 (x13 ) ω1 ω2 ω3 ,
2 3
(5.11)
W (2) (x1 , ω1 , x2 , ω2 , x3 , ω3 )
µ ν µ ν µ ν
= Rµ1 µ2 (x12 )Rν1 µ3 (x13)Rν2 ν3 (x23 )ω1 1 1 ω2 2 2 ω3 3 3 , (5.12)
where Rµν and X12 3 are defined in (3.8) and (3.9).
The expression (5.10) can, sure, be used to produce a non-zero 3-point function of
L (1.3). Thus, the second condition (5.1) is not, at first sight, an automatic consequence
of (1.3), conformal invariance and locality, Hodge duality providing an independent
restriction on correlation functions. It turns out, however, that the combination of (5.8)
and (5.10) implies the existence of a field I in the OPE of F (x1 ) ⊗ F (x2 ) transforming
under an indecomposable representation of C (which would severely complicate the use of

conformal invariance).
This statement follows from the observation that the 3-point function (5.10) does not
satisfy the free Maxwell equation d ∗ F = 0 for any non-zero choice of N1 and N2 while
the 2-point function does:
0|d ∗ F (x1 ) ⊗ F (x2 )|0 = 0,

0|d ∗ F (x1 ) ⊗ F (x2 ) ⊗ F (x3 )|0 = 0. (5.13)
It follows that the OPE of F (x2 ) ⊗ F (x3 ) should involve a local field I that is not
orthogonal to d ∗ F . I cannot be a derivative of F because of the first equation (5.13). On the
other hand, I cannot be an elementary conformal field transforming under an inequivalent
representation of C since then it should again be orthogonal to F , and hence to d ∗ F thus
violating the second equation (5.13).
Thus the vanishing of odd point correlation functions of L is a natural property of the
Lagrangian (1.3) if the local observable algebra is spanned by elementary conformal fields
(and their derivatives).
Remark 5.2. Let Aaµ (x), a = 1, 2, 3 be three commuting purely longitudinal gauge
potentials, i.e., generalized free fields such that
1 δ ab
0|Aaµ (x1 )Abν (x2 )|0 = δ ab rµν (x12)(12) = Rµν (x12 ),
2 8π 2
∂µ Aaν (x) = ∂ν Aaµ (x). (5.14)
Then both the 2-point function (5.8) and the 3-point function (5.12) are reproduced by the
corresponding su(2) Yang–Mills curvature tensor
a
Fµν (x) = ∂µ Aaν (x) − ∂ν Aaµ (x) − gεabc :Abµ (x)Acν (x):
= −gεabc :Abµ (x)Acν (x):. (5.15)
This example is already excluded, however, by our requirement that no d = 2 scalar field
appears in the OPE of two L’s (neither does F (5.15) satisfy the Yang–Mills equation with
connection A). Even for the more general d = 4 composite field
2 µ
Lξ η (x) = ξ : Aaµ (x)Aµ
a (x) : − η:Aµ (x)Ab (x)Aν (x)Aa (x):
a b ν
(5.16)
(which includes (1.3) with F given by (5.15) for ξ = η = g 2 /4) we find that the leading
term in the OPE of two L’s,

Lξ η (x1 )Lξ η (x2 ) ≈ 4 14ξ 2 − 16ξ η + 11η2 (12)3 Aaµ (x1 )r µν (x12 )Aaν (x2 ) , (5.17)
involves a scalar field of dimension d = 2 (with the same coefficient as the 2-point function
of Lξ η —which is non-zero for any not simultaneously zero real ξ, η).
5.3. Concluding remarks
We have presented in the preceding sections two intertwined developments: (i) a sys-
tematic study of the theory of a GCI local scalar field ψ of any (integer) dimension d in
terms of bilocal fields Vκ (x1 , x2 ) of dimension (κ, κ) appearing in the OPE ψ ∗ (x1 )ψ(x2 )
(3.1), (3.5); (ii) first steps in an attempt to construct in a non-perturbative, axiomatic ap-
proach a conformally invariant fixed point of a gauge field theory, formulated entirely in
terms of gauge invariant local observables of dimension 4: the Lagrangian density L(x) and
the stress–energy tensor Tµν (x) or, rather, its polarized bilocal counterpart that determines
V1 (x1 , x2 ) according to (3.32).
(i) The use of bilocal fields simplifies substantially the analysis of the operator content
of GCI correlation functions. In particular, V1 (x1 , x2 ), defined as the (infinite)
sum of twist 2 conserved symmetric tensor fields, satisfies as a consequence the
d’Alembert equation (3.20). This allows to compute the contribution of the correlator
0|V1 (x1 , x2 )V1 (x3 , x4 )|0 to the (truncated) 4-point function w4t of L. (Its expression
for the d = 2 neutral scalar field φ has been computed in [20].) A minimal model
for the connected part of the 4-point function of a general neutral scalar field of
dimension 4 is given by F1 (4.16) which is determined by its harmonic projection.
A general procedure is outlined—based on the Dolan–Osborn formula (4.3)—for
computing the expectation values (3.16) of Vκ . The case κ = 1 is distinguished by
the fact that the conformal harmonic function f1 (s, t) (and hence, the corresponding
symmetrized contribution F1 ) is rational. For κ > 1 fκ (s, t) are linear combinations
(with rational function coefficients) of
log(1 − u) and log(1 − v) for s = uv, s + 1 − t = u + v. (5.18)
We have displayed these functions (taking into account also the additional term
b1 Q1 + b2 Q2 in (4.28)) and the associated (infinite) OPE expansions in terms of
symmetric traceless tensor fields of twist 2κ for κ = 1, 2, 3 (see Eqs. (3.31), (4.31) and
(4.32), (4.35) and (4.36), respectively). The same procedure applies to higher κ as well,
when fκ also depend on the sum of products of 2-point functions 1324 + 1423
(which introduces one more parameter, the normalization N of the 2-point function).
The resulting expressions (together with other computer aided results—concerning 6-
point functions) will be presented in a forthcoming publication, in collaboration with
K.-H. Rehren.
(ii) Conditions (5.1) exclude contributions of scalar fields of dimensions 2 and 4 in
the OPE of L(x1 )L(x2 ) leaving us with a 3-parameter family of truncated 4-point
functions. We assert that L(x) then has the properties of the Lagrangian density of
a gauge field curvature (without matter fields). The study of Wightman positivity for
twists 4 and 6 contributions leads to a rather strong constraint (5.5) for the remaining
parameters. Should, in particular, the analysis of the 2n-point function of L for n 3
yield the constraint a = 0 Eq. (5.5) would also imply b = 0 and leave us with a
multiple of the 4-point function of the Lagrangian of a free abelian gauge field.
This would confirm the general belief (see, e.g., [12,30]) that a (pure) non-abelian
gauge theory necessarily involves a mass gap, thus violating conformal invariance. By
contrast, if we do not impose (5.1), i.e., if we allow for the presence of (at least) a scalar
field in the Lagrangian, a full account of Wightman positivity for the 4-point function
appears to allow for an open set of the 5-dimensional (projective) parameter space.
The evidence that we are displaying a gauge invariant 4-point function in a (non-
abelian) gauge theory is rather indirect. One verifies on a case by case basis that any
other renormalizable Lagrangian would involve fields of d < 4 in the OPE and that is
excluded by condition (5.1). The difficulty in identifying the theory in terms of basic (gauge
a (briefly reviewed in Section 5.2) lies in the fact that the model
dependent) fields like Fµν
we are trying to construct is necessarily non-perturbative (if it exists at all). From this point
of view our model is not incompatible with the N = 4 supersymmetric Yang–Mills theory
(see, e.g., [1,3,4,11] and references therein) for sufficiently large value of the coupling
constant g, such that the anomalous dimension of the Konishi field (which appears also in
the OPE of two fields of the supermultiplet of the stress–energy tensor) is a positive even
integer.
Acknowledgements
The authors thank Dirk Kreimer for a stimulating discussion and Karl-Henning Rehren
for an enlightening correspondence. N.N. and I.T. acknowledge the hospitality of the
Erwin Schrödinger International Institute for Mathematical Physics (ESI) as well as partial
support by the Bulgarian National Council for Scientific Research under contract F-828.
The research of Ya.S. was supported in part by INFN, by the EC contracts HPRN-CT-
2000-00122 and -00148, by the INTAS contract 99-0-590 and by the MURST-COFIN
contract 2001-025492. All three authors acknowledge partial support by the Research
Training Network within the Framework Programme 5 of the European Commission under
contract HPRN-CT-2002-00325 and by a NATO linkage grant PST.CLG.978785. I.T.
thanks l’Institut des Hautes Etudes Scientifiques (Bures-sur-Yvette), the Theory Division
of CERN and Section de Mathématiques, Université de Genève for hospitality during the
final stage of this work.
Appendix A. Vacuum representation of the algebra generated by the harmonic

bilocal field V1 . The case d = 2
There are two main ingredients in the proof of the central result, Theorem 5.1 of [20]
to be reviewed in the two sections of this appendix. The first is a (computer aided)
study of the (5- and) 6-point function of the basic field φ(x) of dimension 2, combined
with the expansion (3.3) (for κ = 1) and the conservation law for the twist two fields
T# (x, ζ ) (# = 2, 4, . . .). Here we sum up a modified version of the argument which uses
Proposition 3.1. The (technical) difficulty of such an analysis increases drastically with
increasing the dimension of the underlying scalar field and it has not been completed for
d = 4. The second ingredient is quite general: it uses the discrete mode expansion of V1
with respect to the conformal Hamiltonian which can be carried out for any d.
A.1. Analysis of 5- and 6-point functions for d = 2
The general GCI and crossing symmetric 5-point function of φ(x) (for dφ = 2) involves
two independent terms: the sum of twelve 1-loop graphs
c
w(1) = (1σ2 )σ2 σ3 . . . σn−1 σn (1σn ) for n = 5 (A.1)
2
σ ∈Perm(2,...,n)
where (ij ) is defined in (2.9) and

σi σj = min(σi , σj ), max(σi , σj ) , (A.2)
and a sum, w(2) , of 10 products of 7 factors each:

w(2) = λ ρij ikkj. (A.3)
1i<j 5 1k5
i=k=j
A rather nasty (computer aided) calculation shows that the 5-point function of three φ’s
and a V1 satisfies the d’Alembert equation in the last two arguments
✷j 0|φ(x1 )φ(x2 )φ(x3 )V1 (x4 , x5 )|0 = 0 for j = 4, 5 (A.4)
iff λ = 0. In view of Proposition 3.1 (A.4) is equivalent to demanding the infinite set of
conservation laws
∂2
µ 0|φ(x1 )φ(x2 )φ(x3 )T2# (x4 , ζ )|0 = 0. (A.5)
∂x4 ∂ζµ
Similarly, only the crossing symmetric sum of ( 12 × 5! = 60) 1-loop graph contributions
to the truncated 6-point function is consistent with T2# conservation. The 1-loop expression
for the 6-point function allows to prove that the limit

V (x1 , x2 ) = lim (2π)4 ρ13 ρ23 (φ(x1 )φ(x2 )φ(x3 )
ρ13 →0
ρ23 →0

− (13)φ(x2 ) − (23)φ(x1) − 123) (A.6)
exists, does not depend on x3 and defines a harmonic in each argument bilocal field
V1 (x1 , x2 ); moreover, the truncated n-point functions of φ will be given by (A.1) for all n
(see [20, Proposition 2.3]).
Remark A.1. The above sketched analysis for the 5-point function has been carried out
in the d = 4 case, too, with the following results. There are 37 independent GCI and
crossing symmetric 5-point functions of L (compared to 2 for d = 2(!)). After imposing
conservation of the twist 2 tensors T2# (x, ζ ) there remain only 8, just 1 among them
actually contributing to the twist 2 part of the OPE. The latter is non-zero only if the
condition (5.1) is violated and then it yields a2 = 0.
The analysis of the general 6-point function has to deal with 31990 (instead of 8 for
d = 2) independent structures (most of them consisting of 6! = 720 terms each).
A.2. Analytic compact picture fields. Conformal Hamiltonian. Mode expansions
Compactified Minkowski space

= S1 × S3 /Z2
M (A.7)
(see [8]) has a convenient realization in terms of (euclidean) complex 4-vectors [26]:

M = zµ = e2πiζ uµ , µ = 1, 2, 3, 4; ζ ∈ R, u ∈ S3 , i.e., u2 := u2 + u24 = 1 . (A.8)

Real Minkowski space M is mapped on an open dense subset of M:
1 − x2
M ! x 0 , x → z = ω−1 (x)x, z4 = ,
2ω(x)
1 + x2
ω(x) = − ix 0 . (A.9)
2
can be obtained by adding to the image of M the 3-cone at infinity:
M

K∞ = z ∈ M; 1 + 2z4 + z2 = 2(u4 + cos 2πζ )e2πiζ = 0 . (A.10)
Note that Eq. (A.9) can be interpreted as the Cayley map from the Lie algebra u(2)( R4 )
to the group U (2), cf. [28].
We note that the map (A.9) extends to complex arguments x → x + iy and is regular in
the forward tube (2.1) which is mapped (for D = 4) on the domain

D
2 2
D 2
T+ = z ∈ C : z < 1, 2|z| = 2 2
|zµ | < 1 + z ,
2
(A.11)
µ=1
where
2 2 −1
z2 = 1 − y 0 + ix 0 + (x + iy)2 1 + y 0 − ix 0 + (x + iy)2 .
The maximal compact subgroup K(= K(4)) acts by linear homogeneous transformations
on zµ : SU(2) × SU(2) is represented by real SO(4) rotations of the vector (zµ ) while the
U (1) factor acts by phase transformations: zµ → eiα zµ . Thus the point z = 0 ∈ T+ (the
image of (x + iy) = (i, 0 ) ∈ T+ ) is left invariant by K.
A scalar M-space field φM (x) of dimension d is related to its z-picture counterpart,
φ(z) (for z = z(x) given in (A.9)) by

φM (x) = [2πω(x)]−d φ z(x) . (A.12)
The numerical factor 2π in (A.12) is chosen for convenience so that the free massless
scalar propagator assumes a simple z-picture form
1
(12) = 2
. (A.13)
z12
A z-picture scalar field φ(z) of dimension d transforms in such a way that the
form φ(z)(dz2 )d/2 remains invariant. (The transformation properties of the Lagrangian
L(z), for d = 4, can be read off, alternatively, from the invariance of the volume form
L(z) dz1 ∧ dz2 ∧ dz3 ∧ dz4 , cf. (1.6).) This implies, in particular, that correlation functions
are invariant under complex euclidean transformations. The conformal Hamiltonian H
(that generates translations in the conformal time ζ ) acts on φ(z) and on the bilocal fields
Vκ (z1 , z2 ) according to the law

∂
[H, φ(z)] = d + z φ(z),
∂z

∂ ∂
[H, Vκ (z1 , z2 )] = 2κ + z1 + z2 Vκ (z1 , z2 ). (A.14)
∂z1 ∂z2
Here Vκ are again related to φ by an expansion of type (3.1) with (12) substituted by its
z-picture counterpart (A.13).
The mode expansion of V1 (z1 , z2 ) is, in particular, an expansion in homogeneous
harmonic polynomials:

V1 (z1 , z2 ) = Vm−1n−1 (z1 , z2 ),
n,m∈Z
Vm−1n−1 (z1 , z2 )
Vm−1n−1 (λ1 z1 , λ2 z2 ) = for λ1 , λ2 ∈ C∗ , (A.15)
λm
1 λ2
n
∂ 2
∆1 Vmn (z1 , z2 ) = 0 = ∆2 Vmn (z1 , z2 ) for ∆a = µ . (A.16)
µ=1
∂za
For a hermitian scalar field φ the modes Vmn have the following simple conjugation
property. For m, n 0
µ ...µ ν ...ν
V−m−1,−n−1 (z, w) = V−m−1,−n−1
1 m 1 n
zµ1 . . . zµm wν1 . . . wνn , (A.17)
µ ...µ ν ...ν
1 m 1 n
where V−m−1−n−1 are symmetric traceless tensors in µ1 , . . . , µm and in ν1 , . . . , νn
(separately) which are mapped under hermitian conjugation into tensors with the same
properties,
µ1 ...µm ν1 ...νn ∗ µ1 ...µm ν1 ...νn
V−m−1,−n−1 = Vm+1n+1 (A.18)
so that
1 zµm wν1
µ ...µ ν ...νn zµ1 wν
Vm+1n+1 (z, w) = 1 m 1
Vm+1n+1 ···
· · · 2n . (A.19)
z2 w 2 2z2
z w 2 w
We are interested in the vacuum representation of the algebra of V1 modes for which
Vmn |0 = 0 = 0|V−m−n if either m 0 or n 0. (A.20)
For d = 2 the algebra generated by Vnm (z, w) coincides with a central extension

sp(∞, R) of the infinite symplectic Lie algebra. This is particularly simple to see when
the (invariant under rescaling) structure constant is a natural number:
121323
c := 8 =N ∈N (A.21)
(123)2
and

N

V1 (z1 , z2 ) = :ϕj (z1 )ϕj (z2 ): V1 (z, z) = 2φ(z) , (A.22)
j =1
where ϕj (z) are free (commuting for different j ) massless scalar fields. The modes ϕn (z)
(n ∈ Z) of each ϕ(z) generate the infinite Heisenberg algebra:
ϕn (z) = e−2πi(n+1)ζ ϕn (u), [ϕn (u), ϕm (v)] = δn,−m ε(n)C|n|−1
1
(uv) (A.23)
where ε(n) = sign n(= n/|n|, ε(0) = 0) Cn1 are the Gegenbauer polynomials generated by
the 2-point function of ϕ:
∞
sin 2πnα 1 1 1 n 1
1
Cn−1 (cos 2πα) = , 2
= 1 − 2xz + z 2
= z Cn (x),
sin 2πα z12 z12 z12 n=0
z22
z= , x = uv. (A.24)
z12
It is well known that the quadratic combination of the generators of a Heisenberg algebra
give rise to a symplectic Lie algebra. The central extension comes, as usual, from the
normal products.

The algebra sp(∞, R) of Vnm (u, v) has a simple diagonal subalgebra, generated by
vnm := Vnm (u, u), u ∈ S3 ,
[vn1 m1 , vn2 m2 ] = cn1 m1 (δn1 ,−n2 δm1 ,−m2 + δn1 ,−m2 δm1 ,−n2 )
+ n1 (δn1 ,−n2 vm1 m2 + δn1 ,−m2 vm1 n2 )
+ m1 (δm1 ,−n2 vn1 m2 + δm1 ,−m2 vn1 n2 ). (A.25)
The proof of Theorem 5.1 of [20] is now based on the construction of a sequence ∆n |,
n = 1, 2, . . . of vectors (Lemma 5.2)

v11 v12 . . . v1n

1 v21 v22 . . . v2n
∆n | = 0| .. .. .. ..
n! . . . .

vn1 vn2 . . . vnn
such that ∆n |∆n = (n + 1)!c(c − 1) · · · (c − n + 1), (A.26)
implying c ∈ N for unitary vacuum representations of sp(∞, R). Then one deduces that
Wightman positivity implies that V1 has the form (A.22) for some N .
The difficulty in extending this analysis to the case d = 4 again lies in the necessity of
having the 6-point function of L in order to be able to compute the structure constants of
the infinite Lie algebra generated by the modes of V1 (z1 , z2 ).
References
[1] G. Arutyunov, B. Eden, A.C. Petkou, E. Sokatchev, Exceptional non-renormalization properties and OPE
analysis of chiral 4-point functions in N = 4 SYM4 , Nucl. Phys. B 620 (2002) 380–404, hep-th/0103230.
[2] V. Bargmann, I.T. Todorov, Spaces of analytic functions on the complex cone as carriers for the symmetric
tensor representations of SO(n), J. Math. Phys. 18 (1977) 1141–1148.
[3] N. Beisert, C. Kristensen, M. Staudacher, The dilation operator of conformal N = 4 super-Yang–Mills
theory, hep-th/0303060.
[4] M. Bianchi, S. Kovacs, G. Rossi, Ya.S. Stanev, Properties of Konishi multiplet in N = 4 SYM theory,
JHEP 0105 (2001) 042, hep-th/0104016.
[5] N.N. Bogolubov, A.A. Logunov, A.I. Oksak, I.T. Todorov, General Principles of Quantum Field Theory,
Kluwer, Dordrecht, 1990 (Original Russian edition: Nauka, Moscow, 1987).
[6] E. Cunningham, The principle of relativity in electrodynamics and an extension thereof, Proc. London Math.
Soc. 8 (1910) 77–98.
[7] P. Di Francesco, P. Mathieu, D. Senechal, Conformal Field Theories, Springer, Berlin, 1996.
[8] P.A.M. Dirac, Wave equations in conformal space, Ann. Math. 37 (1936) 429–442.
[9] V.K. Dobrev, G. Mack, V.B. Petkova, S.G. Petrova, I.T. Todorov, Harmonic Analysis of the n-dimensional
Lorentz Group and its Application to Conformal Quantum Field Theory, Springer, Berlin, 1977.
[10] F.A. Dolan, H. Osborn, Conformal four point functions and operator product expansion, Nucl. Phys. B 599
(2001) 459–496, hep-th/0011040.
[11] B. Eden, A.C. Petkou, C. Schubert, E. Sokatchev, Partial nonrenormalization of the stress-tensor four-point
function in N = 4 SYM and AdS/CFT, Nucl. Phys. B 607 (2001) 191–212, hep-th/0009106.
[12] L.D. Faddeev, Mass in quantum Yang–Mills theory (comment on a Clay millenium problem), Bull. Braz.
Math. Soc. 33 (2002) 201–212.
[13] S. Ferrara, R. Gatto, A. Grillo, G. Parisi, The shadow operator formalism for conformal algebra vacuum
expectation values and operator products, Nuovo Cimento Lett. 4 (1972) 115–120.
[14] E.S. Fradkin, M.Ya. Palchik, New developments in d-dimensional conformal quantum field theory, Phys.
Rep. 300 (1998) 1–112;
E.S. Fradkin, M.Ya. Palchik, Conformal Quantum Field Theory in D Dimensions, Kluwer, Dordrecht, 1996.
[15] P. Furlan, G.M. Sotkov, I.T. Todorov, Two-dimensional conformal quantum field theory, Riv. Nuovo
Cimento 12 (6) (1989) 1–202.
[16] R. Haag, Local Quantum Physics, Fields, Particles, Algebras, Springer, Berlin, 1992.
[17] G. Mack, All unitary representations of the conformal group SU(2, 2) with positive energy, Commun. Math.
Phys. 55 (1977) 1–28.
[18] G. Mack, K. Symanzik, Currents, stress tensor and generalized unitarity in conformal invariant quantum
field theory, Commun. Math. Phys. 27 (1972) 247–281.
[19] N.M. Nikolov, I.T. Todorov, Rationality of conformally invariant local correlation functions on compactified
Minkowski space, Commun. Math. Phys. 218 (2001) 417–436, hep-th/0009004.
[20] N.M. Nikolov, Ya.S. Stanev, I.T. Todorov, Four-dimensional CFT models with rational correlation functions,
J. Phys. A: Math. Gen. 35 (2002) 2985–3007, hep-th/0110230.
[21] N.M. Nikolov, Ya.S. Stanev, I.T. Todorov, Global conformal invariance and bilocal fields with rational
correlation functions, in: Proceedings of the Third International Sakharov Conference on Physics, Moscow,
Vol. 2, 2002, pp. 256–268, hep-th/0211106.
[22] H. Osborn, A. Petkou, Implications of conformal invariance for field theories in general dimensions, Ann.
Phys. (N.Y.) 231 (1994) 311–362.
[23] I.E. Segal, Causally oriented manifolds and groups, Bull. Amer. Math. Soc. 77 (1971) 958–959.
[24] R.F. Streater, A.S. Wightman, PCT, Spin and Statistics, and All That, Benjamin, 1964, Princeton Univ. Press,
Princeton, NJ, 2000.
[25] I.T. Todorov, Local field representations of the conformal group and their applications, in: L. Streit (Ed.),
Mathematics and Physics, in: Lectures on Recent Results, Vol. 1, World Scientific, Singapore, 1985, pp. 195–
338.
[26] I.T. Todorov, Infinite-dimensional Lie algebras in conformal QFT models, in: A.O. Barut, H.-D. Doebner
(Eds.), Conformal Groups and Related Symmetries, Physical Results and Mathematical Background, in:
Lecture Notes in Physics, Vol. 261, Springer, Berlin, 1986, pp. 387–443.
[27] I.T. Todorov, M.C. Mintchev, V.B. Petkova, Conformal Invariance in Quantum Field Theory, Scuola
Normale Superiore, Pisa, 1978.
[28] A. Uhlmann, Remarks on the future tube, Acta Phys. Pol. 24 (1963) 293;
A. Uhlmann, The closure of Minkowski space, Acta Phys. Pol. 24 (1963) 295–296.
[29] K.G. Wilson, Operator product expansions and anomalous dimensions in the Thirring model, Phys. Rev. D 2
(1970) 1473–1477.
[30] E. Witten, Physical laws and the quest of mathematical understanding, Bull. Amer. Math. Soc. 40 (2003)
21–29.
The algebraic Bethe ansatz for the Izergin–Korepin

model with open boundary conditions
Guang-Liang Li
Department of Applied Physics, Xi’an Jiaotong University, Xi’an 710049, China
Kang-Jie Shi, Rui-Hong Yue

Institute of Modern Physics, Northwest University, Xi’an 710069, China
Received 25 March 2003; received in revised form 2 June 2003; accepted 1 August 2003
Abstract
We present the procedure of exactly solving the Izergin–Korepin model with open boundary
conditions by using the algebraic Bethe ansatz, which include constructing the multi-particle state
and achieving the eigenvalue of the transfer matrix and corresponding Bethe equations. We give a
proof about our conclusions on the multi-particle state based on an assumption. When the model
is Uq (su(2)) quantum invariant, our results agree with that obtained by analytic Bethe ansatz
method.
PACS: 75.10.J; 05.20; 05.30
Keywords: Izergin–Korepin model; Algebraic Bethe ansatz; Open boundary
1. Introduction
The Bethe ansatz solutions to a integrable model will make it possible to study the
thermodynamic properties of the model, such as correlation functions [1], specific heat,
magnetic susceptibility [2,3], and finite size effects [4–6]. At same time, in the procedure of
solving the models, it will also help us to better understand their underlying mathematical
structures.
E-mail address: leegl@mail.xjtu.edu.cn (G.-L. Li).
doi:10.1016/j.nuclphysb.2003.08.001
402 G.-L. Li et al. / Nuclear Physics B 670 [FS] (2003) 401–438
There are several ways to solve a model. One of the powerful mathematical tool
is the quantum inverse scattering method (QISM) [7–11], namely, the algebraic Bethe
ansatz method. Besides describing the spectra of quantum integrable systems, the Bethe
ansatz is also used to construct exact and manageable expressions for correlation
functions [1]. The QISM was first developed for the system with periodic boundary
conditions. It was later generalized to systems with open boundary conditions [11].
Since Sklyanin’s work [11], many works have been done on the integrable models with
open boundary conditions [12–22]. Solving a model with open boundary conditions will
enable us to investigate the boundary effect of the system. It may also be helpful for
us to study the impurities problems which have attracted considerable interest recently
[23–34].
The algebraic Bethe ansatz have been applied to the six-vertex type model successfully.
It is also have been used in other type models with periodic boundary conditions [35–
40] or with open boundary conditions [41,42], such as Izergin–Korepin (IK) model [43].
The R matrix of the IK model is the simplest example of an R matrix of the twisted type
and the model with open boundary conditions has the important physical applications.
For example, it can be related to the loop models [21] and self-avoiding walks at a
boundary [44]. The IK model was solved by using other methods under the periodic
boundary [40,45] or open boundary conditions [21,46–48]. Employing the algebraic Bethe
ansatz method, for the IK model with periodic boundary conditions, Tarasov proposed
the construction of two-particle state for the model and argued that his conclusion can
be generalized to any n-particle state [35]. Relying on the previous work by Tarasov
and Martins [35,36], Fan solved the IK model with open boundary conditions [41]. In
Fan’s work, he give the expression of two-particle state and a conjecture for the n-
particle state. He also checked the Bethe equation at the case n = 1. For the IK model,
it is interesting to find the explicit expression of n-particle state and verify the Bethe
equations in the case n 2, which will also help us to learn about the feasibility of
applying the algebraic Bethe ansatz method to the model. In this paper, our aim is to do
this work.
The paper is organized as following. In Section 2 we introduce Izergin–Korepin model
and the diagonal K± matrices of reflection equation which determines the nontrivial
boundary terms in the Hamiltonian. In Section 3, by direct calculation, we present our
results for the one-particle state, two-particle state and three-particle state and generalize
our results to the case of n-particle state. We then prove our conclusions on n-particle
state based on an assumption. When the model is Uq (su(2)) quantum invariant, our
conclusions recover that obtained by analytic Bethe ansatz method [46]. For the non-
quantum-group invariant cases, there are a little difference between our results and that
proposed by Yung and Batchelor in Ref. [21]. The summary and some discussions of our
main results are included in Section 4. In the appendix, some necessary relations and proofs
are provided.
2. The vertex model and integrable boundary conditions
The R matrix for the Izergin–Korepin model [43] is

G.-L. Li et al. / Nuclear Physics B 670 [FS] (2003) 401–438 403
 
c 0 0 0 0 0 0 0 0
0 b 0 e 0 0 0 0 0
 
0 0 d 0 g 0 f 0 0
 
 0 ē 0 b 0 0 0 0 0
 
R(u) =  0 0 ḡ 0 a 0 g 0 0 (1)
 
0 0 0 0 0 b 0 e 0
 
0 0 f¯ 0 ḡ 0 d 0 0
 
0 0 0 0 0 ē 0 b 0
0 0 0 0 0 0 0 0 c
with
a(u) = sinh(u − 3q) − sinh(5q) + sinh(3q) + sinh(q),

b(u) = sinh(u − 3q) + sinh(3q),
c(u) = sinh(u − 5q) + sinh(q),
d(u) = sinh(u − q) + sinh(q),
e(u) = −2e−u/2 sinh(2q) cosh(u/2 − 3q),
ē(u) = −2eu/2 sinh(2q) cosh(u/2 − 3q),
f (u) = −2e−u+2q sinh(q) sinh(2q) − e−q sinh(4q),
f¯(u) = 2eu−2q sinh(q) sinh(2q) − eq sinh(4q),
g(u) = 2e−u/2+2q sinh(u/2) sinh(2q),
ḡ(u) = −2eu/2−2q sinh(u/2) sinh(2q). (2)
The R matrix satisfies the following properties
regularity: R12 (0) = ρ(0)1/2P12 ,

t t
unitarity: 1 2
R12 (u)R12 (−u) = ρ(u),
t1 t2
PT-symmetry: P12 R12 (u)P12 = R12 (u),
1 1
crossing-symmetry: t2
R12 (u) = V R12 (−u − η)V −1 . (3)
Here P is the exchange operator defined by P(x ⊗ y) = y ⊗ x, ti denotes transposition

1 2
in the space i, V = V ⊗ 1, V = 1 ⊗ V , η is the crossing
√ parameter and V determines
the crossing matrix M ≡ V t V = M t with η = −6q − −1π and M = diag(e2q , 1, e−2q ),
ρ(u) = ((sinh(q) − sinh(5q + u))(sinh(q) − sinh(5q − u))).
The R matrix also fulfil Yang–Baxter equation (YBE) [49]
R12 (u − v)R13 (u)R23 (v) = R23 (v)R13 (u)R12 (u − v), (4)
where R12 (u), R13 (u) and R23 (u) act on C 3 ⊗C 3 ⊗C 3 , with R12 (u) = R(u)⊗1, R23 (u) =
1 ⊗ R(u), etc.
For an N × N square lattice, if we can find K± (u) which satisfy the following reflection
equations [11,50]
1 2
t1 t2
R12 (u − v)K − (u)R12 (u + v)K − (v)
2 1
t t
= K − (v)R12 (u + v)K − (u)R12
1 2
(u − v), (5)
1 1 1 2
R12 (−u + v)K t+1 (u)M −1 R12
t1 t2
(−u − v − 2η)M K t+2 (v)
2 1 1 1
= K t+2 (v)MR12 (−u − v − 2η)M −1 K t+1 (u)R12
t1 t2
(−u + v), (6)
where Eq. (5) is called reflection equation and Eq. (6) is called dual reflection equation,
1 2
K ± (u) = K± (u) ⊗ 1, K ± (u) = 1 ⊗ K± (u), then the transfer matrix t (u) defined as
t (u) = tr K+ (u)U (u) (7)
can constitute a one-parameter commutative family [t (u), t (v)] = 0. Here
U (u) = T (u)K− (u)T −1 (−u), (8)
T (u) = R01 (u)R02 · · · R0N (u), (9)
the space V0 is usually called the auxiliary space, the space V1 ⊗ V2 · · · ⊗ VN is called the
quantum space. The corresponding integrable open chain Hamiltonian takes the form
0

N−1
11 tr K + (0)HN,0
H= Hk,k+1 + K− (0) + , (10)
2 tr K+ (0)
k=1
where Hk,k+1 = Pk,k+1 Rkk+1 (u)|u=0 . From Eqs. (5) and (6), we can see that, given a
solution K− (u) of Eq. (5), the matrix
K+ (u) = K−
t
(−u − η)M (11)
satisfies Eq. (6). The general solution to Eq. (5) of IK model have been obtained in
Ref. [51], here we will choose the diagonal ones. Denote

K− (u) = diag K− (u)1 , K− (u)2 , K− (u)3 ,

K+ (u) = diag K+ (u)1 , K+ (u)2 , K+ (u)3 ,
by Eq. (11), we have
(i):
K− (u) = 1, K+ (u) = M, (12)
(ii):

K− (u) = diag e−u c cosh(q) + sinh(u − 2q) , c cosh(q + u) − sinh(2q),

eu c cosh(q) + sinh(u − 2q) ,

K+ (u) = diag eu−4q c cosh(q) + sinh(u − 4q) , c cosh(u − 7q) + sinh(2q),

e−u+4q c cosh(q) + sinh(u − 4q) (13)
with c2 = c 2 = −1. In case (ii), we omit a factor −1 in K+ (u) for simplicity.
3. The algebraic Bethe ansatz
3.1. The vacuum state and commutation relations
Firstly, we write the double-monodromy matrix as

A(u) B1 (u) F (u)
U (u) = C1 (u) D1 (u) B2 (u) . (14)
G(u) C2 (u) D2 (u)
With the help of Eqs. (4), (5), we can prove that (14) also satisfy the reflecting equation (5)
1 2 2 1
t1 t2 t1 t2
R12 (u − v)U (u)R12 (u + v)U (v) = U (v)R12 (u + v)U (u)R12 (u − v), (15)
1 2
where U (u) = U (u) ⊗ 1, U (u) = 1 ⊗ U (u). Now we introduce the vacuum state,
⊗N

|0 = (1, 0, 0)t , (16)
where t denotes the transposition. Acting the double-row monodromy matrix (14) on the
vacuum state, we can find
Ci (u)|0 = 0, Bi (u)|0 = 0 (i = 1, 2)
G(u)|0 = 0, F (u)|0 = 0. (17)
Considering the definition of U (u) Eq. (8), we have
A(u)|0 = T (u)11 K− (u)1 T −1 (−u)11 |0 + T (u)12 K− (u)2 T −1 (−u)21 |0

+ T (u)13 K − (u)3 T −1 (−u)31 |0, (18)
D1 (u)|0 = T (u)21 K− (u)1 T −1 (−u)12 |0 + T (u)22 K− (u)2 T −1 (−u)22 |0
+ T (u)23 K − (u)2 T −1 (−u)32 |0, (19)
D2 (u)|0 = T (u)31 K− (u)1 T −1 (−u)13 |0 + T (u)32 K− (u)2 T −1 (−u)23 |0
+ T (u)33 K − (u)3 T −1 (−u)33 |0. (20)
In above equations, the first term of Eq. (19) and the previous two terms of Eq. (20)
cannot be calculated directly but it can be worked out by using the following method.
Taking v = −u in the Yang–Baxter equation, we can get
T2−1 (−u)R12 (2u)T1 (u) = T1 (u)R12 (2u)T2−1 (−u). (21)

Taking special indices in this relation and applying both sides of this relation to the vacuum
state, we find:
ē(2u) −1
T (u)21 T −1 (−u)12 |0 = T (−u)11 T (u)11 − T (u)22 T −1 (−u)22 |0,
c(2u)
ē(2u)c(2u) − f¯(2u)e(2u)
T (u)32 T −1 (−u)23 |0 =
a(2u)c(2u) − ē(2u)e(2u)

× T −1 (−u)22 T (u)22 − T (u)33 T −1 (−u)33 |0,
¯
f (2u) −1
T (u)31 T −1 (−u)13 |0 = T (−u)11 T (u)11
c(2u)
a(2u)f¯(2u) − ē(2u)ē(2u)
− T (u)33 T −1 (−u)33
a(2u)c(2u) − ē(2u)e(2u)

ē(2u) ē(2u)c(2u) − f¯(2u)e(2u) −1
− ( )T (−u)2 T (u)2 |0.
2 2
c(2u) a(2u)c(2u) − ē(2u)e(2u)
(22)
After defining two new operators
1 (u) = D1 (u) − f1 (u)A(u),
D (23)
2 (u) = D2 (u) − f2 (u)A(u) − f3 (u)D
D 1 (u), (24)
and substituting Eq. (22) to Eqs. (18)–(20), we obtain
A(u)|0 = K− (u)1 c(u)2N ρ(u)−N |0 = ω1 (u)|0, (25)

1 (u)|0 = K− (u)2 − f1 (u)K− (u)1 b(u)2N ρ(u)−N |0 = ω2 (u)|0,
D (26)

2 (u)|0 = K− (u)3 − f3 (u)K− (u)2 − f4 (u)K− (u)1 d(u)2N ρ(u)−N |0
D
= ω3 (u)|0, (27)
where
ē(2u) f¯(2u)
f1 (u) = , f2 (u) = ,
c(2u) c(2u)
ē(2u)c(2u) − f¯(2u)e(2u)
f3 (u) = ,
a(2u)c(2u) − ē(2u)e(2u)
a(2u)f¯(2u) − ē(2u)ē(2u)
f4 (u) = .
a(2u)c(2u) − ē(2u)e(2u)
Rewriting Eq. (15) in component form
R12 (u− )ac11ca22 U (u)cd11 R21 (u+ )db11 cd22 U (v)db22

= U (v)ac22 R12 (u+ )ac11dc22 U (u)cd11 R21 (u− )db11 db22 , (28)
where the repeated indices sum over 1 to 3, u− = u − v, u+ = u + v, we can obtain the
following fundamental commutation relations
g(u− )b(2v) g(u+ ) 1 (v)

B1 (u)B1 (v) − F (u)A(v) + F (u)D
d(u− )c(2v) b(u+ )

cosh( u2− + q) g(−u− )b(2u) g(u+ )
= B1 (v)B1 (u) − F (v)A(u) + F (v)
D1 (u) ,
cosh( u2− − q) d(−u− )c(2u) b(u+ )
(29)
1 1 1
A(u)B1 (v) = a1 (u, v)B1 (v)A(u) + a2 (u, v)B1 (u)A(v) + a3 (u, v)B1 (u)D1 (v)
+ a41 (u, v)F (u)C1 (v) + a51 (u, v)F (u)C2 (v) + a61 (u, v)F (v)C1 (u),
(30)
1 (u)B1 (v) = a12 (u, v)B1 (v)D
D 1 (u) + a22 (u, v)B1 (u)A(v) + a32 (u, v)B1 (u)D
1 (v)
+ a42 (u, v)B2 (u)A(v) + a52 (u, v)B2 (u)D1 (v) + a62 (u, v)F (u)C1 (v)
+ a72 (u, v)F (u)C2 (v) + a82 (u, v)F (v)C1 (u) + a92 (u, v)F (v)C2 (u),
(31)
D 2 (u) + a23 (u, v)B1 (u)A(v) + a33 (u, v)B1 (u)D
2 (u)B1 (v) = a13 (u, v)B1 (v)D 1 (v)
+ a 3 (u, v)B2 (u)A(v) + a 3 (u, v)B2 (u)D1 (v) + a 3 (u, v)F (u)C1 (v)
4 5 6
+ a73 (u, v)F (u)C2 (v) + a83 (u, v)F (v)C1 (u) + a93 (u, v)F (v)C2 (u),
(32)
A(u)F (v) = b1 (u, v)F (v)A(u) + b2 (u, v)F (u)A(v) + b3 (u, v)F (u)D1 (v)
1 1 1
+ b41 (u, v)F (u)D2 (v) + b51 (u, v)B1 (u)B1 (v) + b61 (u, v)B1 (u)B2 (v),
(33)
1 (u)F (v) = b12 (u, v)F (v)D
D 1 (u) + b22(u, v)F (u)A(v) + b32 (u, v)F (u)D 1 (v)
+ b42 (u, v)F (u)D2 (v) + b52(u, v)B1 (u)B1 (v) + b62 (u, v)B1 (u)B2 (v)
+ b72 (u, v)B2 (u)B1 (v) + b82 (u, v)B2 (u)B2 (v), (34)
2 (u)F (v) = b3 (u, v)F (v)D
D 2 (u) + b3(u, v)F (u)A(v) + b3 (u, v)F (u)D
1 (v)
1 2 3
+ b4 (u, v)F (u)D
3 2 (v) + b (u, v)B1 (u)B1 (v) + b (u, v)B1 (u)B2 (v)
3 3
5 6
+ b7 (u, v)B2 (u)B1 (v) + b83 (u, v)B2 (u)B2 (v),
3
(35)
C1 (u)B1 (v) = c11 (u, v)B1 (v)C1 (u) + c21 (u, v)B1 (v)C2 (u) + c31 (u, v)B1 (u)C2 (u)
+ c41 (u, v)B2 (u)C2 (v) + c51 (u, v)F (v)G(u) + c61 (u, v)A(v)A(u)
+ c71 (u, v)A(v)D1 (u) + c81 (u, v)A(u)A(v) + c91 (u, v)A(u)D1 (v)
+ c10
1 1 (u)A(v) + c11
(u, v)D 1 1 (u)D
(u, v)D 1 (v), (36)
C2 (u)B1 (v) = c12 (u, v)B1 (v)C1 (u) + c22 (u, v)B1 (v)C2 (u) + c32 (u, v)B1 (u)C2 (u)
+ c42 (u, v)B2 (u)C2 (v) + c52 (u, v)F (v)G(u) + c62 (u, v)A(v)A(u)
+ c72 (u, v)A(v)D1 (u) + c82 (u, v)A(v)D 2 (u) + c92 (u, v)A(u)A(v)
+ c10
2
(u, v)A(u)D1 (v) + c11
2 1 (u)A(v)
(u, v)D
1 (u)D
+ c12 (u, v)D
2 1 (v) + c13 (u, v)D
2 2 (u)A(v)
+ c14
2 2 (u)D
(u, v)D 1 (v), (37)
B1 (u)B2 (v) = d11 (u, v)B2 (v)B1 (u) + d21 (u, v)B1 (v)B1 (u) + d31 (u, v)F (v)A(u)
+ d41 (u, v)F (v)D1 (u) + d51 (u, v)F (u)A(v) + d61 (u, v)F (u)D 1 (v)
2 (v),
+ d71 (u, v)F (u)D (38)
F (u)B1 (v) = e11 (u, v)B1 (v)F (u) + e21 (u, v)F (v)B1 (u) + e31 (u, v)F (v)B2 (u), (39)
B1 (u)F (v) = e12 (u, v)F (v)B1 (u) + e22 (u, v)F (v)B2 (u)
+ e32 (u, v)B1 (v)F (u) + e42 (u, v)B2 (v)F (u), (40)
F (u)B2 (v) = e13 (u, v)B2 (v)F (u) + e23 (u, v)B1 (v)F (u)
+ e33 (u, v)F (v)B1 (u) + e43 (u, v)F (v)B2 (u), (41)
B2 (u)F (v) = e14 (u, v)F (v)B2 (u) + e24 (u, v)B1 (v)F (u) + e34 (u, v)B2 (v)F (u), (42)
F (u)F (v) = F (v)F (u), (43)
where the coefficients of Eqs. (30)–(42) are defined in Appendix A.
3.2. The one-particle state
Considering Eqs. (23), (24), the transfer matrix (7) can be rewritten as
1 (u) + w3 (u)D
t (u) = w1 (u)A(u) + w2 (u)D 2 (u), (44)
with
w1 (u) = K+ (u)1 + f1 (u)K+ (u)2 + f2 (u)K+ (u)3 ,
w2 (u) = K+ (u)2 + f3 (u)K+ (u)3 , w3 (u) = K+ (u)3 . (45)
The one-particle state can be constructed as
|Φ1 (v1 ) = B1 (v1 )|0. (46)
Using the commutation relations (30)–(32), we can find
A(u)|Φ1 (v1 ) = ω1 (u)a11 (u, v1 )|Φ1 (v1 ) + ω1 (v1 )a21 (u, v1 )B1 (u)|0
+ ω2 (v1 )a31 (u, v1 )B1 (u)|0, (47)
1 (u)|Φ1 (v1 ) = ω2 (u)a12 (u, v1 )|Φ1 (v1 ) + ω1 (v1 )a22 (u, v1 )B1 (u)|0
D
+ ω2 (v1 )a32 (u, v1 )B1 (u)|0 + ω1 (v1 )a42 (u, v1 )B2 (u)|0
+ ω2 (v1 )a52 (u, v1 )B2 (u)|0, (48)
2 (u)|Φ1 (v1 ) = ω3 (u)a 3 (u, v1 )|Φ1 (v1 ) + ω1 (v1 )a 3 (u, v1 )B1 (u)|0
D 1 2
+ ω2 (v1 )a33 (u, v1 )B1 (u)|0 + ω1 (v1 )a43 (u, v1 )B2 (u)|0
+ ω2 (v1 )a53 (u, v1 )B2 (u)|0. (49)
Then

t (u)|Φ1 (v1 ) = w1 (u)ω1 (u)a11 (u, v1 ) + w2 (u)ω2 (u)a12 (u, v1 )

+ w3 (u)ω3 (u)a13 (u, v1 ) |Φ1 (v1 ) (50)
and all the unwanted terms vanish if the rapidity satisfy the Bethe ansatz equation
ω1 (v1 )
= β(u, v1 ), (51)
ω2 (v1 )
where
e−2q sinh(v − 2q)
β(u, v) = (52)
sinh(v)
corresponding to Eq. (12) and

e−v sinh(v − 2q) cosh( u+v − 3q) − c sinh( u−v − 2q)
β(u, v) = 2 2 (53)

2 − q) − c sinh( 2 − 4q)
sinh(v) cosh( u−v u+v
√
for Eq. (13) with c = ± −1.
3.3. The two-particle state
Considering the commutation relation (29), we define

Φ2 (v1 , v2 ) = B1 (v1 )B1 (v2 ) + F (v1 )A(v2 )ga,2
2 1 (v2 )gd,2
(v1 , v2 ) + F (v1 )D 2
(v1 , v2 )
(54)
with
cosh( v1 −v
2 + q)
2
g(v1 − v2 )b(2v2 )
α(v1 , v2 ) = , 2
ga,2 (v1 , v2 ) = − ,
cosh( v1 −v
2 − q)
2 d(v1 − v2 )c(2v2 )
g(v1 + v2 )
2
gd,2 (v1 , v2 ) = . (55)
b(v1 + v2 )
Then, the two particles state can be constructed as
|Φ2 (v1 , v2 ) = Φ2 (v1 , v2 )|0 (56)
1 (u),
and it has the property |Φ2 (v1 , v2 ) = α(v1 , v2 )|Φ2 (v2 , v1 ). Acting operators A(u), D
2 (u) on it, respectively, and using the corresponding commutation relations Eqs. (29)–
D
(38), we finally obtain the following simplified results
A(u)|Φ2 (v1 , v2 )

2
= ω1 (u) a11 (u, vi )|Φ2 (v1 , v2 )
i=1

+ B1 (u)|Φ1 (v2 ) ω1 (v1 )a21(u, v1 )a11 (v1 , v2 ) + ω2 (v1 )a31 (u, v1 )a12 (v1 , v2 )

+ α(v1 , v2 )B1 (u)|Φ1 (v1 ) ω1 (v2 )a21 (u, v2 )a11 (v2 , v1 )

+ ω2 (v2 )a31(u, v2 )a12 (v2 , v1 )

+ F (u)|0 ω1 (v1 )ω1 (v2 )H1A (u, v1 , v2 ) + ω2 (v1 )ω1 (v2 )H2A (u, v1 , v2 )

+ ω1 (v1 )ω2 (v2 )H3A (u, v1 , v2 ) + ω2 (v1 )ω2 (v2 )H4A (u, v1 , v2 ) ,
(57)
1 (u)|Φ2 (v1 , v2 )
D

2
= ω2 (u) a12 (u, vi )|Φ2 (v1 , v2 )
i=1



+ ω2 (v2 )a32(u, v2 )a12 (v2 , v1 )


+α(v1 , v2 )B2 (u)|Φ1 (v1 ) ω1 (v2 )a42 (u, v2 )a11 (v2 , v1 )

+ ω2 (v2 )a52 (u, v2 )a12 (v2 , v1 )

+ F (u)|0 ω1 (v1 )ω1 (v2 )H1D1 (u, v1 , v2 ) + ω2 (v1 )ω1 (v2 )H2D1 (u, v1 , v2 )

+ ω1 (v1 )ω2 (v2 )H3D1 (u, v1 , v2 ) + ω2 (v1 )ω2 (v2 )H4D1 (u, v1 , v2 ) ,
(58)
2 (u)|Φ2 (v1 , v2 )
D

2
= ω3 (u) a13 (u, vi )|Φ2 (v1 , v2 )
i=1



+ ω2 (v2 )a33(u, v2 )a12 (v2 , v1 )



+ ω2 (v2 )a53(u, v2 )a12 (v2 , v1 )

+ F (u)|0 ω1 (v1 )ω1 (v2 )H1D2 (u, v1 , v2 ) + ω2 (v1 )ω1 (v2 )H2D2 (u, v1 , v2 )

+ ω1 (v1 )ω2 (v2 )H3D2 (u, v1 , v2 ) + ω2 (v1 )ω2 (v2 )H4D2 (u, v1 , v2 ) .
(59)
The coefficients Hi (u, v1 , v2 ), i = 1, 2, 3, 4, x = A, D1 , D2 are defined as
x

H1A (u, v1 , v2 ) = a41 (u, v1 ) c61 (v1 , v2 ) + c81 (v1 , v2 )

+ a51 (u, v1 ) c62 (v1 , v2 ) + c92 (v1 , v2 )
+ b21 (u, v1 )ga,2
2
(v1 , v2 ) + α(v1 , u)a11 (u, v1 )a21 (u, v2 )ga,2
2
(u, v1 ),

H2 (u, v1 , v2 ) = a4 (u, v1 ) c7 (v1 , v2 ) + c10 (v1 , v2 )
A 1 1 1

+ a51 (u, v1 ) c72 (v1 , v2 ) + c11
2
(v1 , v2 )
+ b31 (u, v1 )ga,2
2
(v1 , v2 ) + α(v1 , u)a11 (u, v1 )a21 (u, v2 )gd,2
2
(u, v1 ),
H3A (u, v1 , v2 ) = a41 (u, v1 )c91 (v1 , v2 ) + a51 (u, v1 )c10
2
(v1 , v2 ) + b21 (u, v1 )gd,2
2
(v1 , v2 )
+ α(v1 , u)a11 (u, v1 )a31 (u, v2 )ga,2

2
(u, v1 ),
H4A (u, v1 , v2 ) = a41 (u, v1 )c11
1
(v1 , v2 ) + a51 (u, v1 )c12
2
(v1 , v2 ) + b31 (u, v1 )gd,2
2
(v1 , v2 )
+ α(v1 , u)a11 (u, v1 )a31 (u, v2 )gd,2
2
(u, v1 ), (60)

H1D1 (u, v1 , v2 ) = a62 (u, v1 ) c61 (v1 , v2 ) + c81 (v1 , v2 )

+ a72 (u, v1 ) c62 (v1 , v2 ) + c92 (v1 , v2 )
+ b22 (u, v1 )ga,2
2
(v1 , v2 ) + α(v1 , u)a12 (u, v1 )a22 (u, v2 )ga,2
2
(u, v1 )
+ a12 (u, v1 )a42 (u, v2 )d31 (v1 , u),

H2D1 (u, v1 , v2 ) = a62 (u, v1 ) c71 (v1 , v2 ) + c10
1
(v1 , v2 )

+ a72 (u, v1 ) c72 (v1 , v2 ) + c112
(v1 , v2 )
+ b32 (u, v1 )ga,2
2
(v1 , v2 ) + α(v1 , u)a12 (u, v1 )a22 (u, v2 )gd,2
2
(u, v1 )
+ a12 (u, v1 )a42 (u, v2 )d41 (v1 , u),
D
H3 1 (u, v1 , v2 ) = a62 (u, v1 )c91 (v1 , v2 ) + a72 (u, v1 )c10
2
(v1 , v2 ) + b22 (u, v1 )gd,2
2
(v1 , v2 )
+ α(v1 , u)a12(u, v1 )a32 (u, v2 )ga,2
2
(u, v1 )
+ a12 (u, v1 )a52 (u, v2 )d31 (v1 , u),
D
H4 1 (u, v1 , v2 ) = a62 (u, v1 )c11
1
(v1 , v2 ) + a72 (u, v1 )c12
2
(v1 , v2 ) + b32 (u, v1 )gd,2
2
(v1 , v2 )
+ α(v1 , u)a12(u, v1 )a32 (u, v2 )gd,2
2
(u, v1 )
+ a12 (u, v1 )a52 (u, v2 )d41 (v1 , u), (61)

H1D2 (u, v1 , v2 ) = a63 (u, v1 ) c61 (v1 , v2 ) + c81 (v1 , v2 )
+ a73 (u, v1 )(c62 (v1 , v2 ) + c92 (v1 , v2 ))
+ b23 (u, v1 )ga,2
2
(v1 , v2 ) + α(v1 , u)a13 (u, v1 )a23 (u, v2 )ga,2
2
(u, v1 )
+ a13 (u, v1 )a43 (u, v2 )d31 (v1 , u),

H2D2 (u, v1 , v2 ) = a63 (u, v1 ) c71 (v1 , v2 ) + c10
1
(v1 , v2 )

+ a73 (u, v1 ) c72 (v1 , v2 ) + c112
(v1 , v2 )
+ b33 (u, v1 )ga,2
2
(v1 , v2 ) + α(v1 , u)a13 (u, v1 )a23 (u, v2 )gd,2
2
(u, v1 )
+ a13 (u, v1 )a43 (u, v2 )d41 (v1 , u),
D
H3 2 (u, v1 , v2 ) = a63 (u, v1 )c91 (v1 , v2 ) + a73 (u, v1 )c10
2
(v1 , v2 )
+ b23 (u, v1 )gd,2
2
(v1 , v2 )
+ α(v1 , u)a13(u, v1 )a33 (u, v2 )ga,2
2
(u, v1 )
+ a13 (u, v1 )a52 (u, v2 )d31 (v1 , u),
H4D2 (u, v1 , v2 ) = a63 (u, v1 )c11
1
(v1 , v2 ) + a73 (u, v1 )c12
2
(v1 , v2 ) + b33 (u, v1 )gd,2
2
(v1 , v2 )
+ α(v1 , u)a13(u, v1 )a33 (u, v2 )gd,2

2
(u, v1 )
+ a13 (u, v1 )a53 (u, v2 )d41 (v1 , u) (62)
and we can find that Hix (u, v1 , v2 ) have the following properties
H1x (u, v1 , v2 ) = α(v1 , v2 )H1x (u, v2 , v1 ), (63)

H2x (u, v1 , v2 ) = α(v1 , v2 )H3x (u, v2 , v1 ), (64)
H4x (u, v1 , v2 ) = α(v1 , v2 )H4x (u, v2 , v1 ). (65)
Acting t (u) (44) on |Φ2 (v1 , v2 ) with the help of Eqs. (57)–(59), we get

2
2
t (u)|Φ2 (v1 , v2 ) = w1 (u)ω1 (u) a11 (u, vi ) + w2 (u)ω2 (u) a12 (u, vi )
i=1 i=1

2
+ w3 (u)ω3 (u) a13 (u, vi ) |Φ2 (v1 , v2 ) (66)
i=1
and all the unwanted terms vanish by the Bethe equations
2 2
ω1 (vi ) j =1,=i a1 (vi , vj )
= β(u, vi ) 2 (i = 1, 2). (67)
ω2 (vi ) 1
j =1,=i a1 (vi , vj )
3.4. The three-particle state
Let
Φ3 (v1 , v2 , v3 ) = B1 (v1 )Φ2 (v2 , v3 )

+ F (v1 )Φ1 (v2 ) A(v3 )ga,3
3 1 (v3 )gd,3
(v2 , v3 ) + D 3
(v2 , v3 )

+ F (v1 )Φ1 (v3 ) A(v2 )ga,2
3 1 (v2 )gd,2
(v2 , v3 ) + D 3
(v2 , v3 ) . (68)
Then the three-particle state can be defined as
|Φ3 (v1 , v2 , v3 ) = Φ3 (v1 , v2 , v3 )|0. (69)

Requiring the exchange symmetry
|Φ3 (v1 , v2 , v3 ) = α(v2 , v3 )|Φ3 (v1 , v3 , v2 ), (70)

|Φ3 (v1 , v2 , v3 ) = α(v1 , v2 )|Φ3 (v2 , v1 , v3 ) (71)
and using Eqs. (38), (30), (31), (39), (40), (41), we can get
3
ga,2 (v1 , v2 , v3 ) = a11 (v2 , v3 )ga,2
2
(v1 , v2 ),
3
ga,3 (v1 , v2 , v3 ) = α(v2 , v3 )a11 (v3 , v2 )ga,2
2
(v1 , v3 ),
3
gd,2 (v1 , v2 , v3 ) = a12 (v2 , v3 )gd,2
2
(v1 , v2 ),
3
gd,3 (v1 , v2 , v3 ) = α(v2 , v3 )a12 (v3 , v2 )gd,2
2
(v1 , v3 ). (72)
Before tackling the eigenvalue problem of three-particle state, we have to know the results
when the operators C1 (u) and C2 (u) act on the two-particle state, which will need more
extra commutation relations as listed in Appendix B. Using the commutation relations
(30)–(37), (B.1)–(B.5), and after a tedious calculation, we obtain the following simplified
forms
Cs (u)|Φ2 (v1 , v2 ) = ψ̃c(1)

s
(u, v1 , v2 ) + B1 (u)ψ̃c(2)
s
(u, v1 , v2 ) + B2 (u)ψ̃c(3)
s
(u, v1 , v2 ),
(73)
(1) (2)
where s = 1, 2, ψ̃cs (u, v1 , v2 ) is a linear combination of |Φ1 ’s, ψ̃cs (u, v1 , v2 ) and
ψ̃c(3) (l) (l)
s (u, v1 , v2 ) are coefficients. The ψ̃cs (u, v1 , v2 ) (l = 1, 2, 3) satisfy ψ̃cs (u, v1 , v2 ) =
(l) (l)
α(v1 , v2 )ψ̃cs (u, v2 , v1 ). The explicit expressions of ψ̃cs (u, v1 , v2 ) are rather long and
we omit them here for the sake of simplicity. With the aid of Eqs. (57)–(58), (73) and
commutation relations (30)–(37), (B.1)–(B.5), we finally obtain
A(u)|Φ3 (v1 , v2 , v3 )

3
= ω1 (u) a11 (u, vi )|Φ3 (v1 , v2 , v3 )
i=1

3
i−1
+ α(vj , vi )B1 (u)|Φ2 (v1 , v̌i , v3 )
i=1 j =1

3
3
× ω1 (vi )a21 (u, vi ) a11 (vi , vj ) + ω2 (vi )a31 (u, vi ) a12 (vi , vj )
j =1,=i j =1,=i

2
3
i−1 −1
j
+ α(vk , vi ) α(vl , vj )F (u)|Φ1 (v̌i , v2 , v̌j )
i=1 j =i+1 k=1 l=1=i

× ω1 (vi )ω1 (vj )a11 (vi , vm )a11 (vj , vm )H1A (u, vi , vj )
+ ω2 (vi )ω1 (vj )a12 (vi , vm )a11 (vj , vm )H2A (u, vi , vj )
+ ω1 (vi )ω2 (vj )a11 (vi , vm )a12 (vj , vm )H3A (u, vi , vj )

+ ω2 (vi )ω2 (vj )a12 (vi , vm )a12 (vj , vm )H4A (u, vi , vj ) , (74)
1 (u)|Φ3 (v1 , v2 , v3 )
D

3
= ω2 (u) a12 (u, vi )|Φ3 (v1 , v2 , v3 )
i=1

3
i−1
+ α(vj , vi )B1 (u)|Φ2 (v1 , v̌i , v3 )
i=1 j =1

3
3
j =1,=i j =1,=i

3
i−1
+ α(vj , vi )B2 (u)|Φ2 (v1 , v̌i , v3 )
i=1 j =1

3
3
j =1,=i j =1,=i

2
3
i−1 −1
j
i=1 j =i+1 k=1 l=1=i

× ω1 (vi )ω1 (vj )a11 (vi , vm )a11 (vj , vm )H1D1 (u, vi , vj )
+ ω2 (vi )ω1 (vj )a12 (vi , vm )a11 (vj , vm )H2D1 (u, vi , vj )
D
+ ω2 (vi )ω2 (vj )a12 (vi , vm )a12 (vj , vm )H4 1 (u, vi , vj ) , (75)
2 (u)|Φ3 (v1 , v2 , v3 )
D

3
= ω3 (u) a13 (u, vi )|Φ3 (v1 , v2 , v3 )
i=1

3
i−1
+ α(vj , vi )B1 (u)|Φ2 (v1 , v̌i , v3 )
i=1 j =1

3
3
j =1,=i j =1,=i

3
i−1
+ α(vj , vi )B2 (u)|Φ2 (v1 , v̌i , v3 )
i=1 j =1

3
3
j =1,=i j =1,=i

2
3
i−1 −1
j
i=1 j =i+1 k=1 l=1=i

× ω1 (vi )ω1 (vj )a11 (vi , vm )a11 (vj , vm )H1D2 (u, vi , vj )
D
+ ω2 (vi )ω2 (vj )a12 (vi , vm )a12 (vj , vm )H4 2 (u, vi , vj ) . (76)
Where m = 3 if i = 1, j = 2; m = 2 if i = 1, j = 3; m = 1 if i = 2, j = 3,

Φ2 (v2 , v3 ), i = 1,
Φ2 (v1 , v̌i , v3 ) = Φ2 (v1 , v3 ), i = 2,
Φ2 (v1 , v2 ), i = 3,
Φ (v ), i = 1, j = 2,
1 3
Φ1 (v̌i , v2 , v̌j ) = Φ1 (v2 ), i = 1, j = 3, (77)
Φ1 (v1 ), i = 2, j = 3.
Acting t (u) Eq. (44) on |Φ3 (v1 , v2 , v3 ), we get

2
2
t (u)|Φ3 (v1 , v2 , v3 ) = w1 (u)ω1 (u) a11 (u, vi ) + w2 (u)ω2 (u) a12 (u, vi )
i=1 i=1

2
+ w3 (u)ω3 (u) a13 (u, vi ) |Φ3 (v1 , v2 , v3 ) (78)
i=1
and all the unwanted terms vanish by the Bethe equations

3 2
ω1 (vi ) j =1,=i a1 (vi , vj )
= β(u, vi ) 2 (i = 1, 2, 3). (79)
ω2 (vi ) 1
j =1,=i a1 (vi , vj )
3.5. The n-particle state
Let
Φn (v1 , . . . , vn ) = B1 (v1 )Φn−1 (v2 , . . . , vn )

n
+ F (v1 ) n
Φn−2 (v2 , . . . , v̌i , . . . , vn )A(vi )ga,i (v1 , . . . , vn )
i=2

n
+ F (v1 ) 1 (vi )gd,i
Φn−2 (v2 , . . . , v̌i , . . . , vn )D n
(v1 , . . . , vn ),
i=2 (80)
then the general n-particle state is constructed as
|Φn (v1 , . . . , vn ) = Φn (v1 , . . . , vn )|0. (81)

n n
We can prove that if the coefficients ga,i (v1 , . . . , vn ) and gd,i (v1 , . . . , vn ) are given by (see
Appendix C)

i−1
n
ga,i (v1 , . . . , vn ) = α(vj , vi )Λn−2
1 (vi ; v2 , . . . , v̌i , . . . , vn )ga,2 (v1 , vi ),
2
(82)
j =2

i−1
n
gd,i (v1 , . . . , vn ) = 2 (vi ; v2 , . . . , v̌i , . . . , vn )gd,2 (v1 , vi ),
α(vj , vi )Λn−2 2
(83)
j =2
the |Φn (v1 , . . . , vn ) own the property as below
|Φn (v1 , . . . , vi , vi+1 , . . . , vn ) = α(vi , vi+1 )|Φn (v1 , . . . , vi+1 , vi , . . . , vn ). (84)
Where

n
Λnl (u; v1 , v2 , . . . , vn ) = a1l (u, vi ) (l = 1, 2, 3), (85)
i=1
Λn−1
l (v; vi , {vm }) = Λn−1
l (v; v1 , . . . , v̌i , . . . , vn ) (l = 1, 2), (86)
Λn−2
l (v; vi , vj , {vm }) = Λn−2
l (v; v1 , . . . , v̌i , . . . , v̌j , . . . , vn ) (l = 1, 2). (87)
Denote
i−1
(1)
Ψ (u, vi ; {vm }) = α(vj , vi ) B1 (u)|Φn−1 (v1 , . . . , v̌i , . . . , vn ), (88)
n−1
j =1
i−1
(2)
Ψ (u, vi ; {vm }) = α(vj , vi ) B2 (u)|Φn−1 (v1 , . . . , v̌i , . . . , vn ), (89)
n−1
j =1
i−1 −1

(3) j
Ψ (u, vi , vj ; {vm }) = α(vk , vi ) α(vl , vj ) F (u)
n−2
k=1 l=1=i
× |Φn−2 (v1 , . . . , v̌i , . . . , v̌j , . . . , vn ). (90)
Relying on the previous results of directly calculating the one, two and three-particle state,
for the n-particle state |Φn (v1 , . . . , vn ), we can infer that
A(u)|Φn (v1 , . . . , vn )
= ω1 (u)Λn1 (u; v1 , . . . , vn )|Φn (v1 , . . . , vn )

n
(1)
+ Ψ (u, vi ; {vm })
n−1
i=1

× ω1 (vi )a21 (u, vi )Λn−1 n−1
1 (vi ; vi , {vm }) + ω2 (vi )a3 (u, vi )Λ2 (vi ; vi , {vm })
1

n−1
n
(3)
+ Ψ (u, vi , vj ; {vm })
n−2
i=1 j =i+1

× ω1 (vi )ω1 (vj )H1A (u, vi , vj )Λn−2 n−2
1 (vi ; vi , vj , {vm })Λ1 (vj ; vi , vj , {vm })
+ ω2 (vi )ω1 (vj )H2A (u, vi , vj )Λn−2 n−2

+ ω1 (vi )ω2 (vj )H3A (u, vi , vj )Λn−2
n−2

+ ω2 (vi )ω2 (vj )H4A (u, vi , vj )Λn−2 n−2
(91)
1 (u)|Φn (v1 , . . . , vn )
D
= ω2 (u)Λn2 (u; v1 , . . . , vn )|Φn (v1 , . . . , vn )

n
(1)
+ Ψ (u, vi ; {vm })
n−1
i=1

× ω1 (vi )a22 (u, vi )Λn−1 2 n−1

n
(2)
+ Ψ (u, vi ; {vm })
n−1
i=1

× ω1 (vi )a42 (u, vi )Λn−1 2 n−1

n−1
n
(3)
+ Ψ (u, vi , vj ; {vm })
n−2
i=1 j =i+1

× ω1 (vi )ω1 (vj )H1D1 (u, vi , vj )Λn−2 n−2
+ ω2 (vi )ω1 (vj )H2D1 (u, vi , vj )Λn−2 n−2


(92)
2 (u)|Φn (v1 , . . . , vn )
D
= ω3 (u)Λn3 (u; v1 , . . . , vn )|Φn (v1 , . . . , vn )

n
(1)
+ Ψ (u, vi ; {vm })
n−1
i=1

× ω1 (vi )a23 (u, vi )Λn−1 n−1
3

n
(2)
+ Ψ (u, vi ; {vm })
n−1
i=1

× ω1 (vi )a43 (u, vi )Λn−1 n−1
3

n−1
n
(3)
+ Ψ (u, vi , vj ; {vm })
n−2
i=1 j =i+1

× ω1 (vi )ω1 (vj )H1D2 (u, vi , vj )Λn−2
n−2

+ ω1 (vi )ω2 (vj )H3D2 (u, vi , vj )Λn−2
n−2

(93)
We can prove that our conclusions Eqs. (91)–(93) are still hold at the case n + 1 (see
Appendix D) by using the following assumption.
Assumption. For n variables s1 , s2 , . . . , sn , if

(2) (3) (4)
(1)
ϕm + B1 (s1 )ϕm−1 + B2 (s1 )ϕm−1 + F (s1 )ϕm−2 = 0, (94)
then
(2) (3) (4)
(1)
ϕm = 0, ϕm−1 = 0, ϕm−1 = 0, ϕm−2 = 0, (95)
(1)
where ϕm is a linear combination of B1 (sj1 )|Φm−1 (sj2 , sj3 , . . . , sjm )’s, B2 (sj1 ) ×
|Φm−1 (sj2 , sj3 , . . . , sjm )’s, and F (sj1 )|Φm−2 (sj2 , sj3 , . . . , sjm−1 )’s, with m n − 1, the
spectrum parameters sj1 , sj2 , . . . , sjm−1 , sjm ∈ {s2 , s3 , . . . , sn } and j1 < j2 < · · · < jm .
(2) (3) (4) (1) (l)
The structure of ϕm−1 , ϕm−1 and ϕm−2 are similar to that of ϕm . In ϕm (l = 1, 2, 3, 4),
the variable s1 do not appear in the spectrum parameter of operators B1 , B2 , F and Φk
(k < m), but it can appear in their coefficients. Eqs. (94) and (95) can be viewed as a linear
(1) (2) (3) (4)
independence assumption of ϕm , B1 (s1 )ϕm−1 , B2 (s1 )ϕm−1 , and F (s1 )ϕm−2 .
Using Eqs. (91)–(93) and (44), we get
t (u)|Φn (v1 , . . . , vn )

= w1 (u)ω1 (u)Λn1 (u; v1 , . . . , vn ) + w2 (u)ω2 (u)Λn2 (u; v1 , . . . , vn )

+ w3 (u)ω3 (u)Λn3 (u; v1 , . . . , vn ) |Φn (v1 , . . . , vn ) + u.t., (96)
where u.t. denotes the unwanted terms. We can easily check as we have done on the two-
particle state that if the rapidities satisfy the Bethe equations
ω1 (vi ) Λn−1 (vi ; v1 , . . . , v̌i , . . . , vn )

= β(u, vi ) 2n−1 (i = 1, . . . , n) (97)
ω2 (vi ) Λ1 (vi ; v1 , . . . , v̌i , . . . , vn )
all the unwanted terms vanish. The explicit expression for Eq. (96) is
t (u)|Φn (v1 , . . . , vn )
= w1 (u)w1 (u)c(u)2N ρ(u)−N

n
sinh( u+2ṽi + q) sinh( u−2ṽi + q)
×
i=1 sinh( u+2ṽi − q) sinh( u−2ṽi − q)
+ w2 (u)w2 (u)b(u)2N ρ(u)−N

n
sinh( u+2ṽi − 3q) sinh( u−2ṽi − 3q) cosh( u+2ṽi ) cosh( u−2ṽi )
×
i=1 sinh( u+2ṽi − q) sinh( u−2ṽi − q) cosh( u+2ṽi − 2q) cosh( u−2ṽi − 2q)

n
sinh(cosh( u+2ṽi − 4q) cosh( u−2ṽi − 4q)
+ w3 (u)w3 (u)d(u)2N ρ(u)−N ,
i=1 cosh( u+2ṽi − 2q) cosh( u−2ṽi − 2q)
(98)
where
w1 (u) = 1,
sinh(u)
w2 (u) = e2q ,
sinh(u − 2q)
sinh(u) cosh(u − 5q)
w3 (u) = e2q ,
sinh(u − 4q) cosh(u − 3q)
sinh(u − 6q) cosh(u − q)
w1 (u) = ,
sinh(u − 6q)
w2 (u) = e−2q ,
sinh(u − 4q)
w3 (u) = e−2q (99)
for the case (i) Eq. (12) and
w1 (u) = e−u [c cosh(q) + sinh(u − 2q)],
sinh(u) cosh(u − q)
w2 (u) = c,
sinh(u − 2q)
sinh(u) cosh(u − q)
w3 (u) = eu−4q [c cosh(q) + sinh(u − 4q)],
w1 (u) = eu [c cosh(q) + sinh(u − 2q)],
w2 (u) = c,
sinh(u − 4q)
w3 (u) = e4q−u [c cosh(q) + sinh(u − 4q)] (100)
for the case Eq. (13). The Bethe equations (97) are
sinh( ṽ2i − q)2N
sinh( ṽ2i + q)2N
ṽ +ṽ ṽ −ṽ ṽ +ṽ ṽ −ṽ

n
sinh( i 2 j − 2q) sinh( i 2 j − 2q) cosh( i 2 j + q) cosh( i 2 j + q)
= ṽi +ṽj ṽi −ṽj ṽi +ṽj ṽi −ṽj
j =1=i sinh( 2 + 2q) sinh( 2 + 2q) cosh( 2 − q) cosh( 2 − q)
(i = 1, . . . , n) (101)
and
sinh( ṽ2i − q)2N cosh( u−2ṽi − 2q) − c sinh( u+2ṽi − 3q) cosh(q) − c sinh(ṽi )
sinh( ṽi + q)2N cosh( u+ṽi − 2q) − c sinh( u−ṽi − 3q) cosh(ṽi + q)
2 2 2

n ṽ +ṽ ṽ −ṽ ṽ +ṽ ṽ −ṽ
sinh( i 2 j − 2q) sinh( i 2 j − 2q) cosh( i 2 j + q) cosh( i 2 j + q)
= ṽi +ṽj ṽi −ṽj ṽi +ṽj ṽi −ṽj
j =1=i sinh( 2 + 2q) sinh( 2 + 2q) cosh( 2 − q) cosh( 2 − q)
(i = 1, . . . , n) (102)
for Eqs. (12) and (13), respectively. Here vi = ṽi + 2q, c2 = c 2 = −1.
4. Conclusions
In this paper, we solve Izergin–Korepin model with open boundary conditions by

using the algebraic Bethe ansatz method. Based on the direct calculation for the one-
particle state, two-particle state and three-particle state, we give the explicit expression of
general n-particle state, the eigenvalue of the transfer matrix and the corresponding Bethe
equations. To the Uq (su(2)) quantum invariant case, our results coincide with that obtained
by analytic Bethe ansatz method [46] and [21]. For the non-quantum-group invariant cases,
there are a little difference between our results and that proposed in Ref. [21] by means of
analytic Bethe ansatz method. For example, the spectrum parameter u exists in our Bethe
equations (102), however there is no u in Bethe equations of Ref. [21]. The difference
may be caused by the characters of the two different Bethe ansatz methods. In the frame
of QISM, we construct a multi-particle state with some exchange symmetry and act the
transfer matrix on it to achieve the eigenvalue and Bethe equations. The construction of
the multi-particle state is not unique. Different multi-particle state may lead to different
results. The analytic Bethe ansatz method use the symmetry property and the asymptotic
behavior of the transfer matrix to obtain the eigenvalue and Bethe equations. In Ref. [52],
Nepomechie argued the symmetry property of the transfer matrix of IK model. He proved
the transfer matrix corresponding to Eq. (13) also have Uq (o(3)) symmetry, but with a
nonstandard coproduct, which may lead to the difference between our conclusions and that
in Ref. [21] when the analytic Bethe ansatz method is used.
We propose an assumption by which we show our conclusions for the general n-
particle state with mathematical induction method. In the procedure of proof, the exchange
symmetry property of n-particle state Eq. (84) plays a key role, which ensure some
unreasonable terms do not appear so that we can successfully use the algebraic Bethe
ansatz method. We may apply the above discussion to other similar models, such as all the
nineteen-vertex models or Hubbard-like models. Our assumption hold for n 3 by directly
calculating the two and three-particle state and we have also examined some conclusions
obtained by the assumption for the general n-particle state. It is an open problem for us to
prove the assumption for the general n-particle state.
Using the Bethe ansatz equations and energy spectrum, we can study the boundary
contributions to the thermodynamic quantities and the surface critical behavior of two-
dimensional polymers. We also can use the n-particle state to calculate the corresponding
form factor and correlation functions which have rather important applications in
condensed matter physics.
Acknowledgement
This work is supported by the National Natural Science Foundation of China under
Grant No. 10175050.
Appendix A
b(u+ )c(−u− )
a11 (u, v) = ,
c(u+ )b(−u− )
b(2v)ē(−u− )
a21 (u, v) = − ,
c(2v)b(−u− )
e(u+ )
a31 (u, v) = − ,
c(u+ )
g(u+ )ē(−u− )
a41 (u, v) = − ,
c(u+ )b(−u− )
f (u+ )
a51 (u, v) = − ,
c(u+ )
g(u+ )c(−u− )
a61 (u, v) = , (A.1)
c(u+ )b(−u− )

a(u+ ) e(u+ )ē(u+ ) a(u− ) g(u− )ḡ(u− )
a12 (u, v) = − − ,
b(u+ ) b(u+ )c(u+ ) b(u− ) b(u− )d(u− )
ḡ(u+ )ē(u− )g(u− ) a(u+ )ē(u− )
a22 (u, v) = − f1 (v)
b(u+ )b(u− )d(u− ) b(u+ )b(u− )
+ X1 (u, v)a11 (v, u) − X2 (u, v)a21 (u, v),
a(u+ )ē(u− )
a32 (u, v) = − − X2 (u, v)a31 (u, v),
b(u+ )b(u− )
d(u+ )g(u− ) g(u+ )
a42 (u, v) = − f1 (v) ,
b(u+ )d(u− ) b(u+ )
g(u+ )
a52 (u, v) = − ,
b(u+ )
g(u− )ē(u− )
a62 (u, v) = + X1 (u, v)a61 (v, u) − X2 (u, v)a41 (u, v),
b(u− )d(u− )
e(u+ )ē(u− )
a72 (u, v) = − − X2 (u, v)a51 (u, v),
b(u+ )b(u− )
ḡ(u− ) f¯(u− )g(u− )
a82 (u, v) = − + X1 (u, v)a41 (v, u) − X2 (u, v)a61 (u, v),
b(u− ) b(u− )d(u− )

e(u+ ) a(u− ) g(u− )ḡ(u− )
a92 (u, v) = − + X1 (u, v)a51 (v, u), (A.2)
b(u+ ) b(u− ) b(u− )d(u− )
b(u+ )b(u− ) ḡ(u+ )g(u+ )b(u− )

a13 (u, v) = − ,
d(u+ )d(u− ) b(u+ )d(u+ )d(u− )
ḡ(u+ )ḡ(u+ )ē(u− )b(u− ) ē(u+ )f¯(u− )
a23 (u, v) = − f1 (v) + X3 (u, v)a11 (v, u)
b(u+ )d(u+ )d(u− )d(u− ) d(u+ )d(u− )
− X4 (u, v)a21 (u, v) − X5 (u, v)a22 (u, v),

ē(u+ )f¯(u− )
a33 (u, v) = − − X4 (u, v)a31 (u, v) − X5 (u, v)a32 (u, v),
d(u+ )d(u− )
ḡ(u+ )b(u− )b(u− ) b(u+ )ḡ(u− )
a43 (u, v) = − f1 (v) − X5 (u, v)a42 (u, v),
b(u+ )d(u− )d(u− ) d(u+ )d(u− )
b(u+ )ḡ(u− )
a53 (u, v) = − − X5 (u, v)a52 (u, v),
d(u+ )d(u− )
ḡ(u+ )ē(u− )b(u− )
a63 (u, v) = + X3 (u, v)a61 (v, u)
d(u+ )d(u− )d(u− )
− X4 (u, v)a41 (u, v) − X5 (u, v)a62 (u, v),
c(u+ )f¯(u− )
a73 (u, v) = − − X4 (u, v)a51 (u, v) − X5 (u, v)a72 (u, v),
d(u+ )d(u− )
ḡ(u+ )f¯(u− )b(u− )
a83 (u, v) = − + X3 (u, v)a41 (v, u)
d(u+ )d(u− )d(u− )
− X4 (u, v)a61 (u, v) − X5 (u, v)a82 (u, v),
c(u+ )ē(u− ) ḡ(u+ )e(u+ )ḡ(u− )b(u− )
a93 (u, v) = −
d(u+ )d(u− ) b(u+ )d(u+ )d(u− )d(u− )
+ X3 (u, v)a51 (v, u) − X5 (u, v)a92 (u, v), (A.3)
d(u+ )c(−u− )
b11 (u, v) = ,
c(u+ )d(−u− )
b(u+ )ḡ(−u− )
b51 (u, v) = − ,
c(u+ )d(−u− )
e(u+ )
b61 (u, v) = − ,
c(u+ )
d(u+ )f¯(−u− ) g(u+ )ḡ(−u− ) f (u+ )
b21 (u, v) = − − f1 (v) − f2 (v) ,
c(u+ )d(−u− ) c(u+ )d(−u− ) c(u+ )
g(u+ )ḡ(−u− ) f (u+ )
b31 (u, v) = − − f3 (v) , (A.4)
c(u+ )d(−u− ) c(u+ )
f (u+ )
b41 (u, v) = − ,
c(u+ )

g(u+ )ḡ(u+ ) e(u− )ē(u− )
b12 (u, v) = 1 − 1− ,
b(u+ )b(u+ ) b(u− )b(u− )

ḡ(u+ ) e(u− )ē(u− )
b22 (u, v) = 1− 2
α(v, u)ga,2 (u, v) − X6 (u, v)b21 (u, v)
b(u+ ) b(u− )b(u− )
e(u− )ē(u− ) e(u+ )ē(u− )
+ f1 (v) − f2 (v) ,
b(u− )b(u− ) b(u+ )b(u− )
ḡ(u+ ) e(u− )ē(u− )

b32 (u, v) = (1 − 2
)α(v, u)gd,2 (u, v) − X6 (u, v)b31 (u, v)
b(u+ ) b(u− )b(u− )
e(u− )ē(u− ) e(u+ )ē(u− )
+ − f3 (v) ,
b(u− )b(u− ) b(u+ )b(u− )
e(u+ )ē(u− )
b42 (u, v) = − − X6 (u, v)b41 (u, v),
b(u+ )b(u− )

ḡ(u+ ) e(u− )ē(u− )
b52 (u, v) = 1− α(v, u)
b(u+ ) b(u− )b(u− )
ḡ(u+ ) e(u− )ē(u− )
+ − X6 (u, v)b51 (u, v),
b(u+ ) b(u− )b(u− )
a(u+ )ē(u− )
b62 (u, v) = − − X6 (u, v)b61 (u, v),
b(u+ )b(u− )
d(u+ )e(u− )
b72 (u, v) = ,
b(u+ )b(u− )
g(u+ )
b82 (u, v) = − , (A.5)
b(u+ )

c(u− ) c(u+ ) g(u+ )f¯(u+ )
b13 (u, v) = −
d(u− ) d(u+ ) c(u+ )d(u+ )

ē(u + ) e(u+ ) ē(u+ )f (u+ )
− e−4q − ,
d(u+ ) h(u+ ) h(u+ )c(u+ )

c(u− ) f¯(u+ )c(u− )
b23 (u, v) =
d(u− ) c(u+ )d(u− )

ē(u+ ) e(u+ )d(u+ )c(u− ) ē(u− )e(u+ )
− e−4q − f2 (v)
d(u+ ) h(u+ )c(u+ )d(u− ) e(u− )h(u+ )
¯
c(u+ )f (u− )
− f2 (v) − X7 (u, v)b21 (u, v)
d(u+ )d(u− )
− X8 (u, v)b22 (u, v) − X9 (u, v)α(v, u)ga,2
2
(u, v),
ē(u+ )c(u− ) ē(u− )e(u+ ) c(u+ )f¯(u− )
b33 (u, v) = e−4q f3 (v) − f3 (v)
d(u+ )d(u− ) e(u− )h(u+ ) d(u+ )d(u− )
− X7 (u, v)b31 (u, v) − X8 (u, v)b32 (u, v) − X9 (u, v)α(v, u)gd,2
2
(u, v),
ē(u+ )c(u− ) ē(u− )e(u+ ) c(u+ )f¯(u− )
b43 (u, v) = e−4q −
d(u+ )d(u− ) e(u− )h(u+ ) d(u+ )d(u− )
− X7 (u, v)b41 (u, v) − X8 (u, v)b42 (u, v),
b53 (u, v) = −X7 (u, v)b51 (u, v) − X8 (u, v)b52 (u, v) − X9 (u, v)α(v, u),
ē(u+ )c(u− ) ē(u− )a(u+ ) ē(u+ )f¯(u− )
b63 (u, v) = e−4q −
d(u+ )d(u− ) e(u− )h(u+ ) d(u+ )d(u− )
− X7 (u, v)b61 (u, v) − X8 (u, v)b62 (u, v),
b73 (u, v) = −X8 (u, v)b72 (u, v),

ē(u+ )c(u− ) b(u− )g(u+ ) b(u+ )ḡ(u− )
b83 (u, v) = e−4q − − X8 (u, v)b82 (u, v),
d(u+ )d(u− ) e(u− )h(u+ ) d(u+ )d(u− )
(A.6)
a(u+ )
c11 (u, v) = ,
c(u+ )
g(u+ )e(u− )
c21 (u, v) = ,
c(u+ )b(u− )
g(u+ )ē(u− )
c31 (u, v) = − ,
c(u+ )b(u− )
f (u+ )
c41 (u, v) = − ,
c(u+ )
e(u+ )
c51 (u, v) = ,
c(u+ )
ē(u+ )c(u− ) b(2u)
c61 (u, v) = ,
c(u+ )b(u− ) c(2u)
b(u+ )e(u− )
c71 (u, v) = ,
c(u+ )b(u− )
ē(u− )b(2u)
c81 (u, v) = −f1 (v) ,
b(u− )c(2u)
ē(u− )b(2u)
c91 (u, v) = − ,
b(u− )c(2u)
e(u+ )
1
c10 (u, v) = −f1 (v) ,
c(u+ )
e(u+ )
1
c11 (u, v) = − , (A.7)
c(u+ )

ē(u+ )ḡ(u− ) a(u+ )
c12 (u, v) = 1− ,
b(u+ )d(u− ) c(u+ )
a(u− ) ē(u+ )ḡ(u− )g(u+ )e(u− )
c22 (u, v) = − ,
d(u− ) b(u+ )d(u− )c(u+ )b(u− )
ē(u+ )ḡ(u− )g(u+ )ē(u− ) f¯(u− )
c32 (u, v) = − ,
b(u+ )d(u− )c(u+ )b(u− ) d(u− )

ḡ(u− ) ē(u+ )g(u+ ) e(u+ )
c42 (u, v) = − ,
d(u− ) b(u+ )c(u+ ) b(u+ )

ḡ(u− ) c(u+ ) ē(u+ )e(u+ )
c52 (u, v) = − ,
d(u− ) b(u+ ) b(u+ )c(u+ )
ḡ(2u)ē(u+ )c(u− )
c62 (u, v) = ,
c(2u)c(u+ )b(u− )
ḡ(u+ )a(u− ) ē(u+ )ḡ(u− )e(u− ) d(u+ )g(u− )

c72 (u, v) = − + f3 (u) ,
b(u+ )d(u− ) b(u+ )d(u− )b(u− ) b(u+ )d(u− )
d(u+ )g(u− )
c82 (u, v) = ,
b(u+ )d(u− )
ḡ(2u)ē(u− )
c92 (u, v) = −f1 (v) ,
c(2u)c(u− )
ḡ(2u)ē(u− )
2
c10 (u, v) = − ,
c(2u)c(u− )

g(u+ ) ḡ(u− )a(u+ ) ē(u+ )ḡ(u− )e(u+ )
2
c11 (u, v) = −f1 (v) f3 (u) + − ,
b(u+ ) b(u+ )d(u− ) b(u+ )d(u− )c(u+ )

g(u+ ) ḡ(u− )a(u+ ) ē(u+ )ḡ(u− )e(u+ )
c12 (u, v) = − f3 (u)
2
+ − ,
b(u+ ) b(u+ )d(u− ) b(u+ )d(u− )c(u+ )
g(u+ )
2
c13 (u, v) = − ,
b(u+ )
g(u+ )
2
c14 (u, v) = −f1 (v) , (A.8)
b(u+ )
c(u+ )
d11 (u, v) = ,
h(u+ )

c(u+ ) ḡ(u+ )ē(−u− )
d21 (u, v) = − X10 (u, v)α(u, v) ,
h(u+ ) d(u+ )b(−u− )

c(u+ ) b(u+ )ē(−u− ) ē(u+ )c(−u− )
d3 (u, v) =
1
f1 (u) −
h(u+ ) d(u+ )b(−u− ) c(u+ )b(−u− )

− X10 (u, v)α(u, v)ga,2 (v, u) ,
2

c(u+ ) b(u+ )ē(−u− )
d41 (u, v) = − − X10 (u, v)α(u, v)gd,2
2
(v, u) ,
h(u+ ) d(u+ )b(−u− )

c(u+ ) b(u+ )ē(−u− ) e(u+ ) ē(u+ ) 1
d5 (u, v) = −
1
f1 (v) + f2 (v) + b (u, v)
h(u+ ) d(u+ )b(−u− ) d(u+ ) d(u+ ) 2

+ X10 (u, v)ga,2
2
(u, v) ,

c(u+ ) b(u+ )ē(−u− ) e(u+ ) ē(u+ ) 1
d61 (u, v) = − + f3 (v) + b (u, v)
h(u+ ) d(u+ )b(−u− ) d(u+ ) d(u+ ) 3

+ X10 (u, v)gd,2
2
(u, v) ,

c(u+ ) e(u+ ) ē(u+ ) 1
d71 (u, v) = − + b (u, v) , (A.9)
h(u+ ) d(u+ ) d(u+ ) 4
where Xi (u, v) (i = 1, . . . , 10) and h(u) are defined as below

ē(u+ ) a(u− ) g(u− )ḡ(u− )
X1 (u, v) = − ,
b(u+ ) b(u− ) b(u− )d(u− )
ē(u+ )ē(u− )
X2 (u, v) = f1 (u) + ,
b(u+ )b(u− )
f¯(u+ )ē(u− ) ḡ(u+ )ē(u+ )ḡ(u− )b(u− )
X3 (u, v) = − ,
d(u+ )d(u− ) b(u+ )d(u+ )d(u− )d(u− )
f¯(u+ )f¯(u− ) ḡ(u+ )ḡ(u− )
X4 (u, v) = + f1 (u) + f2 (u),
d(u+ )d(u− ) d(u+ )d(u− )
ḡ(u+ )ḡ(u− )
X5 (u, v) = + f3 (u),
d(u+ )d(u− )
ē(u+ )ē(u− )
X6 (u, v) = + f1 (u),
b(u+ )b(u− )
f¯(u+ )f¯(u− )
X7 (u, v) =
d(u+ )d(u− )

−4q ē(u+ )c(u− ) ē(u+ )ē(u− ) b(u+ )b(u− )
−e + f1 (u)
d(u+ )d(u− ) h(u+ )e(u− ) h(u+ )e(u− )
ḡ(u+ )ḡ(u− )
+ f1 (u) + f2 (u),
d(u+ )d(u− )
ḡ(u+ )ḡ(u− ) ē(u+ )c(u− ) b(u+ )b(u− )
X8 (u, v) = + f3 (u) − e−4q ,
d(u+ )d(u− ) d(u+ )d(u− ) h(u+ )e(u− )
f¯(u+ )c(u− ) b(u+ )ḡ(u− )
X9 (u, v) =
d(u+ )d(u− ) c(u+ )d(u− )

−4q ē(u+ )c(u− ) ḡ(u+ )b(u− ) ē(u+ ) b(u+ )ḡ(u− )
+e − ,
d(u+ )d(u− ) h(u+ )e(u− ) h(u+ ) c(u+ )d(u− )
ḡ(u+ )ē(−u− ) ē(u+ ) 1
X10 (u, v) = − b (u, v),
d(u+ )b(−u− ) d(u+ ) 5

u u
h(u) = 2 sinh − 4q cosh − 3q , (A.10)
2 2
b(u+ )b(u− )
e11 (u, v) = ,
c(u− )d(u+ )
g(u+ )b(u− )
e21 (u, v) = ,
c(u− )d(u+ )
ē(u− )
e31 (u, v) = ,
c(u− )
b(u+ )b(u− )
e12 (u, v) = ,
c(u− )k(u+ )

g(u+ )b(u+ ) e(u− ) ē(−u− )
e22 (u, v) = − ,
d(u+ )k(u+ ) c(u− ) c(−u− )
b(u+ )b(u+ )e(u− ) g(u+ )ḡ(u+ )ē(−u− )
e32 (u, v) = − ,
d(u+ )k(u+ )c(u− ) d(u+ )k(u+ )c(−u− )
g(u+ )b(−u− )
e42 (u, v) = − ,
c(−u− )k(u+ )
d(u+ )b(u− ) ḡ(u+ ) 2
e13 (v, u) = − e (v, u),
c(u− )b(u+ ) b(u+ ) 4
ḡ(u+ )ē(u− ) ḡ(u+ ) 2
e23 (v, u) = − e (v, u),
c(u− )b(u+ ) b(u+ ) 3
ḡ(u+ ) 2
e33 (v, u) = − e (v, u),
b(u+ ) 1
ē(u− ) ḡ(u+ ) 2
e43 (v, u) = − e (v, u),
c(u− ) b(u+ ) 2
b(u+ )b(u− )
e14 (v, u) = ,
c(u− )d(u+ )
ḡ(u+ )b(u− )
e24 (v, u) = ,
c(u− )d(u+ )
e(u− )
e34 (v, u) = ,
c(u− )

u u
k(u) = 2 sinh cosh − 5q . (A.11)
2 2
Appendix B
A(u)B2 (v)
= X11 (u, v)B2 (v)A(u) + X21 (u, v)B1 (v)A(u) + X31 (u, v)B1 (u)A(v)
1 (v) + X51 (u, v)B1 (u)D
+ X41 (u, v)B1 (u)D 2 (v) + X61 (u, v)F (u)C1 (v)
+ X71 (u, v)F (u)C2 (v) + X81 (u, v)F (v)C1 (u), (B.1)
1 (u)B2 (v)
D
1 (u) + X22 (u, v)B1 (v)D
= X12 (u, v)B2 (v)D 1 (u) + X32 (u, v)B1 (u)A(v)
1 (v) + X2 (u, v)B1 (u)D
+ X42 (u, v)B1 (u)D 2 (v) + X2 (u, v)B2 (u)A(v)
5 6
1 (v) + X82 (u, v)B2 (u)D
+ X72 (u, v)B2 (u)D 2 (v) + X92 (u, v)F (u)C1 (v)
+ X10
2
(u, v)F (u)C2 (v) + X11
2
(u, v)F (v)C1 (u) + X12
2
(u, v)F (v)C2 (u), (B.2)
2 (u)B2 (v)
D
2 (u) + X3 (u, v)B1 (v)D
= X13 (u, v)B2 (v)D 2 (u) + X3 (u, v)B1 (u)A(v)
2 3

+ X (u, v)B1 (u)D1 (v) + X (u, v)B1 (u)D2 (v) + X3 (u, v)B2 (u)A(v)
3 3
4 5 6

+ X7 (u, v)B2 (u)D1 (v) + X8 (u, v)B2 (u)D2 (v) + X93 (u, v)F (u)C1 (v)
3 3
+ X10
3
(u, v)F (u)C2 (v) + X11
3
(u, v)F (v)C1 (u) + X12
3
(u, v)F (v)C2 (u), (B.3)
C1 (u)F (v)
= Y11 (u, v)F (v)C1 (u) + Y21 (u, v)F (v)C2 (u) + Y31 (u, v)F (u)C1 (v)
+ Y41 (u, v)F (u)C2 (v) + Y51 (u, v)B1 (v)A(u) + Y61 (u, v)B1 (v)D 1 (u)
+ Y71 (u, v)B2 (v)A(u) + Y81 (u, v)B2 (v)D1 (u) + Y91 (u, v)B1 (u)A(v)
+ Y10
1 1 (v) + Y11
(u, v)B1 (u)D 1 2 (v) + Y12
(u, v)B1 (u)D 1
(u, v)B2 (u)A(v)
+ Y13 (u, v)B2 (u)D
1 1 (v) + Y14 (u, v)B2 (u)D
1 2 (v), (B.4)
C2 (u)F (v)
= Y12 (u, v)F (v)C2 (u) + Y22 (u, v)F (v)C1 (u) + Y32 (u, v)F (u)C1 (v)
+ Y42 (u, v)F (u)C2 (v) + Y52 (u, v)B1 (v)A(u) + Y62 (u, v)B1 (v)D 1 (u)
+ Y72 (u, v)B1 (v)D2 (u) + Y82 (u, v)B2 (v)A(u) + Y92 (u, v)B2 (v)D1 (u)
+ Y10
2 2 (u) + Y11
(u, v)B2 (v)D 2
(u, v)B1 (u)A(v) + Y122 1 (v)
(u, v)B1 (u)D
+ Y13
2 2 (v) + Y14
(u, v)B1 (u)D 2
(u, v)B2 (u)A(v) + Y15
2 1 (v)
(u, v)B2 (u)D
+ Y16
2 2 (v).
(u, v)B2 (u)D (B.5)
Here the explicit form of all the coefficients of Eqs. (B.1)–(B.5) are not presented for
their long and tedious expressions.
Appendix C
Considering the definition of n-particle state Eqs. (81), (80), from

|Φn (v1 , . . . , vi , vi+1 , . . . , vn ) = α(vi , vi+1 )|Φn (v1 , . . . , vi+1 , vi , . . . , vn ) (C.1)
with i = 1, we can easily obtain
n
ga,i+1 (v1 , . . . , vi , vi+1 , . . . , vn ) = α(vi , vi+1 )ga,i
n
(v1 , . . . , vi+1 , vi , . . . , vn ), (C.2)
n
gd,i+1 (v1 , . . . , vi , vi+1 , . . . , vn ) = α(vi , vi+1 )gd,i
n
(v1 , . . . , vi+1 , vi , . . . , vn ). (C.3)
Let i = 1 in Eq. (C.1), we have
|Φn (v1 , v2 , . . . , vn )

= α(v1 , v2 ) B1 (v2 )B1 (v1 )|Φn−2 (v3 , . . . , vn )

n
+ B1 (v2 )F (v1 ) Φn−3 (v3 , . . . , v̌i , . . . , vn )
j =3

n−1
× A(vj )ga,j n−1
−1 (v1 , v3 , . . . , vn ) + D1 (vj )gd,j −1 (v1 , v3 , . . . , vn ) |0

+ F (v2 )Φn−2 (v3 , . . . , vn ) A(v1 )ga,2
n
(v2 , v1 , . . . , vn )

+ D1 (v1 )gd,2
n
(v2 , v1 , . . . , vn ) |0

n
+ F (v2 ) B1 (v1 )Φn−3 (v3 , . . . , v̌i , . . . , vn )
j =3
+ F (v1 ) Φn−4 (v3 , . . . , v̌i , . . . , vˇj , . . . , vn )
j =3,j =i
n−2
× A(vj )ga,j (v1 , v3 , . . . , v̌i , . . . , vn )

1 (vj )g n−2 (v1 , v3 , . . . , v̌i , . . . , vn )
+D d,j

n
× A(vi )ga,i (v2 , v1 , . . . , vn ) + D1 (vi )gd,i (v2 , v1 , . . . , vn ) |0 ,
n
(C.4)
where j = j − 1 whenj < i and j = j − 2 when j > i. Firstly, we substitute com-

mutation relation (29) to the term B1 (v2 )B1 (v1 ) and act the operators A(v2 ), D 1 (v2 ),

A(v1 ), D1 (v1 ), respectively, on |Φn−2 (v3 , . . . , vn ) with the help of the conclusions
(91), (92). Then we compare the coefficients of F (v1 )Φn−2 (v3 , . . . , vn )A(v2 )|0 and
1 (v2 )|0 in the l.h.s. of Eq. (C.4) with that in the r.h.s. of
F (v1 )Φn−2 (v3 , . . . , vn )D
Eq. (C.4), respectively, and we will get the following results immediately
n
ga,2 (v1 , . . . , vn ) = Λn−2
1 (v2 ; v3 , . . . , vn )ga,2 (v1 , v2 ),
2
(C.5)
n
gd,2 (v1 , . . . , vn ) = Λn−2
2 (v2 ; v3 , . . . , vn )gd,2 (v1 , v2 ).
2
(C.6)
From Eqs. (C.2), (C.5) and (C.3), (C.6), we can arrive at

i−1
n
ga,i (v1 , . . . , vn ) = α(vj , vi )Λn−2
1 (vi ; v2 , . . . , v̌i , . . . , vn )ga,2 (v1 , vi ),
2
(C.7)
j =2

i−1
n
gd,i (v1 , . . . , vn ) = α(vj , vi )Λn−2
2 (vi ; v2 , . . . , v̌i , . . . , vn )gd,2 (v1 , vi ).
2
(C.8)
j =2
Now we have to test Eqs. (C.7), (C.8). Putting Eqs. (C.7), (C.8) into Eq. (C.4)
and applying relations (40), (39), (41), (43) to the terms B1 (v2 )F (v1 ), F (v2 )B1 (v1 ),
F (v2 )B2 (v1 ), F (v2 )F (v1 ) in the r.h.s. of Eq. (C.4), respectively, (reminding that the
fist step have been done), we will find that in the r.h.s. of Eq. (C.4), the terms
F (v1 )B2 (v2 )|Φn−3 (v3 , . . . , v̌i , . . . , vn ) and B2 (v1 )F (v2 )|Φn−3 (v3 , . . . , v̌i , . . . , vn ) are
easily cancelled out and the rest terms can be simplified into one term |Φn (v1 , v2 , . . . , vn )
which is just the l.h.s. of Eq. (C.4). We have done this work with the aid of computer.
Appendix D
In this appendix, we only take D 1 (u)|Φn+1 (v1 , . . . , vn+1 ) as an example, because the
similar procedure of calculating the D 1 (u)|Φn+1 (v1 , . . . , vn+1 ) can be applied to other
operators. We now carry on our induction. In order to calculate the operators D 1 (u) acting
on the |Φn+1 (v1 , . . . , vn+1 ), besides Eqs. (91)–(93), we also need to know the results of
Cs (u)|Φn (v1 , . . . , vn ) (s = 1, 2). See Eq. (D.5). Let us assume
Cs (u)|Φn (v1 , . . . , vn ) = ψc(1)

s
+ B1 (u)ψc(2)
s
+ B2 (u)ψc(3)
s
+ F (u)ψc(4)
s
, (D.1)
(1) (2) (3)
where ψcs is a linear combination of |Φn−1 ’s. ψcs , ψcs are respectively the linear
combination of |Φn−2 ’s. ψc(4) s are the linear combination of |Φn−3 ’s. All the ψc(i)
s ,
i = 1, 2, 3, 4 satisfy the following property
ψc(i)
s
(v1 , . . . , vj , vj +1 , . . . , vn ) = α(vj , vj +1 )ψc(i)
s
(v1 , . . . , vj +1 , vj , . . . , vn ) (D.2)
and the explicit expression of ψc(1)

s are given by
ψc(1)
1
(v1 , v2 , . . . , vn )

n
i−1
= α(vj , vi )|Φn−1 (v1 , . . . , v̌i , . . . , vn )
i=1 j =1

× ω1 (u)ω1 (vi )Λn1 (u; vi , {vm })Λn1 (vi ; vi , {vm }) c61 (u, vi ) + c81 (u, vi )
+ ω1 (u)ω2 (vi )Λn1 (u; vi , {vm })Λn2 (vi ; vi , {vm })c91 (u, vi )

+ ω2 (u)ω1 (vi )Λn2 (u; vi , {vm })Λn1 (vi ; vi , {vm }) c71 (u, vi ) + c10
1
(u, vi )
+ ω2 (u)ω2 (vi )Λn2 (u; vi , {vm })Λn2 (vi ; vi , {vm })c11
1
(u, vi ) , (D.3)
ψc(1)
2
(v1 , v2 , . . . , vn )

n
i−1
= α(vj , vi )|Φn−1 (v1 , . . . , v̌i , . . . , vn )
i=1 j =1

× ω1 (u)ω1 (vi )Λn1 (u; vi , {vm })Λn1 (vi ; vi , {vm }) c62 (u, vi ) + c92 (u, vi )
2
(u, vi )
2
+ ω2 (u)ω1 (vi )Λ2 (u; vi , {vm })Λ1 (vi ; vi , {vm }) c7 (u, vi ) + c11
n n 2
(u, vi )
2
(u, vi )

+ ω3 (u)ω1 (vi )Λ3 (u; vi , {vm })Λ1 (vi ; vi , {vm }) c8 (u, vi ) + c13
n n 2 2
(u, vi )
2
(u, vi ) . (D.4)
We can prove that the assumption Eq. (D.1) hold for the case n + 1 by the same method
employed in the following induction procedure. Now let operator D 1 (u) act on the
|Φn+1 (v1 , . . . , vn+1 ), we have
1 (u)|Φn+1 (v1 , . . . , vn+1 )

D

= a12 (u, v1 )B1 (v1 )D 1 (u) + a22 (u, v1 )B1 (u)A(v1 )
1 (v1 ) + a42 (u, v1 )B2 (u)A(v1 ) + a52 (u, v1 )B2 (u)D
+ a32 (u, v1 )B1 (u)D 1 (v1 )
+ a62 (u, v1 )F (u)C1 (v1 ) + a72 (u, v1 )F (u)C2 (v1 ) + a82(u, v1 )F (v1 )C1 (u)
+ a92 (u, v1 )F (v1 )C2 (u) × |Φn (v2 , . . . , vn+1 )

+ b12(u, v1 )F (v1 )D 1 (u)
1 (v1 ) + b42 (u, v1 )F (u)D
+ b22 (u, v1 )F (u)A(v1 ) + b32 (u, v1 )F (u)D 2 (v1 )
+ b52 (u, v1 )B1 (u)B1 (v1 ) + b62 (u, v1 )B1 (u)B2 (v1 ) + b72 (u, v1 )B2 (u)B1 (v1 )

n+1
+ b82 (u, v1 )B2 (u)B2 (v1 ) × |Φn−1 (v2 , . . . , v̌i , . . . , vn+1 )
i=2

× ω1 (vi )ga,i
n+1
(v1 , . . . , vn+1 ) + ω2 (vi )gd,i
n+1
(v1 , . . . , vn+1 ) . (D.5)
After a not very hard deducing by using Eqs. (91)–(93), (D.1), (38)–(43), we find that
Eq. (D.5) can be written as
1 (u)|Φn+1 (v1 , . . . , vn+1 ) = ψ (1) + B1 (u)ψ (2) + B2 (u)ψ (3) + F (u)ψ (4) ,
D (D.6)
where ψ (1) is a linear combination of B1 |Φn ’s and F |Φn−1 ’s, ψ (2) and ψ (3) , respectively,
are the linear combination of B1 |Φn−1 ’s, B2 |Φn−1 ’s and F |Φn−2 ’s and ψ (4) is a linear
combination of B1 |Φn−2 ’s, B2 |Φn−2 ’s and F |Φn−3 ’s. For instance, the expression of
ψ (2) are
ψ (2) (v1 , v2 , . . . , vn+1 )

= ω1 (v1 )a22 (u, v1 )Λn1 (v1 ; v1 , {vm })|Φn (v2 , . . . , vn+1 )
+ ω2 (v1 )a32 (u, v1 )Λn2 (v1 ; v1 , {vm })|Φn (v2 , . . . , vn+1 )

n+1
+ B1 (v1 ) |Φn−1 (v2 , . . . , v̌i , . . . , vn+1 )Kib1 (v1 , vi , {vm })
i=2

n+1
i=2

n
n+1
f
+ F (v1 ) |Φn−2 (v2 , . . . , v̌i , . . . , v̌j , . . . , vn+1 )Kij (v1 , vi , vj , {vm }),
i=2 j =i+1 (D.7)
f
where Kib1 , Kib2 , and Kij are coefficients. Then, by the assumption Eqs. (94), (95) and the
property of Eq. (84), we find all the ψ (i) , i = 1, 2, 3, 4 satisfy
ψ (i) (v1 , . . . , vj , vj +1 , . . . , vn+1 )

= α(vj , vj +1 )ψ (i) (v1 , . . . , vj +1 , vj , . . . , vn+1 ) (D.8)
for the rapidities v1 , . . . , vn+1 . In the following procedure, we will determine the explicit
expressions of ψ (i) with the aid of Eq. (D.8) and Eqs. (94), (95). Without losing of
generality, we take ψ (2) as an example here. From Eqs. (D.7) and (D.8), we have
ψ (2) (v1 , v2 , . . . , vn+1 )

= α(v1 , v2 )ψ (2) (v2 , v1 , . . . , vn+1 )

= α(v1 , v2 ) ω1 (v2 )a22 (u, v2 )Λn1 (v2 ; v2 , {vm })|Φn (v1 , v3 , . . . , vn+1 )
+ ω2 (v2 )a32 (u, v2 )Λn2 (v2 ; v2 , {vm })|Φn (v1 , v3 , . . . , vn+1 )

n+1
i=1=2

n+1
i=1=2

n
n+1
+ F (v2 ) |Φn−2 (v1 , . . . , v̌i , . . . , v̌j , . . . , vn+1 )
i=1=2 j =i+1=2

f
× Kij (v1 , vi , vj , {vm }) . (D.9)
From Eq. (D.9), we can find that in the l.h.s. of Eq. (D.9), there are such terms
B2 (v1 )B1 (v3 )|Φn−2 (v4 , . . . , vn+1 )K2b2 (v1 , v2 , {vm }),

n+1
B2 (v1 )B1 (v2 ) |Φn−2 (v3 , . . . , v̌i , . . . , vn+1 )Kib2 (v1 , vi , {vm }).
i=3
However, in the r.h.s. of Eq. (D.9), we cannot obtain this terms above when we use the
commutation relations Eqs. (29), (38)–(43) and Eqs. (91)–(93) to permute the spectrum
parameters v1 and v2 . Subtracting the r.h.s. of Eq. (D.9) from l.h.s. of Eq. (D.9) and putting
the spectrum parameters v1 to the left side of v2 by necessary commutation relations, we
turn Eq. (D.9) into the form of Eq. (94). Using our assumption Eqs. (94), (95), we can
easily get
b
Ki 2 (v1 , vi , {vm }) = 0, i ∈ [2, . . . , n + 1].
So ψ (2) can be written as

(2) (2)
ψ (2) (v1 , v2 , . . . , vn+1 ) = ψ1 (v1 , v2 , . . . , vn+1 ) + ψ2 (v1 , v2 , . . . , vn+1 ) (D.10)
with
ψ1(2) (v1 , v2 , . . . , vn+1 )

n+1
i−1
= α(vj , vi )|Φn (v1 , . . . , v̌i , . . . , vn+1 )
i=1 j =1
! "
× ω1 (vi )a22 (u, vi )Λn1 (vi ; vi , {vm }) + ω2 (vi )a32 (u, vi )Λn2 (vi ; vi , {vm }) ,
(D.11)
ψ2(2) (v1 , v2 , . . . , vn+1 )

n+1
= B1 (v1 ) |Φn−1 (v2 , . . . , v̌i , . . . , vn+1 )Kib1 (v1 , vi , {vm })
i=2

n
n+1
f
+ F (v1 ) |Φn−2 (v2 , . . . , v̌i , . . . , v̌j , . . . , vn+1 )Kij (v1 , vi , vj , {vm })
i=2 j =i+1

n+1
i−1
− α(vj , vi )ω1 (vi )a22 (u, vi )Λn1 (vi ; vi , {vm })|Φn (v1 , . . . , v̌i , . . . , vn+1 )
i=2 j =1

n+1
i−1
− α(vj , vi )ω2 (vi )a32 (u, vi )Λn2 (vi ; vi , {vm })|Φn (v1 , . . . , v̌i , . . . , vn+1 )
i=2 j =1

n+1
= B1 (v1 ) 1 (v1 , vi , {vm })
|Φn−1 (v2 , . . . , v̌i , . . . , vn+1 )K
b
i
i=2

n
n+1
+ F (v1 ) f (v1 , vi , vj , {vm }),
|Φn−2 (v2 , . . . , v̌i , . . . , v̌j , . . . , vn+1 )Kij
i=2 j =i+1
(D.12)
where K b1 are new coefficients which include K b1 and the contributions from the terms of
i i
f . We can easily check that
|Φn (v1 , . . . , v̌i , . . . , vn+1 ). So do Kij
(2) (2)
ψ1 (v1 , . . . , vj , vj +1 , . . . , vn+1 ) = α(vj , vj +1 )ψ1 (v1 , . . . , vj +1 , vj , . . . , vn+1 ),
(D.13)
therefore, from Eqs. (D.8) and (D.10) we should have
ψ2(2) (v1 , . . . , vj , vj +1 , . . . , vn+1 ) = α(vj , vj +1 )ψ2(2) (v1 , . . . , vj +1 , vj , . . . , vn+1 ).

(D.14)
We will determine the coefficients K in Eq. (D.12) with the aid of Eq. (D.14) in
b1 and K f
i ij
the following step. From
(2) (2)
ψ2 (v1 , v2 , . . . , vn+1 ) = α(v1 , v2 )ψ2 (v2 , v1 , . . . , vn+1 ), (D.15)
we get
b1 (v1 , v2 , {vm })
B1 (v1 )|Φn−1 (v3 , v4 , . . . , vn+1 )K2

n+1
+ B1 (v1 ) b1 (v1 , vi , {vm })
|Φn−1 (v2 , . . . , v̌i , . . . , vn+1 )Ki
i=3

n
n+1
+ F (v1 ) f (v1 , vi , vj , {vm })
|Φn−2 (v2 , . . . , v̌i , . . . , v̌j , . . . , vn+1 )Kij
i=2 j =i+1
b1 (v2 , v1 , {vm })
= α(v1 , v2 ){B1 (v2 )|Φn−1 (v3 , v4 , . . . , vn+1 )K2

n+1
+ B1 (v2 ) b1 (v2 , vi , {vm })
|Φn−1 (v1 , v3 , . . . , v̌i , . . . , vn+1 )Ki
i=3

n
n+1
+ F (v2 ) |Φn−2 (v1 , v3 , . . . , v̌i , . . . , v̌j , . . . , vn+1 )
i=1=2 j =i+1=2
(v2 , vi , vj , {vm })}.
×K
f
(D.16)
ij
It can be seen clearly that in the r.h.s. of Eq. (D.16) the first term will receive no any
contributions from the rest terms in the procedure of permuting the spectrum parameters
v1 and v2 with the help of the commutation relations Eqs. (29), (38)–(43). Using l.h.s. of
Eq. (D.16) minus the r.h.s. of Eq. (D.16) and putting v1 to the left side of v2 by necessary
commutation relations, we change Eq. (D.16) into the form of Eq. (94). With the help of
the assumption Eqs. (94), (95), we easily get
b1 = 0.
K2
So we can get the following result
ψ2(2) (v1 , v2 , . . . , vn+1 )

n+1
= B1 (v1 ) b1 (v1 , vi , {vm })
|Φn−1 (v2 , . . . , v̌i , . . . , vn+1 )Ki
i=3

n
n+1
+ F (v1 ) f (v1 , vi , vj , {vm }).
|Φn−2 (v2 , . . . , v̌i , . . . , v̌j , . . . , vn+1 )Kij
i=2 j =i+1
(D.17)
From
(2) (2)
ψ2 (v1 , v2 , v3 , . . . , vn+1 ) = α(v2 , v3 )ψ2 (v1 , v3 , v2 , . . . , vn+1 ), (D.18)
we get
b1 (v1 , v3 , {vm })
B1 (v1 )|Φn−1 (v2 , v4 , . . . , vn+1 )K3

n+1
+ B1 (v1 ) b1 (v1 , vi , {vm })
|Φn−1 (v2 , v3 , . . . , v̌i , . . . , vn+1 )Ki
i=4

n
n+1
+ F (v1 ) (v1 , vi , vj , {vm })
|Φn−2 (v2 , v3 , . . . , v̌i , . . . , v̌j , . . . , vn+1 )K
f
ij
i=2 j =i+1
b1 (v1 , v2 , {vm })
= α(v2 , v3 ){B1 (v1 )|Φn−1 (v3 , v4 , . . . , vn+1 )K3

n+1
+ B1 (v1 ) b1 (v1 , vi , {vm })
|Φn−1 (v3 , v2 , . . . , v̌i , . . . , vn+1 )Ki
i=4

n
n+1
+ F (v1 ) |Φn−2 (v3 , v2 , . . . , v̌i , . . . , v̌j , . . . , vn+1 )
i=2 j =i+1
f (v1 , vi , vj , {vm })}.
×K (D.19)
ij
Changing Eq. (D.19) into the form of Eq. (94) by subtracting the r.h.s. of Eq. (D.19) from
l.h.s. of Eq. (D.19), we then obtain
b1 = 0
K3
with the help of the assumption Eqs. (94), (95). Repeating the same procedure as above,
we achieve
b1 = K
K b1 = · · · = K
b1 = 0.
2 3 n+1
(2)
At this moment, ψ2 become
(2)
ψ2 (v1 , . . . , vn+1 )

n
n+1
= F (v1 ) f (v1 , vi , vj , {vm }).
|Φn−2 (v2 , . . . , v̌i , . . . , v̌j , . . . , vn+1 )Kij
i=2 j =i+1
(D.20)
From Eq. (D.15), we have

n+1
F (v1 ) (v1 , vi , vj , {vm })
|Φn−2 (v3 , . . . , v̌j , . . . , vn+1 )K
f
2j
j =3

n
n+1
+ F (v1 ) f (v1 , vi , vj , {vm })
|Φn−2 (v2 , . . . , v̌i , . . . , v̌j , . . . , vn+1 )Kij
i=3 j =i+1

n+1
= α(v1 , v2 ) F (v2 ) (v2 , vi , vj , {vm })
|Φn−2 (v3 , . . . , v̌j , . . . , vn+1 )K
f
2j
j =3

n
n+1
+ F (v2 ) |Φn−2 (v1 , v3 , . . . , v̌i , . . . , v̌j , . . . , vn+1 )
i=3 j =i+1

f
× Kij (v1 , vi , vj , {vm }) . (D.21)
Using Eqs. (39), (43) to permute v1 and v2 in the r.h.s. of Eq. (D.21) and changing
Eq. (D.21) into the form of Eq. (94) by l.h.s. of Eq. (D.21) minus the r.h.s. of Eq. (D.21),
we can obtain
f = 0,
K j = 3, 4, . . . , n + 1
2j
based on the assumption Eqs. (94), (95). Carrying on the similar steps as that we determine
b1 , we finally get
the coefficients Ki
f = 0,
K i = 2, 3, . . . , n, j = 3, 4, . . . , n + 1, j > i.
ij
Thus, we have the conclusion
(2)
ψ2 (v1 , v2 , . . . , vn+1 ) = 0 (D.22)
and
(2)
ψ (2) (v1 , v2 , . . . , vn+1 ) = ψ1 (v1 , v2 , . . . , vn+1 ). (D.23)
Now the ψ 2 is totally determined. From above deducing procedure we can see that ψ 2 is a
specifically linear combination of specific |Φn ’s and such linear combination is unique.
Carrying on the similar procedure as we analyse the structure of ψ 2 , we can prove that
ψ , ψ 3 , ψ 4 are only the linear combination of the |Φn+1 , |Φn ’s, |Φn−1 ’s, respectively.
1
It is not very hard to deduce the explicit expressions of ψ 1 , ψ 3 , ψ 4 . They are

ψ 1 (v1 , v2 , . . . , vn+1 ) = ω2 (u)Λn+1
2 (u; {vm })|Φn+1 (v1 , . . . , vn+1 ), (D.24)
(3)
ψ (v1 , v2 , . . . , vn+1 )

n+1
i−1
= α(vj , vi )[ω1 (vi )a42 (u, vi )Λn1 (vi ; vi , {vm })
i=1 j =1
+ ω2 (vi )a52 (u, vi )Λn2 (vi ; vi , {vm })]|Φn (v1 , . . . , v̌i , . . . , vn+1 ), (D.25)
(4)
ψ (v1 , v2 , . . . , vn+1 )

n
n+1
i−1 −1
j
= α(vk , vi ) α(vl , vj )|Φn−1 (v1 , . . . , v̌i , . . . , v̌j , . . . , vn+1 )
i=1 j =i+1 k=1 l=1=i

× H1D1 (u, vi , vj )Λn−1
1 (vi ; vi , vj , {vm })
× Λn−1
1 (vj ; vi , vj , {vm })ω1 (vi )ω1 (vj )
+ H2D1 (u, vi , vj )Λn−1
2 (vi ; vi , vj , {vm })
× Λn−1
1 (vj ; vi , vj , {vm })ω2 (vi )ω1 (vj )
D
+ H3 1 (u, vi , vj )Λn−1
1 (vi ; vi , vj , {vm })
× Λn−1
2 (vj ; vi , vj , {vm })ω1 (vi )ω2 (vj )
+ H4D1 (u, vi , vj )Λn−1
2 (vi ; vi , vj , {vm })
× Λn−1
2 (vj ; vi , vj , {vm })ω2 (vi )ω2 (vj ) . (D.26)
In Eq. (D.26), the procedure of obtaining the coefficients of |Φn−1 (v2 , . . . , v̌j , . . . , vn+1 )
is the same as that we have carried out in calculating the two-particle state. At this step, all
the ψ (i) are determined. We can see that Eq. (D.6) can be obtained from Eq. (92) with n
replaced by n + 1, which means that our conclusion Eq. (92) still hold at the case n + 1.
Although we cannot prove our assumption Eqs. (94), (95) so far, we can examine
the conclusions obtained by using the assumption. We verify that the coefficients Kib2 in
Eq. (D.9) are zeroes indeed. Verifying K b1 , Kf in Eq. (D.12) to be zeroes can be simplified
i ij
into the same work as we have done in directly calculating the three-particle state, in which
we have used the explicit expressions of ψc(2) (3)
s and ψcs inferred from Eq. (73). The ψ can
1
be easily verified by six kinds of simple identities. Such terms B2 (v1 )|Φn−1 in ψ 3 are
cancelled out by the similar identities to Kib2 in Eq. (D.9) and such terms B2 (v1 )|Φn−2 in
ψ 4 Eq. (D.26) are also cancelled out by the same identities used in the three-particle state.
The whole procedure above can be applied to the operators A(u), D 2 (u) acting on the
(n + 1)-particle state, respectively. Similarly, given
G(u)|Φn (v1 , . . . , vn ) = ψg(1) + B1 (u)ψg(2) + B2 (u)ψg(3) + F (u)ψg(4) , (D.27)
where ψg(1) is the linear combination of |Φn−2 ’s, ψg(2) and ψg(3) are respectively the
linear combination of |Φn−3 ’s, ψg(4) is a linear combination of |Φn−4 ’s, and all
(i) (i)
the ψg , i = 1, 2, 3, 4 satisfy the following property ψcs (v1 , . . . , vj , vj +1 , . . . , vn ) =
α(vj , vj +1 )ψc(i)
s (v1 , . . . , vj +1 , vj , . . . , vn ), we can prove that Eqs. (D.1), (D.27) still hold
at the n + 1. Here we omit this steps for the sake of simplicity.
References
[1] V.E. Korepin, N.M. Bogoliubov, A.G. Izergin, Quantum Inverse Scattering Method and Correlation
Function, Cambridge Univ. Press, Cambridge, 1993.
[2] C.N. Yang, C.P. Yang, Phys. Rev. 150 (1966) 321;
C.N. Yang, C.P. Yang, Phys. Rev. 150 (1966) 332;
C.N. Yang, C.P. Yang, Phys. Rev. 151 (1966) 258.
[3] M. Takahashi, Prog. Theor. Phys. 46 (1971) 1388.
[4] A. Klümper, R.Z. Bariev, Nucl. Phys. B 458 (1996) 623.
[5] P.A. Pearce, A. Klümper, Phys. Rev. Lett. 66 (1991) 623.
[6] C. Destri, H.J. de Vega, Nucl. Phys. B 438 (1995) 413.
[7] L.D. Faddeev, E.K. Sklyanin, L.A. Takhtajian, Teor. Mat. Fiz. 40 (1979) 194.
[8] P.P. Kulish, E.K. Sklyanin, in: Lecture Notes in Physics, Vol. 151, Springer, Berlin, 1982, p. 61.
[9] H.B. Thacker, Rev. Mod. Phys. 53 (1982) 253.
[10] I.V. Cherednik, Theor. Math. Phys. 17 (1983) 77;
I.V. Cherednik, Theor. Math. Phys. 61 (1984) 911.
[11] E.K. Sklyanin, J. Phys. A 21 (1988) 2375.
[12] H.J. de Vega, Int. J. Mod. Phys. A 4 (1989) 2371.
[13] V. Pasquier, H. Saleur, Nucl. Phys. B 316 (1990) 523.
[14] B.Y. Hou, K.J. Shi, Z.X. Yang, R.H. Yue, J. Phys. A 24 (1991) 3825.
[15] H.J. de Vega, E. Lopes, Phys. Rev. Lett. 67 (1991) 489.
[16] A. Foerster, M. Karowski, Nucl. Phys. B 396 (1993) 611;
A. Foerster, M. Karowski, Nucl. Phys. B 408 (1993) 512.
[17] C. Destri, H.J. de Vega, Nucl. Phys. B 361 (1992) 361;

C. Destri, H.J. de Vega, Nucl. Phys. B 374 (1992) 692.
[18] S.M. Fei, R.H. Yue, J. Phys. A 27 (1994) 3715.
[19] H.J. de Vega, A. Gonzalez-Ruiz, Nucl. Phys. B 417 (1994) 553.
[20] A. Gonzalez-Ruiz, Nucl. Phys. B 424 (1994) 468.
[21] C.M. Yung, M.T. Batchelor, Nucl. Phys. B 435 (1995) 430.
[22] R.H. Yue, H. Fan, B.Y. Hou, Nucl. Phys. B 462 (1996) 167.
[23] N. Andrei, H. Johannesson, Phys. Lett. A 100 (1984) 108.
[24] P.A. Bares, cond-mat/9412011.
[25] G. Bedürfitg, F.H.L. Eßler, H. Frahm, Phys. Rev. Lett. 77 (1996) 5098.
[26] G. Bedürfitg, F.H.L. Eßler, H. Frahm, Nucl. Phys. B 489 (1997) 697.
[27] P. Schlottmann, A.A. Zvyagin, Phys. Rev. B 55 (1997) 5027.
[28] A. Foerster, J. Links, A. Tonel, Nucl. Phys. B 552 (1999) 707.
[29] J. Links, A. Foerster, J. Phys. A 32 (1999) 147.
[30] H. Frahm, A.A. Zvyagin, J. Phys. C 9 (1997) 9939.
[31] Y.P. Wang, Phys. Rev. B 56 (1997) 14045.
[32] Y.P. Wang, J.H. Dai, Z.N. Hu, F.C. Pu, Phys. Rev. Lett. 79 (1997) 1901.
[33] Z.N. Hu, F.C. Pu, Y. Wang, J. Phys. A 31 (1998) 5241.
[34] G. Bedürfitg, H. Frahm, J. Phys. A 32 (1999) 4585.
[35] V.A. Tarasov, Theor. Math. Phys. 56 (1988) 793.
[36] M.J. Martins, Nucl. Phys. B 450 (1995) 768.
[37] P.B. Ramos, M.J. Martins, Nucl. Phys. B 474 (1996) 678.
[38] M.J. Martins, P.B. Ramos, Nucl. Phys. B 500 (1997) 579.
[39] M.J. Martins, P.B. Ramos, Nucl. Phys. B 522 (1998) 413.
[40] A. Lima-Santos, J. Phys. A 32 (1999) 1819.
[41] H. Fan, Nucl. Phys. B 488 (1997) 409.
[42] X.W. Guan, J. Phys. A 33 (2000) 5391.
[43] A.G. Izergin, V.E. Korepin, Commun. Math. Phys. 79 (1981) 303.
[44] M.T. Batchelor, C.M. Yung, Phys. Rev. Lett. 74 (1995) 2026.
[45] N.Yu. Reshetikhin, Lett. Math. Phys. 7 (1983) 303.
[46] L. Mezincescu, R. Nepomechie, Nucl. Phys. B 372 (1992) 597.
[47] W.L. Yang, Y.Z. Zhang, Nucl. Phys. B 596 (2001) 495.
[48] E.C. Fireman, A. Lima-Santos, W. Utiel, Nucl. Phys. B 626 (2002) 435.
[49] R.J. Baxter, Exactly Solved Models in Statistical Mechanics, Academic Press, London, 1982.
[50] L. Mezincescue, R.T. Nepomechie, J. Phys. A 24 (1991) L17.
[51] H. Fan, B.Y. Hou, G.L. Li, K.J. Shi, J. Phys. A 32 (1999) 6021.
[52] R.I. Nepomechie, J. Phys. A 33 (2000) L21.
The N = 4 SYM integrable super spin chain

Niklas Beisert, Matthias Staudacher
Max-Planck-Institut für Gravitationsphysik, Albert-Einstein-Institut, Am Mühlenberg 1,
D-14476 Golm, Germany
Received 30 July 2003; accepted 12 August 2003
Abstract
Recently it was established that the one-loop planar dilatation generator of N = 4 super-Yang–
Mills theory may be identified, in some restricted cases, with the Hamiltonians of various integrable
quantum spin chains. In particular Minahan and Zarembo established that the restriction to scalar
operators leads to an integrable vector so(6) chain, while recent work in QCD suggested that
restricting to twist operators, containing mostly covariant derivatives, yields certain integrable
Heisenberg XXX chains with non-compact spin symmetry sl(2). Here we unify and generalize these
insights and argue that the complete one-loop planar dilatation generator of N = 4 is described by
an integrable su(2, 2|4) super spin chain. We also write down various forms of the associated Bethe
ansatz equations, whose solutions are in one-to-one correspondence with the complete set of all
one-loop planar anomalous dimensions in the N = 4 gauge theory. We finally speculate on the non-
perturbative extension of these integrable structures, which appears to involve non-local deformations
of the conserved charges.
PACS: 11.15.-q; 11.25.Hf; 11.25.Tq; 02.30.Ik; 75.10.Pq
1. Introduction and overview
Of all four-dimensional gauge field theories the one with the maximum number N = 4
[1] of supersymmetries appears to be very special in many ways. At first sight this model
appears to be incredibly complicated due to the very large number of fields of different
kinds, but upon closer inspection many hidden simplicities appear. It is believed that the
model is exactly conformally invariant at the quantum level [2] and thus a non-trivial
example (as opposed to free massless field theory) of a four-dimensional conformal field
E-mail addresses: nbeisert@aei.mpg.de (N. Beisert), matthias@aei.mpg.de (M. Staudacher).
doi:10.1016/j.nuclphysb.2003.08.015
440 N. Beisert, M. Staudacher / Nuclear Physics B 670 [FS] (2003) 439–463
theory (CFT). Given the huge successes in understanding CFT’s in two dimensions, one
might hope that at least some of the aspects allowing their treatment in D = 2 might
fruitfully reappear in D = 4. One of the many intriguing features of two-dimensional
CFT’s is that they are intimately connected to integrable (2 + 0)-dimensional lattice
models in statistical mechanics, or, equivalently, to (1 + 1)-dimensional quantum spin
chains. Integrable models are even more important for the study of integrable massive
deformations of these CFT’s. Thus, an optimistic mind could hope that integrability might
also play a rôle in the putative simplicity of a theory such as N = 4 Yang–Mills. Excitingly,
a number of recent discoveries lend much support to this idea. One might wonder about
standard, naive no-go theorems that seem to suggest that integrability can never exist
above D = 2. These may be potentially bypassed by the fact that there appears to be a
hidden “two dimensionality” in SU(N) SYM4 when we look at it at large N , i.e., when
we introduce N1 as an additional coupling constant in the theory, aside from the ’t Hooft
coupling λ = gYM 2 N.
In fact, one might interpret the AdS/CFT correspondence [3], believed to be exact in
the N = 4 case, as one important indication of the validity of this idea: again following
’t Hooft, the worldsheet of the string is a clear candidate for explaining the hidden two
dimensions. Interestingly, first indications that the world sheet theory, highly non-trivial
due to the curved AdS5 × S 5 background, might be integrable have recently appeared [4,5]
(for the simpler but related case of pp-wave backgrounds see also [6]).
However, let us get back to the gauge theory and review how integrable structures
have recently emerged directly, without resorting to the correspondence. Following the
seminal BMN paper [7] it became clear that, despite much work on SYM4 , not much was
known about how to efficiently calculate the anomalous dimensions of local, non-protected
conformal operators composed of an arbitrary number of elementary fields. In particular,
in [8–10] effective vertex techniques were developed in order to efficiently calculate one-
loop two-point functions of operators composed of an arbitrary numbers of scalars. From
these two-point correlators the scaling dimensions could then be (indirectly) extracted.
The main focus was on the case where the classical dimension of the operators tended
to infinity, but the technique was applicable to the finite case as well. In particular, [9]
contained the full so(6) invariant effective vertex which was used in [11] to find the exact
finite-dimensional generalizations of the BMN operators. A significant simplification took
place when it was realized [12,13] that the effective vertex could be directly interpreted
as a Hamiltonian acting on a Hilbert space formed by all possible scalar operators of
fixed classical dimension. Therefore it was no longer necessary to extract the anomalous
dimensions from a two-point function; instead one could directly go about diagonalizing
the Hamiltonian. This was done in [13] for a Hilbert space containing multi-trace gauge-
invariant scalar operators; a natural split between a planar “free” part and a non-planar
trace-splitting and trace-joining “interaction” part appeared, where the interaction is purely
cubic with coupling constant N1 . This split turned out to be very reminiscent of similar
structures appearing in cubic string field theories. In turn, Minahan and Zarembo focused
on the planar part of the Hamiltonian and noticed the remarkable fact that it is identical
to the one of an so(6) invariant integrable spin chain. The integrablity leads, via a hidden
Yang–Baxter symmetry, to the appearance of an infinite number of charges commuting
with the Hamiltonian. This allows to write down Bethe equations whose solution furnish
N. Beisert, M. Staudacher / Nuclear Physics B 670 [FS] (2003) 439–463 441
the spectrum, i.e., the planar anomalous dimensions. For low-dimensional operators (i.e.,
those containing only a few elementary fields) the approach is not necessarily easier than
direct, brute force diagonalization of the Hamiltonian (however, it does lead to a very
transparent and short derivation of the finite J BMN operators found in [11]). Its real
power lies in the fact that it allows to obtain anomalous dimensions in the limit of a large
number of constituent fields.
This power was very recently illustrated in a study of scalar operators belonging to the
so(6) representations [J2 , J1 − J2 , J2 ] with ∆0 = J1 + J2 in the limit where both charges
J1 , J2 become large [14]. The ground state energy in this representation is argued to be
related to the minimal energy solution of a string rotating in two planes in S 5 , which can
be calculated exactly [15,16]. The result for this energy is given by inverting certain elliptic
functions, it is a non-trivial function of the parameters J1 , J2 , and agrees on both sides. On
the gauge theory side, we believe it can only be obtained by using the Bethe ansatz. This
is arguably the most subtle dynamical, quantitative test of AdS/CFT existing to date.
Integrable spin chains had appeared before in gauge theories through the pioneering
work of Lipatov on high energy scattering in planar QCD [17]. The model was
subsequently identified as a XXX Heisenberg sl(2) spin chain of non-compact spin
zero [18]. More recently, and physically closely related to the present study, further
integrable structures were discovered by Belitsky, Braun, Derkachov, Korchemsky and
Manashov in the computation of planar one-loop anomalous dimensions of various types
of quasi-partonic operators in QCD1 [19]. Subsequently, Kotikov and Lipatov [20] applied
the corresponding integrable structure, originally found in the QCD context, to the
computation of anomalous dimensions of twist-two operators (i.e., operators containing
two scalar fields and an arbitrary number of symmetrized, traceless covariant derivatives
acting on them) in SYM4 . The result for the anomalous dimension is given in terms of
harmonic numbers and was verified by Dolan and Osborn using totally different methods
[21]. This result was very useful in (qualitatively) comparing certain string states in
AdS5 to twist-two operators with a large spin in the gauge theory [22]. Very recently
Belitsky, Gorsky and Korchemsky [23] considered QCD composite light-cone operators,
corresponding to operators of arbitrarily high twist, and discussed the hidden integrability
of the associated dilatation operator.
Let us ask the question whether the integrable structures appearing in the planar gauge
theory, when one investigates the first radiative corrections to anomalous dimensions, are
accidental or rather more generic and thus indicative of a deeper underlying structure.
The preliminary evidence from the correspondence with strings [4] certainly points to the
second possibility. And indeed a first study of the two-loop corrections to the integrable
Hamiltonian yields intriguing evidence that these do not break the integrability property
[24]. One indirect way to detect the hidden commuting charges is to look for unexpected
degeneracies in the energy spectrum. We found that these degeneracies exist in planar
SYM4 , can be explained at one loop by the hidden integrability, and are not broken at the
1 While QCD is surely not a conformally invariant quantum field theory, it still behaves like one as far as one-
loop anomalous dimensions are concerned. We thank A. Belitsky, A. Gorsky and G. Korchemsky for pointing
this out to us. In one-loop QCD, integrability is not complete; in particular it requires aligned helicities of the
partonic degrees of freedom.
two-loop level. It is important to note that, while the one-loop Hamiltonian for scalar fields
is exactly integrable [12], the combined one- and two-loop dilatation operator is not exactly
integrable and requires the inclusion of three-loop terms, and so on [24]. Interpreting
these corrections in the spin chain language one finds that the higher-loop deformations
of the charges are in a certain sense non-local, reminiscent of [4]. The integrable system
corresponding to the all-loop planar dilatation operator of N = 4 Yang–Mills, if it exists,
should be some kind of long-range spin chain and has yet to be identified. In particular it
is unclear how to extend the Bethe ansatz of [12] to include the higher loop effects.
All this suggests that, maybe, planar N = 4 Yang–Mills is integrable. If this is the
case, the hidden symmetries should extend the known symmetries of SYM4 . The full
symmetry algebra of SYM4 is neither so(6) nor sl(2), but the full superconformal algebra
psu(2, 2|4).2 If the integrable structures discovered to date are not accidental, we should
expect that the so(6) one-loop results of Minahan–Zarembo and the sl(2) results suggested
from one-loop QCD [19,23] (see also [20]) can be combined and “lifted” to a full
psu(2, 2|4) super spin chain, as first conjectured3 in [24]. In this paper we will argue that
this is indeed the case. In a companion paper, one of us worked out the complete one-
loop dilatation operator of N = 4 Yang–Mills theory [25]. That is, the effective vertex (not
necessarily restricted to the planar case) allowing the computation of mixing matrices of
composite operators containing arbitrary sequences of elementary fields (scalars, covariant
derivatives, field strengths, and fermions) was derived. Upon diagonalization of this matrix
(which in practice is however only possible for states of small classical dimension) its
eigensystem yields the eigenoperators and their anomalous dimensions. Here we will
establish that the restriction of this complete one-loop dilatation operator to the planar
case yields the promised su(2, 2|4)4 spin chain. Our arguments contain some assumptions
which should be filled in by experts on integrability; in particular, while we propose an
R-matrix, we will assume its existence and uniqueness. At any rate we then proceed to
write down the Bethe ansatz equations expected to describe our super spin chain. Again,
it is not clear to us whether the validity of the ansatz has been rigorously proven in the
spin chain literature. However, by analyzing the first few states in our superchain we find
complete agreement with the gauge theory results. Therefore we are confident that the
details of our Bethe ansatz are indeed correct. A subtlety, to be discussed below, is that for
superalgebras one can write down several Bethe ansätze, which appear to be different and
possess different (pseudo)vacuum states. They nevertheless yield an identical spectrum.
We will discuss two specific ansätze which are both natural from different points of view.
The outline of this paper is as follows. We begin with a short review/preview on Bethe
ansatz equations, and how they are determined by the Dynkin diagram of the symmetry
algebra of the chain in conjunction with the highest weight representation of the degrees of
freedom distributed along the chain. We will also introduce some useful, compact notation
2 The conjugation properties of the algebras are not relevant here, we will always refer to complex algebras,
e.g., sl(4|4) = su(2, 2|4), so(6) = sl(4), etc.
3 In [23] it was conjectured that the sl(2) and so(6) spin chains combine into a so(2, 4) × so(6) = su(2, 2) ×
su(4) chain. As we shall see, we need the full psu(2, 2|4) algebra in order to achieve this “grand unification”.
4 At the one-loop level the superconformal algebra psu(2, 2|4) may be extended by a gl(1) charge to the
semi-direct product su(2, 2|4).
for writing the Bethe equations. We explain how to employ the latter in order to derive
anomalous dimensions in N = 4 Yang–Mills. Next we present our arguments, via the
existence of a unique R-matrix, why one expects the complete one-loop dilatation operator
of [25] to become integrable in the large N limit. The procedure naturally incorporates
the sl(2) (twist operators) and the so(6) (Minahan–Zarembo) integrable structures. For
super chains we can write seemingly different sets of Bethe ansatz equations. We will pick
two examples: the first (which we call “Beauty”) does not correspond to the standard way
of choosing the fermionic root(s) of the Dynkin diagram. However, this formulation is
particularly useful since the vacuum states of the chain correspond to the half-BPS states
of SYM4 . To illustrate the freedom of choosing various forms of the Bethe equations we
then present a second version (“the Beast”) which employs the standard “distinguished”
choice of one fermionic root in the middle of the diagram. As we will see, this shifts the
vacuum to a different, high energy state containing only field strengths. It obscures certain
well known results (e.g., it becomes hard to see that half-BPS states have vanishing energy)
but instead leads to some further interesting spectroscopic results. We end with an outlook
on interesting applications and extensions of this work.
2. Anomalous dimensions, (super) spin chains, and Bethe ansätze
Let us review how anomalous dimensions in YM4 are computed by employing the
algebraic Bethe ansatz, once it has been established that the anomalous part of the dilatation
operator is identical to the Hamiltonian of an integrable quantum spin chain [12]. Let
us explain this in the simplest case of an sl(2) chain, the so-called XXXs/2 Heisenberg
chain. (For a very pedagogical introduction, see [26].) The proposal is that the energy
eigenvalues E of the Hamiltonian are proportional to the anomalous dimensions δ∆ of
conformal scaling operators:
2 N
gYM
δ∆ = E. (2.1)
8π 2
For spin 2s = 12 these scaling operators are, in the planar limit, linear combinations of
a single trace of sequences of two of the three complex scalars of SYM4 , say Z and φ:
Tr(ZZZφZφφ . . .). (2.2)
The single trace operator is then interpreted as a chain of spins, where Z and φ represent,
respectively, up and down spins. The total number L of complex scalars inside the trace
corresponds to the length of the chain. Each eigenstate (alias scaling operator) of the
Hamiltonian is uniquely characterized by a set of Bethe roots uj , j = 1, . . . , n, and the
energy E of the state is given by

n
s
E=± (2.3)
u2 + 14 s 2
j =1 j
(where for spin 2s = 12 we should pick the plus sign in front of the sum). The Bethe roots
are found by solving the Bethe equations for j = 1, . . . , n

uj + 2i s L n
uj − uk + i
= . (2.4)
uj − 2i s uj − uk − i
k=1
k =j
The vacuum state of energy E = 0 is the half-BPS state

|0
= Tr Z L , (2.5)
i.e., it is interpreted as the ferromagnetic ground state (all spins are up, no Bethe roots) of
the chain. The excitation number n, giving the total number of roots, counts the number of
φ’s or down spins along the chain; it is naturally bounded by n L. There is an additional
constraint on the Bethe roots:

n
uj + 2i s
1= . (2.6)
j =1
uj − 2i s
For the spin chain, it means that we have periodic boundary conditions and we are only
looking for zero-momentum states. In the gauge theory interpretation (2.2) it expresses the
cyclicity of the trace.
The second simplest example concerns twist operators, closely related to quasi-partonic
operators in one-loop QCD [19,20,23]. Just as the fields in (2.2) form a closed subsector
(i.e., they do not mix with any other types of fields) the following operators form another
such subsector (see also [25] for further details):

Tr Dm1 Z Dm2 Z · · · DmL Z , (2.7)
where D is the “light-cone” covariant derivative D1+i2 (as in [23]). This is still a spin chain
of length (twist) L, with non-compact spin 2s = − 12 , as will be proven in the next section.
Here the spins at each lattice site k may take any value mk = 0, 1, 2, . . . , as we have an
infinite s = −1 representation of sl(2). Also the total excitation number n = mk is not
bounded as in the above example. The vacuum is still Tr Z L . Note that in this example the
length of the chain is not equal to the classical dimension L + n of the operator. Again,
the anomalous dimensions of the scaling operators formed from the states (2.7) are given
via Eqs. (2.1), (2.3), (2.4), (2.6)! (Here we have to choose the negative sign in the energy
formula (2.3).)
In the above example the algebra is sl(2) and thus of rank one. There is a beautiful
extension of the Bethe equations to any algebra and any representation of the elementary
constituents of the chain at each lattice site [27,28]. This general form easily extends to
the case of super algebras as well, see [29] and references therein. This is what we need
when we are looking for a spin chain describing SYM4 at one loop, where we expect
Bethe equations for the super algebra sl(4|4). This general equation may be written in the
following compact notation, based on knowing the Dynkin diagram of the algebra. The
Dynkin diagram of sl(4|4) contains seven dots corresponding to a choice of seven simple
roots. Consider a total of n excitations. For each of the corresponding Bethe roots ui ,
i = 1, . . . , n, we specify which of the seven simple roots is excited by kj = 1, . . . , 7. The
Bethe equations for j = 1, . . . , n can then be written in the compact form

L
n
uj + 2i Vkj uj − ul + 2i Mkj ,kl
= . (2.8)
uj − 2i Vkj l=1
uj − ul − 2i Mkj ,kl
l =j
Here, Mkl is the Cartan matrix of the algebra and Vk are the Dynkin labels of the
spin representation. Furthermore, we still consider a cyclic spin chain with zero total
momentum. This gives the additional constraint

n
uj + 2i Vkj
1= . (2.9)
u − 2i Vkj
j =1 j
The energy of a configuration of roots that satisfies the Bethe equations and constraint is
now given by5
n

i i
E = cL ± − . (2.10)
j =1
uj + 2i Vkj uj − 2i Vkj
It is easily seen that restricting these equations to the Dynkin diagram of the algebra so(6)
reproduces the Bethe equations of [12]. It will turn out, see below, that these general
equations, which are well known in the literature on integrable spin chains, indeed solve
the entire problem of computing planar anomalous dimensions in SYM4 , once we (i)
identify the correct representations of the fundamental fields on the lattice sites, and (ii)
after resolving certain subtleties concerning Dynkin diagrams for superalgebras. However,
let us first show that we expect planar one-loop SYM4 to be completely integrable indeed.
3. The R-matrix
In this section we derive the R-matrix for the integrable spin chain considered in
this work. For this we make use of a special subsector of the spin chain with residual
sl(2) symmetry and show how to lift the universal sl(2) R-matrix to an sl(4|4) invariant
R-matrix. The derived Hamiltonian is shown to agree with the complete one-loop planar
dilatation generator of N = 4 SYM, thus proving the integrability of the latter.
The planar one-loop dilatation operator The complete one-loop planar dilatation
generator of N = 4 SYM acting on two fields is given by [25]
∞

H12 = 2h(J12 ) := 2h(j )P12,j . (3.1)
j =0
5 The Bethe equations determine the energy only up to scale and a shift cL where c is a constant to be fixed
such that the energy corresponds to planar anomalous dimensions of N = 4 SYM via (2.1).
Here, h(j ) is the j th harmonic number defined alternatively by a series or by the digamma
function Ψ (x) = ∂ log Γ (x)/∂x

n
1
h(n) = = Ψ (n + 1) − Ψ (1). (3.2)
k
k=1
The operator P12,j projects the spins at sites 1, 2 to the module Vj . Each spin belongs to
the ‘fundamental’ module VF , consequently the modules Vj arise in the tensor product of
two fundamental modules
∞

VF × VF = Vj . (3.3)
j =0
The primary weights of these modules are given by (see (4.4) for details)
wF = [1; 0, 0; 0, 1, 0; 0, 1], w0 = [2; 0, 0; 0, 2, 0; 0, 2],

w1 = [2; 0, 0; 1, 0, 1; 0, 2], wj = [j ; j − 2, j − 2; 0, 0, 0; 0, 2]. (3.4)
The sl(2) subsector A crucial point in the derivation of the full Hamiltonian density [25]
was the restriction to a subsector. This subsector is obtained by restricting the allowed
weights of states to
w = [L + s; s, s; 0, L, 0; 0, L]. (3.5)
For the construction of such states we must restrict all spins to the form
1 n
zn := D Z, with D := D1 + iD2 , Z := Φ5 + iΦ6 . (3.6)
n!
The residual symmetry algebra within the subsector is sl(2) which can be represented by
the generators
1
J− = ∂, J3 =
+ z∂, J+ = z + z2 ∂, (3.7)
2
where ∂ is the derivative with respect to z. It was then shown that the modules VF , Vj of
the full theory correspond to the sl(2) modules VF , Vj with highest weights
wF = [−1], wj = [−2 − 2j ]. (3.8)

Here, the weights are described by twice the spin. The Hamiltonian density (3.1) restricts
within this subsector to
∞

H12 = 2h(J12 ) := 2h(j )P12,j , (3.9)
j =0

where P12,j projects states of the tensor product VF × VF to the module Vj . Interestingly,
this is the Hamiltonian of the XXX−1/2 Heisenberg spin chain. In other words it is an
integrable spin chain with unbroken sl(2) symmetry, where each site transforms in a spin
wF /2 = − 12 representation.
Integrability within the sl(2) subsector To prove this statement we make use of the
universal R-matrix of sl(2) spin chains. This sl(2) invariant operator can be decomposed
into its irreducible components corresponding to the modules Vj
∞

R12 (u) = Rj (u)P12,j

. (3.10)
j =0
The eigenvalues Rj (u) of the sl(2) universal R-matrix were determined in [30]. In a spin
wj /2 = −1 − j representation the eigenvalue is
Γ (−j − cu)
Rj (u) = (−1)j +1 f (cu). (3.11)
Γ (−j + cu)
The arbitrary function f (u) and normalization constant c reflect the trivial symmetries
of the Yang–Baxter equation. The induced eigenvalues of the Hamiltonian density are
obtained as the logarithmic derivative of Rj (u) at u = 0 times the permutation operator
P12

∂ log Rj (u)

H12 Vj = P12 V . (3.12)
∂u j
u=0
We note that for even (odd) j the composite module Vj is a (anti)symmetric combination
of two VF , consequently the permutation acts as
P12 Vj = (−1)j Vj . (3.13)
We choose the function and constant to be
Γ (−cu)
f (cu) = − , c = 1. (3.14)
Γ (cu)
Using the fact that j is exactly integer we find the eigenvalues of the Hamiltonian density

∂ log Rj (u)
(−1)j = 2Ψ (j + 1) − 2Ψ (1) = 2h(j ). (3.15)
∂u
u=0
This proves that the Hamiltonian density (3.9) is integrable.
Integrability of the full sl(4|4) R-matrix To derive an R-matrix for the full sl(4|4) spin
chain we will assume that for given representations of the symmetry algebra there exists a
unique R-matrix which satisfies the Yang–Baxter equation (modulo the symmetries of the
YBE). We are not aware whether this claim [30] has been proven. Let R12 be this R-matrix
for the sl(4|4) integrable spin chain. The R-matrix is an invariant operator, thus it can be
reduced to its irreducible components corresponding to the modules Vj
∞

R12 (u) = Rj (u)P12,j . (3.16)
j =0
The restriction R of the R-matrix to the sl(2) sector must also satisfy the Yang–Baxter
equation. The unique solution for the eigenvalues of R is (3.11). Due to the one-to-one
correspondence of modules Vj and Vj ,
Vj ⊂ Vj (3.17)
the eigenvalues of the unique sl(4|4) R-matrix (3.16) must be
Γ (−j − cu)
Rj (u) = (−1)j f (cu). (3.18)
Γ (−j + cu)
As in (3.15) this R-matrix yields H12 = 2h(J12 ) in agreement with (3.1). This in turn
shows that the planar one-loop dilatation generator of N = 4 is integrable. Note however,
that this proof is based on the assumption of the existence of a unique R-matrix.
The Minahan and Zarembo chain As an application let us investigate the restriction of
the R-matrix to the so(6) subsector investigated by Minahan and Zarembo [12]
w = [L; 0, 0; q1, p, q2 ; 0, L], (3.19)
i.e., where the spins are given by the N = 4 SYM scalars Φm . Two scalars Φp , Φq
can transform in three different irreducible representations of so(6), symmetric-traceless,
antisymmetric and singlet. These correspond to the modules V0 , V1 , V2 , respectively.
We set the normalization function and constant to
Γ (3 + cu)
f (u) = , c=1 (3.20)
2Γ (1 − cu)
and compute the eigenvalues of the R-matrix (3.18) for V0 , V1 , V2
1
R12,0 (u) = (u + 1)(u + 2),
2
1
R12,1 (u) = (u − 1)(u + 2),
2
1
R12,2 (u) = (u − 1)(u − 2). (3.21)
2
We note the projectors to V0 , V1 , V2 in the scalar sector
1 1 1 1 1 1
P12,0 = + P12 − K12 , P12,1 = − P12 , P12,2 = K12 , (3.22)
2 2 6 2 2 6
where K12 is the so(6)-trace operator defined by K12 Φ1,m Φ2,n = δmn Φ1,p Φ2,p . Assem-
bling (3.21) and (3.22) we obtain the R-matrix operator (3.16)
1 1 1
R12 (u) = (u + 2)P12 + u(u + 2) − uK12 . (3.23)
2 2 2
This is precisely the R-matrix of the Minahan and Zarembo chain [12] which has
previously been found by Reshetikhin [27].
4. Beauty . . .
In the last section we have established that the planar one-loop dilatation operator of
N = 4 YM4 is integrable. We therefore expect the general Bethe ansatz equations (2.8) to
hold. However, for them to be useful, we still need to specify the Dynkin labels, the Cartan
matrix and precise form of the energy (2.10). Furthermore, we will perform a check of the
validity of this sl(4|4) Bethe ansatz which goes beyond the so(6) subsector.
Representations First, we need to specify the Cartan matrix, determined by the Dynkin
diagram, and the Dynkin labels of the spin representation corresponding to the module
VF . For classical semi-simple Lie algebra the Dynkin diagram is unique. In the case of
superalgebras, however, there is some freedom to distribute the simple fermionic roots. In
the context of N = 4 SYM one particular choice of Dynkin diagram turns out to be very
useful [31]:6
(4.1)
On top of the Dynkin diagram we have indicated the Dynkin labels of the spin
representation. We write the Cartan matrix corresponding to this choice of Dynkin diagram
and the representation vector as7
   
−2 +1 0
   
 +1 −1  0
   
 −1 +2 −1  0
   
 
M = −1 +2 −1 , V = 1 .
 (4.2)
  
 −1 +2 −1   0
   
  0
 −1 +1 
+1 −2 0
The energy corresponding to a solution to the Bethe equations is

n
i i
E= − . (4.3)
j =1
Excitation numbers Finally, we need to obtain the number of excitations nk , k = 1, . . . , 7

of the individual simple roots for a state with a given weight
w = [∆0 ; s1 , s2 ; q1 , p, q2 ; B, L]. (4.4)
A weight of (the classical) sl(4|4) is described by various labels. The (bare) dimension is
indicated by ∆0 . The sl(4) and sl(2) × sl(2) Dynkin labels are given by [q1 , p, q2 ] and
6 We thank V. Dobrev for this hint, and for informing us about Ref. [31].
7 In fact, the Cartan matrix is obtained from this by inverting some lines. The Bethe equations are invariant
under the inversion and it is slightly more convenient to work with a symmetric matrix M.
[s1 , s2 ], respectively, where s1 , s2 equal twice the spin. Furthermore, B describes the gl(1)
hypercharge or chirality in the semi-direct product sl(4|4) = gl(1) psl(4|4). Finally, L,
which is not a label of sl(4|4), gives the length or number of spins. In N = 4 SYM the
chirality and length of an operator are not good quantum numbers, they are broken at
higher loops by the Konishi anomaly. Nevertheless, at one-loop order they are conserved
and their leading order values can be used to describe a state.
This is most easily seen in the oscillator picture in [25] using the physical vacuum |Z
.
We write down the action of the generators corresponding to the simple roots in terms of
creation and annihilation operators
(4.5)
It is now clear that n1 = na† , n2 = na† + na† and so on. Using the formulas in [25] we write
2 1 2
down the corresponding excitation numbers of the simple roots
1 
2 ∆0 − 2 (L − B) − 2 s1
1 1
 ∆0 − (L − B) 
 
 ∆0 − 1 (L − B) − 1 p − 3 q1 − 1 q2 
 2 2 4 4 
 
nk =  ∆0 − p − 12 q1 − 12 q2  . (4.6)
 
 ∆0 − 1 (L + B) − 1 p − 1 q1 − 3 q2 
 2 2 4 4 
 ∆0 − (L + B) 
2 ∆0 − 2 (L + B) − 2 s2
1 1 1
Not all excitations of the simple roots correspond to physical states. Obviously, the
excitation numbers of the oscillators must be non-negative, this gives the bounds
0 n1 n2 n3 n4 n5 n6 n7 0. (4.7)
Furthermore, each fermionic oscillator cannot be excited more than once, this gives the
bounds8
n2 + 2L n3 + L n4 n5 + L n6 + 2L. (4.8)
Certainly, we should obtain the so(6) = sl(4) subsector studied by Minahan and
Zarembo [12] when we remove the outer four simple roots from the Dynkin diagram in
(4.1). When we restrict to the states (3.19) of this subsector the number of excitations (4.6)
of the outer four roots is trivially zero. They become irrelevant for the Bethe ansatz and
can be discarded. Thus all solutions to the so(6) Bethe equations are also solutions to the
sl(4|4) Bethe equations. What is more, we can apply this Bethe ansatz to a wider range of
operators, in fact, to all single-trace operators of N = 4 SYM.
8 Superconformal primaries reside in the fundamental Weyl chamber defined by the bounds −2n + n > −1,
1 2
n2 − 2n3 + n4 > −1, n3 − 2n4 + n5 + L > −1, n4 − 2n5 + n6 > −1, n6 − 2n7 > −1. Together with (4.7) this
implies, among other relations, (4.8). Solutions of the Bethe equations outside the fundamental domain apparently
correspond to mirror images of primary states due to reflections at the chamber boundaries. (We thank J. Minahan
and K. Zarembo for this insight.)
Multiplet splitting Now we can write down and try to solve the Bethe equations for any
state in N = 4 SYM. Note, however, that the Bethe equations need to be solved only for
highest weight states. All descendants of a highest weight state are obtained by adding
Bethe roots at infinity, ui = ∞. In other words, the solutions to the Bethe equations
corresponding to highest weight states are distinguished in that they have no roots ui at
infinity. Nevertheless, there is one subtlety related to this point which can be used to our
advantage. Namely this is multiplet splitting at the unitarity bounds [32]. We assume that
the spin chain of L sites transforms in the tensor product of L spin representations. The
corrections δ∆ to the scaling dimension induced by the Hamiltonian H are not included in
this picture. Thus, in terms of the spin chain only the classical sl(4|4) algebra applies where
the scaling dimension is exactly ∆0 . Long multiplets of the interacting sl(4|4) algebra close
to one or both of the unitarity bounds
I: ∆0 = 2 + s1 + p + 32 q1 + 12 q2 ,
II: ∆0 = 2 + s2 + p + 12 q1 + 32 q2 (4.9)
split up into several semi-short or BPS multiplets. The weights of the additional primary
states are offset by

[+0.5; −1, 0; +1, 0, 0; −0.5, +1], for s1 > 0,
δwI =
[+1.0; 0, 0; +2, 0, 0; 0.0, +1], for s1 = 0,

[+0.5; 0, −1; 0, 0, +1; +0.5, +1], for s2 > 0,
δwII = (4.10)
[+1.0; 0, 0; 0, 0, +2; 0.0, +1], for s2 = 0.
The unitarity bounds can also be expressed in terms of excitations of simple roots, we find
I: n1 + n3 = n2 + 1,
II: n7 + n5 = n6 + 1. (4.11)
The corresponding offsets translate into

(0, −1, −1, 0, 0, 0, 0), for n2 > n1 ,
δnI =
(0, 0, −1, 0, 0, 0, 0), for n2 = n1 ,

(0, 0, 0, 0, −1, −1, 0), for n6 > n7 ,
δnII = (4.12)
(0, 0, 0, 0, −1, 0, 0), for n6 = n7 ,
together with an increase of the length L by one. We thus see that in the case of multiplet
shortening the primaries of higher submultiplets have less excitations. In a calculation this
may reduce the complexity of the Bethe equations somewhat as we shall see in an example
below.
Multiplet splitting is an extremely interesting issue from the point of view of
integrability. Let us consider some operator acting on a spin chain. Assume the operator
is invariant under the classical algebra sl(4|4). In the most general case, this operator can
assign a different value to all irreducible multiplets of states. In particular this is so for the
submultiplets of a long multiplet at the unitarity bound. Now, if we impose integrability
on the operator all these submultiplets become degenerate.9 A priori, this seems like a
miracle. Why should integrability imply this degeneracy? It almost seems as if integrability
selects that scalar operator which is suitable as a consistent deformation of the dilatation
generator! Then, clearly the miracle would turn into the condition for integrability. If this
can be made sense of, maybe it also helps to understand higher loop corrections in the light
of integrability [24].
Example Let us apply the Bethe ansatz to the twist-two operator with primary weight
w = [4; 2, 2; 0, 0, 0; 0, 2]. (4.13)

Using (4.6) we find the excitation numbers and length
n0,k = (0, 2, 3, 4, 3, 2, 0), L0 = 2. (4.14)

This weight is on both unitarity bounds, (4.11), the excitation numbers of the highest
submutiplet, (4.12), are
nk = (0, 1, 2, 4, 2, 1, 0), L = 4. (4.15)

We therefore configure the simple roots as follows
kj = (2, 3, 3, 4, 4, 4, 4, 5, 5, 6). (4.16)

Now we note that the twist-two operators are unpaired states. Therefore, the configurations
of Bethe roots must be invariant under the symmetry uj → −uj of the Bethe equations.
This tells us
u1 = u10 = 0, u2 = −u3 , u4 = −u5 , u6 = −u7 ,

u8 = −u9 ,
(4.17)
which automatically satisfies the momentum constraint (2.9). Furthermore the excitations
(4.14) are invariant under flipping the Dynkin diagram, nk → n8−k . This can be used to set
u2 = u8 . (4.18)
The Bethe equations (2.8) are then solved exactly by
√
5 65 ± 4 205
u2 = , u4,6 = , (4.19)
7 140
which yields the energy (4.3)
25 g2 N 25gYM2 N
E= , δ∆ = YM2 E = . (4.20)
3 8π 24π 2
This is the energy of the twist-two state at dimension four [33], see also Table 1.
9 This is because the unique integrable operator is equivalent to the one-loop planar correction to the dilatation
operator of N = 4 SYM. In N = 4 SYM the submultiplets must rejoin into a long multiplet which would be
inconsistent if their one-loop anomalous dimensions were different.
Table 1
All energies E of primary states with ∆0 4, see [25]. The parity P is related to parity under inversion of the
spin chain, parity ± indicates a degenerate pair. The label ‘+conj.’ indicates conjugate states with sl(2)2 , sl(4)
labels reversed and opposite chirality B
∆0 sl(2)2 sl(4) B L EP
2 [0, 0] [0, 2, 0] 0 2 0+
[0, 0] [0, 0, 0] 0 2 6+
3 [0, 0] [0, 3, 0] 0 3 0−
[0, 0] [0, 1, 0] 0 3 4−
4 [0, 0] [0, 4, 0] 0 4 0+ √
[0, 0] [0, 2, 0] 0 4 (5 ± 5 )+
[0, 0] [1, 0, 1] 0 4 6− √
[0, 0] [0, 0, 0] 0 4 1 (13 ± 41 )+
2
[2, 0] [0, 0, 0] 1 3 9− + conj.
[1, 1] [0, 1, 0] 0 3 15 ±
2
[2, 2] [0, 0, 0] 0 2 25 +
3
5. . . . & the Beast
The alert reader will have noticed that the Dynkin diagram in (4.1) is not the standard
one found in textbooks. In fact, for superalgebras there are alternative choices for the
Dynkin diagram. In this subsection we will consider the ‘distinguished’ Dynkin diagram
with one simple fermionic root
(5.1)
Note, that for a different choice of Dynkin diagram, we get different Dynkin labels for the
spin representation. What is more, the highest weight has actually changed as well as we
will see below. The Cartan matrix and the representation vector read
 
+2 −1  
  0
 −1 +2 −1   −3 
   
 −1 +2 −1   2
   
   
M = −1 +1 , V =  0 . (5.2)
   
 +1 −2 +1   
   0
   0
 +1 −2 +1 
0
+1 −2
The Bethe equations and momentum constraint are still given by (2.8), (2.9), while the
energy of a solution to the Bethe equations is now
n
i i
E = 3L − − . (5.3)
j =1
In this form of the Bethe equations, the vacuum has an energy, it is a pseudovacuum and
not the ground state of the theory. The vacuum state is therefore not the half-BPS state
Tr Z L in (2.5), but instead
|0
= Tr F L , (5.4)
i.e., a composite operator of classical dimension ∆0 = 2L and energy E = 3L [25] built
from the field strength component F = F1+i2,3+i4 . This vacuum configuration is the major
difference to the Bethe ansatz discussed in the previous section. What used to be a rather
trivial state in one Bethe ansatz, becomes a highly excited state in the other. Nevertheless
there is a duality between both sets of Bethe equations in that they must both yield the
same spectra of energies. This remarkable fact can be made use of in the usual sense of
dualities. To determine the energy of some set of states we apply that Bethe ansatz which
seems most suitable for the problem. For example, the Bethe ansatz of the previous section
is useful for states with a low energy. In contrast, the Bethe ansatz discussed in this section
seems most suitable for states of a large chirality B or of an energy around 3L.
The number of excitations can be determined as above. We note the action of the simple
roots in the oscillator picture of [25]
(5.5)
The vacuum corresponds to the configuration a†1 a†1 |0

. The numbers of excitations for a
given weight are
 
2 (L − B) + 2 ∆0 − 2 s2 − L
1 1 1
2 
 (L − B) + ∆0 − 2L 
2 
3 
 2 (L − B) + 12 ∆0 − 12 s1 
 

nk =  2 (L − B)
4 . (5.6)

3 
 2 (L − B) − 2 p − 4 q1 − 4 q2 
1 1 3
 
 2 (L − B) − p − 1 q − 1 q 
2 2 1 2 2 
2 (L − B) − 2 p − 4 q1 − 4 q2
1 1 3 1
We note the bounds
0 n1 n2 , n3 n4 n5 n6 n7 0, (5.7)
as well as
n3 − n2 2L, n4 − n5 , n5 − n4 , n6 − n7 , n7 L. (5.8)
We defer the discussion of highest weight states and multiplet splitting in the context of
the ‘distinguished’ choice of Dynkin diagram to Appendix A. There we also compare the
energies of all dimension four operators to the results of SYM.
Example We see that the vacuum of this choice of Bethe equations has energy 3L.
Clearly, this is not the ground state, the ground state is half-BPS and has zero energy. This
situation is similar to the antiferromagnetic spin chain, where the ground state is described
by a highly excited state (in terms of Bethe roots). In this spin chain the first half-BPS
multiplet with primary weight
w0 = [2; 0, 0; 0, 2, 0; 0, 2] (5.9)
has the highest weight (in the sense of highest chirality B)
w = [4; 0, 0; 0, 0, 0; 2, 2]. (5.10)
This corresponds to the excitations and length
nk = (0, 0, 2, 0, 0, 0, 0), kj = (3, 3), L = 2. (5.11)
The Bethe equations and momentum constraint are solved exactly by
i
u1,2 = ± √ . (5.12)
3
Reassuringly, the energy due to the excitations precisely cancels off the vacuum energy
E = 0. (5.13)
Spin waves For a low number of excitations but arbitrary length of the spin chain, the
Bethe ansatz can quickly provide good approximations to the energies. Especially in the
case of two excitations it is often trivial to solve the Bethe equations exactly. We will now
discuss these spin wave solutions, which correspond to small fluctuations of the vacuum.
The major difference to the previous investigation of spin waves in the context of N = 4
SYM [12] is the modified vacuum. The solution is quite similar and we will follow along
the lines of [12].
Let us consider the following excitations of simple roots
nk = (0, 1, 1, 0, 0, 0, 0), kj = (2, 3). (5.14)
The momentum constraint implies
3
u1 = u2 (5.15)
2
and the two Bethe equations then collapse into

i + u2 L−1
= −1. (5.16)
i − u2
This equation has L − 1 independent solutions
πn L−1 L−1
u2 = cot , − <n (5.17)
L−1 2 2
with energy
2 2 πn
E = 3L − sin . (5.18)
3 L−1
Note that the symmetry uj → −uj translates into n → −n. The states n and −n therefore
have degenerate energies and form a pair unless n = 0 or n = (L − 1)/2. Then the solutions
u1,2 = ∞ or u1,2 = 0, respectively, are invariant and do not pair up.
We can also consider the thermodynamic limit of a large L. In leading order we can
approximate the positions of the Bethe roots and the corresponding energy. An excitation
of simple root 2 yields
3L 4π 2 n2
u2,n ≈ , δE2,n = , (5.19)
2πn 3L2
whereas for an excitation of simple root 3 we find
−2L 4π 2 n2
u3,n ≈ , δE3,n = − . (5.20)
2πn 2L2
The sum of all mode numbers n must vanish due to the momentum constraint. Using this
we quickly find an approximation to the above exactly solved spin wave.
3L −2L
u1 ≈ , u2 ≈ ,
2πn 2π(−n)
4π 2 n2 4π 2 n2 2π 2 n2
E ≈ 3L + − = 3L − . (5.21)
3L2 2L2 3L2
It would be interesting if some string theory solution describing this Bethe vacuum could
be found. The spectrum of fluctuations, somewhat reminiscent of the plane wave spectrum,
could then be compared.
6. Outlook
We are confident that our proposal furnishes the tool for very comprehensive
“spectroscopic” studies of N = 4 Yang–Mills theory. The recent progress with semi-
classical approaches on the string side [16,22,34] allows for very interesting dynamical
comparisons with gauge theory. In particular our superchain and its associated Bethe
equations should be very useful for string motions that also involve the AdS part of the
background, or even joint motion in AdS and on the five sphere. In all these applications
string theory is able, in principle, to predict anomalous dimensions for large dimension
gauge theory operators, to which the Bethe ansatz is ideally suited: on the level of the spin
chain these situations correspond to the thermodynamic limit of the chain, in which the
Bethe equations often become tractable [14].
It would be helpful to classify all possible ways to write Bethe equations for the above
sl(4|4) super spin chain, and identify the corresponding pseudovacua. Correspondingly one
should then find further spin wave solutions, describing the simplest fluctuations around
the pseudovacua. It would be interesting to see whether the large dimension limits of the
pseudovacuum states and their elementary excitations play a special rôle in string theory.
One puzzling aspect of the observed integrable structures is that they really correspond
to “hidden” symmetries appearing in large N SYM4 . Clearly we would like to understand
their presence from the point of view of the gauge theory. Are they a technical consequence
of what one could call “planar supersymmetry”, or is there a deeper reason, possibly related
to string theory?
However, the most exciting open problem remains to identify the integrable deformation
of the super spin chain, with deformation parameter λ = gYM2 N , that correctly includes the
effects of all quantum loops. This presumably would be a psl(4|4) super spin chain with
long-range interactions. A related question concerns the closure of the interacting algebra
and the consistency requirements imposed by it. How are these related to integrability and
can they be used to fix higher-loop corrections?
Acknowledgements
We would like to thank Gleb Arutyunov, Vladimir Dobrev, Stefano Kovacs, Joe
Minahan, Jan Plefka and Kostya Zarembo for interesting discussions, and A. Belitsky,
V. Braun, A. Gorsky and G. Korchemsky for email correspondence concerning the
integrable structures previously found in QCD. N.B. dankt der Studienstiftung des
Deutschen Volkes für die Unterstützung durch ein Promotionsförderungsstipendium.
Appendix A. Dissection of the Beast
In this appendix we present some more details on the Bethe ansatz of Section 5. We will
start by an investigation of highest weight states. Next we will apply this to the Konishi
state and its descendants, for which we then solve the Bethe equations. Finally, we present
the set of all solutions corresponding to states up to classical dimension four.
A.1. Highest weights
A primary state is defined as a state annihilated by the boosts K, S, Ṡ. This condition is
identical to highest weight condition in the sense of Section 4. However, when we change
the Dynkin diagram of the superalgebra, we also change the notion of positive roots and
highest weights. For the ‘distinguished’ Dynkin diagram of Section 5 not all of the boosts
K, S, Ṡ can be positive roots simultaneously. Therefore primary weights (where primary
weight shall refer to highest weight in the sense of Section 4) are not highest weights.
Instead, a highest weight state is annihilated by K, S, Q̇.
For a given primary weight we need to find the corresponding highest weight. For
a generic long multiplet with primary weight w this is achieved by adding
δw = [4; 0, 0; 0, 0, 0; 4, 0]. (A.1)

It corresponds to acting with Q̇8 on the primary state. A further application of Q̇, as well
as K, S which commute with Q̇8 , will therefore kill the state.
In the classical sl(4|4) there are various shortening conditions that affect the highest
weight. For a half-BPS state we add
δw1/2 = [2; 0, 0; 0, −2; 0; 2, 0] (A.2)

instead. The fundamental multiplet VF with primary weight [1; 0, 0; 0, 1, 0; 0, 1] is subject

to an additional shortening condition. The highest weight is [2; 2, 0; 0, 0, 0; 1, 1] and
corresponds to the field F1+i2,3+i4 of N = 4. The vacuum is constructed from exactly
this field.
There are two shortening conditions for unprotected multiplets, see (4.9). When
shortening condition I applies, the multiplet splits in two with highest weights
w + δw + δwI , w + δwI+ . (A.3)
where
δwI = [0; 0, 0; 0, 0, 0; −1, +1],


 [+3.5; +1, 0; −1, 0, 0; +3.5, 0],

 [+3.0; +2, 0; 0, −1, 0; +3.0, 0],
δwI+ = (A.4)

 [+2.5; +3, 0; 0, 0, −1; +2.5, 0],


[+2.0; +4, 0; 0, 0, 0; +2.0, 0].
For δwI+ it is understood that the topmost line applies for which the resulting labels are
positive. When shortening condition II applies, the highest weights are
w + δw, w + δwII + δw (A.5)
with

[+0.5; 0, −1; 0, 0, +1; +0.5, +1],
δwII = (A.6)
[+1.0; 0, 0; 0, 0, +2; 0.0, +1].
Finally, if both shortening conditions apply, the four highest weights are
w + δw + δwI ,
w + δwI+ ,
w + δwII + δw + δwI ,
w + δwII + δwI+ . (A.7)
Note that to determine the last highest weight, it is necessary to choose the correct δwI+
for w + δwII . The submultiplets join in the interacting theory. In each of the three cases
(A.3), (A.5), (A.7) we have listed the highest weight of the long multiplet first.
A.2. Konishi submultiplets
The Konishi state has primary weight

w0 = [2; 0, 0; 0, 0, 0; 0, 2]. (A.8)
This is at both unitarity bounds. According to the rules of the previous section we find the
highest weights of the four submultiplets
w1 = [4.0; 4, 0; 0, 0, 0; 2.0, 2],
w2 = [6.0; 0, 0; 0, 0, 0; 3.0, 3],
w3 = [5.5; 3, 0; 0, 0, 1, 2.5, 3],

w4 = [7.0; 0, 0; 0, 0, 2; 3.0, 4]. (A.9)
The excitation numbers are determined by (5.6)
n1,k = (0, 0, 0, 0, 0, 0, 0), L = 2,
n2,k = (0, 0, 3, 0, 0, 0, 0), L = 3,
n3,k = (0, 0, 2, 1, 0, 0, 0), L = 3,
n4,k = (0, 0, 5, 2, 0, 0, 0), L = 4. (A.10)
The first configuration is the vacuum, we find the energy E = 3L = 6 straight away.
For the second configuration with kj = (3, 3, 3) we note that the roots must be invariant
under the map uj → −uj because the Konishi state is unpaired. Therefore one root is zero
while the other two sum to zero. The only way to satisfy the momentum constraint is to
pick the singular configuration
u1,2 = ±i, u3 = 0. (A.11)
This has to be regularized in the usual way, we again find the energy E = 6. As an
aside, this state has the highest weight for the interacting multiplet. At leading order, it
corresponds to the operators εβγ εδε εηα Tr Fαβ Fγ δ Fεη of N = 4 SYM.
For the third configuration with kj = (3, 3, 4) it is straightforward to find and verify
1
u1,2 = ± √ , u3 = 0 (A.12)
3
as a solution to the Bethe equations and momentum constraint with energy E = 6.
For the last configuration with kj = (3, 3, 3, 3, 3, 4, 4) we again require a singular
solution. Taking the roots (A.11) as an ansatz, the remaining roots can be found

1 7
u1,2 = ±i, u3 = 0, u4,5 = ± √ , u6,7 = ± . (A.13)
3 20
They satisfy the momentum constraint and yield, not surprisingly, E = 6.
We have seen that all submultiplet of the Konishi multiplet have the same energy. This
is somewhat remarkable, as they have rather different configurations of roots.
A.3. Bethe roots for dimension four states
Here we will present the Bethe roots for operators up to dimensions four. This
demonstrates that also the ‘distinguished’ choice of Dynkin diagram with a pseudovacuum
leads to the correct results. We present the complete list of operators and anomalous
dimensions in Table 1. The Bethe roots which solve the Bethe equations and momentum
constraint are given below. The energies agree with Table 1, we do not list them again.
[2; 0, 0; 0, 2, 0; 0, 2] kj = (3, 3), L = 2. See (5.12).

i
u1,2 = ± √ . (A.14)
3
[2; 0, 0; 0, 0, 0; 0, 2] kj = ( ), L = 2. Trivial.
[3; 0, 0; 0, 3, 0; 0, 3] kj = (3, 3, 3, 3, 4, 4, 5), L = 3.

√
−3 ± i 15 5
u1,2,3,4 = ± , u5,6 = i , u7 = 0. (A.15)
12 12
[3; 0, 0; 0, 1, 0; 0, 3] kj = (3, 3), L = 3.

i
u1,2 = ± √ . (A.16)
5
[4; 0, 0; 0, 4, 0; 0, 4] kj = (3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 5, 5), L = 4. This is left as an

exercise for the reader. Good luck!
[4; 0, 0; 0, 2, 0; 0, 4] kj = (3, 3, 3, 3, 4, 4, 5) L = 4. Two states distinguished by ± .

√ √
± 5 ± i(5 ± 3 5)
u1,2,3,4 = ± ,
15
√
−15 ∓ 2 5
u5,6 = ± , u7 = 0. (A.17)
36
[4; 0, 0; 1, 0, 1; 0, 4] kj = (3, 3, 3, 3, 4), L = 4.

±1 ± i
u1,2,3,4 = √ , u5 = 0. (A.18)
2 3
[4; 0, 0; 0, 0, 0; 0, 4] Two states distinguished by ± . kj = (3, 3, 3, 3), L = 4.

√ √

−9 ∓ 41 173 ± 33 41
u1,2,3,4 = ± ± . (A.19)
12 360
[4; 2, 0; 0, 0, 0; 1, 3] kj = ( ), L = 3. Trivial.
[4; 0, 2; 0, 0, 0;-1, 3] kj = (2, 2, 3, 3, 3, 3), L = 3.

√
17 −39 ± 993
u1,2 = ± , u3,4,5,6 = ± . (A.20)
60 30
[4; 1, 1; 0, 1, 0; 0, 3] kj = (2, 3, 3), L = 3. A pair of states distinguished by ± .

√
5 1 ± i 399
u1 = ± , u2,3 = ∓ . (A.21)
28 30
[4; 2, 2; 0, 0, 0; 0, 2] kj = (2, 2), L = 2.

9
u1,2 = ± . (A.22)
28
References
[1] F. Gliozzi, J. Scherk, D.I. Olive, Supersymmetry, supergravity theories and the dual spinor model, Nucl.
Phys. B 122 (1977) 253;
L. Brink, J.H. Schwarz, J. Scherk, Supersymmetric Yang–Mills theories, Nucl. Phys. B 121 (1977) 77.
[2] M.F. Sohnius, P.C. West, Conformal invariance in N = 4 supersymmetric Yang–Mills theory, Phys. Lett.
B 100 (1981) 245;
P.S. Howe, K.S. Stelle, P.K. Townsend, Miraculous ultraviolet cancellations in supersymmetry made
manifest, Nucl. Phys. B 236 (1984) 125;
L. Brink, O. Lindgren, B.E.W. Nilsson, N = 4 Yang–Mills theory on the light cone, Nucl. Phys. B 212
(1983) 401.
[3] J.M. Maldacena, The large N limit of superconformal field theories and supergravity, Adv. Theor. Math.
Phys. 2 (1998) 231, hep-th/9711200;
E. Witten, Anti-de Sitter space and holography, Adv. Theor. Math. Phys. 2 (1998) 253, hep-th/9802150;
S.S. Gubser, I.R. Klebanov, A.M. Polyakov, Gauge theory correlators from non-critical string theory, Phys.
Lett. B 428 (1998) 105, hep-th/9802109.
[4] G. Mandal, N.V. Suryanarayana, S.R. Wadia, Aspects of semiclassical strings in AdS5 , Phys. Lett. B 543
(2002) 81, hep-th/0206103;
I. Bena, J. Polchinski, R. Roiban, Hidden symmetries of the AdS5 × S 5 superstring, hep-th/0305116.
[5] B.C. Vallilo, Flat currents in the classical AdS5 × S 5 pure spinor superstring, hep-th/0307018.
[6] J. Maldacena, L. Maoz, Strings on pp-waves and massive two-dimensional field theories, JHEP 0212 (2002)
046, hep-th/0207284;
J.G. Russo, A.A. Tseytlin, A class of exact pp-wave string models with interacting light-cone gauge actions,
JHEP 0209 (2002) 035, hep-th/0208114;
I. Bakas, J. Sonnenschein, On integrable models from pp-wave string backgrounds, JHEP 0212 (2002) 049,
hep-th/0211257.
[7] D. Berenstein, J.M. Maldacena, H. Nastase, Strings in flat space and pp waves from N = 4 super-Yang–
Mills, JHEP 0204 (2002) 013, hep-th/0202021.
[8] C. Kristjansen, J. Plefka, G.W. Semenoff, M. Staudacher, A new double-scaling limit of N = 4 super-Yang–
Mills theory and PP-wave strings, Nucl. Phys. B 643 (2002) 3, hep-th/0205033;
N.R. Constable, D.Z. Freedman, M. Headrick, S. Minwalla, L. Motl, A. Postnikov, W. Skiba, PP-wave string
interactions from perturbative Yang–Mills theory, JHEP 0207 (2002) 017, hep-th/0205089.
[9] N. Beisert, C. Kristjansen, J. Plefka, G.W. Semenoff, M. Staudacher, BMN correlators and operator mixing
in N = 4 super-Yang–Mills theory, Nucl. Phys. B 650 (2003) 125, hep-th/0208178.
[10] N.R. Constable, D.Z. Freedman, M. Headrick, S. Minwalla, Operator mixing and the BMN correspondence,
JHEP 0210 (2002) 068, hep-th/0209002.
[11] N. Beisert, BMN operators and superconformal symmetry, Nucl. Phys. B 659 (2003) 79, hep-th/0211032.
[12] J.A. Minahan, K. Zarembo, The Bethe-ansatz for N = 4 super-Yang–Mills, JHEP 0303 (2003) 013, hep-
th/0212208.
[13] N. Beisert, C. Kristjansen, J. Plefka, M. Staudacher, BMN gauge theory as a quantum mechanical system,
Phys. Lett. B 558 (2003) 229, hep-th/0212269.
[14] N. Beisert, J.A. Minahan, M. Staudacher, K. Zarembo, Stringing spins and spinning strings, hep-th/0306139,
JHEP, in press.
[15] S. Frolov, A.A. Tseytlin, Multi-spin string solutions in AdS5 × S 5 , hep-th/0304255.
[16] S. Frolov, A.A. Tseytlin, Quantizing three-spin string solution in AdS5 × S 5 , JHEP 0307 (2003) 016, hep-
th/0306130;
S. Frolov, A.A. Tseytlin, Rotating string solutions: AdS/CFT duality in non-supersymmetric sectors, hep-
th/0306143.
[17] L.N. Lipatov, High-energy asymptotics of multicolor QCD and exactly solvable lattice models, JETP
Lett. 59 (1994) 596, hep-th/9311037.
[18] L.D. Faddeev, G.P. Korchemsky, High-energy QCD as a completely integrable model, Phys. Lett. B 342
(1995) 311, hep-th/9404173.
[19] V.M. Braun, S.E. Derkachov, A.N. Manashov, Integrability of three-particle evolution equations in QCD,
Phys. Rev. Lett. 81 (1998) 2020, hep-ph/9805225;
A.V. Belitsky, Fine structure of spectrum of twist-three operators in QCD, Phys. Lett. B 453 (1999) 59,
hep-ph/9902361;
V.M. Braun, S.E. Derkachov, G.P. Korchemsky, A.N. Manashov, Baryon distribution amplitudes in QCD,
Nucl. Phys. B 553 (1999) 355, hep-ph/9902375;
A.V. Belitsky, Integrability and WKB solution of twist-three evolution equations, Nucl. Phys. B 558 (1999)
259, hep-ph/9903512;
A.V. Belitsky, Renormalization of twist-three operators and integrable lattice models, Nucl. Phys. B 574
(2000) 407, hep-ph/9907420;
S.E. Derkachov, G.P. Korchemsky, A.N. Manashov, Evolution equations for quark gluon distributions in
multi-color QCD and open spin chains, Nucl. Phys. B 566 (2000) 203, hep-ph/9909539.
[20] A.V. Kotikov, L.N. Lipatov, NLO corrections to the BFKL equation in QCD and in supersymmetric gauge
theories, Nucl. Phys. B 582 (2000) 19, hep-ph/0004008;
A.V. Kotikov, L.N. Lipatov, DGLAP and BFKL evolution equations in the N = 4 supersymmetric gauge
theory, hep-ph/0112346;
A.V. Kotikov, L.N. Lipatov, DGLAP and BFKL equations in the N = 4 supersymmetric gauge theory,
hep-ph/0208220;
A.V. Kotikov, L.N. Lipatov, V.N. Velizhanin, Anomalous dimensions of Wilson operators in N = 4 SYM
theory, Phys. Lett. B 557 (2003) 114, hep-ph/0301021.
[21] F.A. Dolan, H. Osborn, Superconformal symmetry, correlation functions and the operator product expansion,
Nucl. Phys. B 629 (2002) 3, hep-th/0112251.
[22] S.S. Gubser, I.R. Klebanov, A.M. Polyakov, A semi-classical limit of the gauge/string correspondence, Nucl.
Phys. B 636 (2002) 99, hep-th/0204051.
[23] A.V. Belitsky, A.S. Gorsky, G.P. Korchemsky, Gauge/string duality for QCD conformal operators, hep-
th/0304028.
[24] N. Beisert, C. Kristjansen, M. Staudacher, The dilatation operator of N = 4 conformal super-Yang–Mills
theory, Nucl. Phys. B 664 (2003) 131, hep-th/0303060.
[25] N. Beisert, The complete one-loop dilatation operator of N = 4 super-Yang–Mills theory, hep-th/0307015.
[26] L.D. Faddeev, How algebraic Bethe ansatz works for integrable model, hep-th/9605187.
[27] N.Y. Reshetikhin, A method of functional equations in the theory of exactly solvable quantum system, Lett.
Math. Phys. 7 (1983) 205;
N.Y. Reshetikhin, Integrable models of quantum one-dimensional magnets with O(N ) and Sp(2K)
symmetry, Theor. Math. Phys. 63 (1985) 555.
[28] E. Ogievetsky, P. Wiegmann, Factorized S matrix and the Bethe ansatz for simple Lie groups, Phys. Lett.
B 168 (1986) 360.
[29] H. Saleur, The continuum limit of sl(N/K) integrable super spin chains, Nucl. Phys. B 578 (2000) 552,
solv-int/9905007.
[30] P.P. Kulish, N.Y. Reshetikhin, E.K. Sklyanin, Yang–Baxter equation and representation theory: I, Lett. Math.
Phys. 5 (1981) 393.
[31] V.K. Dobrev, V.B. Petkova, Group theoretical approach to extended conformal supersymmetry: function
space realizations and invariant differential operators, Fortschr. Phys. 35 (1987) 537.
[32] V.K. Dobrev, V.B. Petkova, All positive energy unitary irreducible representations of extended conformal
supersymmetry, Phys. Lett. B 162 (1985) 127;
L. Andrianopoli, S. Ferrara, E. Sokatchev, B. Zupnik, Shortening of primary operators in N -extended SCFT 4
and harmonic-superspace analyticity, Adv. Theor. Math. Phys. 3 (1999) 1149, hep-th/9912007.
[33] D. Anselmi, The N = 4 quantum conformal algebra, Nucl. Phys. B 541 (1999) 369, hep-th/9809192.
[34] S. Frolov, A.A. Tseytlin, Semiclassical quantization of rotating superstring in AdS5 × S 5 , JHEP 0206 (2002)
007, hep-th/0204226;
J.G. Russo, Anomalous dimensions in gauge theories from rotating strings in AdS5 × S 5 , JHEP 0206 (2002)
038, hep-th/0205244;
J.A. Minahan, Circular semiclassical string solutions on AdS5 × S 5 , Nucl. Phys. B 648 (2003) 203, hep-
th/0209047;
A.A. Tseytlin, Semiclassical quantization of superstrings: AdS5 × S 5 and beyond, Int. J. Mod. Phys. A 18
(2003) 981, hep-th/0209116.
Finite-volume form factors in semi-classical

approximation
G. Mussardo a,b , V. Riva a,b,1 , G. Sotkov c
a International School for Advanced Studies, Via Beirut 1, 34100 Trieste, Italy
b Istituto Nazionale di Fisica Nucleare, Sezione di Trieste, Trieste, Italy
c Instituto de Fisica Teorica, Universidade Estadual Paulista, Rua Pamplona 145, 01405-900 Sao Paulo, Brazil
Received 24 July 2003; accepted 13 August 2003
Abstract
A semi-classical approach is used to obtain Lorentz covariant expressions for the form factors
between the kink states of a quantum field theory with degenerate vacua. Implemented on a cylinder
geometry it provides an estimate of the spectral representation of correlation functions in a finite
volume. Illustrative examples of the applicability of the method are provided by the sine-Gordon and
the broken φ 4 theories in 1 + 1 dimensions.
PACS: 11.10.Kk; 11.15.Kc; 11.27.+d
Keywords: Form factors; Kink solutions in finite volume; Spectral density in finite volume
1. Introduction
The problem of understanding quantum field theories (QFT) on a finite volume is

of great importance both for theoretical interests and for many realistic applications. In
particular, it is crucial to explore the possibility of defining a “form factor representation”
for the correlation functions, in analogy to the infinite volume case [1]. There are reasons to
expect, in fact, a fast convergent behaviour of these series also for finite volume correlators,
as it happens in infinite volume (see, for instance, [2]). If this would be indeed the
case, accurate estimates of finite volume correlators and other related physical quantities,
could be obtained by just using few exact terms of their spectral representations, having
E-mail address: riva@he.sissa.it (V. Riva).

1 Postal address: Valentina Riva, International School for Advanced Studies, via Beirut 4, 34014 Trieste, Italy.
doi:10.1016/j.nuclphysb.2003.08.017
G. Mussardo et al. / Nuclear Physics B 670 [FS] (2003) 464–478 465
consequently a great simplification of the problem. This observation makes clear that it is
worth pursuing the research on this topic and looking, in particular, for an efficient scheme
of approximation.
It is well known that finite volume quantum field theories are also related to field
theories at finite temperature. In 1 + 1 space–time dimensions, this gives the possibility
of viewing a QFT defined on a finite volume of a cylinder geometry in the alternative
picture, in which the space variable is infinite and the (Euclidean) time variable is instead
compactified on a circle of radius R = 1/T , where T is the temperature of the system
(the so-called Matsubara imaginary time formalism). In the case of integrable theories,
promising exact results have been obtained in [3], thanks to the thermodynamic Bethe
ansatz (TBA) technique, which relates the thermodynamical properties of the system to
its exact S-matrix [4]. As shown in [3], the finite temperature correlation functions can be
expressed in terms of the usual form factors, relative to the infinite volume Hilbert space,
dressed with the so-called “filling-fractions”, which encode the thermodynamics of the
system. Within this approach, it has been possible to fully understand the finite temperature
one-point functions, while a further analysis seems still necessary for the two-point and
higher correlation functions.
Coming back to the picture where the time evolution is the usual one while the space
variable is compactified, the form factors have to be seen as the matrix elements of local
fields between eigenstates of the finite volume Hamiltonian. In this case there is still a poor
understanding of the correlation functions behaviour, even in the integrable cases. It must
be stressed that the reason is not only the present undeveloped and unsatisfactory theory for
computing finite volume matrix elements. In fact, contrary to the infinite volume situation,
the only knowledge of the form factors is not enough to obtain the correlators in a finite
volume since for their computation the energy eigenvalues of the finite volume hamiltonian
are also needed. For integrable models, these extra data may be obtained along the lines
discussed in [5], while for nonintegrable theories one has necessarily to rely on numerical
methods, as the one proposed in [6], for instance.
Leaving apart the above problem of energy eigenvalues and concerning instead about
the status of the form factors in a finite volume, there are so far only semi-classical results
relative to a conformal theory [7] as well as exact calculations relative to the Ising model
[8] (for a Bethe ansatz approach see, however, [9]). Although these findings are very
interesting, the techniques employed in the above papers are however strictly related either
to the specific integrable structure of the considered models or to the free nature of the
Majorana fermion field of the Ising model. In this paper we are going to use, instead, a
semi-classical approach which does not require integrability and it may be then of more
general applicability, of course within its range of approximation. As shown below, this
approach gives in particular the possibility of checking some of the results of our analysis
against the exact known quantities in the case of integrable theories in infinite volume,
while it permits to have new predictions on the semi-classical regime of the nonintegrable
ones. As illustrative examples of this method, we will present its application to two
significant models: the integrable QFT of the sine-Gordon model
m2
V (φ) = (1 − cos β φ), (1.1)
β2
466 G. Mussardo et al. / Nuclear Physics B 670 [FS] (2003) 464–478
and the nonintegrable QFT with a φ 4 interaction in the broken Z2 symmetry phase
λ 4 m2 2 m4
V (φ) = φ − φ + . (1.2)
4 2 4λ
Both these theories display degenerate minima, that can be chosen as the constant
classical solution representing the vacuum state around which one decides to quantize the
theory. However, the most interesting classical backgrounds are the kink-type ones, which
interpolate between degenerate minima and give rise to a nontrivial quantization scheme.
Our aim is to define this kind of backgrounds in finite volume and to estimate the relative
form factors in an appropriate limit. At the same time, doing so, we will have the possibility
of better exploring the properties of the same form factors in the infinite volume case.
2. Classical solutions and form factors
The semi-classical quantization of a field theory defined by a potential V (φ) is based on

the identification of a classical background φcl (x) which satisfies the equation of motion
∂µ ∂ µ φcl + V (φcl ) = 0. (2.3)

In the infinite volume, this can be performed with various well established techniques,
like the path integral formalism [10] or the solution of the field equations in classical
background [11], usually called the DHN method (for a systematic review, see [12]).
The procedure is particularly simple and interesting if one considers classical field
solutions φcl (x) in 1 + 1 dimensions which are static “kink” configurations interpolating
between degenerate minima of the potential, and whose quantization gives rise to a particle-
like spectrum. The kink solutions are obtained, firstly, by integrating the first order equation
related to (2.3)

1 ∂φcl 2
= V (φcl ) + C, (2.4)
2 ∂x
and, secondly, imposing that φcl (x) reaches two different minima of the potential as
x → ±∞. In the infinite volume case, these boundary conditions correspond to fixing
the constant of integration C equal to zero. As we will show below, finite volume case can
be described, on the contrary, by a nonvanishing value of C which can be directly related
to the size of the system. In the following we will be interested to discuss QFT defined on
a cylinder geometry where the time variable is infinite while the space one is compactified
on a circle of circumference L, with twisted boundary conditions, i.e.,
φ(x + L) = 2π − φ(x) for sine-Gordon,

φ(x + L) = −φ(x) for φ 4 theory.
A very interesting result, due to Goldstone and Jackiw [13] and well explained in [12],
is that the classical background φcl (x) has the quantum meaning of Fourier transform of
the form factor of the basic field φ(x) between kink states.2 The technique to derive this
result relies on the basic hypothesis that the kink momentum is very small compared to
its mass, which is indeed inversely proportional to the coupling constant, considered small
in the semi-classical regime. Calling p1 and p2 the momenta of the in and out one-kink
states, at the leading order in the coupling constant one obtains

p2 |φ(0)|p1
= da ei(p1 −p2 )a φcl (a). (2.5)
This important result has, unfortunately, two serious drawbacks: expressing the form factor
as a function of the difference of momenta, Lorentz covariance is lost and moreover, the
antisymmetry under the interchange of momenta makes problematic any attempt to go in
the crossed channel and obtain the matrix element between the vacuum and a kink–antikink
state.
In order to overcome these problems, we need to refine the method proposed in [13].
This can be done by using the rapidity variable θ of the kink (and considering it as very
small), instead of the momentum. For example, in the φ 4 theory (1.2), where the kink mass
M is of order 1/λ, we will work under the hypothesis that θ is of order λ. In this way we
get consistently
E ≡ M cosh θ M, p ≡ M sinh θ Mθ M. (2.6)
We can now define the form factor between kink states as the Fourier transform with respect
to the Lorentz invariant rapidity difference θ ≡ θ1 − θ2 :

p1 |φ(0)|p2
≡ f (θ ) ≡ M da eiMθa fˆ(a), (2.7)
with the inverse Fourier transform defined as

ˆ dθ −iMθa
f (a) ≡ e f (θ ). (2.8)
2π
Following the same procedure used by Goldstone and Jackiw [13], i.e., starting from the
Heisenberg equation of motion for the field φ(x, t)
2
∂t − ∂x2 φ(x, t) = −V φ(x, t) , (2.9)
and taking the matrix elements of both sides

−(p1 − p2 )µ (p1 − p2 )µ e−i(p1 −p2 )µ x p1 |φ(0)|p2

= −e−i(p1 −p2 )µ x p1 |V φ(0) |p2
,
µ
(2.10)
it is easy to show that, at leading order in λ, the function fˆ(a) obeys the same differential
equation satisfied by the kink solution, i.e.,
d2 ˆ
2
f (a) = V fˆ(a) . (2.11)
da
2 The classical solution can also be directly seen as the matrix element of φ between asymptotic states in the
space coordinates representation: x2 |φ(0)|x1
= δ(x1 − x2 )φcl (x1 ).
This means that we can take fˆ(a) to be equal to φcl (a), i.e., fˆ(a) = φcl (a) and adjusting its
boundary conditions by an appropriate choice for the value of the constant C in Eq. (2.4).
With the above considerations, it is now possible to express the crossed channel form
factor through the variable transformation θ → iπ − θ :
F2 (θ ) ≡ 0|φ(0)|p1 , p̄2
= f (iπ − θ ). (2.12)
The analysis of this quantity in infinite volume allows us, in particular, to get information
about the spectrum of the theory. Its dynamical poles, in fact, located at θ ∗ = i(π − u)
with 0 < u < π , coincide with the poles of the kink–antikink S-matrix [1], and the relative
bound states masses can be then expressed as
u
m(b) = 2M sin . (2.13)
2
It is worth to note that this procedure for extracting the semi-classical bound states masses
is remarkably simpler than the standard DHN method of quantizing the corresponding
classical backgrounds, because in general these solutions depend also on time and have a
much more complicated structure than the kink ones. Moreover, in nonintegrable theories
these backgrounds could even not exist as exact solutions of the field equations: this
happens for example in the φ 4 theory, where the DHN quantization has been performed on
some approximate backgrounds [11].
Once the matrix elements (2.12) are known, one can estimate the leading behaviour in
λ of the spectral function in a regime of the momenta dictated by our assumption of small
kink rapidity. In infinite volume the spectral function ρ(p2 ) is defined as

d 2 p 2 ip·x
0|φ(x)φ(0)|0
≡ ρ p e , (2.14)
(2π)2
and has the form factor expansion
1
ρ(p2 ) = 2π dΩ1 · · · dΩn δ p0 − E1 − · · · − En δ p1 − p1 − · · · − pn
n
n!
2
× 0|φ(0)|n
, (2.15)
with dΩ ≡ 2πdp2E = 4π dθ
. The delta functions in the above expression make meaningful
the use of our form factors, derived in the small θ approximation, only if we consider
a regime in which p0 M and p1 M and, from now on, we will always understand
this restriction. The leading O(1/λ) contribution3 to the spectral function, denoted in
the following by ρ̂(p2 ), is given by the trivial vacuum term plus the kink–antikink
3 About the orders in λ of the various form factors, we refer to the complete discussion in [12,13]. Our
formalism is slightly different because the M factor in front of (2.7), which is a natural consequence of considering
the rapidity as the basic variable, increases by 1/λ the order of all form factors in the kink sector with respect to
[13]; however, since θ O(λ), in the final expression for the spectral function all orders recombine consistently.
contribution:
2
ρ̂ p2 = 2πδ p0 δ p1 0|φ(0)|0

p0
π δ M −2 dθ1 p1 2
+ F 2θ + iπ − , (2.16)
2π M
2 1
4 M2
where the range of integration of the above quantity is of order p1 /M (note that, being
p1 /M 1, the integral can be roughly estimated by evaluating |F2 |2 at θ1 = 0: this is
what we will do in the next sections).
The application of this procedure to the finite volume case is straightforward, thanks
to the possibility of choosing fˆ(a) as a solution of Eq. (2.4) with any constant C. As
explicitly shown by the examples discussed in the next sections, this is equivalent to define
a kink solution configuration on a finite volume, with the constant C directly related to
the size of the system. We have now to consider the matrix elements of φ(0) between
two eigenstates |pn1
and |pn2
of the finite volume Hamiltonian. These states can be
naturally labelled with the so-called “quasi-momentum” variable pn , which corresponds to
the eigenvalues of the translation operator on the cylinder (multiples of π/L), and appears
in the space dependent part of Eq. (2.10) in the case of finite volume. The TBA equations
[4], valid for large L, are exactly a relation between this variable and the free momentum
p∞ of the infinite volume asymptotic states, through a phase shift δ(p∞ ) which encodes
the information about the interaction:
δ(pn∞ ) 2nπ
pn∞ + = ≡ pn . (2.17)
L L
Defining θn as the “quasi-rapidity” of the kink states by
pn = M(L) sinh θn M(L)θn , (2.18)

we can now write the form factor at a finite volume by replacing the Fourier integral
transform with a Fourier series expansion:
L/2
f (θn ) ≡ M(L) da eiM(L)θna fˆ(a), (2.19)
−L/2
∞
1
fˆ(a) ≡ e−iM(L)θn a f (θn ), (2.20)
LM(L) n=−∞
where
(2n1 − 1)π (2n2 − 1)π 2nπ
M(L)θn pn1 − pn2 = − ≡ . (2.21)
L L L
Since the energy eigenvalues of the finite volume Hamiltonian cannot be related to the
quasi-rapidity as in Eq. (2.6), in principle we are not allowed to express the crossed channel
form factor F2 (θ ) via the change of variable θ → iπ − θ . However, it is easy to show that
in our regime of approximations the deviations of the kink energy from (2.6) are of higher
order in the coupling and can be neglected at this stage.
On the cylinder, the spectral function can be expressed as a series expansion on the form
factors:
1 1 1
ρ(Ek , pk ) = 2π
n
n! (2L) n Ek1 Ek2 · · · Ekn
k1 ,...,kn
× δ(Ek − Ek1 − · · · − Ekn )δ(pk − pk1 − · · · − pkn )
2
× 0|φ(0)|n
, (2.22)
where pki are the quasi-momenta of the intermediate states, and Eki are the finite volume
energy eigenvalues, to be determined by other means (see the comment in Section 1). In
our semi-classical regime, however, this is not necessary: in fact, in order to evaluate the
leading contribution ρ̂(Ek , pk ), we can consistently approximate the kink energies with
their classical values (of order 1/λ), which can be exactly computed as a function of the
volume. We then have
2
ρ̂(Ek , pk ) = 2πδ(Ek )δ(pk ) 0|φ(0)|0

E
π δ Mk − 2 pk 2
+ F
2 2θ + iπ − . (2.23)
M
k1
4 M2
θk1
As in the infinite volume case, the consistency of the semi-classical approximation selects
as the relevant values of the above series those with θk 0 and therefore it can be roughly
estimated by simply evaluating |F2 |2 at θk1 = 0.
It is now interesting to apply these general considerations to the analysis of the form
factors of the fundamental field φ(x) both for an integrable and a nonintegrable QFT.
3. Sine-Gordon model
3.1. Infinite volume
The sine-Gordon model, defined in (1.1), is an integrable quantum field theory, for
which the infinite volume form factors are exactly known [1]. The (anti)soliton background
is given by
4
φcl (x) = arctan e±mx . (3.24)
β
Its classical energy is Ecl = 8 βm2 , and the first quantum corrections are known to be of
higher order in β 2 [11]. Hence we can consistently approximate the mass M∞ with this
value and assume that the (anti)soliton rapidity will be of order β 2 . The semi-classical form
factor (2.7) is explicitly given by
∞
4M∞ 2πi 1 2π 2
f (θ ) = da ei(M∞ θ)a arctan ema = 4π + 3 δ θ/β 2 .
β β θ cosh 2 θ β
−∞ β
(3.25)
It is an interesting check to see whether our formulation of the semi-classical form factor
in terms of the rapidity variable, i.e., the expression F2 (θ ), matches with the semi-classical
limit of the exact one. In doing this check, the only thing to take into account is that, in
the definition of the exact form factor of the fundamental field φ(x), the asymptotic two-
particle state is actually given by the antisymmetric combination of soliton and antisoliton.
Since at our level the form factor between antisoliton states is simply
∞
4M∞
p̄1 |φ(0)|p̄2
= da ei(M∞ θ)a arctan e−ma = f (−θ ), (3.26)
β
−∞
we finally obtain
4πi 1
F2 (θ ) = , (3.27)
β (iπ − θ ) cosh 4π2 (iπ − θ )
β
which indeed coincides with the exact result [1] in the regime4 iπ − θ O(β 2 ).
Furthermore, we can also check that the dynamical poles of this quantity, located at

β2 1 1 4π
θn = iπ 1 − (2n + 1) , − < n < − + 2 , (3.28)
8π 2 2 β
consistently reproduce the odd part of the well-known semi-classical breathers spectrum
[11]

2

(2n+1) β (2n + 1)2 4
mb = 2M∞ sin (2n + 1) = (2n + 1)m 1 − β + · · · . (3.29)
16 3 · 83
Since in the vacuum sector 0|φ|0
= 0, in this model the 1/β 2 leading contribution to the
spectral function takes the form:

p0 1
ρ̂(p2 ) = 4π 3 δ −2 p1
. (3.30)
M β 2 (p1 )2 cosh2 π2 m
3.2. Finite volume
In order to consider the effects of a cylindrical geometry, let us integrate Eq. (2.4) with
a nonzero constant C. Rescaling for convenience the variables as
β2
φ̄ = βφ, x̄ = mx, C̄ = C, (3.31)
m2
4 The classical background (3.24) has been first used in [14] to derive a Lorentz covariant extrapolation of the
form factor (2.5). However, as we explained in Section 2, the comparison of this quantity with the semi-classical
limit of the exact result is successful only with some restrictions on the momenta, which can be unambiguously
formulated in terms of the rapidity variable.
Fig. 1. Sine-Gordon antisoliton in finite volume.
a solution of Eq. (2.4), with −2 < C̄ < 0, can be expressed as

C̄ + 2
φ̄cl (x̄) = 2 arccos sn ±x̄, k ,
2
(3.32)
2
where sn(x̄, k 2 ) is the Jacobi elliptic function with modulus k 2 = C̄+2

2 (for its properties,
see [15]). The plot of this function is drawn in Fig. 1. For a given value of C̄, this solution
oscillates with a period 4K(k 2 ) between φ̄0 and 2π − φ̄0 , where φ̄0 is defined by the
condition V (φ̄0 ) = −C̄, and K(k 2 ) is the complete elliptic integral of the first kind.
The solution (3.32) has been proposed in Ref. [16] as a model of a crystal of solitons and
antisolitons in the sine-Gordon theory in infinite volume.5 In our finite volume case, the
solution (3.32) has to be interpreted, instead, as a single (anti)soliton defined on a cylinder
of circumference

1 C̄ + 2
L = 2K . (3.33)
m 2
Within this interpretation, the periodic oscillations of the solution represent the soliton
circling around the cylinder. Eq. (3.33) is the explicit relation between the size of the
system and the integration constant C̄; one can consistently recover the infinite volume
limit for C̄ → 0: in this limit L goes to infinity and the function (3.32) goes to the standard
(anti)soliton solution (3.24).
A strong motivation for the choice of this function as the (anti)soliton background on the
cylinder can be obtained by analyzing the so-called “classical energy per kink” computed
5 It is worth mentioning the existence of some impressive old papers [17] where the basic ideas discussed in
[16] were already present.
in [16]:
L/2

1 ∂φcl 2 m2
Ecl (L) = dx + 2 (1 − cos β φcl )
2 ∂x β
−L/2

m 2 1 2
=8 2 E k − 1−k K k ,
2
(3.34)
β 2
where E(k 2 ) is the complete elliptic integral of the second kind. In the L → ∞ limit (which
corresponds to k → 0, with (k )2 ≡ 1 − k 2 = − C̄2 ), Ecl (L) approaches exponentially the
value M∞ ; the expansions of E and K for small k [15], compared with
1 2
e−mL = (k ) + · · · ,
16
lead to the following expansion for large L of the classical energy
m
Ecl (L) = M∞ − 32 2 e−mL + O e−2mL . (3.35)
β
By using the theory of finite volume corrections discussed in [18,19], the behaviour (3.35)
can be put in correspondence with the scattering data of the QFT in infinite volume:6
1
M(L) − M∞ = − g 2 e−mb L + · · · ,
2 k k̄b
(3.36)
8mb M∞
where gk k̄b is the 3-particle on-shell coupling of kink, antikink and elementary boson,
whose mass in infinite volume is mb = m. Using the known expression for the sine-Gordon
kink-breather S-matrix [20] and evaluating the limit β → 0 of its residue, we exactly get
the coefficient reported in (3.35).
We can now write the finite volume form factor (2.19) in terms of the antisoliton
background (3.32):
L/2
2M C̄ + 2
f (−θn ) = da e iMθn a
arccos sn(ma)
β 2
−L/2
2 iMθn L L

=− e 2 log(k + ik ) − e −iMθn 2 log(−k + ik )
βθn
2πi 1
− , (3.37)
β θn cosh K Mθn
m
√
where k = 1 − k 2 and K (k 2 ) = K(k 2 ). In order to obtain this result one has to use the
relation
1
arccos k sn(ma) = log k sn(ma) + i dn(ma) , (3.38)
i
6 In Eq. (3.36), we have only written the term that is relevant in our semi-classical limit, in which the leading
contribution to the mass is given by Ecl (L).
and, after an integration by parts, finally compare the inverse Fourier transform (2.20) with
the expansion [15]
∞ (2n−1)π
2π 1 cos a
cn(ma) = L
(2n−1)π K
. (3.39)
k mL cosh
n=1 L m
The form factor (3.37) has the correct IR limit,7 and leads to the following expressions
for F2 (θ ) and for the spectral function:

4πi 1
F2 (θn ) = M
β(iπ − θn ) cosh m K (iπ − θn )

2 k L
+ −1 + arctan cos M(iπ − θn ) , (3.40)
π k 2

En 1
ρ̂(En , pn ) = 4π 3 δ −2 2
M β (pn )2

1 2 k L 2
× + −1 + arctan cos pn . (3.41)
cosh Km pn π k 2
Note that the finite volume dependence of both the form factor (3.40) and the spectral
function (3.41) is not restricted to the second term only. The M(L)K (k 2 ) factor in the first
term carries the main L-dependence, although it is not manifest but implicitly defined by
Eq. (3.33).
4. φ 4 field theory in the broken symmetry phase
The semi-classical analysis performed on the sine-Gordon model can be repeated very
similarly for the quantum field theory defined by the potential (1.2). There is, however, the
important conceptual difference that this QFT is nonintegrable: as we are going to show,
this gives us the possibility of estimating quantities of this quantum field theory which
were unknown even in the infinite volume case.
4.1. Infinite volume
The standard (anti)kink background is given in this case by

m mx
φcl (x) = (±) √ tanh √ , (4.42)
λ 2
√
2 2 m3
with classical energy M∞ = 3 λ .
7 The function e−ixL/2 can be shown to tend to −iπ δ(x) in the distributional sense for L → ∞, and in the
x
same way one can show that cos(xL/2)
x tends to zero.
With our formulation in terms of the rapidity, we can write an unambiguous8 Lorentz
covariant expression for the form factor (2.7)
∞
m ma
p2 |φ(0)|p1
= M∞ √ da eiM∞ θa tanh √
λ 2
−∞
3
4 m 1
= iπ √ 2 m2 . (4.43)
3 λ sinh 3 π λ θ
It is of great interest to analyze in this case the dynamical poles of F2 (θ ) for extracting
information about the spectrum of this theory. They are located at

3 λ 2π m2
θn = iπ 1 − 2
n , 0<n< , (4.44)
2π m 3 λ
and the corresponding bound states masses are given by

(n) 3 λ √ 3 λ2 2
mb = 2M∞ sin n = n 2m 1 − n + ··· . (4.45)
4 m2 32 m4
√
Note that the leading term is consistently given by multiples of 2m, which is the known
mass of the elementary boson. Contrary to the sine-Gordon model, we now have all integer
multiples of this mass and not only the odd ones: this is because we are in the broken phase
of the theory, where the invariance under φ → −φ is lost. Furthermore, this spectrum
exactly coincides with the one derived in [11] by building approximate classical solutions
to represent the “breathers”.
Another important information can be extracted from the residue of F2 (θ ) on the pole
corresponding to the lightest bound state b(1) . This quantity, indeed, has to be proportional
to the one-particle form factor 0|φ|b(1)
through the semi-classical 3-particle on-shell
coupling of kink, antikink and elementary boson gk k̄b , a quantity so far unknown in this
nonintegrable theory:
g
Resθ=θ1 F2 (θ ) = i √ k k̄b (1) 0|φ b(1) . (4.46)
2 2 M∞ mb
√
Since the one-particle form factor takes the constant value 1/ 2, at leading order in the
coupling we get
32m5
gk k̄b = . (4.47)
3λ3/2
Finally, the 1/λ leading contribution of the spectral function is given in this case by
1
2π p0 p π 3 p0 1
ρ̂(p2 ) = δ δ + δ −2 1
. (4.48)
λ m m 2λ M sinh2 √π pm
2
8 This has to be contrasted with the tentative covariant extrapolation discussed in [13].
4.2. Finite volume
Our proposal to describe the finite volume (anti)kink is to use the following solution of
the differential equation (2.4)

φ̄0
φ̄cl (x̄) = (±) 2 − φ̄0 sn √ x̄, k ,
2 2
(4.49)
2
√
with k 2 = 2
− 1, V (φ0 ) = −C and 1 < φ̄0 < 2, where we have rescaled the variables as
φ̄02
√
λ
φ̄ = φ, x̄ = mx. (4.50)
m
This function oscillates with period
√
2 2
2L̄ = 4 K k
φ̄0

between the two values − 2 − φ̄02 and 2 − φ̄02 , and satisfies the antiperiodic boundary
condition φ(x + L) = −φ(x). Moreover, it goes to the standard (anti)kink solution (4.42)
for φ̄0 → 1, i.e., when L̄ → ∞.
For the “classical energy per kink” in this case we find
√
m3 2 1 1 K(k 2 )
Ecl (L) = − φ̄04 K k 2 + φ̄02 2E k 2 − K k 2 + , (4.51)
λ φ̄0 6 3 2
which for L → ∞ indeed reproduces M∞ . Taking into account the k → 1 (k → 0)
expansions of E and K [15] and noting that
√ 1 4
e− 2 mL
= (k ) + · · · ,
256
we derive the following asymptotic expansion of Ecl for large L:
√ m3 −√2 mL √
Ecl (L) = M∞ − 8 2 e + O e−2 2 mL , (4.52)
λ
√
(note that in this theory the mass of the elementary boson is mb = 2 m). By using
Eq. (3.36), it is easy to see that this expansion exactly reproduces the value (4.47) for the
3-particle coupling gk k̄b , previously obtained by looking at the residue of the form factor
in infinite volume.
Comparing the inverse Fourier transform (2.20) with the expansion [15]
∞ (2n−1)π
π sin 2K u
sn(u) = , (4.53)
kK sinh (2n−1)π K
n=1 2K
we obtain for the form factor in a finite volume the following expression
L/2
m φ̄0
F2 (θn ) = M √ 2 − φ̄0
2 da e iM(iπ−θn )a
sn √ ma
λ 2
−L/2

2 1
= iπ √
λ sinh 2 K M(iπ − θn )
M , (4.54)
mφ̄0
so that, the 1/λ leading contribution to the spectral function is given by

2π En pn π 3 En 1
ρ̂(En , pn ) = δ δ + δ −2 √2 . (4.55)
λ m m 2λ M sinh mφ̄ K pn
2
0
Again, as in the sine-Gordon case, the finite volume dependence of these quantities comes
from the factor M(L)K (k 2 ), where M(L) is the kink mass given by Eq. (4.51).
5. Further developments and open problems
Although the form factors provide an important piece of information about the quantum
theory, a more complete understanding of the semi-classical behaviour of sine-Gordon and
φ 4 models requires an extension of the DHN procedure to finite volume. This would permit
to analyze the energy levels of the kink “particles” and to compute the quantum corrections
to their masses. A detailed discussion of these topics will be presented somewhere else
[21].
Furthermore, in order to systematically analyze the spectra of the proposed theories
on the cylinder, it seems necessary to find the nontrivial periodic classical solutions in the
vacuum sector and apply the DHN procedure to them. As a matter of fact, the identification
the kink–antikink bound states masses in infinite volume from the dynamical poles of
F2 (θ ) is a very powerful tool, whose application to other theories is a topic interesting in
itself, but cannot be directly implemented on the cylinder.
Finally, an important open problem which also deserves further investigation is the
evaluation of the higher loop corrections to the semi-classical energies (masses), form
factors and Green functions. For the infinite volume case, relevant loop calculations around
the nonperturbative kink solutions have been performed for the sine-Gordon and φ 4 models
in the papers [22]. Whether one can extend these loop calculations to the case of finite
volume (cylinder geometry), in order to estimate the loop corrections to the semi-classical
scaling functions, form factors, etc., remains to be seen. This is particularly interesting in
the sine-Gordon case, because it would permit to understand whether also in finite volume
the semi-classical results can be related to the exact ones by a simple redefinition of the
coupling.
Acknowledgements
The authors thank G. Delfino for stimulating discussions. G.S. acknowledges UNESP,
FAPESP and SISSA for financial support, which makes this collaboration possible. G.M.
and V.R. are grateful to IFT—Sao Paulo for the warm hospitality.
References
[1] M. Karowski, P. Weisz, Nucl. Phys. B 139 (1978) 455;

F.A. Smirnov, Form Factors in Completely Integrable Models of Quantum Field Theory, in: Advanced Series
in Mathematical Physics, Vol. 14, World Scientific, Singapore, 1992.
[2] J.L. Cardy, G. Mussardo, Nucl. Phys. B 410 (1993) 451;
V.P. Yurov, Al.B. Zamolodchikov, Int. J. Mod. Phys. A 6 (1991) 3419;
Al.B. Zamolodchikov, Nucl. Phys. B 358 (1991) 524;
G. Delfino, G. Mussardo, Phys. Lett. B 324 (1994) 40;
G. Delfino, G. Mussardo, Nucl. Phys. B 455 (1995) 724.
[3] A. LeClair, G. Mussardo, Nucl. Phys. B 552 (1999) 624.
[4] Al.B. Zamolodchikov, Nucl. Phys. B 342 (1990) 695.
[5] V.V. Bazhanov, S.L. Lukyanov, A.B. Zamolodchikov, Nucl. Phys. B 489 (1997) 487.
[6] V.P. Yurov, Al.B. Zamolodchikov, Int. J. Mod. Phys. A 5 (1990) 3221.
[7] F.A. Smirnov, Quasi-classical study of form factors in finite volume, hep-th/9802132.
[8] P. Fonseca, A. Zamolodchikov, Ising field theory in a magnetic field: analytic properties of the free energy,
hep-th/0112167;
A.I. Bugrij, The correlation function in two-dimensional Ising model on the finite size lattice. I, hep-
th/0011104;
A.I. Bugrij, Form factor representation of the correlation function of the two-dimensional Ising model on a
cylinder, hep-th/0107117.
[9] V.E. Korepin, N.A. Slavnov, Int. J. Mod. Phys. B 13 (1999) 2933.
[10] C.G. Callan, D.J. Gross, Nucl. Phys. B 93 (1975) 29;
J.L. Gervais, A. Jevicki, B. Sakita, Phys. Rev. D 12 (1975) 1038;
J.L. Gervais, A. Jevicki, Nucl. Phys. B 110 (1976) 113.
[11] R.F. Dashen, B. Hasslacher, A. Neveu, Phys. Rev. D 10 (1974) 4130;
R.F. Dashen, B. Hasslacher, A. Neveu, Phys. Rev. D 11 (1975) 3424.
[12] R. Rajaraman, Solitons, Instantons, North-Holland, Amsterdam, 1982.
[13] J. Goldstone, R. Jackiw, Phys. Rev. D 11 (1975) 1486.
[14] P.H. Weisz, Phys. Lett. B 67 (1977) 179.
[15] I.S. Gradshteyn, I.M. Ryzhik, Table of Integrals, Series, and Products, Academic Press, New York, 1980.
[16] K. Takayama, M. Oka, Nucl. Phys. A 551 (1993) 637.
[17] A. Seeger, A. Kochendörfer, Z. Phys. 130 (1951) 321;
A. Seeger, H. Donth, A. Kochendörfer, Z. Phys. 134 (1953) 173.
[18] M. Lüscher, Commun. Math. Phys. 104 (1986) 177.
[19] T.R. Klassen, E. Melzer, Nucl. Phys. B 362 (1991) 329.
[20] A.B. Zamolodchikov, Al.B. Zamolodchikov, Ann. Phys. 120 (1979) 253.
[21] G. Mussardo, V. Riva, G. Sotkov, in preparation.
[22] H.J. de Vega, Nucl. Phys. B 115 (1976) 411;
J. Verwaest, Nucl. Phys. B 123 (1977) 100.
Infinite products of large random matrices

and matrix-valued diffusion
Ewa Gudowska-Nowak a,b , Romuald A. Janik b , Jerzy Jurkiewicz b ,
Maciej A. Nowak a,b
a Gesellschaft für Schwerionenforschung, Planckstrasse 1, D-64291 Darmstadt, Germany
b M. Smoluchowski Institute of Physics, Jagellonian University, Reymonta 4, PL-30-059 Kraków, Poland
Received 13 May 2003; received in revised form 22 July 2003; accepted 8 August 2003
Abstract
We use an extension of the diagrammatic rules in random matrix theory to evaluate spectral
properties of finite and infinite products of large complex matrices and large Hermitian matrices.
The infinite product case allows us to define a natural matrix-valued multiplicative diffusion process.
In both cases of Hermitian and complex matrices, we observe the emergence of a “topological phase
transition”, when a hole develops in the eigenvalue spectrum, after some critical diffusion time τcrit is
reached. In the case of a particular product of two Hermitian ensembles, we observe also an unusual
localization–delocalization phase transition in the spectrum of the considered ensemble. We verify
the analytical formulas obtained in this work by numerical simulation.
PACS: 05.40.+j; 05.45.+b; 05.70.Fh; 11.15.Pg
Keywords: Non-Hermitian random matrix models; Diagrammatic expansion; Products of random matrices
1. Introduction
Applications of random matrix theory (RMT) in physics range from interpretation of

complex spectra of energy levels in atomic and nuclear physics [1], studies of disordered
systems [2,3], chaotic behavior [4] to quantum gravity [5–7]. Use of progressively
developed methods of RMT has also proved to be of particular interest in other sciences
E-mail addresses: gudowska@th.if.uj.edu.pl (E. Gudowska-Nowak), ufrjanik@if.uj.edu.pl (R.A. Janik),

jurkiewicz@th.if.uj.edu.pl (J. Jurkiewicz), nowak@th.if.uj.edu.pl (M.A. Nowak).
doi:10.1016/j.nuclphysb.2003.08.012
480 E. Gudowska-Nowak et al. / Nuclear Physics B 670 [FS] (2003) 479–507
such as meteorology [8], image processing [9], population ecology [10] or economy
[11,12].
In this work, we will be concerned with infinite products of random matrices (PRM)
sampled from general Gaussian ensembles. Products of that type appear in various
branches of physics and interdisciplinary applications [13] where system dynamics is
described in terms of evolution operators with random coefficients. Perhaps the best
known examples are those related to thermal properties of disordered magnetic systems,
in particular, to the localization of electronic wave functions in random potentials
[2,14]. Attention has also been focused on applications of PRM to the analysis of
chaotic dynamical systems [13,15–17], where stability of trajectories and their sensitive
dependence on initial conditions are measured in terms of characteristic Lyapunov
exponents. Moreover, there are several results exploring the use of PRM formalism in
applied physics and interdisciplinary research, ranging from the studies of stability of large
eco- and social-systems [10], adaptive algorithms and the analysis of system performance
under the influence of external noises [18] to image compression [19] and communication
via antenna arrays [20,21]. However, despite the richness of the potential applications, the
number of analytical results available in the PRM theory is still rather limited.
In this paper we develop effective calculational techniques for studying products of
random matrices in the large N limit, and use these tools to derive the properties of
the natural matrix valued generalization of the geometric (multiplicative) diffusion type
process. The scalar versions of these processes, leading to log-normal distributions are
ubiquitous in various fields [22] and we expect the matrix valued generalizations to find
numerous applications.
Let us briefly review the conventional (scalar) multiplicative (geometric) random walk
in one dimension in the presence of some external drift force. The process belongs to the
class of (Markovian) Ito diffusion processes, whose stochastic variable s undergoes an
evolution
ds
= µ(s, t) dt + σ (s, t) dWt . (1)
s
In the above stochastic differential equation (SDE), part of the evolution is driven by a
variable dWt that represents the Wiener process (integral of the Gaussian white-noise),
respecting

dWt = 0, dWt2 = dt. (2)
For simplicity we will limit ourselves to the constant drift µ and constant variance σ . Finite
time evolution of the system could be viewed as a string of increments

T T T T
s(T ) = lim 1+µ +σ x1 1 + µ + σ x2 · · ·
M→∞ M M M M

T T
× 1+µ +σ xM s(0), (3)
M M
where M stands for the number of steps in the discretization of the time interval. After
averaging s(T ) over the independent identical distributions (iid) of the Gaussian variables
E. Gudowska-Nowak et al. / Nuclear Physics B 670 [FS] (2003) 479–507 481
xi and taking the limit M → ∞, we recover1 the well-known solution for the probability
density of s(T ):

1 (log(s/s0 ) − µT + 12 σ 2 T )2
p(s, T |s0 , 0) = √ exp − (4)
s 2πσ 2 T 2σ 2 T
with s restricted to positive values.
The aim of this paper is to study a similar construction in the space of matrices taking
the defining equation (3) as a guideline for the generalization. We are therefore interested
in properties of the matrix-valued evolution operator defined as

T T T T
Y (T ) = lim 1+µ +σ X1 1 + µ + σ X2 · · ·
M→∞ M M M M

T T
× 1+µ +σ XM , (5)
M M
where µ is some deterministic “drift” matrix and the stochastic matrices Xi belong to
identical independent random matrix ensembles. In particular we will be studying the
eigenvalue distribution of Y (T ):
1 (2)
ρT (z) = tr δ z − Y (T ) , (6)
N
where the average is taken over the stochastic N by N matrices Xi appearing in the
definition of Y (T ).
Contrary to several standard multidimensional extensions of Brownian walks, we
concentrate here on studying how the full spectrum of operator Y evolves with time T .
Let us note some key features of the operator Y . First, since the matrices Xi in
general do not commute, we are dealing with a ‘path ordered product’. Second, even
if the matrices Xi are Hermitian, their product is not, i.e., the spectrum in general
disperses into the complex plane, showing—as pointed out in this paper—some rather
unusual feature described as a “topological phase transition”. Namely the support of the
eigenvalue distribution changes from a simply-connected two-dimensional island to a two-
dimensional island with a hole.
We note that the time evolution of some initial vector |0 under Y forms a very general
setup for several multivariate stochastic evolutions of the type
|τ = Y (τ )|0. (7)
Analytical results for such processes are however scarce. In a recent study, Jackson et
al. [23] solved the matrix-valued evolution driven by Gaussian noise of the type considered
above in the case of 2 by 2 Hermitian matrices. There are also some known results
for the diffusion of the norm of vector |τ , since this problem can be linked to the
1 The log-normal distribution s(T ) is easy to infer looking at the form of the product (3). Taking the
logarithm, expanding and using the central limit theorem we immediately see, that the r.h.s. tends to the Gaussian
distribution.
Lyapunov exponents of the Hermitian matrix Y † Y . We briefly comment on this issue

in Section 9.
The main result of this work is the derivation of exact (finite M) and asymptotic
(M → ∞) spectral formulas for the evolution operators Y , in the limit where the dimension
N of Y tends to infinity. For simplicity, we set to zero the deterministic entries µ—although
one can easily include them using our formalism. In this paper we also limit ourselves
to ensembles of the Gaussian type. Generalizations to non-Gaussian measures (involving
finite and infinite spectral supports) will be investigated elsewhere. We discuss here two
cases depending on whether the Xi ’s are (i) Hermitian Gaussian matrices (GUE) or (ii)
complex (arbitrary) Gaussian matrices (GCE). Similar constructions could be used for real
and quaternionic matrices.
The main motivation for this work is to formulate a new and relatively simple theoretical
framework for studying spectra of the diffusive-like evolution operators with their further
applications to various branches of the complex systems analysis in physics, mathematics
and in interdisciplinary science.
In particular, the formalism outlined here could be used as a starting point for matrix
versions of super- or sub-diffusions, generated by, e.g., matrix Lévy analogs of the stable
power-tailed distributions. It will be used also for the study of spectral statistics of infinite
products of unitary operators (e.g., Polyakov lines) [24].
Most of the derivations presented in this paper will be based on diagrammatic methods.
In order to make the paper self-contained, in the first two sections we briefly present
the formalism, first for Hermitian RMM (Section 2) and then in Section 3, we recall the
extension of diagrammatic methods to the non-Hermitian case [27,28], by using complex
Gaussian non-Hermitian random matrix model as an example. In Section 4 we move to the
main part of our paper and introduce our method for the treatment of products of random
matrices by performing calculations for a product of two matrices. Then (Section 5) we
move toward the products of M complex matrices, and we perform the limit M → ∞. We
present the final formulas for the spectral density and the equations defining the boundary
of the two-dimensional support of the eigenvalues. In Sections 6 and 7 we consider the
case of Hermitian matrices which, surprisingly, is more difficult than the previous one. We
analyze first (Section 6) the product of two matrices. This is interesting on its own, since it
exhibits quite singular behavior—for evolution times smaller than a certain critical value,
the a priori complex eigenvalues are “frozen” to the real axis. In Section 7 we move toward
the general case. For products of matrices involving more than two Hermitian matrices,
the eigenvalues always form two-dimensional islands. We solve the spectral problem in
the large M limit. In Section 8 we show the numerical analysis, confirming our analytical
predictions and the existence of “topological phase transitions”, corresponding to structural
changes of the spectrum.
Finally, in Section 9 we briefly discuss the problems of stability of such products and
the relation of our results to the standard Lyapunov exponents. Section 10 concludes the
paper.
2. Hermitian random ensembles
A key problem in random matrix theories is to find the distribution of eigenvalues λi , in

the large N (size of the matrix H ) limit, i.e.,

N
1
ρ(λ) = δ(λ − λi ) , (8)
N
i=1
where the averaging · · · is done over the ensemble of N × N random Hermitian matrices
generated with probability
P (H ) ∝ e−N Tr V (H ) . (9)
The eigenvalues of course lie on the real axis. By introducing the resolvent (Green’s
function)

1 1
G(z) = Tr (10)
N z−H
with z = z1N and by using the standard relation
1 1
= P.V. ∓ iπδ(λ) (11)
λ ± i λ
the spectral function (8) can be derived from the discontinuities of the Green’s function
(10)
1 1
lim G(λ − i) − G(λ + i) = Tr δ(λ − H )
2πi →0 N

N
1
= δ(λ − λi ) = ρ(λ). (12)
N
i=1
There are several ways of calculating Green’s functions for HRMM [1,3,7]. We will
follow the diagrammatic approach introduced by [25]. A starting point of the approach is
the expression allowing for the reconstruction of the Green’s function from all the moments
Tr H n ,

1 1 1 1 1 1 1 1 1
G(z) = Tr = Tr + H + H H + · · ·
N z−H N z z z z z z
1 1
= Tr H n . (13)
N n zn+1
The reason why the above procedure works correctly for Hermitian matrix models is the
fact that the Green’s function is guaranteed to be holomorphic in the whole complex plane
except at most on one or more 1-dimensional intervals. We will use the diagrammatic
method to evaluate efficiently the sum of the moments on the right-hand side. We will now
restrict ourselves to the well-known case of a random Hermitian ensemble with Gaussian
distribution.
1 b a c 1
δ , Hb Hd = δbc δda
z a N
Fig. 1. Large N “Feynman” rules for “quark” and “gluon” propagators.
Fig. 2. Diagrammatic expansion of the Green’s function up to the O(H 4 ) terms.
The first step is to introduce a generating function with a matrix-valued source J :

N
Z(J ) = dH e− 2 (Tr H )+Tr H ·J ,
2
(14)
where we integrate over all N 2 elements of the matrix H . All the moments follow directly
from Z(J ) through the relation
n
1 ∂
Tr H =
n
Tr Z(J ) (15)
Z(0) ∂J J =0
and are straightforward to calculate, since in the Gaussian case the partition function (14)
reads Z(J ) = exp 2N
1
Tr J 2 . Accordingly, the lowest non-zero expectation value is

a c ∂ 2 Z(J ) 1 ∂Jdc Z(J ) 1
Hb Hd = = = δbc δda (16)
b d
∂Ja ∂Jc J =0 N ∂Ja b N
J =0
and the next non-vanishing expectation value reads
g 1 g
Hba Hdc Hfe Hh = 2 δbc δda δf δhe + δha δgb δfc δed + δfa δeb δhc δgd . (17)
N
The key idea in the diagrammatic approach is to associate to the expressions for the
moments, like the above, a graphical representation following from a simple set of rules.
The power of the approach is that it enables to perform a resummation of the whole power
series (13) through the identification of the structure of the relevant graphs.
We depict the “Feynman” rules in Fig. 1, similar to the standard large N diagrammatics
for QCD [26]. The 1/z = 1/zδab in (13) is represented by a horizontal straight line. The
propagator (16) is depicted as a double line.
The diagrammatic expansion of the Green’s function is visualized in Fig. 2, where
one connects the vertices with the double line propagators in all possible ways. Each
“propagator” brings a factor of 1/N , and each loop a factor of δaa = N . From the three
terms, corresponding to (17) contributing to tr H 4 only the first two are presented in
Fig. 2 (the third and the fourth diagram). The diagram corresponding to the third term in
(17) represents a non-planar contribution which is suppressed as 1/N 2 and hence vanishes
when N → ∞. In general, only planar graphs survive the large N limit.
Fig. 3. Schwinger–Dyson equation for rainbow diagrams.
The resummation of (13) is done by introducing the self-energy Σ comprising the sum
of all one-particle irreducible graphs (rainbow-like). Then the Green’s function reads
1
G(z) = . (18)
z − Σ(z)
In the large N limit the equation for the self energy Σ, follows from resumming the
rainbow-like diagrams of Fig. 2. The resulting equation (“Schwinger–Dyson” equation of
Fig. 3) encodes pictorially the structure of these graphs and reads
1 N
Σ= Tr G1 = G = G. (19)
N N
Eqs. (18) and (19) give immediately G(z − G) = 1 which can be solved to yield
1
G∓ (z) = z ∓ z2 − 4 . (20)
2
Only the G− is a normalizable solution, with the proper asymptotic behavior G(z) → 1/z
in the z → ∞ limit. From the discontinuity (cut), using (12), we recover Wigner’s
semicircle [29] for the distribution of the eigenvalues for Hermitian random matrices
1
ρ(λ) = 4 − λ2 . (21)
2π
3. Non-Hermitian random ensembles
The main difficulty in the treatment of non-Hermitian random matrix models is the fact
that now the eigenvalues accumulate in two-dimensional domains in the complex plane
and the Green’s function is no longer holomorphic. Therefore the power series expansion
(13) no longer captures the full information about the Green’s function. In particular the
eigenvalue distribution is related to the non-analytic (non-holomorphic) behaviour of the
Green’s function:
1
∂z̄ G(z) = ρ(z). (22)
π
This phenomenon can be easily seen even in the simplest non-Hermitian ensemble—the
Ginibre–Girko one [30,31], with non-Hermitian matrices X, and measure
P (X) = e−N Tr XX .
†
(23)
It is easy to verify that all moments vanish tr Xn
= 0, for n > 0 so the expansion (13)
gives the Green’s function to be G(z) = 1/z (diagrammatically this follows from the fact
that the propagator Xba Xdc vanishes and hence the self-energy Σ = 0). The true answer
is, however, different. Only for |z| > 1 one has indeed G(z) = 1/z. For |z| < 1 the Green’s
function is non-holomorphic and equals G(z) = z̄.
In spite of the above difficulties a diagrammatic approach can be used for recovering
the full information including the eigenvalue density ρ(z). The first step is to define
a regularized Green’s function from which one can extract ρ(z). This has been done
by exploiting the analogy to two-dimensional electrostatics [4,31,32]. The resulting
regularized Green’s function is however difficult to calculate. The second step is to give
an equivalent linearized form which can form a basis for a diagrammatic calculation [27].
We will now briefly review the above developments.
Let us define the “electrostatic potential”
1
F = Tr ln (z − M)(z̄ − M† ) + 2 . (24)
N
Then

∂ 2 F (z, z̄) 1 2 π (2)
lim = lim Tr = δ (z − λi ) ≡ πρ(x, y)
→0 ∂z∂ z̄ →0 N (|z − M|2 + 2 )2 N
i
(25)
represents Gauss law, where z = x + iy. The last equality follows from the representation
of the complex Dirac delta
2
πδ (2) (z − λi ) = lim . (26)
→0 ( 2 + |z − λi |2 )2
In the spirit of the electrostatic analogy we can define the Green’s function G(z, z̄), as an
“electric field”

∂F 1 z̄ − X†
G≡ = lim Tr . (27)
∂z N →0 (z̄ − X† )(z − X) + 2 )
Then Gauss law reads
∂z̄ G = πρ(x, y) (28)
in agreement with (22).
As mentioned above, instead of working ab initio with the object (27), and in view of
applying diagrammatic methods it is much more convenient to proceed differently— as we
will now discuss.
Following [27] we define the matrix-valued resolvent through

−1
1 z−X i
Ĝ = TrB2
N i z̄ − X†

1 A B G11 G11̄
= TrB2 ≡ (29)
N C D G1̄1 G1̄1̄
with
z̄ − X† −i
A= , B= ,
(z̄ − X† )(z − X) + 2 (z − X)(z̄ − X† ) + 2
−i z−X
C= , D= (30)
(z̄ − X† )(z − X) + 2 (z − X)(z̄ − X† ) + 2
and where we introduced the ‘block trace’ defined as

A B Tr A Tr B
TrB2 ≡ . (31)
C D 2N×2N Tr C Tr D 2×2
Then, by definition, the upper-right component G11 , is equal to the Green’s function (27).
The block approach has several advantages. First of all it is linear in the random
matrices X allowing for a simple diagrammatic calculational procedure. Let us define 2N
by 2N matrices

z i1 X 0
Z= , H= . (32)
i1 z̄ 0 X†
Then the generalized Green’s function is given formally by the same definition as the usual
Green’s function G,

1 1
G= TrB2 . (33)
N Z −H
What is more important, also in this case the Green’s function is completely determined by
the knowledge of all matrix-valued moments

TrB2 Z −1 HZ −1 H · · · Z −1 . (34)
This last observation allows for a diagrammatic interpretation. The Feynman rules are
analogous to the Hermitian ones, only now one has to keep track of the block structure
of the matrices, e.g., single straight lines will now be associated with a matrix factor Z −1 .
We will now demonstrate the above procedure by solving diagrammatically the complex
Gaussian random matrix model.
3.0.1. Example: the Girko–Ginibre ensemble

The ensemble is defined by the measure
P (X) ∝ e−N Tr XX .
†
(35)
In this case the double line propagators are
a c †a †c a †c †a c 1 a b
Xb Xd = X b X d = 0, Xb X d = X b Xd = δd δc . (36)
N
As previously, we introduce the self-energy Σ̃, (but which is now matrix-valued), in terms
of which we get a two by two matrix expression
G = (Z − Σ̃)−1 , (37)
where the 2 by 2 matrix Z reads

z i
Z= . (38)
i z̄
The resummation of the rainbow diagrams for Σ̃ is more subtle, but follows easily from
the structure of the propagators (36). The analogue of (19) is now:

Σ11 Σ11̄ 0 G11̄
Σ̃ ≡ = . (39)
Σ1̄1 Σ1̄1̄ G1̄1 0
The two by two matrix equations (37)–(39) completely determine the problem of finding
the eigenvalue distribution for the Girko–Ginibre ensemble. Inserting (39) into (37) we get:

G11 G11̄ 1 z̄ G11̄
= 2 · . (40)
G1̄1 G1̄1̄ |z| − G11̄ G1̄1 G1̄1 z
Note that at this moment we can safely put to zero the regulators . Looking at the off-
diagonal equation
G11̄
G11̄ = (41)
|z|2 − G11̄ G1̄1
we see that there are two solutions: one with G11̄ = 0, and another with G11̄ = 0. The first
one leads to a holomorphic Green’s function, and a straightforward calculation gives
1
G(z) = . (42)
z
The second one is non-holomorhic, imposing the condition
|z|2 − b2 = 1, (43)
where we denoted G11̄ G1̄1 ≡ b2 , hence
G(z, z̄) = z̄ (44)

which leads, via the Gauss law, to
1 ∂ 1
ρ(x, y) = G(z, z̄) = . (45)
π ∂ z̄ π
Both solutions match at the boundary b2 = 0, which in this case reads zz̄ = 1. In such a
simple way we recovered the results of Ginibre and Girko for the complex non-Hermitian
ensemble. The eigenvalues are uniformly distributed on the unit disk |z|2 < 1.
This example illustrates more general properties of the matrix valued generalized
Green’s function. Each component of the matrix carries important information about
the stochastic properties of the system. There are always two solutions for G11 , one
holomorphic, another non-holomorphic. The second one leads, via Gauss law, to the
eigenvalue distribution. The shape of the “coastline” bordering the “sea” of complex
eigenvalues is determined by the matching conditions for the two solutions, i.e., it is
determined by imposing on the non-holomorphic solution for b2 the equation b2 = 0. For
other important features of the explained above diagrammatic method and more complex
examples see [27,33].
4. Product of two complex matrices
We will now move toward the main goal of this paper, namely the evaluation of the
spectral properties of the random matrix products of the form (5)—with zero drift and
normalized variance:

T T T
Y (T ) = lim YM ≡ lim 1+ X1 1+ X2 · · · 1 + XM .
M→∞ M→∞ M M M
(46)
In principle the methods of Section 3 could be used to study such random matrices,
however the non-linear product structure would make the resulting diagrammatic rules
exceedingly difficult to control. Fortunately, for the purposes of calculating the Green’s
function, and hence the eigenvalue density, the problem may be linearized at the cost of
increasing the size of the matrices. In this section we will show how the method works
for Y2 —the product of two matrices, while in Section 5 we will consider the general case
including the limit M → ∞. We therefore have to calculate

τ τ
Y2 = 1 + X1 1 + X2 ≡ A1 A2 , (47)
2 2
where X1 and X2 belong to identical independent Gaussian complex ensembles (iiGCE =
Girko–Ginibre). Since the matrices are non-Hermitian we have to use the formalism of
Section 3 with a conjugate ‘antiholomorphic’ copy. From now on in the rest of the paper
we set the regulator to zero. In order to linearize the product structure we use a trick: we
introduce a further doubling, thus constructing an auxiliary 4N by 4N Green’s function
   0 A −1
w 0 0 0 1 0 0
 0 w 0 0   A2 0 0 0  
G(w) =  
 0 0 w̄ 0  −  0 0 0

A†2 
0 0 0 w̄ 0 0 A†1 0 4N×4N
   −1
w −1 0 0 0 X1 0 0
 −1 w " X 0  
0 0  0 0
=  − 2  † 
τ 2
 0 0 w̄ −1  0 0 0 X2  
0 0 −1 w̄ 0 0 X1† 0 4N×4N
√ −1
≡ W − τ/2 X . (48)
In the above we first separated the “deterministic” part from the “random one” (second
line) and in the last equality we introduced the notation similar to the Hermitian and non-
Hermitian cases analyzed in the previous sections.
In analogy to (31), we define now a block-trace operation trB4 , where we trace each N
by N block of the 4N by 4N matrix G(w) separately. In such a way, we obtain a 4 by 4
auxiliary Green’s function
 
g11 g12 g11̄ g12̄
 g21 g22 g21̄ g22̄  1
g(w) ≡ 
g

 = trB4 G(w). (49)
1̄1 g1̄2 g1̄1̄ g1̄2̄ N
g2̄1 g2̄2 g2̄1̄ g2̄2̄ 4×4
The advantage of this function lies in the fact, that the eigenvalues of the product A1 A2
coincide with the squares of the eigenvalues of the 2N × 2N block matrix

0 A1
. (50)
A2 0
This is due to the off-diagonal structure of the sub-blocks in (50). For a general discussion
for an arbitrary product see Section 5.
As before we define a 4 by 4 matrix of self-energies Σ̃ij
g(w) = (W − Σ̃)−1 , (51)

where W is the result of block-tracing ‘bold’ W.
Similarly as in the Girko–Ginibre case, we can now diagrammatically analyze the
content of the matrix Σ̃. We see that only the “double line” propagators X1 X1† X1 and
X2 X2† X2 are different from zero. Therefore from all of the 16 elements of matrix Σ̃ only
four are non-zero. Moreover, due to identical measures of X1 and X2 variables, we get
Σ11̄ = Σ22̄ = αg1̄1 = αg2̄2 , Σ1̄1 = Σ2̄2 = αg11̄ = αg22̄ , (52)
where the factor α = τ/2 comes from the propagator in the corresponding Schwinger–
Dyson equation of the type discussed in the previous chapter.
Filling the matrix Σ̃ with entries (52) we arrive at the 4 by 4 matrix equation for the
elements of the Green’s function:
 
g11 g12 g11̄ g12̄
 g21 g22 g21̄ g22̄ 
 
g g g g 
1̄1 1̄2 1̄1̄ 1̄2̄
g2̄1 g2̄2 g2̄1̄ g2̄2̄ 4×4
   −1
w −1 0 0 0 0 αg1̄1 0
 −1 w 0 0   0 0 0 αg1̄1 
= 
 0
−  . (53)
0 w̄ −1   αg11̄ 0 0 0 
0 0 −1 w̄ 0 αg11̄ 0 0
As in the Ginibre case, we solve it for g1̄1 and g11̄ . Then, using the solution, we
calculate g11 . Coming back to original variables with w2 = z we set G(z, z̄) = g11 (w).
The eigenvalue distribution then follows from
1w
ρ(x, y) = ∂z̄ G(z, z̄). (54)
π z
This is a consequence of the relation between the eigenvalues of the block matrix (50)
and the eigenvalues of the product A1 A2 . A general derivation is outlined in the following
section.
A little algebra (basically an inversion of the 4 by 4 matrix) shows, that, as before, we
get two solutions for b2 = g11̄ g1̄1 = g22̄ g2̄2 : (i) the holomorphic one (corresponding to
b2 = 0, i.e., all Σij equal to zero); (ii) the non-holomorphic one, given by the equation
τ 2
|w| − α 2 b2 + 1 = α 2 b2 − 1 + w2 α 2 b2 − 1 + w̄2 − α 2 b2 (w + w̄)2 . (55)
2
Fig. 4. Evolution of the contour of the non-holomorphic domain in the eigenvalue spectrum of the product of
two Ginibre–Girko ensembles as a function of evolution times τ = 0.1, 0.25, 0.5, 1, 1.5, 2, 3, 4, 5, 6, rising from
inside to outside.
The boundary of the eigenvalue support is given, as before, by the above equation with b2
set to 0. It is therefore a fourth order conchoid-like planar curve, which in polar coordinates
is given by
τ
(1 + r) = r 2 + 1 − 2r cos φ. (56)
2
On Fig. 4 we show the shape of the non-holomorphic island for several sample
“evolution times” τ .
The Green’s function for the product follows from
w(w̄2 − 1) − α 2 b2 w̄
g11 (w) = , (57)
det
where

det = α 2 b2 − 1 + w2 α 2 b2 − 1 + w̄2 − α 2 b2 (w + w̄)2 . (58)
5. Product of arbitrary number of complex matrices
We consider now the product of an arbitrary number of matrices. To see the pattern, let
us look briefly at the case of three matrices

τ τ τ
Y3 = 1 + X1 1 + X2 1 + X3 ≡ A1 A2 A3 , (59)
3 3 3
where X1 , X2 , X2 again belong to the Girko–Ginibre ensemble.
Our approach again follows from a very simple exact relation between the eigenvalues
of a product of N × N matrices A1 A2 A3 and the eigenvalues of a block matrix
 
0 A1 0
B3 =  0 0 A2  . (60)
A3 0 0 3N×3N
Indeed, if {λi } are the eigenvalues of A1 A2 A3 then the eigenvalues of the block matrix
1/3 1/3 2π i 1/3 4π i
are {λi , λi · e 3 , λi · e 3 }. This is an exact relation for any N and follows from the
relation between the resolvents (here the matrices Ai are of finite size and fixed, i.e., no
averaging)
1 1 1 1
GB3 (w) = tr , GA1 A2 A3 (z) = tr , (61)
3N w − B3 N z − A1 A2 A3
namely

wGB3 (w) = zGA1 A2 A3 w3 ≡ z . (62)
This is due to the cyclic structure of the block matrix (60). Obviously, only multiplicities
of the cubic powers of B3 contribute to the trace.
The relation between the eigenvalues now follows from the location of the finite number
of poles of both functions.
We can thus safely calculate the eigenvalue density for the block matrix B3 using the
diagrammatic methods and then extract the density for the product through
1 1 ww̄
ρA1 A2 ···AM (z, z̄) = ∂z̄ GA1 A2 ···AM = ρBM (w, w̄), (63)
π M zz̄
where M = 3 and wM = z and ρBM (w, w̄) = π1 ∂w̄ GB (w, w̄).
To extract only the moments involving powers of the triples A1 A2 A3 , we construct an
auxiliary 6N by 6N Green’s function, triplicating the A1 A2 A3 products by rewriting them
as cyclic block matrices:
 
w −1 0 0 0 0
 0 w −1 0 0 0 
 
 −1 0 w 0 0 0 
G(w) =  
 0 0 0 w̄ 0 −1 
 
0 0 0 −1 w̄ 0
0 0 0 0 −1 w̄
 −1
0 X1 0 0 0 0
 0 0 X2 0 0 0 
"  X 0 0 0 0 0


− τ3  † 
3
 0 0 0 0 0 X3  , (64)
 
 0 0 0 X1† 0 0 
0 0 0 0 X2† 0 6N×6N
where again we separated the “deterministic” part from the “random one”. Introducing
now the block-trace operation trB6 , we obtain a 6 by 6 auxiliary Green’s function g(w) =
trB6 G(w).
The generalization for arbitrary M is now straightforward. For

τ τ τ
YM = 1 + X1 1 + X2 · · · 1 + XM (65)
M M M
we define an auxiliary 2MN by 2MN Green’s function of the form

# " $−1
W − M τ
X
G(w) = " , (66)
− M τ
X† W†
where the blocks
 w −1 0 · · · 0 
 0 w −1 · · · 0 
 
W =  ··· ··· ··· ··· ··· 
 (67)
 0 0 · · · w −1 
−1 0 ··· 0 w MN×MN
and
 0 X1 0 ··· 0 
 0 0 X2 ··· 0 
 
X =
 ··· ··· ··· ··· ··· 
 (68)
 0 0 ··· 0 XM−1 
XM 0 ··· 0 0 MN×MN
are themselves NM by NM matrices, i.e., each of the listed elements in W and in X is
itself an N by N matrix, either diagonal, denoted by a bold symbol, or a random entry Xi ,
otherwise an N by N block of zeroes.
We take now a block-trace operation trBM , where we trace each N by N block of the
2MN by 2MN matrix G(w) separately. In such a way, we obtain an 2M by 2M auxiliary
Green’s function
 
g11 · · · g1M g11̄ · · · g1M̄
 .. .. .. .. .. .. 
 . . . . . . 
 
g 
 M1 · · · gMM gM 1̄ · · · gM M̄  1
g(w) ≡   = TrBM G(w). (69)
 g1̄1 · · · g1̄M g1̄1̄ · · · g1̄M̄  N
 
 . .. .. .. .. .. 
 .. . . . . . 
gM̄1 · · · gM̄M gM̄ 1̄ · · · gM̄ M̄ 2M×2M
As before we define a 2M by 2M matrix of self-energies Σ̃ij
−1
W 0
g(w) = − Σ̃ , (70)
0 W†
where W is a result of block-tracing W. We can now diagrammatically analyze the content

of the matrix Σ̃ . As previously, only “double line” propagators Xi Xi† Xi for i = 1, . . . , M
are different from zero. Therefore from all of the 4M 2 elements of the matrix Σ̃ only 2M
are different from zero. Due to the symmetries we get
Σ11̄ = · · · = ΣM M̄ = αg1̄1 = αgM̄M ≡ αg,
Σ1̄1 = · · · = ΣM̄M = αg11̄ = αg1M̄ ≡ α g̃, (71)
where the factor α = τ/M comes from the propagator in the corresponding set of
Schwinger–Dyson equations as in the previous section.
Inserting now the matrix Σ̃ with entries (71) we arrive at a 2M by 2M matrix equation
for the elements of the Green’s function. First, we solve it for g and g̃. Second, using the
solutions, we calculate g11 . Third, using (63) we get an explicit equation for the spectral
density
1 1 ww̄
ρ(x, y) = ∂w̄ g11 (w, w̄), (72)
π M zz̄
where z = x + iy. For arbitrary M, the algebra is rather involved. Luckily, both the block
and the cyclic structure of the main entries W and Σ̃ make taking an inverse of the 2M
by 2M matrix possible. The inverse of a cyclic matrix is a cyclic matrix, and its explicit
form can be obtained from the solution of an associated recurrence relation, e.g., using the
transfer matrix techniques which we will now proceed to do.
At this moment, it is useful to introduce the notation
1 + |w|2 − α 2 b2 = 2|w| cosh u, |w|/w = δ. (73)
The inversion needed for the Green’s function reduces to finding explicit forms of the
cyclic M by M matrices C, C˜ defined as
−1 −1
C = W · W † − α2 b2 1 , C˜ = W † · W − α 2 b2 1 , (74)
where b2 = g g̃. C̃ is just the transpose, or equivalently, the complex conjugate of C. The
elements c0 , . . . , cM−1 of the first row of the cyclic matrix C fulfill the recurrence relation

ck+1 2δ cosh u −δ 2 ck
= (75)
ck 1 0 ck−1
with the ‘boundary conditions’

c1 2δ cosh u −δ 2 cM 1 1
= − . (76)
c0 1 0 cM−1 w 0
The above recurrence can be easily solved using transfer matrix techniques, i.e., by
diagonalizing the 2 by 2 transfer matrix in (75) and returning at the end of the calculations
to the original basis. An explicit solution for the cyclic string ck , for k = 0, 1, . . . , M − 1
reads

1 Λk+ Λk−
ck = − , (77)
w(Λ+ − Λ− ) ΛM + −1 ΛM− −1
where Λ± = δe±u denote the eigenvalues of the transfer matrix.

We may now express our key quantities for arbitrary M in terms of (77):
∗
g11 (w) = [W † · C]11 ≡ w̄ · c0 − cM−1 ,
g = αg[C]˜ 11 ≡ αg · c∗ , g̃ = α g̃[C]11 ≡ α g̃ · c0 , (78)
0
where ‘∗ ’ denotes complex conjugation. As always in the case of the complex spectra, we
have two solutions: trivial (holomorphic), corresponding to g = g̃ = 0, and the non-trivial
(non-holomorphic), corresponding to the case b2 ≡ g g̃ = 0.
Inserting the non-holomorphic solution into g11 , redefining G(z, z̄) = g11 (w) with
wM = z, and using (63) yields the eigenvalue density for YM . Since ∂w̄ z̄ = M z̄/w̄, the
result for the final spectral function can be further simplified
1w
ρ(x, y) = ∂z̄ G(z, z̄). (79)
π z
Explicit formulas for arbitrary M could be easily constructed, e.g., by running any
symbolic program using our solution. The shape of the boundary is given by the condition
that the non-holomorphic solution meets the holomorphic one on the complex plane.
After solving the problem for arbitrary M, we can now address the original problem,
i.e., the solution for a continuous product of matrices, corresponding to the limit M → ∞.
This means, that we have to perform a careful limiting procedure in the above formulas.
Introducing U = Mu, expanding in M, and using the polar parametrization z = r exp iφ,
we get the solution
1 1 iU sin φ
G(z, z̄) = log r + − , (80)
τ 2 τ sinh U
"
where U = log2 r − τ 2 b2 with b2 fulfilling the equation
τ
cos φ = cosh U − sinh U. (81)
2U
The set of Eqs. (80), (81) completes the solution of the continuous product of large complex
matrices with Gaussian disorder. We would like to note, that due to the limiting procedure
M → ∞ differentiation of the Green’s functions has to be done with care. Taking the large
M limit of (79) yields
11
ρ(x, y) = ∂z̄ G(z, z̄). (82)
πz
As a cross-check of our calculations we verified that the spectral function obtained in this
way is real.
Substituting b2 = 0 in (81), we get the equation for the shape of the support of
the eigenvalues for arbitrary τ . Note that the shape of the boundary depends on the
variable log |z|, reflecting in the spectrum the multiplicative (geometric) feature of the
stochastic process induced by the multiplication of random matrices. Explicit solution for
the boundary reads

τ r2 − 1
= r 2 + 1 − 2r cos φ. (83)
2 ln r
It is rewarding to compare solution (83) to the solution of the conchoid-like boundary in

the case of two complex matrices (56). In the limit of infinitely many products, the shape
of the boundary acquires a new symmetry—(83) is invariant under inversion operation
r → 1/r. It is visible also from the general form (81), since the radial dependence for
b 2 = 0 is a function of | log r| only. This symmetry is responsible for the appearance of
a structural change of the spectrum—provided τ is sufficiently large, two boundaries,
related by inversion in the radial variable, appear. We describe this structural change
of the complex spectrum as a “topological phase transition”. In Section 8, we study
numerically the spectral density predicted by our formulas, in particular, the “diffusion”
of the shape of the boundary caused by evolution with the “time” parameter τ . We also
confirm numerically the critical behavior for certain τcrit .
6. Product of two Hermitian matrices
In this section we will consider the spectral distribution of the product

τ τ
Y2 = 1 + H1 1 + H2 , (84)
2 2
where H1 , H2 are not arbitrary complex matrices, but they are identical independent
Hermitian Gaussian matrices (iiGUE). Contrary to naive expectations, the above problem
is more subtle then the analogous problem of the two complex Gaussian matrices. First, let
us note, that the product of two Hermitian matrices is not, in general, a Hermitian matrix,
so the spectrum of the strings of Hermitian matrices develops complex eigenvalues.
Second, the case of only two Hermitian matrices of the considered type turns out to
be in some sense singular—for certain values of the evolution parameter τ , the a priori
complex eigenvalues get localized on the real axis. This phenomenon reminds a bit
the non-Hermitian localization in the Hatano–Nelson model [38]. Third, the problem is
algebraically more involved since there will be more non-zero propagators than in the
complex case.
Since even a product of two Hermitian matrices can develop a priori a complex
spectrum, we have to use the same type of auxiliary Green’s function as in the complex
Gaussian case. The construction parallels exactly the scheme for two complex matrices
considered previously, with the matrices Xi and Xi† replaced both by Hi . The first new
feature appears when we construct the “Schwinger–Dyson” equations for the 4 by 4
self-energy matrix Σ̃ . Since the matrices are Hermitian (Hi = Hi† ), additional non-zero
contractions appear in the matrix Σ̃
   
0 g12 g11̄ 0 0 h g 0
 g21 0 0 g22̄   
Σ̃ = α   ≡ α h 0 0 g, (85)
g 0 0 g   g̃ 0 0 h̃ 
1̄1 1̄2̄
0 g2̄2 g2̄1̄ 0 0 g̃ h̃ 0
where again the factor α = τ/2 comes from the propagator, h and h̃ are the new entries
and we have exploited the symmetries of the problem (e.g., g11̄ = g22̄ , etc.).
The solution of the 4 by 4 matrix equation for the 4 by 4 Green’s function
G(w) = (W − Σ̃)−1 (86)

is more involved, since we have to solve it not only for the “gaps” g and g̃ but also for
additional gaps h and h̃. Only then we can calculate g11 , to recover the wanted eigenvalue
distribution in the same way as in the complex case (54).
Introducing d 2 = (1 + αh)(1 + α h̃) and b 2 = g g̃, we obtain two (modulo their tilded
partners) gap equations
αg 2 2
g=− α b − |w|2 − d 2 ,
det
1 2 2
h= α b (1 + α h̃) + (1 + αh) w̄2 − (1 + α h̃)2 (87)
det
coupled to the needed g11
1 2 2
g11 = −α b w̄ + w w̄2 − (1 + α h̃)2 , (88)
det
where det is the determinant of the matrix W − Σ̃ and reads explicitly

det = α 4 b4 + w2 − (1 + αh)2 w̄2 − (1 + α h̃)2 − 2α 2 b2 d 2 + |w|2 . (89)
Numerical simulations show that the eigenvalue distribution for the product of two matrices
is concentrated on an interval on the real axis for τ < 1/2 and only then when the edge
of the interval reaches the origin, do the eigenvalues start to spread into the complex plane
(see Figs. 5 and 6). Let us show how to find the critical value τ = 1/2 from the above
formulas based on the above numerical observation.
We thus have to find the point when the boundary of the two-dimensional island hits the
origin w = 0. Then Eqs. (87) and (89) simplify to
−1
h= , (90)
1 + αh
1 = αhh̃, (91)
where we set = g g̃ = 0. Then it is easy to verify that this set of equations is satisfied for
b2
α = 1/4 which is equivalent to τ = 1/2.
Fig. 5. Complex eigenvalues of 10 products of two 200 × 200 Hermitian matrices before and at the transition
√
( τ/2 = 0.4 and 0.5). The x and y axis correspond to λ and λ, respectively.
√
Fig. 6. Eigenvalues of 10 products of two 200 × 200 Hermitian matrices after the transition ( τ/2 = 0.7 and
1.0). The axis are defined as in the previous figure.
7. Products of arbitrary number of Hermitian matrices
Finally, we consider an arbitrary number of matrices belonging to iiGUE.

τ τ τ
YM = 1 + H1 1 + H2 · · · 1 + HM . (92)
M M M
As before, the construction of the auxiliary Green’s function parallels the complex case,
up to the point when we have to construct explicitly the matrix of self-energies Σ̃ ,
corresponding to the set of 4M 2 Schwinger–Dyson equations. Generalization of the results
from the previous sections leads to the explicit form of the auxiliary Green’s function.
Introducing
−1
W S
G(w) = (93)
S̃ W †
with blocks
 w −1 − αh 0 ··· 0 
 0 w −1 − αh ··· 0 
 

W =  ··· ··· ··· ··· ···  (94)

 0 0 ··· w −1 − αh 
−1 − αh 0 ··· 0 w M×M
and
 g 0 0 ··· 0 
 0 g 0 ··· 0 
 
S = −α 
 · · · · · · · ·· ··· ··· 
 (95)
 0 0 ··· g 0 
0 0 ··· 0 g M×M
the consistent set of equations comes as
g11 = G1,1 , h = G1,M , h̄ = GM+1,M+2 ,
g = G1,M+1 , ḡ = GM+1,1 , (96)
i.e., requires an explicit inversion of the 2M by 2M matrix. The algebra is more

complicated since, like in the case of two Hermitian matrices, we have now two types
of gap equations (for g and h and their tilded partners). Using the solutions, we calculate
g11 , then, we get an explicit equation for the spectral density (63). Finally, we perform the
limiting procedure. Technically, one has to repeat the transfer matrix technique introduced
in Section 5 for the complex case. The algebra is lengthy, but due to the cyclic nature of the
blocks of matrix R the results form a rather surprisingly simple set of equations. Returning
to the original variable z = r exp(iφ) we get
1 3 iU sin ψ
G(z, z̄) = log r + − (97)
2τ 4 τ sinh U
and the gap equations read
1 1 iU sin ψ
h= log r − − =G−1 (98)
2τ 4 τ sinh U
and
τ
cos ψ = cosh U − sinh U, (99)
2U
"
where ψ = φ + U sinh sin ψ
U and U = ( log2 r + τ/4)2 − τ 2 g g̃. The above set of equations
completes the solution of the continuous product of matrices with Gaussian Hermitian
disorder. Spectral density follows from (82). Note, that the dependence on log |z| reflects
the multiplicative character of the Brownian motion. Again, the boundary has also an
inversion-type symmetry, but now it is of the form |z| → exp(−τ )/|z|. Due to the non-zero
gap h, the boundary in the Hermitian case is given by the pair of coupled transcendental
equations and can be presented numerically only. The appearance of the aforementioned
symmetry is responsible for the “topological phase transition” in the spectrum of the
product of infinitely many matrices, similar to the one observed for the product of infinitely
many complex matrices.
In Section 8, we study numerically the spectral density predicted by our formulas, in
particular, the “diffusion” of the shape of the boundary caused by evolution with the “time”
parameter τ . We confirm the appearance of a “phase transition” for certain τcrit in the
Hermitian case as well.
8. Numerical tests
In this section we present the results of numerical simulations, confirming our analytical
predictions. We start from the complex case. Fig. 7 shows the evolution of the boundary
for several sample evolution times τ = 0.5, 1, 1.5, 2, 3, 4, 5, 6, 8, 12. To present the whole
set of the boundaries on the same figure, each boundary is rescaled by a corresponding
factor exp(−τ/2). We would like to note here that the effect of rescaling is equivalent to
the non-rescaled process with drift µ = 1. One can see how the original ellipse-like shape
(innermost figure) evolves through a twisted-like shape to the set of double-ring structures.
The inner ring, always containing the origin, is so small on the scale of Fig. 7 that is not
Fig. 7. Evolution of the rescaled (see text) contour of non-holomorphic domain in the eigenvalue spectrum of the
Ginibre–Girko multiplicative diffusion as a function of several times τ .
visible. At τ = 4, corresponding to the curve with an inner loop, we observe a topological

phase transition. The support of the spectrum is no longer simply connected, for τ 4 it is
annulus-like, i.e., eigenvalues are expelled from the central region. For even larger times,
the outside rim of the annulus approaches the circle, and the inner one shrinks to the point
z = 0 in the τ = ∞ limit. Indeed for large τ the radius of the inner ring behaves like
τ
r(τ ) ∼ e− 2 . (100)
The outer boundary then forms approximately a circle with the radius eτ/2 . To visualize the
repulsion of the eigenvalues from the region around z = 0 we performed higher statistics
simulations for τ = 4 and τ = 5, corresponding to Figs. 9, 10, respectively. Note that the
eigenvalue distributions are quite high for small z and the size of the inner ring is indeed
very small. Therefore it is quite difficult to observe numerically the exact exclusion of
eigenvalues from the marked circle.
Fig. 8 shows the comparison of numerically generated spectra versus the analytical
prediction of the shape of the support of eigenvalues of the “evolution operator”. Again,
the same rescaling by exp(−τ/2) was applied.
Fig. 11 shows the evolution of the boundary for several sample evolution times
τ = 0.5, 1, 1.5, 2, 3, 4, 5, 6, 8, 12 in the case of GUE. Again, to present the whole set of
boundaries on the same figure, each boundary was rescaled by a corresponding factor
exp(−τ/2). One can observe how the almost real (for small times τ ) long-cigar shaped
spectrum evolves into a broader shape, developing again singular behavior at τ = 4 at
the leftmost-edge of the spectrum. Again, the spectrum develops a topological transition,
although it is not as clearly visible in the figure as in the complex case. The reason is that
inversion symmetry has an additional exponential suppression factor, i.e., r → e−τ /r. For
Fig. 8. Comparison of the analytical contour for the eigenvalue spectrum of Ginibre–Girko multiplicative
diffusion at evolution time τ = 1, versus the numerical simulation of the spectrum. The generated ensemble
consisted of 60 matrices, each for N = M = 100. Note that the vertical axis is located at x = 1 and not at x = 0.
The origin lies outside the figure.
diffusion at evolution time τ = 4, versus the high-statistics numerical simulation of the spectrum.
very large times, the outside rim of the spectrum approaches the circle, and the inner one
shrinks to a point in the τ → ∞ limit.
As in the complex case, here we also show a sample simulation (Fig. 12) versus the
analytical prediction for the shape of the island.
Finally, in Figs. 5 and 6 above, we demonstrated the “localization–delocalization” phase
transition in the case of the product of two Hermitian matrix models of the considered type.
diffusion at evolution time τ = 5, versus the numerical simulation of the spectrum. The presence of few
eigenvalues inside the inner contour is caused by numerics, i.e., finite N and finite M effects.
Fig. 11. Evolution of the contour of non-holomorphic domain in the eigenvalue spectrum of the GUE
multiplicative diffusion as a function of evolution times τ = 0.5, 1, 1.5, 2, 3, 4, 5, 6, 8, 12, from inside to outside,
respectively.
Below τcrit , a priori complex spectra condense on the positive real axes. At τ = 1/2, the
eigenvalues start to diffuse onto the x 0 half-plane. For very large times τ , eigenvalues
start to appear on both sides of the x = 0 axis.
Fig. 12. Comparison of the analytical contour for the eigenvalue spectrum of GUE multiplicative diffusion at
evolution time τ = 1, versus the numerical simulation of the spectrum. The ensemble consists of 60 matrices,
each obtained as a string of M = 100 of N by N matrices, where N = 100.
9. Stability and Lyapunov exponents
In this section, we briefly discuss the relation between the spectrum of the “evolution
operator” YM and Lyapunov spectra. In the studies of products of random matrices one of
the central issues is the use of limit theorems that would provide an information about the
asymptotic limit of the rates of growth and of the spectrum of the product for a sequence
of matrices. In particular the limit
1
lim ln |YM | ≡ λ1 (101)
M→∞ M
defines the maximum Lyapunov characteristic exponent. The existence of this limit is
assured for an infinite product of matrices by the Furstenberg [13] theorem
%
M
YM = X(i) (102)
i=1
for independent random matrices X(i) characterized by the probability distribution

function dµ[X] = P[X] d[X]. It can be shown under very general assumptions that λ1
is a non-random, self-averaging quantity, so that the brackets · · · in the above expression
can be dropped in almost any realization. The best known example is the use of λ1 in the
description of the DC conductivity of a one-dimensional disordered system coupled to two
electron reservoirs. The conductivity κ of the system, given by the Landauer formula yields
for large M
κ ≈ e−2λ1 M , (103)
where λ1 is the maximal Lyapunov exponent of the product YM . By the Borland conjecture,
the inverse of λ1 gives the localization length measuring the decay of an eigenvector of the
system’s Hamiltonian over the chain composed of M units.
It turns out that the formalism of generalized Green functions (Sections 2 and 3)
is particularly suitable for deriving an analytical formula [34,35] for λ−1 1 for the
one-dimensional disordered wires within the tight-binding Hamiltonian with a site
diagonal disorder. Extensions to generalized Lyapunov exponents L(q) [13] related to the
exponential growth rates of the moments of the matrix product
1
L(q) = lim ln |YM |q (104)
M→∞ M
have been also reported for systems described by Markovian evolutionary operators [36]
used in the study of DC conductance statistics of the random dimer model introduced
for explaining the exceptionally high electronic transport properties of some conjugated
polymer chains.
†
The eigenvectors of the product YM YM form an orthonormal set of vectors (Lyapunov
basis) corresponding to characteristic Lyapunov exponents whose exponentials, exp(λi )
†
are eigenvalues of the matrix |YM YM |1/2M in the limit of large M (Oseledec’s theorem).
The Lyapunov exponents are closely related to the linear stability analysis of the
dynamic system ẋ = F (x) for which small perturbations y around a given solution x(t)
evolve in time according to y(t2 ) = K(t2 , t1 )y(t1 ). They are defined as
1 & &
λi = lim log&K(t2 , t1 )fi (t1 )& (105)
t2 →∞ t2 − t1
and for ergodic systems their spectrum describes global properties of the system’s attractor
(does not depend on the initial point). In contrast to the Lyapunov spectra, the Lyapunov
basis for the ergodic sequence of matrices describes local properties of the PRM system.
Here we would like to underline, that Lyapunov exponents are in general different from
the spectrum of eigenvalues obtained in this paper. Since the first Lyapunov exponent is
defined as a property of the norm of the operator Y , it is related to the spectral properties of
the Y † Y operator. This is why the Lyapunov exponents are always real. The spectra studied
in this paper correspond to so-called stability exponents, which are directly related to the
eigenvalues of the Y operator itself. As we have demonstrated above, the eigenvalues of
YM are in general complex numbers and so are the stability exponents:
1
αi = lim ln λi . (106)
M→∞ M
Their real parts coincide with the Lyapunov exponents λi = Re(αi ) only if the matrix YM
can be written in a diagonal form.
The relation between the stability and Lyapunov exponents in a broad class of dynamical
systems has been investigated by Goldhirsch et al. [37]. Motivated by mostly numerical
results, the authors have conjectured that Lyapunov exponents equal stability ones for
the ergodic dynamical systems whose spectrum is non-degenerate and the attractor is
bounded. The geometrical argument behind that statement is that generically, stretching
and shrinking of the relative separation of “trajectories”, starting at a given initial point
happen along fixed directions in the phase space. Taken from that perspective further
elucidation would be required to clarify the relation between the Lyapunov stability and
spectral properties of the generalized stability matrix investigated in this work.
10. Summary
We have introduced a natural generalization of the concept of geometric random walk

(‘geometric diffusion’) in the space of large, non-commuting matrices. In particular, we
proposed a simple diagrammatic method, allowing to calculate spectral properties of the
resulting evolution operators as functions of the “evolution” time τ . The construction was
explicitly presented in the case of complex Gaussian ensembles and GUE ensembles.
Our method was based on an exact relation between the eigenvalues of a product matrix
and the eigenvalues of a certain block matrix. The essential simplification is the fact that
the random matrices appear linearly in the block matrix. Although we used the above
identification only in the large N limit, it could a priori be used to study the properties of the
spectrum of the product matrix also in the microscopic limit, i.e., on the scale of eigenvalue
spacing. But then different, non-diagrammatic, methods would have to be employed.
Despite the simplicity of the random matrix models chosen as the building blocks of
the matrix-valued diffusion, the spectrum develops a surprising feature, namely a region
without eigenvalues appears within the spectrum, thus changing its topological properties.
We can thus describe this behavior as a topological phase transition. This points at the
appearance of particular repulsion mechanism for sufficiently large evolution times. Our
results were based on the simulation based on Gaussian ensembles. We are however
tempted to conjecture, that unfolding the spectrum in the vicinity of the point where
the topological change occurs may unravel a possible novel universal spectral behavior.
We would like to mention, that the critical behavior described here as a “topological
phase transition” appeared only in the limit M → ∞, as a consequence of the inversion
symmetry, acquired in this limit, of the equation describing the boundary of the two-
dimensional support of eigenvalues. We would like to mention, that similar, qualitative
effect of expulsion of complex eigenvalues from the certain region of complex plane was
observed by Feinberg and Zee (first reference in [28]). They considered however, a single
random complex ensemble with non-Gaussian measure and they imposed the rotational
symmetry of the spectrum. In our case, we considered the infinite product representing
the Ginibre–Girko (therefore, Gaussian) multiplicative diffusions, and we obtained the
structural change of the spectrum for the general case, with non-trivial dependence on the
polar variables. It is an intriguing problem for further studies, to what extent these results
reflect the general “single ring hypothesis”.
We pointed out, in agreement with [37], that stability exponents considered in our
paper carry much richer information about the properties of the system than the Lyapunov
exponents.
As a by product of our analysis, we found a previously unknown (as far as we know)
critical behavior in the case of the particular product of two Hermitian ensembles. The
a priori complex spectra of the product are nevertheless real for evolution times below a
certain critical value τcrit , and migrate into the complex plane only for τ > τcrit . A similar
in spirit localization–delocalization phase transition was recently observed for a class of

non-Hermitian Anderson Hamiltonians [38].
One of the motivations for this work was to propose a general formalism, which
can provide a straightforward method of analyzing spectral properties of multivariate
diffusion-like processes, with the idea that our method could be used in different branches
of theoretical physics and interdisciplinary applications. Some particular applications
will be presented elsewhere. Finally, we would like to mention, that the presented
formalism allows also to establish a direct link to the diffusion processes based on Free
Random Variables techniques [39–41], in particular it can demonstrate the emergence
of the complex Burgers equations governing the evolution of the spectral generating
functions [24].
Acknowledgements
This work was partially supported by the Polish State Committee for Scientific Research
(KBN) grants 2P03B 09622 (2002–2004), 2P03B08225 (2003–2006) and the EU Center
of Excellence in Information Society Technologies “COPIRA”. J.J. and M.A.N. would like
to thank the Niels Bohr Institute, where a part of this work has been completed, for the
hospitality. Their stay at NBI was supported by “MaPhySto”, the Center of Mathematical
Physics and Stochastics, financed by the National Danish Research Foundation.
The authors would like to thank Angelo Vulpiani for interesting discussions. We are
also very grateful to Roland Speicher and Piotr Śniady for comments after submitting this
paper to math-ph and for valuable remarks in relation to Free Random Variable calculus.
References
[1] See, e.g., M.L. Mehta, Random Matrices, Academic Press, New York, 1991;
C.E. Porter, Statistical Theories of Spectra: Fluctuations, Academic Press, New York, 1969.
[2] N. Dupuis, G. Montambaux, Phys. Rev. B 43 (1991) 14390.
[3] T. Guhr, A. Mueller-Groeling, H.A. Weidenmueller, Phys. Rep. 299 (1998) 189, and references therein.
[4] F. Haake, et al., Z. Phys. B 88 (1992) 359;
N. Lehmann, D. Saher, V.V. Sokolov, H.-J. Sommers, Nucl. Phys. A 582 (1995) 223.
[5] J. Ambjørn, J. Jurkiewicz, Yu.M. Makeenko, Phys. Lett. B 251 (1990) 517.
[6] J. Ambjørn, L. Chekhov, C.F. Kristjansen, Yu. Makeenko, Nucl. Phys. B 404 (1993) 127.
[7] See P. Di Francesco, P. Ginsparg, J. Zinn-Justin, Phys. Rep. 254 (1995) 1, and references therein.
[8] M.S. Santhanam, P.K. Patra, Phys. Rev. E 64 (2001) 016102.
[9] A.M. Sengupta, P.P. Mitra, Phys. Rev. E 60 (1999) 3389.
[10] H. Caswell, Matrix Population Models, Sinauer Ass. Inc., Sunderland, MA, 2001.
[11] R. Mantegna, H. Stanley, An Introduction in Econophysics, Cambridge Univ. Press, Cambridge, 2000.
[12] J.P. Bouchaud, M. Potters, Theory of Financial Risks, Cambridge Univ. Press, Cambridge, 2001.
[13] A. Crisanti, G. Paladin, A. Vulpiani, Products of Random Matrices in Statistical Physics, Springer-Verlag,
Berlin, 1993.
[14] C.W.J. Beenakker, in: H. Fukuyama, T. Ando (Eds.), Transport Phenomena in Mesoscopic Systems, in:
Springer Series Solid-State Sciences, Vol. 109, Springer, Berlin, 1992.
[15] X.R. Wang, J. Phys. A 29 (1996) 3053.
[16] F.K. Diakonos, D. Pingel, P. Schmelcher, Phys. Rev. E 62 (2000) 4413.
[17] M.J. de Oliveira, A. Petri, Phys. Rev. E 53 (1996) 2960.

[18] D. Tse, O. Zeitouni, IEEE Trans. Inform. Theory 46 (2000) 172.
[19] M.F. Barnsley, in: H.O. Peitgen, D. Saupe (Eds.), The Science of Fractal Images, Springer-Verlag, Berlin,
1988.
[20] I.E. Telatar, Eur. Trans. Telecommun. 10 (1999) 585.
[21] R.R. Müller, IEEE Trans. Inform. Theory 48 (2002) 2495.
[22] H.M. Taylor, S. Karlin, An Introduction to Stochastic Modeling, 3rd Edition, Academic Press, New York,
1998.
[23] A.D. Jackson, B. Lautrup, P. Johansen, M. Nielsen, Phys. Rev. E 66 (2002) 066124, physics/0202037.
[24] In preparation.
[25] E. Brézin, A. Zee, Phys. Rev. E 49 (1994) 2588;
E. Brézin, A. Zee, Nucl. Phys. B 453 (1995) 531;
E. Brézin, S. Hikami, A. Zee, Universal correlations for deterministic plus random Hamiltonians, hep-
th/9412230.
[27] R.A. Janik, M.A. Nowak, G. Papp, J. Wambach, I. Zahed, Phys. Rev. E 55 (1997) 4100;
R.A. Janik, M.A. Nowak, G. Papp, I. Zahed, Nucl. Phys. B 501 (1997) 603;
A different, but equivalent approach was also developed by [28].
[28] J. Feinberg, A. Zee, Nucl. Phys. B 501 (1997) 643;
J. Feinberg, A. Zee, Nucl. Phys. B 504 (1997) 579;
J.T. Chalker, Z. Jane Wang, Phys. Rev. Lett. 79 (1997) 1797;
Y.V. Fyodorov, H.-J. Sommers, unpublished.
[29] E. Wigner, Canad. Math. Congr. Proc. Univ. Toronto Press p.174 and other papers reprinted;
C.E. Porter (Ed.), Statistical Theories of Spectra: Fluctuations, Academic Press, New York, 1965;
M.L. Mehta, Random Matrices, Academic Press, New York, 1991.
[30] J. Ginibre, J. Math. Phys. 6 (1965) 440.
[31] V.L. Girko, Spectral Theory of Random Matrices, Nauka, Moscow, 1988 (in Russian), and references
therein.
[32] Y.V. Fyodorov, H.-J. Sommers, J. Math. Phys. 38 (1997) 1918;
Y.V. Fyodorov, B.A. Khoruzhenko, H.-J. Sommers, Phys. Lett. A 226 (1997) 46;
H.-J. Sommers, A. Crisanti, H. Sompolinsky, Y. Stein, Phys. Rev. Lett. 60 (1988) 1895.
[33] R.A. Janik, W. Noerenberg, M.A. Nowak, G. Papp, I. Zahed, Phys. Rev. E 60 (1999) 2699.
[34] E. Gudowska-Nowak, G. Papp, J. Brickmann, Chem. Phys. 232 (1998) 247.
[35] E. Gudowska-Nowak, G. Papp, J. Brickmann, J. Phys. Chem. A 102 (1998) 9554.
[36] M.J. de Oliveira, A. Petri, Phys. Rev. E 53 (1996) 2960.
[37] I. Goldhirsch, P.-L. Sulem, S.A. Orszag, Physica D 27 (1987) 311.
[38] N. Hatano, D.R. Nelson, Phys. Rev. Lett. 77 (1996) 570.
[39] D. Voiculescu, Invent. Math. 104 (1991) 201;
D.V. Voiculescu, K.J. Dykema, A. Nica, Free Random Variables, American Mathematical Society,
Providence, RI, 1992.
[40] R. Speicher, Math. Ann. 298 (1994) 611.
[41] Ph. Biane, R. Speicher, Ann. Inst. H. Poincaré PR 37 (5) (2001) 581.
CUMULATIVE AUTHOR INDEX B661–B670
Abel, S.A. B663 (2003) 197 Becker, M. B666 (2003) 144

Acquaviva, V. B667 (2003) 119 Beenakker, W. B667 (2003) 359
Adams, D.H. B662 (2003) 220 Beisert, N. B664 (2003) 131
Aglietti, U. B668 (2003) 3 Beisert, N. B670 (2003) 439
Akbar, M.M. B663 (2003) 215 Belitsky, A.V. B667 (2003) 3
Akemann, G. B664 (2003) 457 Bellucci, S. B663 (2003) 605
Alexandrov, S.Yu. B667 (2003) 90 Bellucci, S. B665 (2003) 402
Alishahiha, M. B661 (2003) 174 Bena, I. B664 (2003) 45
ALPHA Collaboration B663 (2003) 3 Benakli, K. B662 (2003) 40
ALPHA Collaboration B669 (2003) 173 Benson, D. B665 (2003) 367
Álvarez, E. B663 (2003) 365 Berezin, V. B661 (2003) 409
Álvarez-Gaumé, L. B668 (2003) 293 Berg, M. B669 (2003) 3
Antoniadis, I. B662 (2003) 40 Bernabéu, J. B669 (2003) 255
Aranda, A. B670 (2003) 90 Bertola, M. B669 (2003) 435
Armoni, A. B664 (2003) 233 Bhattacharyya, T. B668 (2003) 415
Armoni, A. B667 (2003) 170 Bigi, I.I. B665 (2003) 367
Armoni, A. B670 (2003) 148 Bijnens, J. B669 (2003) 341
Arnaudon, D. B668 (2003) 469 Bilal, A. B663 (2003) 343
Arutyunov, G. B663 (2003) 163 Blažek, T. B662 (2003) 359
Blumenhagen, R. B663 (2003) 319
Arutyunov, G. B665 (2003) 273
Boer, D. B667 (2003) 201
Arutyunov, G. B670 (2003) 103
Bonciani, R. B661 (2003) 289
Astefanesei, D. B665 (2003) 594
Bonciani, R. B668 (2003) 3
Avan, J. B668 (2003) 469
Bonelli, G. B669 (2003) 159
Axenides, M. B662 (2003) 170
Boonekamp, M. B669 (2003) 277
Bourbonnais, C. B663 (2003) 568
Bais, F.A. B666 (2003) 243
Bouttier, J. B663 (2003) 535
Balázs, C. B670 (2003) 90 Brandenburg, A. B667 (2003) 394
Balog, J. B668 (2003) 506 Brax, P. B667 (2003) 149
Banerjee, R. B668 (2003) 179 Brignole, A. B666 (2003) 105
Barbieri, R. B663 (2003) 141 Brower, R.C. B661 (2003) 344
Barbieri, R. B668 (2003) 273 Brower, R.C. B662 (2003) 393
Bardakci, K. B661 (2003) 235 Bruckmann, F. B666 (2003) 197
Bartolo, N. B667 (2003) 119 Buchbinder, I.L. B665 (2003) 402
Baseilhac, P. B669 (2003) 417 Buchmüller, W. B665 (2003) 445
Basu-Mallick, B. B668 (2003) 415
Batchelor, M.T. B669 (2003) 385 Cacciari, M. B664 (2003) 299
Beccaria, M. B663 (2003) 394 Campanario, F. B663 (2003) 280
Becchi, C. B664 (2003) 371 Campanario, F. B670 (2003) 331
Becker, K. B666 (2003) 144 Cao, J. B663 (2003) 487
0550-3213/2003 Published by Elsevier B.V.

doi:10.1016/S0550-3213(03)00808-3
510 Nuclear Physics B 670 (2003) 509–514
Casas, J.A. B666 (2003) 105 Freidel, L. B662 (2003) 279

Chakraborty, B. B668 (2003) 179 Freitas, A. B666 (2003) 305
Chandrasekharan, S. B662 (2003) 220 Frolov, S. B668 (2003) 77
Chapovsky, A.P. B667 (2003) 359 Fursaev, D.V. B664 (2003) 403
Chernodub, M.N. B669 (2003) 233 Fyodorov, Y.V. B664 (2003) 457
Chetyrkin, K.G. B666 (2003) 289
Chitov, G.Y. B663 (2003) 568 Ganjali, M.A. B661 (2003) 174
Choi, K.-S. B662 (2003) 476 Gannon, T. B670 (2003) 335
Conde, J. B663 (2003) 365 García Canal, C.A. B662 (2003) 334
Crampé, N. B668 (2003) 469 Gardi, E. B664 (2003) 299
Cvetič, M. B662 (2003) 89 Geyer, B. B662 (2003) 531
Ghilencea, D.M. B670 (2003) 183
Daleo, A. B662 (2003) 334 Ghodsi, A. B661 (2003) 174
Darriulat, P. B661 (2003) 3 Ghosh, S. B670 (2003) 359
Dasgupta, K. B666 (2003) 144 Giannakis, I. B669 (2003) 462
de Azcárraga, J.A. B662 (2003) 185 Giedt, J. B668 (2003) 138
de Boer, J. B665 (2003) 545 Giudice, G.F. B663 (2003) 377
de Haro, S. B664 (2003) 45 Giusto, S. B664 (2003) 371
D’Elia, M. B661 (2003) 139 Goldstein, K. B669 (2003) 325
Demasure, Y. B661 (2003) 153 Golec-Biernat, K. B668 (2003) 345
Denner, A. B662 (2003) 299 Gomis, J. B665 (2003) 49
Derkachov, S.É. B661 (2003) 533 González, J. B663 (2003) 605
Deser, S. B662 (2003) 379 Gorsky, A.S. B667 (2003) 3
Díaz-Cruz, J.L. B670 (2003) 90 Gracey, J.A. B662 (2003) 247
Di Bari, P. B665 (2003) 445 Gracey, J.A. B667 (2003) 242
Di Francesco, P. B663 (2003) 535 Graham, N. B665 (2003) 623
Dinh, P.N. B661 (2003) 3 Groot Nibbelink, S. B663 (2003) 60
Dobashi, S. B665 (2003) 94 Groot Nibbelink, S. B665 (2003) 236
Doikou, A. B668 (2003) 447 Grozin, A.G. B663 (2003) 280
Doikou, A. B668 (2003) 469 Grozin, A.G. B666 (2003) 289
Dolan, F.A. B665 (2003) 273 Grozin, A.G. B670 (2003) 331
Dorey, P. B661 (2003) 425 Guadagnoli, D. B670 (2003) 264
Dorey, P. B661 (2003) 464 Guan, X.-W. B669 (2003) 385
Dorogovtsev, S.N. B666 (2003) 396 Gudowska-Nowak, E. B670 (2003) 479
Dotsenko, V.S. B664 (2003) 477 Guitter, E. B663 (2003) 535
Dubovsky, S.L. B664 (2003) 407 Gustavsson, A. B667 (2003) 111
Dung, N.T. B661 (2003) 3
Dunne, G.V. B670 (2003) 307 Haack, M. B669 (2003) 3
Haba, N. B669 (2003) 381
Emmanuel-Costa, D. B661 (2003) 62 Hagiwara, K. B668 (2003) 364
Engquist, J. B664 (2003) 439 Hall, L.J. B663 (2003) 141
Espinosa, J.R. B666 (2003) 105 Harada, M. B669 (2003) 381
Etesi, G. B662 (2003) 511 Harmark, T. B662 (2003) 3
Evlampiev, K. B662 (2003) 120 Hebecker, A. B670 (2003) 3
Heitger, J. B669 (2003) 173
Faisst, M. B665 (2003) 649 Herdeiro, C.A.R. B665 (2003) 189
Falkowski, A. B667 (2003) 149 Hernández, L. B663 (2003) 365
Farhi, E. B665 (2003) 623 Hieu, B.D. B661 (2003) 3
Feng, B. B661 (2003) 113 Hikida, Y. B669 (2003) 57
Feverati, G. B663 (2003) 409 Hiller, J.R. B661 (2003) 99
Floratos, E. B662 (2003) 170 Hofmann, R. B668 (2003) 151
Foerster, A. B669 (2003) 385 Holland, K. B668 (2003) 207
Forger, M. B667 (2003) 435 Hollik, W. B666 (2003) 305
Frappat, L. B668 (2003) 469 Honecker, G. B666 (2003) 175
Nuclear Physics B 670 (2003) 509–514 511
Hosotani, Y. B669 (2003) 381 Kumar, K. B668 (2003) 179

Hou, B.Y. B663 (2003) 467 Kuroki, T. B664 (2003) 185
Huber, P. B665 (2003) 487 Kurth, M. B669 (2003) 173
Huber, S.J. B666 (2003) 269 Kurylov, A. B667 (2003) 321
Hwang, K. B662 (2003) 476 Kutasov, D. B666 (2003) 56
Ichinose, I. B663 (2003) 520 Lalak, Z. B667 (2003) 149

Imai, T. B665 (2003) 520 Laliena, V. B668 (2003) 403
Imbimbo, C. B664 (2003) 371 Larosa, M. B667 (2003) 261
Intriligator, K. B667 (2003) 183 Laugier, A. B662 (2003) 40
Isaev, A.P. B662 (2003) 461 Lee, H.K. B665 (2003) 153
Ito, M. B668 (2003) 322 Lee, K. B665 (2003) 179
Ivanov, N.Ya. B666 (2003) 88 Lehners, J.-L. B661 (2003) 273
Izquierdo, J.M. B662 (2003) 185 Leonhardt, T. B667 (2003) 413
Li, G.-L. B670 (2003) 401
Jack, I. B662 (2003) 63 Li, Y.Q. B666 (2003) 337
Jacobsen, J.L. B664 (2003) 477 Lin, H.-Q. B663 (2003) 487
Jaffe, R.L. B665 (2003) 623 Lin, H.Q. B666 (2003) 337
Janik, R.A. B661 (2003) 153 Lindner, M. B665 (2003) 487
Janik, R.A. B670 (2003) 479 Lindström, U. B662 (2003) 147
Janssen, B. B669 (2003) 363 Lipatov, L.N. B661 (2003) 19
Jones, D.R.T. B662 (2003) 63 Livine, E.R. B663 (2003) 231
Jugeau, F. B670 (2003) 221 Louapre, D. B662 (2003) 279
Jung, E. B669 (2003) 306 Lowe, D.A. B667 (2003) 55
Jurkiewicz, J. B670 (2003) 479 Lowe, D.A. B669 (2003) 325
Lozano, Y. B669 (2003) 363
Kamimura, K. B662 (2003) 491 Lü, H. B662 (2003) 89
Kaminsky, K. B663 (2003) 33 Lü, H. B668 (2003) 237
Kanai, T. B670 (2003) 289 Ludwig, A.W.W. B661 (2003) 577
Kanaki, A. B667 (2003) 359 Lüst, D. B663 (2003) 319
Kawai, H. B664 (2003) 185 Lynker, M. B667 (2003) 484
Kawamura, Y. B669 (2003) 381
Kazakov, V.A. B667 (2003) 90 Maillard, T. B662 (2003) 40
Kehagias, A. B662 (2003) 170 Majumdar, P. B664 (2003) 213
Khater, W. B661 (2003) 209 Manashov, A.N. B661 (2003) 533
Khemani, V. B665 (2003) 623 Mannel, T. B663 (2003) 280
Kim, J.E. B662 (2003) 476 Mannel, T. B670 (2003) 331
Kim, S.-H. B669 (2003) 306 Mannel, Th. B665 (2003) 367
Kimura, Y. B664 (2003) 512 Manvelyan, R. B667 (2003) 413
King, S.F. B662 (2003) 359 Marandella, G. B663 (2003) 141
Kitazawa, Y. B665 (2003) 520 Marandella, G. B668 (2003) 273
Klebanov, I.R. B664 (2003) 3 Marchi, M. B665 (2003) 425
Kleinert, H. B666 (2003) 361 Martin, L.C. B668 (2003) 335
Knechtli, F. B663 (2003) 3 Martucci, L. B666 (2003) 230
Koizumi, K. B669 (2003) 417 Masina, I. B661 (2003) 365
Korchemsky, G.P. B661 (2003) 533 Mastrolia, P. B661 (2003) 289
Korchemsky, G.P. B667 (2003) 3 Mastrolia, P. B664 (2003) 341
Körs, B. B669 (2003) 3 Matarrese, S. B667 (2003) 119
Kostov, I.K. B667 (2003) 90 Mathur, S.D. B661 (2003) 344
Kotikov, A.V. B661 (2003) 19 Mawatari, K. B668 (2003) 364
Kraus, E. B661 (2003) 83 Meggiolaro, E. B665 (2003) 425
Kristjansen, C. B664 (2003) 131 Melles, M. B662 (2003) 299
Krykhtin, V.A. B665 (2003) 402 Melnikov, K. B662 (2003) 409
Kühn, J.H. B665 (2003) 649 Mendes, J.F.F. B666 (2003) 396
Metzger, S. B663 (2003) 343 Periwal, V. B667 (2003) 484

Meyer, H.B. B668 (2003) 111 Peschanski, R. B669 (2003) 277
Minkowski, P. B668 (2003) 207 Petcov, S.T. B669 (2003) 255
Moore, G. B670 (2003) 27 Phuong, P.T. B661 (2003) 3
Moore, J.E. B661 (2003) 514 Picón, M. B662 (2003) 185
Morita, T. B664 (2003) 185 Pijlman, F. B667 (2003) 201
Moriyama, S. B665 (2003) 49 Pinsky, S.S. B661 (2003) 99
Morozov, A. B666 (2003) 311 Pittau, R. B667 (2003) 359
Mulders, P.J. B667 (2003) 201 Plümacher, M. B665 (2003) 445
Mülsch, D. B662 (2003) 531 Pocklington, A. B661 (2003) 425
Mussardo, G. B670 (2003) 464 Pocklington, A. B661 (2003) 464
Pons, J.M. B665 (2003) 129
Nadolsky, P.M. B666 (2003) 3 Pope, C.N. B662 (2003) 89
Nadolsky, P.M. B666 (2003) 31 Pope, C.N. B668 (2003) 237
Naón, C.M. B663 (2003) 591 Pozzorini, S. B662 (2003) 299
Nason, P. B667 (2003) 394 Pradisi, G. B667 (2003) 261
Nastase, H. B667 (2003) 55 Prokushkin, S. B666 (2003) 144
Navarro, I. B666 (2003) 105
Niarchos, V. B666 (2003) 56 Radu, E. B665 (2003) 594
Nicolai, H. B668 (2003) 167 Ragoucy, E. B668 (2003) 469
Niemi, A.J. B666 (2003) 311 Ramgoolam, S. B667 (2003) 55
Nikolov, N.M. B670 (2003) 373 Ramsey-Musolf, M.J. B667 (2003) 321
Nilles, H.P. B665 (2003) 236 Ratz, M. B670 (2003) 3
Nishigaki, S.M. B670 (2003) 307 Ravindran, V. B665 (2003) 325
Nógrádi, D. B666 (2003) 197 Remiddi, E. B661 (2003) 289
Nogueira, F.S. B666 (2003) 361 Remiddi, E. B664 (2003) 341
Nomura, Y. B663 (2003) 141 Ren, H.-C. B669 (2003) 462
Nowak, M.A. B670 (2003) 479 Renard, F.M. B663 (2003) 394
Nyawelo, T.S. B663 (2003) 60 Rey, S.-J. B669 (2003) 57
Riccioni, F. B663 (2003) 60
Okawa, Y. B663 (2003) 33 Riotto, A. B667 (2003) 119
Okui, T. B663 (2003) 141 Riva, V. B670 (2003) 464
Oleari, C. B667 (2003) 394 Roček, M. B662 (2003) 147
Olechowski, M. B665 (2003) 236 Rodríguez-Gómez, D. B669 (2003) 363
Oliver, S.J. B663 (2003) 141 Roiban, R. B664 (2003) 45
Onorato, P. B663 (2003) 605 Roiban, R. B665 (2003) 211
Ooguri, H. B663 (2003) 33 Royon, C. B669 (2003) 277
Oriti, D. B663 (2003) 231 Rühl, W. B667 (2003) 413
Osborn, H. B665 (2003) 273 Rupp, C. B661 (2003) 83
Oshimo, N. B668 (2003) 258 Russo, R. B669 (2003) 207
Osland, P. B661 (2003) 209
Owen, A.W. B663 (2003) 197 Sakaguchi, M. B662 (2003) 491
Saleur, H. B663 (2003) 443
Paccetti Correia, F. B668 (2003) 151 Salvay, M.J. B663 (2003) 591
Palomares-Ruiz, S. B669 (2003) 255 Samtleben, H. B668 (2003) 167
Papadopoulos, C.G. B667 (2003) 359 Samukhin, A.N. B666 (2003) 396
Papucci, M. B663 (2003) 141 Santachiara, R. B664 (2003) 477
Papucci, M. B668 (2003) 273 Santambrogio, A. B670 (2003) 103
Park, D.K. B669 (2003) 306 Sasaki, R. B663 (2003) 467
Park, J. B665 (2003) 49 Sassot, R. B662 (2003) 334
Parvizi, S. B661 (2003) 174 Saulina, N. B670 (2003) 27
Pearce, P.A. B663 (2003) 409 Savoy, C.A. B661 (2003) 365
Penati, S. B670 (2003) 103 Sazdjian, H. B670 (2003) 221
Pepe, M. B668 (2003) 207 Schimmrigk, R. B667 (2003) 484
Nuclear Physics B 670 (2003) 509–514 513
Schmidt, M.G. B668 (2003) 151 Takahashi, T. B670 (2003) 161

Schubert, C. B668 (2003) 335 Takayama, Y. B665 (2003) 520
Schwetz, T. B665 (2003) 487 Takayanagi, T. B662 (2003) 3
Sciuto, S. B669 (2003) 207 Takeda, K. B663 (2003) 520
Scrucca, C.A. B669 (2003) 128 Talavera, P. B665 (2003) 129
Segal, A.Y. B664 (2003) 59 Talavera, P. B669 (2003) 341
Seidensticker, T. B665 (2003) 649 Tamai, K. B668 (2003) 385
Seki, S. B661 (2003) 257 Tan, C.-I. B661 (2003) 344
Serbo, V.G. B662 (2003) 409 Tan, C.-I. B662 (2003) 393
Serone, M. B669 (2003) 128 Tanaka, T. B662 (2003) 413
Sezgin, E. B664 (2003) 439 Tatar, R. B665 (2003) 211
Sezgin, E. B668 (2003) 237 Tateo, R. B661 (2003) 425
Sfetsos, K. B669 (2003) 103 Tateo, R. B661 (2003) 464
Shafi, Q. B665 (2003) 469 Tavartkiladze, Z. B665 (2003) 469
Sharpe, E. B664 (2003) 21 Tavartkiladze, Z. B668 (2003) 151
Shi, K.-J. B663 (2003) 487 Taylor, M. B665 (2003) 3
Shi, K.-J. B670 (2003) 401 Taylor, T.R. B663 (2003) 319
Shifman, M. B664 (2003) 233 Teper, M.J. B668 (2003) 111
Shifman, M. B667 (2003) 170 Thao, N.T. B661 (2003) 3
Shifman, M. B670 (2003) 148 Thieu, D.Q. B661 (2003) 3
Shik, H.Y. B666 (2003) 337 Thorn, C.B. B661 (2003) 235
Shimada, H. B665 (2003) 94 Thuan, V.V. B661 (2003) 3
Shin, H. B669 (2003) 78 Tobe, K. B663 (2003) 123
Todorov, I.T. B670 (2003) 373
Sibiryakov, S.M. B664 (2003) 407
Tomino, D. B665 (2003) 520
Sibold, K. B661 (2003) 83
Tonel, A.P. B669 (2003) 385
Siegel, W. B665 (2003) 179
Trittmann, U. B661 (2003) 99
Silva, P.J. B666 (2003) 230
Tseytlin, A.A. B664 (2003) 247
Silvestrini, L. B669 (2003) 128
Tseytlin, A.A. B668 (2003) 77
Simula, S. B670 (2003) 264
Tsujikawa, S. B670 (2003) 289
Sin, S.-J. B667 (2003) 310
Tsutsui, I. B662 (2003) 447
Singh, H. B661 (2003) 394
Skenderis, K. B665 (2003) 3
Uchino, T. B662 (2003) 447
Smith, J. B665 (2003) 325
Uraltsev, N. B665 (2003) 367
Sokatchev, E. B663 (2003) 163
Sokatchev, E. B665 (2003) 273 van Baal, P. B666 (2003) 197
Sokatchev, E. B670 (2003) 103 van Holten, J.W. B663 (2003) 60
Solodukhin, S.N. B665 (2003) 545 van Neerven, W.L. B665 (2003) 325
Sommer, R. B669 (2003) 173 van Nieuwenhuizen, P. B662 (2003) 147
Sotkov, G. B670 (2003) 464 Varela, O. B662 (2003) 185
Stanev, Ya.S. B670 (2003) 373 Vázquez-Mozo, M.A. B668 (2003) 293
Staśto, A.M. B668 (2003) 345 Veneziano, G. B667 (2003) 170
Staudacher, M. B664 (2003) 131 Veretin, O. B665 (2003) 649
Staudacher, M. B670 (2003) 439 Verzegnassi, C. B663 (2003) 394
Stefański Jr., B. B666 (2003) 71 Villanueva Sandoval, V.M. B668 (2003) 335
Stelle, K.S. B661 (2003) 273
Stelle, K.S. B662 (2003) 89 Walcher, J. B665 (2003) 211
Striet, J. B666 (2003) 243 Waldron, A. B662 (2003) 379
Strumia, A. B663 (2003) 377 Walter, M.G.A. B665 (2003) 236
Su, S. B667 (2003) 321 Walter, W. B666 (2003) 305
Sudbø, A. B666 (2003) 361 Wang, W. B667 (2003) 349
Sugawara, Y. B661 (2003) 191 Wang, Y. B663 (2003) 487
Sugiyama, K. B669 (2003) 78 Wecht, B. B667 (2003) 183
Sundell, P. B664 (2003) 439 Wehefritz-Kaufmann, B. B663 (2003) 443
Weigel, H. B665 (2003) 623 Yoneya, T. B665 (2003) 94

Weiglein, G. B666 (2003) 305 Yoshida, K. B669 (2003) 78
Weisz, P. B668 (2003) 506 Yuan, C.-P. B666 (2003) 3
Wells, J.D. B663 (2003) 123 Yuan, C.-P. B666 (2003) 31
Wiese, K.J. B661 (2003) 577 Yue, C. B667 (2003) 349
Wiese, U.-J. B668 (2003) 207 Yue, R.-H. B670 (2003) 401
Wiesenfeldt, S. B661 (2003) 62 Yung, A. B662 (2003) 120
Winter, W. B665 (2003) 487
Winterhalder, A. B667 (2003) 435 Zakharov, V.I. B669 (2003) 233
Witten, E. B664 (2003) 3 Zeuthen–Rome (ZeRo) Col-
Wolff, U. B663 (2003) 3 laboration B664 (2003) 276
Wu, J.-B. B663 (2003) 79 Zheng, Z.-J. B663 (2003) 79
Wu, J.-B. B663 (2003) 95 Zheng, Z.-J. B663 (2003) 95
Wu, X. B665 (2003) 153 Zhou, H.-Q. B669 (2003) 385
Zhu, C.-J. B663 (2003) 79
Yang, W.-L. B663 (2003) 467 Zhu, C.-J. B663 (2003) 95
Yılmaz, N.T. B664 (2003) 357 Znojil, M. B662 (2003) 554
Yokoya, H. B668 (2003) 364 Zong, H. B667 (2003) 349

Nucl - Phys.B v.670

Enviado por

Dados do documento

Descrição original:

Título original

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Nucl - Phys.B v.670

Enviado por

Direitos autorais:

Formatos disponíveis

Nuclear Physics B 670 (2003) 3–26

Group-theoretical aspects of orbifold

E-mail address: mratz@mail.desy.de (M. Ratz).

2. Basics of group theory

tr(H i H j ) = λδij , (2.1)

[E α , E β ] = Nα,β E α+β , (2.3)

α−β >0 ⇔ first non-vanishing component of α − β > 0. (2.4)

Name Real algebra Extended Dynkin diagram

with integer coefficients k i . Motivated by this, a basis

µ(i) · e(j ) = δji . (2.7)

3. Obtaining all regular subgroups by orbifolding

Orbifold GUTs [10–15] are based on a gauge theory on R4 × M, where M is a manifold

the requirement of K invariance, a gauge theory on a manifold with singularities, M/K,

We can also choose to write P = exp(−2πiξ T ), where T is a normalized Lie algebra

3.1. Orbifold-breaking to any maximal regular subgroup

P E α(i) P −1 = e2πi/n E α(i) ,

ad G → ad H ⊕ 1(0) ⊕ R 1 (q1 ) ⊕ R 2 (q2 ) ⊕ · · · . (3.7)

distributed among {0, 2π/n, 4π/n, . . ., (n − 1)2π/n} where an excess at 0 is possible if

3.2. Some examples from the series SO(10) ⊂ E6 ⊂ E7 ⊂ E8

45 → (1, 1)(6,4) ⊕ (3, 1)(−4,4) ⊕ (3, 2)(1,4) ⊕ (3, 2)(−5,0)

4. Rank reduction and non-Abelian twists

Pj −1 hPj = cos(4πξ )h − i sin(4πξ )(E α(j) − E −α(j) ), (4.3)

Pj −1 E α Pj = cos(2πNα,α(j) ξ )E α + i sin(2πNα,α(j) ξ )E α+α(j) , (4.4)

ad G → ad H ⊕ 1(0) ⊕ R(q) ⊕ R(−q) ⊕ · · · , (4.5)

5.1. Geometry and gauge symmetry breaking

In 5 dimensions, the geometry is very constraining. Up to isomorphism, the only smooth

Fig. 3. Construction of a compact manifold with singularities from two triangles.

We now turn to the possibilities of geometric gauge symmetry breaking on conifolds.

5.2. Generating chiral matter

In general, compactification on a non-flat manifold can provide chiral matter if the

5 For more details see, e.g., [42] as well as [4,43].

6.1. E7 → SU(5) × SU(3)F × U(1)

Consider a SYM theory based on an E7 gauge group. E7 contains SU(5) × SU(3)F ×

6.2. E8 → GPS × U(1)3

Under E8 → SO(10) × SU(4), the adjoint representation of E8 decomposes like

(15, 1, 1; 10 ): 0 (1, 1, 1; 3−4/3 ): 2/3 (4, 1, 2; 3−1 ): 11/12

6.3. E8 → GPS × SU(3)F × U(1)

We have explored some of the group-theoretical possibilities in orbifold GUTs. In

We would like to thank Fabian Bachmaier, Wilfried Buchmüller, John March-Russell,

[1] H. Georgi, S. Glashow, Phys. Rev. Lett. 32 (1974) 438.

[10] Y. Kawamura, Prog. Theor. Phys. 105 (2001) 999, hep-ph/0012125.

T-duality, and the K-theoretic partition function

1. Introduction and summary

Duality symmetries, such as the U-duality symmetry of toroidally compactified M-

E-mail address: saulina@princeton.edu (N. Saulina).

1.1. The effective eight-dimensional supergravity, and its partition function

gmn = M(τ ), g mn = M(τ )−1 , Gαβ = M(ρ) (1.6)

1.2. The semiclassical expansion

4 By “flat” we mean the DeRham representative of the relevant fieldstrength is zero.

(a) the K-theoretic sum over RR fluxes of g(4) in the presence of F ;

1.3. The K-theoretic RR partition function

1.4. T-duality transformations

Θ(T · F , ρ + 1) = µ(T )Θ(F , ρ),

and hence µ is nontrivial if χ is not divisible by 12. In particular, a homogeneous

should be regarded as a gauge symmetry of M-theory. Accordingly, we turn to the

1.5. Inclusion of 1-loop effects

ZB+F (F , ρ) := DetB DetF e−SB (F ) Θ(F

ZB+F (T · F , ρ + 1) = µ(T )ZB+F (F , ρ),

where η(ρ) is the Dedekind function. The final combination

Z(F , ρ) = e−Squant ZB+F (F , ρ) (1.32)

1.7. U-duality and M-theory

2. Review of T-duality invariance in the standard formulation of type IIA

Pj −1 hPj = cos(4πξ )h − i sin(4πξ )(E α(j) − E −α(j) ), (4.3)

Pj −1 E α Pj = cos(2πNα,α(j) ξ )E α + i sin(2πNα,α(j) ξ )E α+α(j) , (4.4)

ZB+F (F , ρ) := DetB DetF e−SB (F ) Θ(F

Θ(T · F , ρ) = Θ(F , ρ),