Nucl - Phys.B v.688

Nuclear Physics B 688 (2004) 3–69
www.elsevier.com/locate/npe
Two-loop superstrings
on orbifold compactifications ✩
Kenichiro Aoki a , Eric D’Hoker b , D.H. Phong c
a Hiyoshi Department of Physics, Keio University, Yokohama 223-8521, Japan
b Department of Physics and Astronomy, University of California, Los Angeles, CA 90095, USA
c Department of Mathematics, Columbia University, New York, NY 10027, USA
Received 29 March 2004; accepted 1 April 2004
Abstract
The two-loop chiral measure for superstring theories compactified on Z2 reflection orbifolds
is constructed from first principles for even spin structures. This is achieved by a careful
implementation of the chiral splitting procedure in the twisted sectors and the identification of
a subtle worldsheet supersymmetric and supermoduli dependent shift in the Prym period. The
construction is generalized to compactifications which involve more general NS backgrounds
preserving worldsheet supersymmetry. The measures are unambiguous and independent of the gauge
slice.
Two applications are presented, both to superstring compactifications where 4 dimensions are Z2 -
twisted and where the GSO projection involves a chiral summation over spin structures. The first
is an orbifold by a single Z2 -twist; here, orbifolding reproduces a supersymmetric theory and it is
shown that its cosmological constant indeed vanishes. The second model is of the type proposed by
Kachru–Kumar–Silverstein and additionally imposes a Z2 -twist by the parity of worldsheet fermion
number; it is shown here that the corresponding cosmological constant does not vanish pointwise on
moduli space.
 2004 Elsevier B.V. All rights reserved.
PACS: 11.25.Db; 11.25.Hf
✩
Research supported in part by a Grant-in-Aid from the Ministry of Education, Science, Sports and Culture
and grants from Keio University (K.A.), and by National Science Foundation grants PHY-01-40151 (E.D.) and
DMS-02-45371 (D.P.).
E-mail addresses: ken@phys-h.keio.ac.jp (K. Aoki), dhoker@physics.ucla.edu (E. D’Hoker),
phong@cpw.math.columbia.edu (D.H. Phong).
0550-3213/$ – see front matter  2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.nuclphysb.2004.04.001
4 K. Aoki et al. / Nuclear Physics B 688 (2004) 3–69
1. Introduction
Over the past few years, important progress has been made on two-loop superstring
perturbation theory. A formula for the two-loop even spin structure superstring measure
has been constructed which is well-defined and independent of any choice of gauge
slice [1–5]. The independence of gauge slice choices is essential, since without it the
superstring scattering amplitudes would be ambiguous. The formula has been applied to
the evaluation of the two-loop N -point function for N 4 massless external NS–NS string
states, propagating in flat Minkowski space–time. In particular, the cosmological constant
for flat Minkowski space–time had been shown to vanish point by point on moduli space.
The original derivation of these results in [4] was carried out using a gauge where the
worldsheet gravitini are supported at two points q1 , q2 which obey the gauge condition,
Sδ (q1 , q2 ) = 0, where Sδ is the Szegö kernel (or fermion propagator) for even spin
structure δ. Recently, the results of [3] (which were expressed in a general gauge where
the measure is formulated in terms of meromorphic functions and forms) were used to
evaluate the superstring measure in a hyper-elliptic gauge by Zheng, Wu and Zhu in
[6,7], and the vanishing of the N 3-point functions was confirmed. The cancellations
of these amplitudes are examples of non-renormalization conjectures suggested by space–
time supersymmetry [8]. A bibliography of earlier work on this topic may be found in [4].
The purpose of the present paper is to derive the superstring measure (and the
cosmological constant) for orbifold compactification backgrounds produced by general
Z2 reflections. Certain orbifold models of this type provide some of the simplest
solvable examples of superstring theories with broken supersymmetry and are therefore of
considerable interest in particle physics. In particular, Kachru, Kumar and Silverstein [9]
have proposed Z2 × Z2 orbifolds of type II superstring theory with broken supersymmetry
(see also [10]), whose 1-loop cosmological constant vanishes by construction. Clearly, the
construction of any superstring theory model with broken supersymmetry and vanishing
(or very small) cosmological constant would be of great phenomenological value (see also
[11]).
In [9], arguments were presented in favor of the vanishing of the cosmological constant
to two-loop order (and higher), using the earlier string measure suggested by the picture
changing operator and BRST formalism [12,13]. Shortly thereafter, arguments were given
against the vanishing of the two-loop cosmological constant in the same model [14], also
using the picture changing operator and BRST formalism. In both approaches, special
choices for the worldsheet gravitini insertion points were made, the first in unitary gauge,
the second following [15]. By now, however, the picture changing operator and BRST
formalism is well known to be gauge slice dependent [13] (see also [16–19]) a fact that
casts doubts over the conclusions of both [9,14].
A reliable calculation of the cosmological constant requires a gauge slice independent
measure such as the one obtained in [1–4] for flat Minkowski space–time.
In the present paper, the two-loop chiral measure for superstring theories compactified
on Z2 reflection orbifolds is constructed from first principles for even spin structures. This
is achieved by a careful implementation of the chiral splitting procedure in the twisted
sectors and the identification of a subtle worldsheet supersymmetric and supermoduli
K. Aoki et al. / Nuclear Physics B 688 (2004) 3–69 5
dependent shift in the Prym period. The limits of the measure are checked under various
degenerations and are found to agree with the behavior expected on physical grounds.
The construction is generalized to compactifications which involve more general NS
backgrounds, and preserve worldsheet supersymmetry. The measures are shown to be
unambiguous and independent of the gauge slice. Our original primary concern was with
the study of the models of [9]. The results obtained in this paper for the Z2 -twisted chiral
measure, however, will be indispensable ingredients in the construction of the Z2 -twisted
conformal field theory and string amplitudes to two-loop order in any orbifold model based
on Z2 reflections.
Two applications are presented, both to superstring compactifications where 4 dimen-
sions are Z2 -twisted and where the GSO projection involves a chiral summation over spin
structures, i.e., a summation carried out independently on left and right movers. The first
is an orbifold by a single Z2 twist, which is known to reproduce a supersymmetric the-
ory; it is shown here that its cosmological constant vanishes. The second model is of the
type proposed by Kachru, Kumar and Silverstein [9] and additionally imposes a Z2 twist
by the parity of worldsheet fermion number; it is shown here that the corresponding chiral
measure is non-vanishing. Therefore, the associated cosmological constant does not vanish
pointwise on moduli space. It is logically possible, though we shall argue it is unlikely, that
despite this result, a cancellation will occur for the cosmological constant when left and
right chiral blocks are assembled and the integration over moduli is carried out.
1.1. Outline of the construction
In [1–4], a detailed method was laid out for determining the two-loop superstring
measure on general space–times. Explicit formulas were obtained in those papers for the
case of flat Minkowski space–time M; they involve the matter and ghost supercurrents,
stress tensors, and partition functions.
All compactifications discussed in this paper will only affect the matter fields and will
leave the ghost fields unchanged. In orbifold compactifications C of flat Minkowski space–
time, the expressions for the matter supercurrent and stress tensor (in terms of the fields x µ
µ
and ψ± ) are identical to those of flat Minkowski space–time and will be denoted by SMm
µ
and TMm (or simply Sm and Tm when no confusion is possible). The fields x µ and ψ± will
be summed over untwisted and twisted sectors, dependent on C. Therefore, the partition
function, and more generally any vacuum expectation value will depend on C; they will
be referred to by ZC and · · ·C . (On a more general compactification background C, the
µ
expression for the supercurrent and stress tensor (in terms of x µ and ψ± ) will generally be
different from their expressions in flat space–time and will be denoted by SCm and TCm .)
Thus, the main new ingredients are the construction of the propagators for the Z2 -
µ
twisted fields x µ and ψ± , and the identification of the corresponding changes in the string
measure. These changes reflect the global geometry of the worldsheet with a quadratic
branch cut, and must be treated with care. We summarize here the main steps in our
derivation.
• The first step is to determine the Z2 -twisted bosonic propagator ∂z x(z)∂w x(w)ε for
given twist ε. Its explicit expression will be given in terms of the familiar Szegö kernel,
as well as the Prym period τε and the Prym differential ωε for twist ε. The propagator
for the Z2 -twisted fermionic field ψ+ was identified in [9].
• The next step is to carry out chiral splitting for the Z2 -twisted matter superfields. Since
the Z2 -twisted propagators are different from those of flat space–time, the effective
rules, derived in [20,21] for Minkowski space–time, will not apply directly in this
case. The Z2 -twisted case has to be worked out in its own right, and its own global
geometric features carefully accounted for. In fact, it is here that a subtle and crucial
supersymmetric correction ∆τε to the super Prym period makes its appearance.
• An important check is the independence of the gauge-fixed measure from the choice of
gauge slice. In fact, a formula will be derived for the two-loop measure in terms of the
partition function, supercurrent and stress tensor correlators for any compactification
that does not affect the ghosts, preserves worldsheet supersymmetry and has matter
central charge c = 15. The resulting measure will be shown to be unambiguous and
slice independent under these assumptions. This result was announced in [1].
• The next task is to express the Z2 -twisted superstring measure in terms of ϑ-constants.
The procedure follows [1–4], but here there is a significant new difficulty. A new,
internal momentum pε -dependent, contribution arises from the correction ∆τε to
the super Prym period. The explicit evaluation, in terms of ϑ-constants, of this
contribution is one of the more arduous steps in the paper. The resulting chiral
superstring measure depends on both the spin structure δ and the twist ε and is denoted
dµC [δ; ε](pε ).
• The limit of the chiral measure dµC [δ; ε](pε ) is computed for a variety of degener-
ations and checked versus the limiting behavior expected on physical grounds. The
limits will also be needed to investigate the behavior of the cosmological constant.
• Next, the behavior of the chiral measure dµC [δ; ε](pε ) under modular transformations
is derived. The full genus 2 modular group Sp(4, Z) will act on the twist cycle
and relate the measures for various twists. In carrying out the GSO projection, a
summation over all spin structures is to be carried out, for fixed twist. Therefore, it
is the subgroup Hε of all elements in Sp(4, Z) which leave ε invariant which will
be of greatest importance in determining all possible GSO projections consistent with
modular covariance of the chiral measure. The group Hε is studied in detail; it is found
to be generated by 5 independent elements (while Sp(4, Z) is generated by 6), the
various orbits under Hε for twists and spin structures are calculated, and the behavior
of dµC [δ; ε](pε ) is obtained.
• As a general principle, the GSO projection is carried out by summing over the
spin structures δ with suitable phases. In practice, for each specific physical model,
the conformal blocks entering the chiral string measure have to be identified and
summed over δ consistently with modular covariance. For the simplest Z2 -twisted
theories, where the orbifold group is exactly Z2 acting by reflection, the conformal
blocks are just dµC [δ; ε](pε ). However, for more complicated orbifold groups such
as Z2 × Z2 in the models of [9], the conformal blocks can be more subtle, and have
to be determined with some care. The cosmological constant for various asymmetric
Z2 -orbifold models is then determined by summing over the spin structures δ. In
particular, the cosmological constant in the models of Kachru, Kumar and Silverstein
[9] is shown to be non-vanishing pointwise on moduli space.
Fig. 1. Twisted cycles Aε and Bε and the single internal loop momentum pε .
1.2. Organization and summary of main formulas
In Section 2, standard facts about conformal field theory on genus 1 and genus 2
Riemann surfaces as well as the partition function of a Z2 -twisted boson [22] are reviewed.
A self-contained construction is presented of the Prym differential ωε (z), the Prym period
τε and the Schottky relations.1 The key result in Section 2 is for the Z2 -twisted boson
propagator on a genus 2 Riemann surface Σ. In a canonical homology basis with cycles
AI , BI , I = 1, 2, the 16 possible twists may be described in terms of half-characteristics
ε = (εI ,
I ). The bosonic field x(z) in twist sector ε satisfies the following boundary
conditions,

x(z + AI ) = (−)2εI x(z), x(z + BI ) = (−)2εI x(z). (1.1)
The corresponding Z2 -twisted propagator is characterized by the same boundary condi-
tions in both variables z and w, and the short distance singularity (z − w)−2 . It is obtained
in terms of the Szegö kernel Sδ for characteristic δ, and the normalized Prym differential
ωε ,

∂z x(z)∂w x(w) ε = Sδ + (z, w)Sδ − (z, w) − 4πi∂τε ln ϑi (0, τε )ωε (z)ωε (w). (1.2)
i i
The notation here is as follows. Given a twist ε = 0, six of the ten even spin structures δ are
such that δ + ε is also an even spin structure. These six spin structures may be parametrized
by δi± , with i = 2, 3, 4 and δi− = ε + δi+ and δi± modulo ε clearly map onto the 3 even spin
structures of genus 1. The ϑi are the corresponding genus 1 ϑ-functions for even spin
structures. The precise normalization of ωε (z) is derived in Section 2.6.
In Section 3, the chiral splitting procedure is carefully implemented in the twisted
sectors of the Z2 -twisted superstring theory. Because of the Z2 -twisting, only a single
internal loop momentum survives (instead of the two internal loop momenta in the
untwisted sector). This is illustrated in Fig. 1, where a quadratic branch cut is applied
along the cycle Cε producing double-valued behavior along cycle Dε . The only internal
loop momentum left crosses the remaining independent cycle Aε .
A subtle worldsheet supersymmetric and supermoduli dependent shift in the Prym
period is identified. The results are as follows: let AC [δ] be the left-right symmetric
compactified amplitude for fixed spin structure δ and n directions twisted by Z2 . For
1 Detailed discussions of genus 1 ϑ -functions and genus 2 spin structures, ϑ -functions, modular transforma-
tions and Thomae-type formulas are collected in Appendices A and B, respectively.
simplicity, only the 0-point function is considered here. It is well known (cf. [23, p. 967])
that the gauge-fixed expression for AC [δ] as an integral over supermoduli space is

A 2 2
AC [δ] = dm D(XB B̄C C̄) δ HA |B e−Im −Igh . (1.3)
s M2 A A
Here mA are parameters for a (3|2)-dimensional slice for the even and odd supermoduli,
HA is the corresponding super Beltrami differentials, X, B, B̄, C, C̄ are respectively the
matter and ghost superfields, and Im , Igh are respectively the matter and the ghost actions.
This expression is non-chiral, but one of its fundamental properties is that it can be
split as an integral over internal momenta of the norms squared of a chiral amplitude,
AC [δ, ε](pε ),

A 2 µ µ
AC [δ] = d 10−n pI d n pε dm AC [δ, ε](pε )2 eiπpI Ω̂I J pJ 2 .
ε
s M2 A (1.4)
µ
The chiral amplitude AC [δ, ε](pε ) is superholomorphic on supermoduli space sM2 and
may be computed in terms of the chiral partition functions ZC and ZM of the compactified
and Minkowskian theories respectively, following the method of [1–4]. Parametrizing the
genus 2 supermoduli space by the super period matrix Ω̂I J as bosonic coordinates and the
two Grassmann coordinates ζ 1 , ζ 2 as fermionic coordinates, one obtains,
ZC [δ; ε]
AC [δ, ε](pε ) = AM [δ] exp iπ(τ̂ε + ∆τε )pε2

ZM [δ]

1
× 1− d z d 2 w χz̄+ χw̄+ Sm (z)Sm (w) C
2
8π 2

− Sm (z)Sm (w) M

1
+ d z µ̂(z) Tm (z) C − Tm (z) M .
2
(1.5)
2π
Here, the expression AM [δ] is the chiral superstring measure of the Minkowski space
theory, which was evaluated in [1,4]. The expressions ZM [δ] and ZC [δ,
] are the matter
partition functions for the flat Minkowski space theory (M) and the Z2 -twisted theory (C).
The expressions Sm and Tm are the matter supercurrent and stress tensor respectively. The
bosonic Prym period τε is related to the bosonic period matrix Ω, a relation that we shall
denote by τε = Rε (Ω), and is given explicitly by (for any i, j = 2, 3, 4)
ϑi (0, τε )4 ϑ[δi+ ](0, Ω)2ϑ[δi− ](0, Ω)2

τε = Rε (Ω) ⇔ = . (1.6)
ϑj (0, τε )4 ϑ[δj+ ](0, Ω)2ϑ[δj− ](0, Ω)2
The super-Prym period τ̂ε and the super-period matrix Ω̂I J are related via this same
equation τ̂ε = Rε (Ω̂) and both are invariant under local worldsheet supersymmetry. The
quantity ∆τε is a subtle correction which arises from chiral splitting and is given by2

i
∆τε = − d 2 z d 2 w χz̄+ Sδ (z, w)χw̄+ ωε (z)ωε (w) − ωI (z)ωJ (w)∂I J Rε (Ω̂) .

8π
(1.7)
It is locally supersymmetric and non-vanishing. In the expression for AC [δ; ε](pε ), all
dependence on the original period matrix ΩI J has been eliminated in favor of Ω̂I J .
Henceforth, we view Ω̂I J as the true moduli, and write it simply as ΩI J .
In Section 4, a generalization is given for the chiral measure in an arbitrary background
C with worldsheet supersymmetry and matter central charge 15. In the split gauge defined
by Sδ (q1 , q2 ) = 0, this expression is evaluated for each chiral sector, denoted by λ, and
takes the simple form,

ZC [δ; λ] ζ 1ζ 2
AC [δ; λ] = AM [δ] 1− Z SC 1 C 2 Cλ ,
(q )S (q ) (1.8)
ZM [δ] 4π 2
where ZC [δ; λ] is the matter partition function of the compactified theory (C) for spin
structure δ and chiral sector λ, while SC is the supercurrent of the compactified theory.
This last formula was announced in [1]. It shows that the superstring measure for an
arbitrary background reduces to the evaluation of the matter partition function ZC and
the supercurrent correlator SC (q1 )SC (q2 )Cλ in sector λ.
In Section 5, the two-loop Z2 -twisted blocks AC [δ; ε](pε ) are computed in terms of
ϑ-functions, and a final expression for the twisted chiral measure is obtained in terms of
the super-period matrix Ω̂I J (recall that it is now denoted simply by ΩI J ; similarly, the
super-Prym period τ̂ε is denoted simply by τε ). Defining3

dµC [δ; ε](pε ) ≡ d Ω d 2 ζ AC [δ; ε](pε ),
3

Γ [δ; ε] ≡ Z dζ 2 dζ 1 ∆τε (1.9)
we find
dµC [δ; ε](pε )

iπτε pε2 ZC [δ, ε] Ξ6 [δ]ϑ[δ]
4
=e + iπpε − n∂τε ln ϑi (0, τε ) Γ [δ; ε] d 3 Ω.
2
ZM [δ] 16π 6 Ψ10
(1.10)
The expression Ξ6 [δ](Ω) is the modular covariant form found in [1–4]. The key
new expression here is Γ [δ; ε], which requires a substantial calculation, especially the
determination of its overall sign. The final result is

± i ϑi4 ϑ[δi± ]2 ϑj4
Γ [δi , ε] = ±ν0 |µi . (1.11)
(2π)7 η12 ϑ[δi∓ ]2 ϑ[δj+ ]2 ϑ[δj− ]2
2 Throughout, we shall use the notation ∂ ≡ ∂/∂Ω , ∂

II I I I J ≡ (1/2)∂/∂ΩI J when I = J . The heat-kernel
equation for the ϑ -function is then simply 4π i∂I J ϑ[δ](ζ, Ω) = ∂I ∂J ϑ[δ](ζ, Ω), where ∂I ≡ ∂/∂ζ I .
3 The volume element is defined as follows d 3 Ω = dΩ dΩ dΩ .
11 12 22
Here, the ± signs are correlated with one another on both sides of the equation. The ϑi ’s
are genus 1 ϑ-constants with respect to τε , while the ϑ[δ]’s are genus 2 ϑ-constants with
characteristic δ. The above expression for Γ [δi± , ε] is independent of j .
In Section 6, modular properties in the presence of a twist are studied. The subgroup Hε
that fixes a twist ε is determined: it has 48 elements, and is of the form

Hε = I, T2 , S2 , S2 T2 , S22 , S22 M1 T2 × Hε0 . (1.12)

The matrices M1 , M2 , M3 , S2 , T2 are given in Appendix B.3; and Hε0 is the Abelian sub-
group consisting of the elements Hε0 = {I, M1 , M2 , M3 , M1 M2 , M1 M3 , M2 M3 , M1 M2 M3 }.
The orbits under this group of both even spin structures and twists are worked out.
In Section 7, some of the degeneration limits of the measure are worked out explicitly
and checked versus the results expected on physical grounds.
In Section 8, applications to physically interesting models are studied. These models are
obtained as orbifolds of flat Minkowski space–time with symmetric or asymmetric orbifold
groups G. The orbifold groups considered here are simply generated by reflections and
shifts. The procedure of chiral splitting for symmetric orbifolds is reviewed.
A prescription is given for the construction of the string amplitudes for asymmetric
orbifolds in terms of the chiral blocks obtained by chirally splitting symmetric orbifold
amplitudes. The prescription proceeds as follows. For each model, the set of chiral
blocks {dµL [δL , λL ](pL )}, for all left chiral sectors λL , is determined from the symmetric
orbifold theory obtained from the group generated by all symmetric elements of the
form (fL , fL ), where fL runs over left elements of group elements f = (fL , fR ) ∈ G.
Similarly, the set of antiholomorphic blocks {dµR [δR , λR ](pR )} for all right chiral sectors
λR , is determined from the symmetric orbifold group generated by all (fR , fR ). The GSO
projection is always assumed to be implemented by carrying out a chiral summation over
all spin structures independently for left and right movers,

dµL [λL ](pL ) = ηL [δL ; λL ] dµL [δL , λL ](pL ) (1.13)
δL
for suitable phases ηL [δL ; λL ]. The partition function for the asymmetric orbifold model
G is then given by

ZG = (det Im Ω)−5+n/2 dµL [λL ](pL ) ∧ dµR [λR ](pR )
pL ,pR λL ,λR
× K(λL , λR ; pL , pR ) (1.14)
where n is the number of compactified dimensions, and K(λL , λR ; pL , pR ) are suitable
constant coefficients, depending only on the labels λL , λR , pL , pR for left and right chiral
blocks. The determination of K is a difficult issue, and we shall discuss it further in a
forthcoming publication [24]. In this paper, we shall examine only the chiral models, that
is, only the issue of vanishing of the chiral blocks dµL [λL ](pL ) and dµR [λR ](pR ).
We carry out this procedure for the simplest Z2 models, where n = 4 and G is given
by Z2 reflections. The chiral blocks of this theory are of course just the conformal blocks
dµC [δ; ε](pε ), with index λ = λL = ε, of the Z2 theory we have just derived in (1.10). It
is shown that there is a unique choice of phases for each ε that is consistent with modular
covariance, given by
η[δ; ε] = 1. (1.15)
Upon carrying out the GSO projection, the term involving Γ [δ; ε] cancels due to the genus
1 Riemann identities and the chiral block reduces to
+ 2 − 2
eiπτε pL ϑ[δj ] ϑ[δj ] 3
2
dµC [ε](pL ) = d Ω Ξ6 [δ]ϑ[δ]2 ϑ[δ + ε]2 . (1.16)

16π 6 Ψ10 ϑj4 δ
This expression can be shown to vanish identically, by using identities which involve
Ξ6 [δ] and which were established in [4]. Thus the cosmological constant vanishes point
by point on moduli space, in agreement with the fact that the theory just reproduces type II
superstrings.
The more complicated Kachru–Kumar–Silverstein models are constructed in terms of
an asymmetric orbifold group, whose point group PG is isomorphic to Z2 × Z2 . It is
shown that the chiral blocks dµC [δ; λ](pL ) of such theories are indexed by two twists
λ = λL = (ε, α), and given explicitly by
dµC [δ; λ](pL ) = α|δ dµC [δ; ε](pL ) (1.17)
where the expression dµC [δ; ε](pL ) on the right-hand side are the Z2 blocks of (1.10).
It is shown that modular covariance requires that the GSO projection phases be given by
ηL [δ; λ] = α|δ, that is, the chiral blocks of the theory are given by

dµC [λ](pL ) = α|δ dµC [δ; ε](pL)
δ
+ 2 − 2
eiπτε pL ϑ[δj ] ϑ[δj ] 3
2
= d Ω α|δΞ6 [δ]ϑ[δ]2ϑ[δ + ε]2 .

16π 6 Ψ10 ϑj4 δ
(1.18)
For each ε, the label α in λ can be classified into orbits of the modular group Hε leaving ε
invariant. There are 4 such orbits.
We find that the contributions of two orbits vanish identically, the contribution of a third
orbit vanishes of order τ 4 along the divisor of separating nodes,4 but the contribution of
the remaining orbit vanishes only to order τ 2 along the same divisor. More precisely, we
find

α|δΞ6 [δ]ϑ[δ]2 ϑ[δ + ε]2
δ

= −256π 2τ 2 η(τ1 )12 η(τ2 )12 ϑ38 − ϑ48 (τ1 )ϑ24 ϑ32 ϑ42 (τ2 ) + O τ 4 (1.19)
for ε and α given by
1
0 0 0 2
ε= , α= . (1.20)
0 12 0 0
Thus the cosmological constant in the KKS models does not vanish point by point on
moduli space.
4 For brevity, the notation Ω = τ , Ω = τ and Ω = τ will be used throughout.

11 1 22 2 12
2. Z2 orbifold conformal field theory
The main goal of this section is to derive elements from the theory of Z2 -twisted fields
which we need. For this, we require some background on spin structures, ϑ-functions,
double covers, and Prym differentials, summarized in Sections 2.1 and 2.2. The partition
function for Z2 -twisted bosonic scalar fields has been obtained by Dijkgraaf, Verlinde,
and Verlinde [22]. It is recalled in Section 2.3. For our purpose, the main objects of
interest are the propagators. The fermionic one is easy to find, so we concentrate on the
bosonic propagator ∂z x(z)∂w x(w)ε . As a warm-up, it is derived for the case of genus 1
in Section 2.4. The rest of the section is devoted to extending this method to the case of
genus h = 2. The Schottky relations between ϑ-functions and the Prym differential are
also derived.
2.1. Spin structures and ϑ-functions
Let Σ be an orientable compact genus h Riemann surface without boundary on which

a canonical homology basis for H1 (Σ, Z) of cycles AI , BI , I = 1, . . . , h, is chosen,
#(AI , AJ ) = #(BI , BJ ) = 0, #(AI , BJ ) = −#(BJ , AI ) = δI J . (2.1)
Holomorphic Abelian differentials ωI may be normalized on A-cycles and their integral
along B-cycles yields the period matrix ΩI J ,

ωJ = δI J , ωJ = ΩI J . (2.2)
AI BI
Given the intersection numbers (2.1), the basis is unique up to modular transformations
belonging to the modular group Sp(2h, Z). The structure of the modular group, its action
on spin structures, twists and ϑ-functions is summarized in Appendix B.3.
A spinor field on Σ requires the assignment of a spin structure on Σ, corresponding
to an element of H1 (Σ, Z2 ). In a given homology basis, a spin structure may be labeled
by a half-integer characteristics κ ≡ (κ |κ ) where κ and κ , each valued in {0, 1/2}h ,
and specifies the monodromies around A- and B-cycles respectively. The parity of the spin
structure κ is that of the integer 4κ · κ , and is modular invariant. For genus 1, the 4 spin
structures separate into 3 even and 1 odd (respectively denoted µ and ν0 ). For genus 2, the
16 spin structures separate into 10 even and 6 odd spin structures (respectively denoted by
δ and ν). An important pairing between characteristics is given by the signature,

κ|ρ ≡ exp 4πi(κ · ρ − ρ · κ ) = ±1. (2.3)

The ϑ-function with characteristics κ = (κ |κ ) is defined by

ϑ[κ](ζ, Ω) ≡ exp πi(n + κ )Ω(n + κ ) + 2πi(n + κ )(ζ + κ ) . (2.4)

n∈Zh
It satisfies a number of monodromy relations, which are listed in Appendix B. For
genus 1, the relations between the ϑ-functions with characteristics and the standard Jacobi
ϑ-functions are as follows. For odd spin structure, we have ϑ1 (z|τ ) ≡ ϑ[ν0 ](z|τ ), while
for even spin structures, we have ϑi (z|τ ) ≡ ϑ[µi ](z|τ ), with the corresponding even spin
structure characteristics given by

µ2 = 12 0 , µ3 = (0|0), µ4 = 0 12 . (2.5)
Two fundamental building blocks for the theory of functions and forms on a Riemann
surface5 are the prime form E(z, w) and the Szegö kernel Sδ (z, w) (which we shall only
need for even spin structure δ),
z z
ϑ[ν]( w ωI , Ω) ϑ[δ]( w ωI , Ω)
E(z, w) = , Sδ (z, w) = . (2.6)
hν (z)hν (w) ϑ[δ](0, Ω)E(z, w)
The holomorphic 1/2 form hν (z) for odd spin structure ν obeys hν (z)2 = ∂I ϑ[ν](0)ωI (z).
The prime form is a holomorphic −1/2 form in z and w on the universal cover of Σ, which
behaves as z − w when z ∼ w, and is independent of ν. The Szegö kernel for even spin
structure δ is a meromorphic 1/2 form in z and w on Σ with a simple pole at z = w with
unit residue, The importance of these quantities derives from the fact that Sδ is the chiral
fermion propagator, while ∂z ∂w ln E(z, w) is the scalar propagator on Σ.
2.2. Z2 twisting; unramified double covers
The configurations of a scalar x(z) (respectively spinor ψ(z)) field on Σ with a Z2 -

twist correspond to functions (respectively sections of a spin bundle with characteristics κ)
that are double-valued on Σ. These configurations fall into 16 distinct topological sectors
corresponding to the elements of H1 (Σ, Z2 ), and in a given homology basis may be labeled
by half characteristics ε ≡ (ε |ε ), where ε and ε are each valued in {0, 1/2}2 (see [22,
27]). For the scalar field x(z), and the spinor field ψ(z) with spin structure κ, we have

x(z + AI ) = (−)2εI x(z), ψ(z + AI ) = −(−)2εI +2κI ψ(z),

x(z + BI ) = (−)2εI x(z). ψ(z + BI ) = −(−)2εI +2κI ψ(z). (2.7)
A Z2 -twisted scalar (respectively spinor) field may be viewed as a field defined on
the surface Σ with a quadratic branch cut along a cycle Cε . The Z2 -twisted field is then
double-valued around a conjugate cycle Dε , which we parametrize as follows,

Dε = (2εI AI + 2εI BI ). (2.8)
I =1,2
The remaining generators of H 1 (Σ, Z) will be denoted by Aε and Bε , and have intersection
numbers, #(Aε , Bε ) = #(Cε , Dε ) = 1, while all others vanish. Under the modular group,
all twists transform in two orbits, one consisting of ε = 0 and the other of all ε = 0. It
will often be convenient to use the action of the modular group on the twists to choose a
standard reference twist. A simple choice consists of taking

0 0
ε= ⇔ Aε = A1 , Bε = B1 , Cε = A2 , Dε = B2 . (2.9)
0 12
It will often be helpful to keep the notation Aε , Bε , Cε , Dε for the sake of generality.
5 For reviews of string perturbation theory, see [23,25]; for recent results on Green functions on twisted
surfaces, see [26]; for the original mathematical results, see [27].
Fig. 2. The genus 2 Riemann surface and its unramified branched double cover.
Double valued scalar fields on a surface Σ may also be viewed as single valued scalar
functions of the double cover Σ̂ of the original surface Σ. (Similarly, double valued spinor
fields may be viewed as well-defined sections of a spin bundle of the cover Σ̂.) If a function
is double valued along the cycle B2 , and has a quadratic branch cut across the conjugate
cycle A2 , the surface Σ̂ is obtained by cutting open Σ along the cycle A2 and gluing the
two boundaries of this surface to the boundaries of its image under the reflection involution
ι, as shown schematically in Fig. 2. Given any twist ε = 0, the double cover Σ̂ , as defined
above, is unique.
The double cover Σ̂ is a surface of genus 3, whose complex structure is induced by
the complex structure of Σ. Functions and forms on Σ that are Z2 -twisted (i.e., double-
valued) along the cycle B2 = Dε , become functions and forms on Σ̂ that are odd under
the reflection involution ι. Of special importance is the Prym differential, ωε , which
is a holomorphic 1-form, odd under the reflection ι. For genus 2, ωε is unique up to
normalization. By construction, its periods along the cycles A2 = Cε and B2 = Dε vanish.
Normalizing the period of ωε to 1 along A1 = Aε , its integral along B1 = Bε yields the
Prym period τε ,

ωε = 1, ωε = τε . (2.10)
Aε Bε
The Prym period is a complex number satisfying Im τε > 0.

Specification on Σ of a twist ε = 0, canonically separates the even spin structures δ
into two groups, according to whether δ + ε is even or odd. This separation is clearly
modular covariant. The group for which δ + ε is even consists of 6 elements which will be
denoted δi+ and δi− , i = 2, 3, 4. The group for which δ + ε is odd consists of the remaining
4 elements. Without loss of generality, we may assume that δi− = δi+ + ε. For the twist ε
of (2.9), for example, we have

µi µi
δi+ = , δi− = , (2.11)
0|0 0| 12
where µi are the genus 1 even spin structure assignments of (2.5). Clearly, a genus 2
surface with a twist ε = 0 has a canonically associated group of 6 even spin structures δi± ,
which in turn map 2 to 1 onto the 3 genus 1 even spin structures.
The period matrix ΩI J of the genus 2 surface uniquely determines the Prym period τε ,
up to the action of generators in the Torelli group;6 in this case, up to the transformations
τε → τε + 4 (see [22]). The relation between ΩI J and τε is given by the Schottky relations,
already introduced in (1.6). They may be expressed in an equivalent form as follows, for
any i = j and i, j = 2, 3, 4,
+ −
ϑ[δi+ ](0, Ω)2ϑ[δi− ](0, Ω)2 ϑ[δj ](0, Ω) ϑ[δj ](0, Ω)
2 2
= . (2.12)
ϑi (0, τε )4 ϑj (0, τε )4
Here, δi± are the group of 6 even spin structures introduced in the preceding paragraph
and ϑi are the genus 1 ϑ-functions associated with the genus 1 spin structures µi onto
which δi± map. Clearly, all three such relations are equivalent to one another. The Schottky
relations for genus 2 will be proven and their origin will be explained in Section 2.6.
2.3. The Dijkgraaf–Verlinde–Verlinde partition function
The first key result of [22] of interest to us is the formula for the partition function for
a single Z2 -twisted boson on a circle of radius R. In a sector of twist ε, the worldsheet
instanton sum is given by
2
Z[ε](Ω, R) = Z qu [ε](Ω) exp πi pL τε − pR

2
τ̄ε , (2.13)
(pL,pR )∈ΓR
where the momentum sum is taken over the lattice

n n
ΓR = (pL , pR ) = + mR, − mR . (2.14)
2R 2R
The oscillator part of the partition function is independent of the radius and may be
computed by equating the partition functions of the Z2 -twisted
√ model with the partition
function of the circle theory at the self-dual radius R = 1/ 2. The result may be expressed
in terms of the spin structures δi± , introduced in the preceding subsection,

ϑ[δi+ ](0, Ω)ϑ[δi− ](0, Ω)
Z [ε](Ω) =
qu . (2.15)
Z(Ω)2 ϑi (0, τε )2
6 The Torelli group is the quotient of the full mapping class group, consisting of all equivalence classes of
disconnected diffeomorphisms of the surface Σ by the modular group.
Here, Z(Ω) is the inverse of the chiral bosonic partition function of the uncompactified
circle theory. In view of the Schottky relations (2.12), this expression is independent of the
choice of i and therefore only depends upon ε.
2.4. Genus one Z2 -twisted propagators and partition functions
On the torus Σ = C/Z + τ Z, with modulus τ , a scalar field x(z) and a fermion field
ψ(z) with even spin structure δ, twisted by ε, obey the following monodromy conditions,

x(z + 1) = (−)2ε x(z), ψ(z + 1) = −(−)2δ (−)2ε ψ(z),

x(z + τ ) = (−)2ε x(z), ψ(z + τ ) = −(−)2δ (−)2ε ψ(z). (2.16)
The corresponding propagators satisfy the same monodromy conditions as the fields in
both z and w. The bosonic propagator Bε (z, w) is symmetric under interchange of z and
w, and has a double pole at z = w, while the chiral fermion propagator Sδ+ε (z, w) is
antisymmetric under interchange of z and w, and has a simple pole at z = w. Matching
monodromies and poles yields the following propagators,
ϑ1 (z − w|2τ )
Bε (z, w) = ∂x(z)∂x(w) = ∂z ∂w ln ,
ϑ1 (z − w + τ |2τ )
ϑ[δ + ε](z − w, τ )ϑ1 (0, τ )
Sδ+ε (z, w) = ψ+ (z)ψ+ (w) = . (2.17)
ϑ[δ + ε](0, τ )ϑ1 (z − w, τ )
The chiral fermion propagator is just the Szegö kernel with spin structure shifted to δ + ε.
It is convenient to recast the twisted boson propagator in terms of the Szegö kernel,
since it will turn out that it is these formulas which will lend themselves to generalization
to higher genus. Using the doubling formulas for genus 1, as well as the formulas for z-
derivatives (see Appendix A), the following relation is established (valid now for arbitrary
twist ε = 0)
Bε (z, w) = Sµ (z, w)Sµ+
(z, w), (2.18)
where µ is any even spin structure such that µ + ε is also even. There are two possible
choices for µ for any given ε = 0; e.g., if ε = (0| 12 ), one can have µ = (0|0) or µ = (0| 12 ).
Theories with worldsheet supersymmetry will be of special interest. The worldsheet
supercurrent formed of a free scalar x and a chiral fermion ψ+ with spin structure δ is
given by S = −(1/2)ψ+ ∂z x and is invariant under the Z2 twist ε. The supercurrent two-
point function is given by
1 1
S(z)S(w) = Bε (z, w)Sδ+
(z, w) = Sµ (z, w)Sµ+
(z, w)Sδ+
(z, w). (2.19)
4 4
Using the OPE for the two supercurrents yields the stress tensor for the system,
ĉ/4 T (w)/2
S(z)S(w) = + + reg . (2.20)
(z − w)3 z−w
Expanding the Szegö kernels, one find ĉ = 1, as expected. Using the heat equation,
∂z2 ϑ[µ](z|τ ) = 4πi∂τ ϑ[µ](z|τ ), the remaining contribution may be recast as follows,

ϑ[µ]ϑ[µ + ε]ϑ[δ + ε]
T ε,δ = πi∂τ ln . (2.21)
ϑ[0|0]ϑ[ 12 |0]ϑ[0| 12 ]
The corresponding partition function for one boson and one fermion of spin structure δ,
√
both twisted by ε, are obtained using the relation T ε,δ = 4π/ gδ ln Zε,δ /δg zz and are

ϑ[δ + ε](0, τ ) 1/2
Zε,δ = (2.22)
ϑ[ν + ε](0, τ )
with ν = ( 12 | 12 ) in agreement with known results [28,29].
2.5. Genus two Z2 -twisted propagators
Inspection of the monodromy equations (2.7) reveals that implementing a twist ε on the
fermion field ψ+ (z) and on its propagator simply results in replacing the spin structure δ
by δ + ε. Of interest here are only the cases where δ + ε is itself an even spin structure.7
The propagator for the twisted fermion is thus the Szegö kernel for δ + ε,

ψ+ (z)ψ+ (w) ε = Sδ+ε (z, w). (2.23)
Viewed on the surface Σ, this quantity is a double-valued section of a spin bundle for spin
structure δ; it becomes a single-valued section of a spin bundle on Σ̂.
Implementing a twist ε on a scalar field x(z) is achieved by imposing the monodromy
conditions of (2.7). It may be constructed in terms of the prime form for the covering
surface Σ̂ as was done in [22]. Actually, we shall only need the propagator for the twisted
field ∂z x,

Bε (z, w) ≡ ∂z x(z)∂w x(w) ε . (2.24)
It obeys the same monodromy conditions as x does in both z and w and has a double pole
at z = w with unit residue.
By analogy with the case of genus 1, the propagator Bε (z, w) may be more simply
constructed in terms of products of the Szegö kernels for even spin structures δi+ and δi−
of (2.11). Indeed, the products
Bi (z, w) = Sδ + (z, w)Sδ − (z, w), i = 2, 3, 4 (2.25)

i i
readily produce the required double pole. Given the fact that δi+ + δi− = ε modulo integral
periods, they also manifestly exhibit the required monodromy by ε.
The differences Bi (z, w) − Bj (z, w) are holomorphic 1-forms in both z (as well as in
w), and have the same non-trivial monodromy as ∂z x. On a genus 2 surface, such a one-
form must be proportional to the Prym differential ωε (z). Since ωε is unique, there must be
one linear relation between the three Bi . Taking this into account, the twisted propagator
Bε may be expressed via any of the following three expressions,
Bε (z, w) = Sδ + (z, w)Sδ − (z, w) + bi ωε (z)ωε (w). (2.26)

i i
A more detailed discussion of the Prym differential and period will be given later.
7 The cases where δ + ε is and odd spin structure will not contribute to the cosmological constant for the
orbifold models we shall consider later on.
The calculation of bi
Below, the coefficient bi will be computed using the partition function for the twisted
boson theory, derived in [22] and quoted in Section 2.3. The result, in the basis of reference
(2.11), is given by
bi = −4πi∂τε ln ϑi (0, τε ). (2.27)
The detailed calculation of bi proceeds as follows. The stress tensor for the twisted boson
x is computed in two different ways; first, from the propagator Bε ,

1 1
Tzz = lim Bε (z, w) − (2.28)
2 w→z (z − w)2
and second from the variation of the partition function Z qu [ε] of the twisted x field,
4π δ
Tzz = δzz ln Z qu [ε], δzz ≡ √ . (2.29)
g δg zz
The partition function Z qu [ε], was computed in [22], and was presented in the notation of
(2.11) in (2.15). Equating (2.28) and (2.29) yields the following results. In the variation
(2.29) of (2.15), the variation of Z and of the ϑ-functions ϑ[δi+ ]ϑ[δi− ] precisely matches
the contribution to (2.28) of the first term in (2.26). The remaining variation gives
1
bi ωε (z)2 = −δzz ln ϑi (0, τε ) = −(δzz τε )∂ ln ϑi (0, τε ). (2.30)
2
The variation of τε with respect to the metric is analogous to the variation of the
genus 2 period matrix, δzz ΩI J = 2πiωI (z)ωJ (z). In the next subsection (where a detailed
construction of the Prym differential will also be given), we shall derive the following
expression,
δzz τε = 2πiωε (z)2 . (2.31)
Combining these results readily yields (2.27).
2.6. The Prym differential and the Schottky relations
The construction of the normalized Prym differential, the Schottky relations and the
variational equation (2.31) for τε require a more concrete understanding of the Prym
differential. It is convenient to obtain the Prym differential from its relation to the Szegö
kernel,

ωε (z)ωε (w) ∼ Sδ + (z, w)Sδ − (z, w) − Sδ + (z, w)Sδ − (z, w) . (2.32)
i i j j
The constants of proportionality in this relation will be determined below.

To do so, the hyperelliptic parametrization of genus 2 Riemann surfaces is used.8 The

branch points are denoted by p1 , . . . , p6 and the corresponding hyperelliptic curve is

6
s2 = (z − pa ). (2.33)
a=1
Recall that the identification between even and odd spin structures and branch points
proceeds as follows,
νa odd ↔ branch point pa ,

δ even ↔ partition A ∪ B, A = {pa1 , pa2 , pa3 }, B = {pb1 , pb2 , pb3 }. (2.34)
The twist will be left general. To this end, spin structures and twists will be parametrized by
odd spin structures, {νa , νb , νc , νd , νe , νf }, where (abcdef ) is a permutation of (123456).
The 6 C2 = 15 (= 16 − 1) non-zero twists ε can be uniquely labeled by two distinct odd
spin structures νa,b , such that ε = νa + νb ≡ 0. The same twist can also be characterized
as a sum of two even spin structures, ε = −δi+ + δi− . There are 3 distinct choices for these
sets which we label here as
δi+ = νa + νc + νd = −(νb + νe + νf ),
δi− = νb + νc + νd = −(νa + νe + νf ),
δj+ = νa + νd + νe = −(νb + νc + νf ),
δj− = νb + νd + νe = −(νa + νc + νf ),
δk+ = νa + νc + νe = −(νb + νd + νf ),
δk− = νb + νc + νe = −(νa + νd + νf ). (2.35)
We note that a, b cannot belong to the same set in the partition associated with the spin
structure δ since then, δ + ε would be odd.
Let δ be the even spin structure corresponding to the partition of the 6 branch points
into two sets A, B of 3 branch points each, as in (B.18). In the hyperelliptic representation,
the Szegö kernel for the spin structure δ is given by [27],

1 sA (z)sB (w) + sA (w)sB (z) dz 1/2 dw 1/2
Sδ (z, w) = , (2.36)
2 z−w s(z) s(w)
where the following notation is used,
sA (z)2 = (z − pa1 )(z − pa2 )(z − pa3 ),

sB (z)2 = (z − pb1 )(z − pb2 )(z − pb3 ),
s(z)2 = (z − p1 )(z − p2 )(z − p3 )(z − p4 )(z − p5 )(z − p6 ). (2.37)
8 A detailed discussion of the map between the ϑ -function and hyperelliptic formulations of genus 2 Riemann
surfaces was given in [4]; see Appendix B of this paper for a summary of the relevant facts.
The combinations of Szegö kernels that enter into (2.32) may be expressed in this
hyperelliptic formulation,
(pc − pd )(pf − pe )
Sδ + (z, w)Sδ − (z, w) − Sδ + (z, w)Sδ − (z, w) = dz dw,
j j k k 4scdef (z)scdef (w)
(pc − pf )(pe − pd )
Sδ + (z, w)Sδ − (z, w) − Sδ + (z, w)Sδ − (z, w) = dz dw,
k k i i 4scdef (z)scdef (w)
(pc − pe )(pd − pf )
Sδ + (z, w)Sδ − (z, w) − Sδ + (z, w)Sδ − (z, w) = dz dw. (2.38)
i i j j 4scdef (z)scdef (w)
The hyperelliptic form of the Prym differential is thus given by
dz
ωε (z) ∼ . (2.39)
scdef (z)
This quantity involves only 4 branch points, and by SL(2, C)-invariance depends on a
single modulus, which is the Prym period τε .
The Schottky relations

The Schottky relations arise from relating the modulus of the genus 1 curve with branch
points pc , pd , pe , pf to the moduli of the full genus 2 curve with all 6 branch points.
Without loss of generality, the following identification (ej , ek , ei , ∞) = (pc , pd , pe , pf )
can be made, which allows us to put the elliptic curve in standard from,
y 2 = 4(x − e2 )(x − e3 )(x − e4 ) = 4(x − pc )(x − pd )(x − pe ). (2.40)
The genus 1 Thomae formulas then yield, (see [30, p. 361, Eq. (7)])
π2 4
ej − ek = pc − pd = σ (µj , µk ) ϑ ,
4ω2 i
π2
ek − ei = pd − pe = σ (µk , µi ) 2 ϑj4 ,
4ω
π2
ei − ej = pe − pc = σ (µi , µj ) 2 ϑk4 . (2.41)
4ω
The function σ is antisymmetric, σ (µi , µj ) = −σ (µj , µi ) and is normalized by
σ (µ2 , µ3 ) = σ (µ3 , µ4 ) = σ (µ2 , µ4 ) = +1. (2.42)
The consistency of these relations follows from the genus 1 Jacobi identity,
σ (µi , µj )ϑk4 + σ (µj , µk )ϑi4 + σ (µk , µi )ϑj4 = 0. (2.43)
Furthermore, 2ω is the Aε period of the elliptic curve. Recall from [4] that, in terms of the
modular object Mνa νb ,
Mνa νb ≡ ∂1 ϑ[νa ](0, Ω)∂2ϑ[νb ](0, Ω) − ∂2 ϑ[νa ](0, Ω)∂1ϑ[νb ](0, Ω),

6
M2νa νb = π 4 ϑ[νa + νb + νk ](0, Ω)2 (2.44)
k=a,b
we have a refined form of the genus 2 Thomae identities,

(pa − pb )(pc − pd ) Mνa νb Mνc νd
= . (2.45)
(pa − pc )(pb − pd ) Mνa νc Mνb νd
The identification leads to the following relations between the three cross ratios that
may be constructed out of the 4 points c, d, e, f ,

ej − ek 2 Mcd Mef ϑ 4 [δi+ ]ϑ 4 [δi− ]
2 2
ϑi8
(0, τ ε ) = = = (0, Ω). (2.46)
ϑj8 ek − ei M2cf M2de ϑ 4 [δj+ ]ϑ 4 [δj− ]
In particular, the relation between the separating degeneration limit of the genus 2 periods
and the Prym period is ϑ[µi ](0, Ω11)8 = ϑ[µi ](0, τε )8 so that τε = Ω11 + O(Ω12 ) up to
the possible addition of an integer multiple of 8.
The Prym differential in terms of Szegö kernels

It is now straightforward to construct the normalized Prym differential. Still in the
conventions of [30, pp. 330, 331], the Aε cycle is chosen so that it produces the period 2ω.
This is achieved by taking Aε to be the cycle around the branch cut between the branch
points ek and ei , while the Bε cycle goes from ej to ek ,
ei ek
dx dx
ω= , ω = . (2.47)
y(x) y(x)
ek ej
Therefore, the normalized Prym differential is simply

1 dx
ωε (x) ≡ . (2.48)
2ω y(x)
With this normalization, the relations between the normalized Prym differential and the
Szegö kernel are completely fixed and are found to be
Sδ + (z, w)Sδ − (z, w) − Sδ + (z, w)Sδ − (z, w)
i i j j
= −σ (µi , µj )π ϑk (0, τε ) ωε (z)ωε (w).

2 4
(2.49)
Clearly, the sum of the left terms vanishes and so does the sum of the right terms in view
of the genus 1 Jacobi identity (2.43).
The square of the Prym differential may be deduced from the expression in terms of the
product of Szegö kernels, evaluated as w → z,
1
ωε (z)2 = −σ (µi , µj ) lim Sδ + (z, w)S − (z, w) − S + (z, w)S − (z, w)
δ δ δ
π 2 ϑk4 w→z i i j j

1 + −
ϑ[δi ]ϑ[δi ]
= −σ (µi , µj ) ωI (z)ωJ (z)∂I ∂J ln . (2.50)
4
2π 2 ϑk I,J ϑ[δj+ ]ϑ[δj− ]
Differentiating relation (2.46), with respect to ΩI J , we have

ϑi 1 ϑ[δi+ ]ϑ[δi− ]
(∂I J τε )∂τε ln (0, τε ) = ∂I J ln . (2.51)
ϑj 2 ϑ[δj+ ]ϑ[δj− ]
Using the heat equation 4πi∂I J ϑ = ∂I ∂J ϑ, the relation δzz ΩI J = 2πiωI (z)ωJ (z) and
ϑi π
∂τε ln (0, τε ) = i σ (µi , µj )ϑk (0, τε )4 (2.52)
ϑj 4
yields the equation announced in (2.31).
3. Chiral splitting for Z2 orbifold theories
Chiral splitting is the process of decomposition of a chirally symmetric amplitude into

a sum of terms, each of which may be written as the absolute value square of a chiral
block which is meromorphic on supermoduli space and has meromorphic dependence
on the vertex insertion data [20,21]. Each chiral block corresponds to a single N = 1
superconformal family. The chiral blocks thus obtained may be used to carry out chiral
GSO projections [31], which are required for the construction of both the type II [32]
and heterotic [33] superstring theories. The precise manner in which chirally symmetric
amplitudes are chirally split was worked out long ago for flat space–time [21] (see also
[20,23]). In the present section, chiral splitting will be carried out in detail in the Z2 -
twisted sectors for orbifolds of flat space–time involving Z2 twistings. For definiteness,
the restriction to two-loops is made from the outset.
3.1. Chirally symmetric amplitudes
The compactified space–time C will be viewed as an N = 1 super conformal field

theory on a worldsheet Σ of genus 2, coupled to a two-dimensional supergeometry with
superframe EM A and U (1) superconnection ΩM , [23,34]. The supergeometry (EM A , ΩM )
is subject to the Wess–Zumino constraints, indicated here by the delta function δ(T ).
In Wess–Zumino gauge, we have Em a = em a + θ γ a χm , omitting the auxiliary field,
em a is a frame for the worldsheet metric gmn = em a en b δab , and χm α = (χz̄+ , χz− ) is the
worldsheet gravitino field. The definition of a supergeometry requires a spin structure δ on
the worldsheet.
For fixed spin structure δ, superstring scattering amplitudes are built out of chirally
symmetric correlations functions of C coupled to two-dimensional supergeometry,

N
AC [δ] = A
DEM DΩM δ(T ) DX µ
Vi e−Im . (3.1)
C i=1
Here, Vi are vertex operators for physical states and the subscript C stands for the
functional integral evaluated in the superconformal field theory associated with C. Upon
factoring out the local symmetries of the theory, the integration over supergeometries
(EM A , ΩM ) reduces to an integral over supermoduli space [20,35], which is defined by

sM2 = EM A , ΩM /{Gauge Symmetries}. (3.2)

This integral can be written explicitly by choosing a (3|2)-dimensional slice S of

supergeometries (EM A , ΩM ) ([2], Eq. (2.11)),
N
A 2 2
AC [δ] = dm D(XB B̄C C̄) δ HA |B Vi e−Im −Igh .
s M2 A A i=1
(3.3)
In this formula, mA = (ma , ζ α ) is a set of local complex coordinates on sM 2 where A runs
over the (3|2) complex dimensions of sM2 and HA are the Beltrami superdifferentials for
the slice chosen to represent sM2 . The superfields B and C represent all the superghosts;
in standard component notation, they are given by B = β + θ b and C = c + θ γ , omitting
their corresponding auxiliary fields.
For orbifold compactifications, the starting point will be the matter action for flat
space–time. The matter superfield has the following component decomposition, Xµ =
µ µ
x µ + θ ψ+ + θ̄ ψ− , up to auxiliary fields. Henceforth, we restrict to the case of Z2 -orbifolds
and we have,

1
Im = d 2|2 z sdet EM A D+ Xµ D− Xµ
4π

1 µ µ µ µ
= d 2 z ∂z x µ ∂z̄ x µ − ψ+ ∂z̄ ψ+ − ψ− ∂z ψ−
4π

1
− χz̄+ Sm − χz− S̄m − χz̄+ χz− ψ+ ψ−
µ µ
(3.4)
4
while the ghost action is given by

1
Igh = d 2|2 z sdet EM A (BD− C + B̄D+ C̄)
2π

1
= d 2 z (b∂z̄ c + β∂z̄ γ + b̄∂z c̄ + β̄∂z̄ γ̄ − χz̄+ Sgh − χz− S̄gh ). (3.5)
2π
The matter and ghost supercurrents are given by
1 µ µ 1 3
Sm = − ψ+ ∂z x+ , Sgh = bγ − β∂z c − (∂z β)c. (3.6)
2 2 2
These formulas are identical in form to those for flat space–time, with the exception
µ
that here, both the boson x µ and fermion ψ± are Z2 -twisted by the same twist ε. This
guarantees in particular that the matter supercurrent and stress tensor are single-valued.
3.2. Chiral splitting of general amplitudes in Z2 -twisted sectors
Let ε = 0 denote the Z2 -twist which is applied to n components of the scalar field x
and n components of the fermion fields ψ± taken to have spin structure δ. This means that
these fields satisfy the twisted boundary conditions of (2.7). The resulting chiral scalar and
chiral fermion propagators were computed in Section 2, and are given by

∂z x+ (z)∂w x+ (w) ε = Bε (z, w),

ψ+ (z)ψ+ (w) ε = Sδ+ε (z, w). (3.7)
The functional integral over n components of twisted ψ± fields is given by the determinant
of the Dirac determinant, which is holomorphically factorized (up to the Belavin–Knizhnik
anomaly [36], whose effects cancel in the full amplitude),

ϑ[δ + ε](0, Ω) n
/ n/2 =
(DetD) .
(3.8)
Z(Ω)
The quantity Z is the inverse of the chiral partition function of a free untwisted scalar;
its expression in terms of ϑ-functions was given in [2], Eq. (4.5) but will not be needed
here. The functional integral over x µ in the twisted sector is simpler than in the untwisted
sector, because no zero-mode occurs in the twisted sectors. By the Belavin–Knizhnik result
[36], the combination of determinants that holomorphically factorizes involves the inner
products of holomorphic one forms. In the twisted sector there is precisely one such form,
namely the Prym differential ωε , normalized as in (2.10). The functional determinant over
n twisted fields x is then given by the following expression,

1 ϑ[δj+ ](0, Ω)ϑ[δj− ](0, Ω) n
(Det ∆)−n/2 = . (3.9)
(2 Im τε )n/2 Z(Ω)2 ϑj (0, τ )2
The obstruction to its holomorphic factorization lies entirely in the prefactor involving
Im τε . Chiral splitting is achieved by introducing a single internal loop momentum pε , in
terms of which also the prefactor may be split,

1 2 2
n/2
= d n pε eiπτε pε . (3.10)
(2 Im τε )
The functional integral (3.3) exhibits further obstructions to chiral splitting. The first is
through the presence of the quartic fermion term χ χ̄ψ+ ψ− in the matter action Im ; the
second is through the non-holomorphic dependence of the full fermion propagator,
2π
∂z x(z)∂w x(w) ε = Bε (z, w) − ωε (z)ωε (w),
Im τε
2π
∂z x(z)∂w̄ x(w) ε = −2πδ(z, w) + ωε (z)ωε (w). (3.11)
Im τε
The first line may be established by comparing the stress tensors computed from a variation
of the scalar determinant (3.9) and from the full x-field propagator. This procedure is the
same as the one used in (2.28) and (2.29), but is now applied to the full twisted scalar
partition function and propagator. The second line in (3.11) is obtained by applying ∂w̄ to
the first line and then integrating in w; the remaining coefficient in front of ωε (z)ωε (w) is
fixed by requiring that the integration versus ωε (w) vanish.
Next, one proceeds in parallel with the proof of chiral splitting for the case of flat space–
time [21,23]. The result may be summarized in terms of a set of effective rules. The final
formula for the integration over the matter fields may be recast in the following form,

A 2 µ µ
AC [δ] = d 10−n pI d n pε dm AC [δ, ε](pε )2 eiπpI Ω̂I J pJ 2 ,
ε
s M2 A

N
AC [δ, ε](pε ) = δ HA |B Vichi
A i=1
2
d z + µ
× exp χ S(z) + pε
µ
dz ∂z x+ . (3.12)
2π z̄ +
Bε
Here, S(z) is the total worldsheet supercurrent, given by S(z) = Sm (z) + Sgh (z). The
variance of the Gaussian in the first line is given by the super-period matrix, which is a
shift of the bosonic period matrix by an even, Grassmann valued, and nilpotent element,
given by [23],

i
Ω̂I J = ΩI J − d 2 z d 2 w ωI (z)χz̄+ Sδ (z, w)χw̄+ ωJ (w). (3.13)
8π
Furthermore, Vichi is the chiral part of the vertex operator Vi . All contractions in this chiral
correlator · · ·+ are to be carried out with the help of the propagators for the chiral fields
x+ and ψ+ given in (3.7). In the present paper, the emphasis will be on the chiral measure
and the cosmological constant. Therefore, the precise form of the operators Vi and their
chiral part Vichi will not be needed and will not be presented here.
3.3. The chiral measure in the Z2 -twisted sectors

µ
In the absence of vertex operator insertions, the pε -dependence of the amplitude arises
µ µ
from the x+ -contractions of the term involving pε with itself and with Sm , yielding
2
d z + µ
exp χ S(z) + pε µ
dz ∂z x+
2π z̄ x+ ,ψ+
Bε

1
= exp iπτε pε2 − pεµ d 2 z χz̄+ ψ+ ωε (z)
µ
2

1 + +
− 2 2
d z d w χz̄ χw̄ S(z)S(w) x+ 1x+ , (3.14)
8π ψ+
where 1x+ denotes the Z2 -twisted chiral boson partition function. Carrying out also the
ψ+ contractions of the pε -dependent terms, the following result is obtained,
2
d z + µ
exp χ S(z) + pε µ
dz ∂z x+
2π z̄ x+ ,ψ+
Bε

1 + +
= 1− d z d w χz̄ χw̄ S(z)S(w) x ZC [δ; ε] exp iπ τ̃ε pε2 . (3.15)

2 2
8π 2 +
Here, the product of the Z2 -twisted chiral scalar and fermion partition functions 1x+
and 1ψ+ is denoted by ZC [δ, ε]. In the above formula, all the pε -dependence may be
regrouped in terms of a Gaussian with the following variance,

i
τ̃ ≡ τ − d 2 z d 2 w χz̄+ ωε (z)Sδ+ε (z, w)χw̄+ ωε (w). (3.16)
8π
This correction has the same origin as the corrections to the period matrix that lead to the
super period matrix in the uncompactified string. Its proper interpretation here is, however,
more subtle, and will be presented in detail later.
The contributions of the ghost partition function and supercurrent correlators are
the same as they were in flat space–time and may be taken from [2]. Assembling all
contributions, the chiral measure is given by the following expression

a b(pa ) α δ(β(qα ))M ZC [δ; ε]
AC [δ; ε](pε ) = ∗ exp iπ τ̃ε pε2

det ΦI J + (pa ) · detHα |Φβ ZM [δ]

1 + +
× 1− d z d wχz̄ χw̄ S(z)S(w)C .
2 2
(3.17)
8π 2
The factor of ZM [δ] stands for the chiral matter partition function of flat space–time;
it is of course independent of the twist ε. This factor must be divided out since
itwas already
included in the definition used in [2] for the matter–ghost correlator
a b(pa ) α δ(β(qα ))M in flat space–time. Detailed definitions and explicit expressions
for the various ingredients in the above formula were given in [2]. Suffice it here to
remind the reader that the ghost insertion points pa and qα , with a = 1, 2, 3 and α = 1, 2
are arbitrary; that ΦI J and ΦB∗ are superholomorphic 3/2 forms and HA is a super
Beltrami differential, all of which are subject to certain normalization conditions, spelled
out respectively in Eqs. (3.18), (3.29) and (3.28) of [2]. Note that all quantities in (3.17)
are expressed with respect to the period matrix ΩI J .
The last step in the derivation of the consistent and slice-independent measure for the
Z2 -twisted theory is the change of variables from the bosonic period matrix ΩI J to the
super-period matrix Ω̂I J . The super-period matrix is invariant under local supersymmetry.
As was shown in [2], this guarantees the existence of consistent bosonic moduli and
permits the consistent integration over odd supermoduli. The reformulation of superstring
amplitudes in terms of the super-period matrix is one of the key insights into two-loop
superstring perturbation theory presented in [1–4], and was built on earlier work in [21,23,
37,38].
To carry out the change of variables from ΩI J to Ω̂I J , one proceeds as follows.
As explained in [2, Sections 3.3–3.5], the choice of Ω̂I J as parameters for the even
supermoduli determines the Beltrami superdifferentials HA in the gauge-fixed formulas
(1.3) and (3.12). The remaining difficulty is that the string amplitude is still expressed in
terms of correlators of conformal field theories with respect to a background metric with
period matrix ΩI J instead of Ω̂I J . Now the difference between the two period matrices is
of order two in χ and thus nilpotent. The expansion in terms of a Beltrami differential µ̂ is
thus exact to first order,

Ω̂I J = ΩI J − i d 2 z µ̂ωI ωJ (z). (3.18)
The process will therefore only affect the terms in (3.17) that are independent of χ . The
effect of this change of variables is a change in the worldsheet metric by means of the
Beltrami differential µ̂, defined above (up to a diffeomorphism). In any correlator, this
change may be implemented via the insertion of the stress tensor, as was explained in [2].
The combination of partition functions and finite-dimensional determinants may be treated
by these methods, and we obtain,

(Ω)
det ΦI J + (pa ) · detHα |Φβ∗ ZM [δ]
2
a b(pa ) α δ(β(qα ))M ZC [δ; ε] d z
= ( Ω̂) 1 + µ̂(z) T (z) .
det ΦI J + (pa ) · detHα |Φβ∗ ZM [δ] 2π C
Substituting this result into (3.17) yields

AC [δ; ε](pε ) = (Ω̂) (Ω̂) exp iπ τ̃ε pε2

det ΦI J + (pa ) · detHα |Φβ∗ ZM [δ]

1
× 1− d z d 2 w χz̄+ χw̄+ S(z)S(w) C
2
8π 2
2
d z
+ µ̂(z) T (z) C . (3.19)
2π
The ghost part of the partition function and of the supercurrent and stress tensor correlators
as well as the finite-dimensional determinants in (3.19) are exactly the same as those for
the uncompactified theory and given by

a b(pa ) α δ(β(qα ))M
AM [δ] =
det ΦI J + (pa ) · detHα |Φβ∗

1
× 1− 2
d 2
z d 2 w χz̄+ χw̄+ S(z)S(w) M
8π
2
d z
+ µ̂(z) T (z) M . (3.20)
2π
Here and above, the subscripts M and C are used on the correlators to indicate whether
they are evaluated in the flat Minkowski theory (M) or in the compactified orbifold theory
(C). The advantage of expressing AC [δ; ε](pε ) in terms of AM [δ] is that the latter has
already been explicitly evaluated in [4]. The result is,
ζ 1 ζ 2 Ξ6 [δ](Ω)ϑ[δ](0, Ω)4
AM [δ](Ω) = Z + · , (3.21)
16π 6 Ψ10 (Ω)
where the normalized partition function Z of chiral matter, ghosts and superghosts on the
bosonic surface with χ = 0 is given by

a b(pa ) α δ(β(qα ))M
Z≡
det ωI ωJ (pa )

ϑ[δ](0)5 ϑ(p1 + p2 + p3 − 3∆) a<b E(pa , pb ) a σ (pa )3
= 15 . (3.22)
Z ϑ[δ](q1 + q2 − 2∆)E(q1, q2 )σ (q1 )2 σ (q2 )2 det ωI ωJ (pa )
All quantities entering this expression were defined in Section 2, except for the chiral scalar
partition function Z −1 and the holomorphic 1-form σ (z), which are defined by
ϑ(z1 + z2 − w0 − ∆)E(z1 , z2 )σ (z1 )σ (z2 )

Z3 = ,
σ (w0 )E(z1 , w0 )E(z2 , w0 ) det ωI (zJ )
σ (z) ϑ(z − z1 − z2 + ∆)E(w, z1 )E(w, z2 )
= . (3.23)
σ (w) ϑ(w − z1 − z2 + ∆)E(z, z1 )E(z, z2 )
In each formula, the points z1 , z2 and w0 are arbitrary.
In terms of AM [δ], the following expression is obtained for the chiral measure,
ZC [δ; ε]
AC [δ, ε] pεµ = AM [δ] exp iπ τ̃ε pε2

ZM [δ]

1
× 1− d 2
z d 2 w χz̄+ χw̄+
8π 2

× Sm (z)Sm (w) C − Sm (z)Sm (w) M

1
+ d z µ̂(z) Tm (z) C − Tm (z) M .
2
(3.24)
2π
Here, we have used the fact that the ghost contributions in the correlators S(z)S(w)M and
S(z)S(w)C as well as in T (z)M and T (z)C are identical and cancel out upon taking
differences, leaving only the matter correlators in (3.24), evaluated in the sector twisted
by ε. It is understood that all parts of (3.24) are expressed with respect to the super-period
matrix Ω̂I J . The only ingredient in (3.24) which needs further clarification is the correction
to the Prym period τ̃ε , to be presented in the subsequent subsection.
3.4. The super-Prym period
In this subsection, the role of the quantity τ̃ε in the exponential involving the internal
loop momentum pε in (3.24) is clarified. By construction, τ̃ε is invariant under local
worldsheet supersymmetry. At first sight, this property would appear to qualify τ̃ε for the
supersymmetric generalization of the Prym period τε , but this hypothesis is invalid for the
following reasons.
Recall the Schottky relations on a bosonic Riemann surface with period matrix ΩI J ,
already presented in (1.6) and reformulated in (2.12). The solution of the Schottky relations
for τε as a function of Ω for a twist ε was denoted by the function τε = Rε (Ω) in (1.6).
Actually, this function will be multi-valued because, for given Ω, the Schottky relations
determine τ only up to a shift τε → τε + 4. This multivaluedness is required in particular
by the fact that the Dehn twist A1 B2 A−1 −1
1 B1 , which does not act on Ω, shifts τε → τε + 4,
as shown in [22].
The implications for the super-Prym period and for the quantity τ̃ε are as follows. The
genus 2 super-Riemann surface, specified by the supermoduli (ΩI J , ζ α ), uniquely projects
to a bosonic Riemann surface with period matrix Ω̂I J . This projection automatically
entails an associated super-Prym period τ̂ε , which is defined through the bosonic Schottky
relations from the super-period matrix Ω̂I J . In summary, we have the relations,

i
τ̃ε − τε = − d z d 2 w ωε (z)χz̄+ Sδ+
(z, w)χw̄+ ωε (w),
2
8π

i
Ω̂I J − ΩI J = − d 2 z d 2 w ωI (z)χz̄+ Sδ (z, w)χw̄+ ωJ (w) (3.25)
8π
as well as the defining relations,

τε = Rε (Ω), τ̂ε = Rε (Ω̂). (3.26)
By their very construction, Ω̂, τ̂ and τ̃ are supersymmetric invariant. As a result, the
difference
∆τε ≡ τ̃ε − τ̂ε = τ̃ε − τε − (Ω̂I J − ΩI J )∂I J Rε (Ω̂) (3.27)
is also a supersymmetric invariant. This invariance may be verified directly, using the above
definitions. Since this expression is bilinear in χ already, only the supersymmetry variation
of χ is required and this is given by δξ χz̄+ = −2∂z̄ ξ + , so that

δξ ∆τε = −i d 2 z ξ + (z)χz̄+ ωε (z)ωε (z) − ωI (z)ωJ (z)∂I J Rε

1
=− d 2 z ξ + (z)χz̄+ [δzz τ − δzz Rε ], (3.28)
2π
which cancels in view of τε = Rε (Ω).
The difference∆τε is non-vanishing. This may be shown by going to split gauge,
defined by χz̄+ = 2α=1 ζ α δ(z, qα ) with Sδ (q1 , q2 ) = 0, in which Ω̂ = Ω, τ̂ε = τε , so that
iζ 1ζ 2
∆τε = − ωε (q1 )Sδ+
(q1 , q2 )ωε (q2 ) (3.29)
4π
but this quantity is manifestly non-vanishing when ε = 0 and Sδ (q1 , q2 ) = 0.
4. The chiral measure for general compactifications
In this section, the chiral superstring measure will be constructed for more general com-
pactifications than those involving Z2 twists. The total space–time for the compactifica-
tions considered here will again be denoted by C. The chiral measure will be evaluated at
fixed even spin structure. As announced in [1], under some basic but mild assumptions, it
will be shown that the chiral blocks are independent of any choices of gauge slice, just as
they were in flat space–time. The assumptions are
(1) the compactification only modifies the matter part of the theory, leaving the
superghost part unchanged;
(2) the compactification respects local worldsheet supersymmetry, so that the super-
Virasoro algebra with matter central charge c = 15 is preserved.
A simple prescription for their calculation will be given first in split gauge and then in
terms of the OPE of two supercurrents.
4.1. The result of chiral splitting
Chiral splitting (the fact that the superstring amplitudes are the norms squared
of supermeromorphic functions on supermoduli space) holds for the contribution of
individual super-conformal families. In the case of the Z2 orbifold measure discussed in

Section 2, super-conformal families were labelled by the spin structure δ, by the twist ε
µ µ
and by the internal momenta, pI in the untwisted sector and pε in the twisted sectors.
For more general compactifications, the super-conformal family structure may be more
complicated. The super-conformal families will be labeled here by the spin structure δ and
the remaining characterization will be summarized by a label λ. The spin structure label
is singled out here because it must coincide with the spin structure of the ghost part of the
measure.
All the serious complications in the derivation of the superstring measure have to do
with the gauge fixing, ghost and finite-dimensional determinant contributions. In view of
our above assumptions, all these contributions are sensibly the same as in flat Minkowski
space–time or in the orbifolded space–times. Thus, the general form of chiral splitting for
the superstring measure for strings moving on the compactified space–time C is readily
adapted from the expression for the orbifold case in (3.12) at two-loops,9

AC [δ] = dmA 2 AC [δ; λ]2 ,
λ sM A
2
2
chi d z +
AC [δ; λ] = δ HA |B Vi exp χz̄ SC (z) . (4.1)
2π λ,+
A i
Here, the subscript λ refers to the fact that the amplitude is evaluated in the sector
associated with the super-conformal family λ. Notice that, compared to (3.12), no
additional factor depending on internal momenta is exhibited. The presence of internal
momenta amongst the labels for super-conformal blocks is indeed model dependent and is
assumed to be part of the definition of · · ·λ,+ .
In terms of AM [δ], the following expression is obtained for the chiral measure,

ZC [δ; λ] 1
AC [δ; λ] = AM [δ] 1− d 2 z µ̂(z) TCm (z) Cλ − TMm (z) M

ZM [δ] 2π

1
− d 2
z d 2 w χz̄+ χw̄+
8π 2

× SCm (z)SCm (w) Cλ − SMm (z)SMm (w) M . (4.2)
Here, we have used the fact that the ghost contributions in the correlators SM (z)SM (w)M
and SC (z)SC (w)Cλ as well as in TM (z)M and TC (z)Cλ are identical and cancel out
upon taking differences, leaving only the matter correlators in (4.2). It is understood that
all parts of (4.2) are expressed with respect to the super-period matrix Ω̂I J .
9 A subscript C has been appended to the supercurrent S because for general compactifications, the
supercurrent may not assume the flat space–time form; the latter will henceforth be denoted by SM .
4.2. Slice independence of the measure for compactifications
The slice independence of AC [δ; λ] may be deduced from the slice independence of
AM [δ], which was already established in [2,3], together with general properties of the
supercurrent and stress tensor correlators which enter into (4.2). Since the ghost parts
have cancelled out of the stress tensor and supercurrent correlators, their singularities with
the ghost insertion points have also cancelled. Therefore, both Tm (z)Cλ and Tm (z)M
are holomorphic and their difference is a holomorphic 2-form that is well-defined on the
Riemann surface, and thus the formula for AC [δ; λ] is independent of the choice of µ̂
within a given super-conformal family λ.
The supercurrent insertions are similarly independent of the points qα . Since the ghost
parts of SC and SM coincide, all the singularities in z and w with the insertion points
pa and qα are identical, and cancel upon taking the difference between the C and M
contributions. Thus, the only possible singularities in the SS correlator is when z → w.
But this singularity is precisely cancelled by the presence of the stress tensor contribution,
as was shown in the flat case in [2,3].
The mutual cancellation of these singularities is also a necessary and sufficient in order
to maintain local worldsheet supersymmetry, as shown in [2]. Indeed, each singularity
presents an obstruction to supersymmetry invariance since the supersymmetry variation
δξ χz̄+ = −2∂z̄ ξ + will pick up non-vanishing contributions at the poles. Just as in [2] for
flat space–time M, the effect of the singularity in the supercurrent correlator at z = w is
precisely cancelled by the variation δξ µ = ξ + χz̄+ of the stress tensor term. In summary,
the insertion of SCm (z)SCm (w)Cλ − SMm (z)SMm (w)M is completely singularity free
and hence AC [δ; λ] is slice independent, just as AM [δ] was.
4.3. The measure for compactifications in split gauge
To evaluate explicitly superstring measure, we now choose pointlike insertions for χ

χz̄+ (z) = ζ 1 δ(z, x1 ) + ζ 2 δ(z, x2 ). (4.3)
As in [3], the slice independence of AC [δ; λ] guarantees well-defined and regular limits as
xα → qα . We obtain

ZC [δ; λ] ζ 1 ζ 2
AC [δ; λ] = AM [δ] 1− SCm (q1 )SCm (q2 ) Cλ
ZM [δ] 4π 2

− SMm (q1 )SMm (q2 ) M

1
+ µa (q1 , q2 ) TCm (pa ) C (f ) − TMm (pa ) M , (4.4)
2π a
where the flat Minkowski space–time objects AM [δ] and Z were given in (3.21) and (3.22).
Here, µa arises from the Beltrami differential µ̂ representing the shift from the period
matrix Ω to the superperiod matrix Ω̂, given by
ζ 1ζ 2
µa (q1 , q2 ) = a (q1 , q2 )Sδ (q1 , q2 ). (4.5)
4π
Here, a (z, w) is the unique form of degree 1 and holomorphic in both z and w such
that the 3 holomorphic 2-forms a (z, z) are normalized by a (pb , pb ) = δab . Explicit
forms were given in [2, Eq. (1.15)]. The difference of the supercurrent and stress tensor
correlators on manifolds C and M, given by

Q(q1 , q2 ) ≡ + SCm (q1 )SCm (q2 ) Cλ − SMm (q1 )SMm (q2 ) M
1
− a (q1 , q2 )Sδ (q1 , q2 ) TCm (pa ) Cλ − TMm (pa ) M (4.6)
2 a
is holomorphic in both q1 and q2 and odd under the interchange of q1 and q2 . Therefore,
its dependence on qα is determined uniquely up to a qα -independent multiplicative factor.
The expression for AC [δ; λ] becomes,

ZC [δ; λ] ζ 1 ζ 2 ϑ[δ](0)4 Ξ6 [δ] ζ 1 ζ 2
AC [δ; λ] = Z+ · − ZQ(q 1 2 .
, q ) (4.7)
ZM [δ] 16π 6 Ψ10 4π 2
A further simplification takes place in the split gauge, defined by the following relation
between the insertion points q1 and q2 , Sδ (q1 , q2 ) = 0. In this gauge, the expression for
AC [δ; λ] simplifies to

ZC [δ; λ] ζ 1 ζ 2 ϑ[δ](0)4 Ξ6 [δ] ζ 1 ζ 2
AC [δ; λ] = Z+ · − Z SCm (q1 )SCm (q2 ) Cλ .
ZM [δ] 16π 6 Ψ10 4π 2
(4.8)
4.4. The measure via a leading supercurrent OPE operator
In this subsection, an alternative formula for the chiral measure is provided in terms
of simple data that may be obtained from the supercurrent OPE. This calculation may be
carried out in terms of the operators OM and OC , defined as follows,
1 TC (z) + TC (w)
SC (z)SC (w) = + (z − w)OC (w) + O(z − w)2 ,
4 z−w
1 TM (z) + TM (w)
SM (z)SM (w) = + (z − w)OM (w) + O(z − w)2 . (4.9)
4 z−w
Notice that the leading cubic singularity cancels since the central charges for M and for
C are assumed to be the same, namely c = 15. A convenient form for the chiral measure
based on the above OPE operators is obtained by letting all points pa collapse to the point
q2 and subsequently letting q1 → q2 .
Using methods similar to those employed in [3, Section 3.4], a limiting formula is
obtained for the summation against µa of the full stress tensor T (pa ) ≡ TCm (pa )C −
TMm (pa )M in terms of holomorphic Abelian differentials ωα∗ with normalization
ωα∗ (qβ ) = δαβ . In the OPE relation of (4.9), it is customary to expand with respect to the
coordinate q2 of the second operator, so we also let pa → q2 for all a = 1, 2, 3. As a result,
1 1 ∂ω2∗ (q2 )
lim a (q1 , q2 )T (pa ) = ∂T (q ) − T (q2 ). (4.10)
2 ∂ω1∗ (q2 ) ∂ω1∗ (q2 )
2
pa →q2
a
Expanding in powers of q2 − q1 up to and including second order, this limit reduces to

1 2 1
T (q2 ) + (q1 − q2 )∂T (q2 ) + (q1 − q2 ) f (q2 )∂T (q2 ) − 3T0 (q2 )T (q2 ) .
2 2
(4.11)
Here, we have introduced the following notations, familiar from [1,2]
f (w) = ωI (w)∂I ln ϑ(2w − w0 − ∆) − ∂w ln E(w, w0 ) + ∂w ln σ (w),
1 1 1
T0 (w) = ωI (w)ωJ (w)∂I ∂J ln ϑ(2w − w0 − ∆) − ∂f (w) + f (w)2 − T1 (w),
2 6 6
1
T1/2 (w) = ωI (w)ωJ (w)∂I ∂J ln ϑ[δ](0) − T1 (w). (4.12)
2
Collecting these results, the expansion up to order O(q1 − q2 )2 of Q is given by

Q(q1 , q2 ) = + SC (q1 )SC (q2 ) C − SM (q1 )SM (q2 ) M
1 T (q2 ) 1
− − ∂T (q2 )
2 q1 − q2 4

1 1
− (q1 − q2 ) (T1/2 − 3T0 )T + f ∂T (q2 ). (4.13)
2 2
It is easy to check that, within this approximation, Q(q1 , q2 ) is indeed a form of
weight (3/2, 3/2), even though individual terms in its expression above do not transform
covariantly under conformal reparametrizations z → z = ϕ(z). To check this, notice that
S(z) and T (z) are tensors of weights 3/2 and 2 respectively, while f , T1 , T1/2 and T0
transform as connections,
3 ϕ (z)
ϕ (z)f (z ) = f (z) − ,
2 ϕ (z)
1 ϕ (z) 1 ϕ (z)
ϕ (z)2 T0 (z ) = T0 (z) − f (z) + ,
3 ϕ (z) 6 ϕ (z)

2 6n2 − 6n + 1 ϕ (z) 3 ϕ (z) 2
ϕ (z) Tn (z ) = Tn (z) + − , n = 1/2, 1.
12 ϕ (z) 2 ϕ (z)
(4.14)
Assuming that the OPE of two supercurrents is as given in (4.9), we get
Q(q1 , q2 ) = (q1 − q2 )Q̂(q2 ) + O(q1 − q2 )2 ,

1 2 1 1
Q̂(w) = OC C − OM M − ∂ T − (T1/2 − 3T0 )T − f ∂T (w).
8 2 4
(4.15)
It remains to evaluate ZQ(q1 , q2 ). Since this quantity is independent of both q’s, we first
let q1 → q2 and then set q2 = p3 = ∆ + ν3 . The quantity may now be evaluated using the
methods developed for flat space–time, and we find
ϑ[δ](0, Ω)8 Q̂(p3 )
ZQ(q1 , q2 ) = · . (4.16)
M ν1 ν2
2 ων1 (p3 )2 ων2 (p3 )2
Substituting this result into the measure factor, we obtain the chiral measure

ZC ζ 1 ζ 2 ϑ[δ]4 Ξ6 [δ] ζ 1 ζ 2 ϑ[δ]8 Q̂(p3 )
AC = Z+ · − · · ,
ZM 16π 6 Ψ10 4π 2 M2ν1 ν2 ων1 (p3 )2 ων2 (p3 )2
(4.17)
as well as the contribution to the cosmological constant from the left-moving sector, by
integrating over ζ α .
5. Calculation of two-loop chiral blocks for Z2 twists
The starting point is the chiral measure of (3.24), evaluated in split gauge defined by
Sδ (q1 , q2 ) = 0. In this gauge, the following simplifications occur: Ω̂I J = ΩI J , and thus
µ̂ = 0 and the matter supercurrent correlator in M vanishes because the fermion propagator
Sδ is evaluated between the points q1 and q2 . The remaining expression is given by,10
ZC [δ; ε] iπ(τε +∆τε )pε2
AC [δ; ε](pε ) = e
ZM [δ]

ζ 1 ζ 2 Ξ6 [δ]ϑ[δ](0)4 ζ 1 ζ 2
× Z+ − Z Sm 1 m 2 ε ,
(q )S (q ) (5.1)
16π 6 Ψ10 4π 2
where Z was defined in (3.22). As this formula is written in split gauge, we have
Ω̂I J = ΩI J , τ̂ε = τε and the expression for ∆τε , derived in (3.29), is
iζ 1ζ 2
∆τε = − ωε (q1 )Sδ+
(q1 , q2 )ωε (q2 ). (5.2)
4π
The focus of this paper will be on the chiral measure and the cosmological constant, both
of which receive contributions only from the top term in ζ 1 ζ 2 . The resulting chiral measure
takes the following form,

dµC [δ; ε](pε ) ≡ dζ 2 dζ 1 AC [δ; ε](pε )

2 ZC [δ, ε] Ξ6 [δ]ϑ[δ](0)4 Z
= eiπτε pε − Sm (q1 )Sm (q2 ) ε
ZM [δ] 6
16π Ψ10 4π 2

+ iπpε2 Γ [δ, ε] , (5.3)
where the following definition has been made,

iZ
Γ [δ; ε] ≡ Z dζ 2 dζ 1 ∆τε = − ωε (q1 )Sδ+
(q1 , q2 )ωε (q2 ). (5.4)
4π
10 For Z -twisting, the supercurrents S and S take on the same functional form (as do the stress tensors T
2 C M C
and TM ); thus, the subscripts M and C will be dropped from the operators. The subscript C on the correlator will
be replaced with the twist ε for each twisted sector.
It remains to calculate the various terms in the above expression in terms of ϑ-functions,
which will be the subject of the remainder of this section.
It will be assumed that n dimensions are being Z2 -twisted, leaving 10 − n dimensions
untwisted, and that n is at most 8. In this case, only even spin structures δ need to be taken
into account. In particular, this will be the case for the models of [9]. The chiral partition
functions ZM and ZC are well-known, and given by

ϑ[δ](0, Ω) 5
ZM [δ] = ,
Z3
+ −
ϑ[δ](0, Ω) 5−n/2 ϑ[δj ](0, Ω)ϑ[δj ](0, Ω)ϑ[δ + ε](0, Ω) n/2
ZC [δ; ε] =
Z3 Z 3 ϑj (0, τε )2
(5.5)
for any pair δj± such that ε = −δj+ + δj− .
5.1. Calculation of the supercurrent correlator
The supercurrent correlator may be calculated in terms of the twisted scalar and fermion
propagators evaluated in Section 2,
n
Sm (q1 )Sm (q2 ) ε = Bε (q1 , q2 )Sδ+ε (q1 , q2 ), (5.6)
4
where the twisted scalar propagator is
Bε (z, w) = Sδ + (z, w)Sδ − (z, w) + bj ωε (z)ωε (w) (5.7)

j j
for any pair δj± such that ε = −δj+ + δj− .

For given ε = 0, non-vanishing contributions will arise only from [δ; ε] where both δ
and δ + ε are even characteristics. Upon choosing δ = δi+ (the choice δ = δi− leads to the
same result), we have
Bε (q1 , q2 ) = Sδ (q1 , q2 )Sδ+ε (q1 , q2 ) + bi ωε (q1 )ωε (q2 ). (5.8)
As q1 , q2 obey the split gauge relation Sδ (q1 , q2 ) = 0, the first term on the rhs above
vanishes and we have Bε (q1 , q2 ) = bi ωε (q1 )ωε (q2 ). Using the explicit expression for bi ,
computed in (2.27), we have

Sm (q1 )Sm (q2 ) ε = −iπnωε (q1 )ωε (q2 )Sδ − (q1 , q2 )∂τε ln ϑi (0, τε ). (5.9)
i
The combination of this correlator with the factor of Z may be re-expressed conveniently
in terms of Γ [δ; ε],

Z Sm (q1 )Sm (q2 ) ε = 4π 2 nΓ [δ; ε]∂τε ln ϑi (0, τε ). (5.10)
This leads to the following formula for the chiral measure in terms of Γ [δ; ε],
2 ZC [δ, ε]
dµC [δ; ε](pε ) = eiπτε pε
ZM [δ]

Ξ6 [δ]ϑ[δ](0)4
× + iπpε
2
− n∂τε ln ϑ i (0, τ ε ) Γ [δ; ε] .
16π 6Ψ10
(5.11)
It only remains to evaluate Γ [δ; ε].
5.2. Calculation of Γ [δ; ε]
In the present subsection, Γ [δ; ε] will be evaluated in terms of ϑ-constants. The

calculation will be carried out in split gauge, just as for the chiral measure in flat space–
time [4]. The key challenge presented by the calculation of Γ [δ; ε] is its overall sign. This
sign is uniquely fixed by the definition of Γ [δ; ε], but during the course of the evaluation,
a number of non-intrinsic signs appear and need to be determined. For example, the ϑ-
constant itself ϑ[κ](0) may change sign when a full period is added to κ.
Careful choices for the twist and spin structure assignments are needed, not just for their
expression congruent mod 1, but including full periods if they arise as well. The choices
made for this calculation are those given in (2.35) where
ε = −νa + νb , δ = δi+ = νa + νc + νd ,
δ + ε = δi− = νb + νc + νd . (5.12)
By choosing (abcdef ) to be an appropriate permutation of (123456), any twist and spin
structure assignment may be reached by these conventions.
The general expression for Γ [δ; ε] was given in (5.4); it involves the partition function
Z which was given in (3.22). Using the expression for Z in split gauge, calculated in
Eq. (3.15) of [4], we obtain
i
Γ [δ, ε] ≡ −Zωε (q1 )Sδ − (q1 , q2 )ωε (q2 ),
4π i
C ϑ[δ] E(pr , ps )4 σ (pr )2 σ (ps )2

5 1
Z =− 2 2 · · , (5.13)
Cr Cs ϑ[δ](q1 + q2 − 2∆)E(q1 , q2 )σ (q1 ) σ (q2 ) M2νr νs
2 2
where pr = νr + ∆ and ps = νs + ∆ are two arbitrary branch points. The exponential

factors were also introduced in [4] and are given by

C = − exp −4πiνl (2Ωνr + 2νr ) = − exp{−8πiνs Ωνr },

Cr,s = exp{−πiνr,s Ωνr,s − 2πiνr,s νr,s }. (5.14)
We shall make use of the following expression, derived in (2.49), for the product of the
Prym differentials at two different points z, w,

1
ωε (z)ωε (w) = −σ (µi , µj ) S (z, w)Sδ (z, w) − Sδ (z, w)Sδ (z, w) .
+ − + −
π 2 ϑk4 δi i j j
(5.15)
In view of the earlier choice δ = δi+ , and the split gauge condition Sδ (q1 , q2 ) = 0, the above
formula simplifies when z = q1 and w = q2 ,
1
ωε (q1 )ωε (q2 ) = σ (µi , µj ) S + (q1 , q2 )Sδ − (q1 , q2 ). (5.16)
π 2 ϑk4 δj j
Expressions in terms of ϑ-constants are most easily obtained by placing insertion points at
branch points. If q1 , q2 are in split gauge, then their limits to branch points are such that
q1 and q2 belong to the same set in the partition of all six branch points associated with δ.
The key difficulty in the evaluation of Γ [δ; ε] is that Z diverges as the points q1 and q2 ,
are taken to branch points, since ϑ[δ](q1 + q2 − 2∆) vanish. It turns out that one of the
three Szegö kernels arising in Γ [δ; ε] also vanishes, thereby making their ratio finite.
5.2.1. Linearization of split gauge around branch points

To circumvent the above problem, we parametrize the points q1 and q2 as follows (here
the argument is similar to [4, Section 3.6.2]),

q1 (t) = pc + t q̇1 + O t 2 ,

q2 (t) = pd + t q̇2 + O t 2 . (5.17)
The split gauge relation between q1 and q2 is clearly obeyed at the point t = 0, since
ϑ[δ](νc − νd ) ∼ ϑ[νa ] = 0. The vanishing factors in the numerator and denominator are
Sδ+
(pc , pd ) ∼ ϑ[δ + ε](νc − νd ) ∼ ϑ[νb ] = 0,
ϑ[δ](pc + pd − 2∆) = ϑ[δ](νc + νd ) ∼ ϑ[νa ] = 0. (5.18)
The split gauge condition to linear order in t yields a non-trivial condition on the growths
q̇1 and q̇2 , given by
q̇1 ωνa (pc ) − q̇2 ωνa (pd ) = 0. (5.19)

Here and below, the following notation is used for holomorphic Abelian differentials with
double zeros (at the branch points),
ων (z) ≡ ωI (z)∂I ϑ[ν](0). (5.20)

This notation was introduced in [4], Eq. (2.40). It was also shown there that ratios of these
differentials for different ν’s evaluated at the same branch point may be expressed in terms
of ϑ-constants, via the relation
ωνi (pk ) M νi νk
= , (5.21)
ωνj (pk ) Mνj νk
where Mνν was introduced in [4] and reproduced in (2.44). It was also shown in [4] that
Mνν has the following expression in terms of ϑ-constants for even spin structures,

M2ν1 ,ν2 = π 4 ϑ[ν1 + ν2 + νk ](0)2 , (5.22)
k=3,4,5,6
where νi , i = 1, . . . , 6 are all six distinct odd spin structures.

5.2.2. Evaluation at the branch points

The ratio of the two vanishing factors may be computed in the limit of vanishing t, with
the following result,
ϑ[δ + ε](q1 − q2 )
ϑ[δ](q1 + q2 − 2∆)
{q̇1ωI (pc ) − q̇2 ωI (pd )}∂I ϑ[δ + ε](νc − νd )
=
{q̇1 ωI (pc ) + q̇2 ωI (pd )}∂I ϑ[δ](νc + νd )
{ωνa (pd )ωI (pc ) − ωνa (pc )ωI (pd )}∂I ϑ[δ + ε](νc − νd )
= . (5.23)
{ωνa (pd )ωI (pc ) + ωνa (pc )ωI (pd )}∂I ϑ[δ](νc + νd )
The passage from the first to the second line in the above formula is made using the relation
between q̇1 and q̇2 of (5.20). Next, we have the following relations,
ωI (z)∂I ϑ[δ + ε](νc − νd ) = K1 ωνb (z),
ωI (z)∂I ϑ[δ](νc + νd ) = K2 ωνa (z), (5.24)
where the exponential factors K1 and K2 will be computed later. In terms of these
quantities, we have

ϑ[δ + ε](q1 − q2 ) K1 ωνb (pc ) ωνb (pd ) K1 Mab Mcd
= − =− . (5.25)
ϑ[δ](q1 + q2 − 2∆) 2K2 ωνa (pc ) ωνa (pd ) 2K2 Mac Mad
To pass to the last line, (5.21) has been used to express the ratio of ω’s in terms of Mνν ’s,
as well as the following algebraic relation, Mbc Mad − Mac Mbd = −Mab Mcd , which
readily follows from the definition of Mνν , given in (2.44).
5.2.3. Expression in terms of ϑ-constants

Assembling the expression in (5.13) and (5.16), Γ [δ; ε] takes the following form
i C ϑ[δ]5 E(pr , ps )4 σ (pr )2 σ (ps )2
Γ [δ; ε] = σ (µi , µj ) · ·
4π 3 ϑk4 Cr2 Cs2 ϑ[δ](q1 + q2 − 2∆)E(q1, q2 )σ (q1 )2 σ (q2 )2
1
× S + (q1 , q2 )Sδ − (q1 , q2 )Sδ − (q1 , q2 ). (5.26)
M2νr νs δj j i
Using (5.25), the limit q1 → pc = pr , q2 → pd = ps may be safely taken. Next, the

expression for the Szegö kernel in terms of ϑ-functions and the prime form is used. All
factors of σ and E(pc , pd ) now cancel and one obtains,
i σ (µi , µj ) CK1 Mab
Γ [δ; ε] =
8π 3 ϑk4 Cc Cd K2 Mac Mcd Mda
2 2
ϑ[δ](0)5 ϑ[δj+ ](νc − νd )ϑ[δj− ](νc − νd )

× . (5.27)
ϑ[δj+ ](0)ϑ[δj− ](0)ϑ[δi− ](0)
It remains to cast this expression in terms of standard ϑ-constants. To this end, we recast
some of the factors using the following relations between ϑ-constants.
ϑ[δj+ ](νc − νd ) = K3 ϑ[δk+ ](0),
ϑ[δj− ](νc − νd ) = K4 ϑ[δk− ](0). (5.28)
The exponential factors K3 and K4 will be evaluated later. The final result for Γ [δ; ε] is
now obtained as follows,
i σ (µi , µj ) ϑ[δ]2
Γ [δ, ε] = · κ · κ · , (5.29)
8π 7 ϑk4 ϑ[δj ]2 ϑ[δj− ]2 ϑ[δ + ε]2
+
where the following definitions have been made,

CK1 K3 K4
κ≡ ,
Cc2 Cd2 K2
π 4 Mab
κ ≡ · ϑ[δ]3 ϑ[δk+ ]ϑ[δk− ]ϑ[δj+ ]ϑ[δj− ]ϑ[δi− ]. (5.30)
Mac Mcd Mda
The combinations κ and κ will be evaluated in the subsequent subsection, resulting in
κ = ±1 and κ = ±1.
5.3. Calculation of the overall sign of Γ [δ; ε]
Neither κ, nor κ is intrinsic, but the product σ (µi , µj )κκ will be, as will be evidenced
by the fact that the final expression for Γ [δ; ε] is in terms of squares of ϑ-functions only.
5.3.1. Calculating κ
Let us summarize the definitions of the factors entering into κ,
C = −exp{−8πiνc Ωνd },
Cc = exp{−πiνc Ωνc − 2πiνc νc },
Cd = exp{−πiνd Ωνd − 2πiνd νd },
K1 ωνb (z) = ωI (z)∂I ϑ[δ + ε](νc − νd ),
K2 ωνa (z) = ωI (z)∂I ϑ[δ](νc + νd ),
K3 ϑ[δk+ ](0) = ϑ[δj+ ](νc − νd ),
K4 ϑ[δk− ](0) = ϑ[δj− ](νc − νd ). (5.31)
The K-factors may be computed starting from the following basic formula of (B.9),

ϑ[δ](z + Ωρ + ρ ) = ϑ[δ + ρ](z) exp −iπρ Ωρ − 2πiρ (z + δ + ρ ) . (5.32)

One finds,

K1 = −exp −iπ(νc − νd ) Ω(νc − νd ) − 2πi(νc − νd ) νb + 4πiνd νc + 4πiνb νc ,

K2 = −exp −iπ(νc + νd ) Ω(νc + νd ) + 2πi(νc + νd ) νa ,

K3 = exp −iπ(νc − νd ) Ω(νc − νd ) + 2πi(νc − νd ) (νb + νd + νf ) ,

K4 = exp −iπ(νc − νd ) Ω(νc − νd ) + 2πi(νc − νd ) (νa + νd + νf ) . (5.33)

Notice that under c ↔ d, K1 → −K1 while K2 → +K2 , even though these symmetries
are not manifest. Multiplying all factors yields
CK1 K3 K4
κ= = exp 4πi(νd νc + νc νd − νd νa + νc νf − νd νf + νb νc ) (5.34)
Cc2 Cd2 K2
which indeed takes the values ±1.
5.3.2. Calculating κ
A direct calculation of κ from its definition
π 4 Mab
κ ≡ ϑ[δ]2 ϑ[δi+ ]ϑ[δi− ]ϑ[δj+ ]ϑ[δj− ]ϑ[δk+ ]ϑ[δk− ] (5.35)
Mac Mcd Mda
involves the non-intrinsic sign factors, and must be computed case by case. To simplify
the process, we shall carry out this calculation for a single non-trivial twist and obtain
the expression for Γ [δ; ε] for all twists by modular invariance. This choice uniquely
determines νa and νb (up to interchange of a and b). The choice made here is the standard
one,

0 0
ε= , νa = ν2 , νb = ν4 . (5.36)
0 12
The even spin structures are fixed through the choice of the odd spin structures and (2.35).
The calculation of κ is started with the evaluation of the product of all 6 even
spin structures, ϑ[δ2+ ]ϑ[δ2− ]ϑ[δ3+ ]ϑ[δ3− ]ϑ[δ4+ ]ϑ[δ4− ]. This object is independent of the
remaining choices of c, d, e, f , since a permutation of these objects simply permutes the
various factors in the product. To evaluate it, we choose (c, d, e, f ) = (1, 3, 5, 6). The
even spin structures are then determined by (B.7) (see also (2.11)), and the ϑ-constants are
expressed in normalized form by
ϑ[δ2+ ] = ϑ[δ7 + 2νd ] = −ϑ[δ7 ],

ϑ[δ2− ] = ϑ[δ8 + 2νd ] = −ϑ[δ8 ],
ϑ[δ3+ ] = ϑ[δ1 + 2δ0 ] = +ϑ[δ1 ],
ϑ[δ3− ] = ϑ[δ2 + 2δ0 ] = +ϑ[δ2 ],
ϑ[δ4+ ] = ϑ[δ3 + 2νe ] = +ϑ[δ3 ],
ϑ[δ4− ] = ϑ[δ4 + 2νe ] = +ϑ[δ4 ]. (5.37)
Therefore, given the choice for ε = −ν2 + ν4 , the following product is the same for any
choices of c, d, e, f = 2, 4, and we have
ϑ[δ2+ ]ϑ[δ2− ]ϑ[δ3+ ]ϑ[δ3− ]ϑ[δ4+ ]ϑ[δ4− ] = +ϑ[δ1 ]ϑ[δ2 ]ϑ[δ3 ]ϑ[δ4 ]ϑ[δ7 ]ϑ[δ8 ]. (5.38)
Next, the expressions for Mab , a, b = 1, . . . , 6 are needed. As was shown in [4], these
objects may be expressed as products of ϑ-constants, times a non-intrinsic sign factor.
Normalizing the ϑ-constants on the even spin structures in canonical form, as given by
Table 1
Calculation of κ
c d δ m2c mcd md2 κ
1 3 δ7 + − − −
1 5 δ3 + + − +
1 6 δ2 + − − −
3 5 δ1 + + − +
3 6 δ4 + − − −
5 6 δ8 + − − −
(B.5), the sign factors mab are tabulated in (B.24). The quantities needed are given by
M24 = −π 2 ϑ[δ0 ]ϑ[δ5 ]ϑ[δ6 ]ϑ[δ9 ],

9
M2c Mcd Md2 = +m2c mcd md2 π ϑ[δ] 6 2
ϑ[δi ]. (5.39)
i=0
Combining all of the above, we have
κ = −m2c mcd md2 . (5.40)

Table 1 below summarizes the results of the case by case calculation of κ .
5.3.3. Calculating the sign of Γ [δ; ε]

A general expression for κ was obtained in (5.34). For our present purposes, with
a = 2, b = 4, this expression becomes,

κ = νc |νd exp 4πi −νd ν2 + (νc + νd )νf + ν4 νc . (5.41)
Notice that κ depends on f as well as on c and d. Thus, κ does not just depend on ε and
δ (i.e., only on νa , νb and νc + νd ), but also on the choice of νf versus νe . Furthermore, κ
is not, in general, symmetric under the interchange of c and d; instead, we have
κ(c, d)
= −νc |νd (5.42)
κ(d, c)
for fixed a, b, e, f . Fortunately, these non-intrinsic dependences of κ are being compen-
sated in the expression for Γ [δ; ε] by the presence of another sign factor σ (µi , µj ). Recall
that for given i, there can be two choices for j . These choices are actually correlated with
the choices for f in κ.
Clearly, f -dependence enters only when (νc + νd )νf = 0, which is the case (here with
a = 2 and b = 4) when c = 1, d = 6 and c = 3, d = 5. To identify the correlation between
the choice of f in κ and that of j in σ (µi , µj ), we need consider only these cases.
• c = 1, d = 6 implies
+ +
δi = δ2 , δj = ν1 + ν4 + νf ,
−
δi = δ1 , δj− = ν1 + ν2 + νf .
Table 2
Calculation of the sign of Γ [δ; ε]
c d f δ µi µj µk κ κ σ (µi , µj ) Sign
1 3 5 δ7 µ2 µ4 µ3 + − + −
6 µ3 µ4 + − +
1 5 3 δ3 µ4 µ2 µ3 + + − −
6 µ3 µ2 + + −
1 6 3 δ2 µ3 µ2 µ4 − − − −
5 µ4 µ2 + − +
3 5 1 δ1 µ3 µ2 µ4 − + − +
6 µ4 µ2 + + +
3 6 1 δ4 µ4 µ2 µ3 + − − +
5 µ3 µ2 + − −
5 6 1 δ8 µ2 µ4 µ3 − − + +
3 µ3 µ4 − − +
For f = 3, we have δj+ = δ8 and therefore j = 2. For f = 5, we have δj+ = δ4 and

therefore j = 4.
• c = 3, d = 5 implies

δi+ = δ1 , δj+ = ν3 + ν4 + νf ,
δi− = δ2 , δj− = ν2 + ν3 + νf .
For f = 1, we have δj+ = δ8 and therefore j = 2. For f = 6, we have δj+ = δ3 and

therefore j = 4.
The corresponding results are summarized in Table 2.

Assembling all the results with the basic expression for Γ [δ; ε], we obtain,
i ν0 |µi ϑ[δ]2

Γ δiσ , ε = σ 7 . (5.43)
8π ϑk ϑ[δj ] ϑ[δj− ]2 ϑ[δ + ε]2
4 + 2
Here, (i, j, k) is a permutation of the genus 1 spin structure indices (2, 3, 4), and δ = δiσ
with σ = ±. A more illuminating expression, whose form is more manifestly covariant
is obtained by factoring out a combination of the Z qu [ε] partition function for 4 twisted
directions. Using the familiar genus 1 relation ϑi ϑj ϑk = 2η3 , this leads to the following
final expression,

i ϑi4 ϑ[δiσ ]2 ϑj4
Γ δiσ , ε = σ ν0 |µi (5.44)
(2π)7 η12 ϑ[δ (−σ ) ]2
i
ϑ[δj+ ]2 ϑ[δj− ]2
which is independent of j . Here, σ = ± according to which spin structure δi± is being

evaluated.
6. Modular transformations and Z2 twisting
The structure of the genus 2 modular group Sp(4, Z) and its action on spin structures,
twists and ϑ-functions is summarized in Appendix B.
6.1. Subgroup preserving a given twist
The subgroup of the modular group Sp(4, Z) which acts on spin structures or on half
characteristics is Sp(4, Z2 ). It is isomorphic to the permutation group of the six branch
points and therefore has 6! = 720 elements. To determine the subgroup Hε of Sp(4, Z2 )
that leaves a given twist ε invariant, we proceed as follows. First, make a definite choice
for a twist, so that the invariance equation becomes

A B 0 0 ε D −C ε
M= , ε= , = . (6.1)
C D 0 12 ε −B A ε
This equation puts the following restrictions on the entries A12 = C12 = C22 = 0 and
A22 = 1 of the matrices. All elements of the group Hε are found by solving the symplectic
relation MJ M T = J under these restrictions. The independent generators for Hε are
generators of Hε = {M1 , M2 , M3 , T2 = ΣT Σ, S2 = SM1 SM1 }. (6.2)
The full group may be parametrized in terms of the Abelian subgroup Mε of “translations”,
Hε0 = {I, M1 , M2 , M3 , M1 M2 , M1 M3 , M2 M3 , M1 M2 M3 },

Hε = I, T2 , S2 , S2 T2 , S22 , S22 M1 T2 × Hε0 . (6.3)

The matrices Mi ’s are given explicitly in Appendix B.3. In total there are 48 elements. This
number is as expected, since all elements of the group Sp(4, Z2 ) may be decomposed as the
product of elements that move the twists from any given twist to any other twist times an
element that preserves the twist. There are 15 non-trivial twists and hence 720 = 15 × 48.
The transformation laws of the even spin structures under the generators of Hε are listed
in Table 3.
Table 3
The modular group Hε acting on even spin structures
δ M1 M2 M3 T2 S2
2 (M1 )
2 (M2 )
2 (M3 )
2 (T2 )
2 (S2 )
δ1 δ3 δ2 δ1 δ1 δ3 1 1 1 1 i
δ2 δ4 δ1 δ2 δ2 δ4 1 1 1 1 i
δ3 δ1 δ4 δ3 δ4 δ7 1 1 1 1 1
δ4 δ2 δ3 δ4 δ3 δ8 1 1 1 1 1
δ5 δ6 δ5 δ6 δ9 δ6 1 i 1 1 i
δ6 δ5 δ6 δ5 δ0 δ9 1 i 1 1 1
δ7 δ7 δ8 δ8 δ7 δ1 i 1 1 1 i
δ8 δ8 δ7 δ7 δ8 δ2 i 1 1 1 i
δ9 δ9 δ9 δ0 δ5 δ5 i i −1 1 i
δ0 δ0 δ0 δ9 δ6 δ0 i i −1 1 −1
6.2. Transformations of the Schottky relations
In (2.46), a Schottky relation is derived, but a sharper form (namely the precise sign
involved in taking the square root of this relation) is needed in the present study, as
indicated in (2.12). The square root may be fixed by checking the consistency of (2.12)
with modular transformations and with degenerations, both of which are carried out in this
subsection.
The crucial ingredients in this calculation are the factors
(δ, M)2 when M ∈ Hε , since
they will determine the transformation rules for the genus 2 ϑ-constants. To compute them,
we make use of the values for
(δ, Mi )2 , i = 1, 2, 3,
(δ, S)2 ,
(δ, T )2 , and
(δ, Σ)2 , given
in (B.16), as well as the following cocycle rule, which may be derived from the definition
of
(δ, M) in terms of the ϑ-constants,

(κ, MM ) =
(κ, M ) ×
(M κ, M). (6.4)
The results are summarized in Table 3, and may be readily applied to the calculation of the
pairwise products
(δj , M)2
(δj +1 , M)2 . The results are given in Table 4.
For the choice of twist made here, (2.12) becomes,
ϑ44 ϑ 2 [δ3 ]ϑ 2 [δ4 ] ϑ24 ϑ 2 [δ7 ]ϑ 2 [δ8 ] ϑ34 ϑ 2 [δ1 ]ϑ 2 [δ2 ]

= , = , = . (6.5)
ϑ24 ϑ 2 [δ7 ]ϑ 2 [δ8 ] ϑ34 ϑ 2 [δ1 ]ϑ 2 [δ2 ] ϑ44 ϑ 2 [δ3 ]ϑ 2 [δ4 ]
By inspecting Table 3, it is manifest that the transformations M2 , M3 and T2 do not act on

these pairs at all. It remains to consider only the actions of M1 and S2 ,

ϑ34 ϑ 2 [δ1 ]ϑ 2 [δ2 ]
ϑ 2 [δ3 ]ϑ 2 [δ4 ] ϑ4
M1 = M1 =+ 2 = + 44 ,
ϑ44 ϑ 2 [δ3 ]ϑ 2 [δ4 ]
ϑ [δ1 ]ϑ [δ2 ]
2 ϑ3
4 2
ϑ ϑ [δ1 ]ϑ 2 [δ2 ] ϑ 2 [δ3 ]ϑ 2 [δ4 ] ϑ44
M1 34 = M1 = − = − ,
ϑ2 ϑ 2 [δ7 ]ϑ 2 [δ8 ] ϑ 2 [δ7 ]ϑ 2 [δ8 ] ϑ24
4 2
ϑ ϑ [δ1 ]ϑ 2 [δ2 ] ϑ 2 [δ3 ]ϑ 2 [δ4 ] ϑ44
S2 34 = S2 = − = − ,
ϑ4 ϑ 2 [δ3 ]ϑ 2 [δ4 ] ϑ 2 [δ7 ]ϑ 2 [δ8 ] ϑ24
4 2
ϑ3 ϑ [δ1 ]ϑ 2 [δ2 ] ϑ 2 [δ3 ]ϑ 2 [δ4 ] ϑ44
S2 = S2 = + = + . (6.6)
ϑ24 ϑ 2 [δ7 ]ϑ 2 [δ8 ] ϑ 2 [δ1 ]ϑ 2 [δ2 ] ϑ34
Table 4
The modular group Hε acting on pairs δ ± of
even spin structures
j M1 M2 M3 T S2
1 1 1 1 1 −1
3 1 1 1 1 1
7 −1 1 1 1 −1
The transformation properties of ϑ-constants for genus 1, under the two canonical
generators, denoted here by T (1) and S (1) are given by

T (1) : τ → τ + 1,
(6.7)
S (1) : τ → −1/τ.
We have

 ϑ2 (τ + 1) = eiπ/4 ϑ2 (τ ),
T (1) ϑ3 (τ + 1) = ϑ4 (τ ),

ϑ4 (τ + 1) = ϑ3 (τ ),
 √
 ϑ2 (−1/τ ) = √−iτ ϑ4 (τ ),
S (1) ϑ3 (−1/τ ) = √−iτ ϑ3 (τ ), (6.8)

ϑ4 (−1/τ ) = −iτ ϑ2 (τ ).
Therefore, it is clear that, signs and all, we have
M1 −→ T (1) ,
S2 −→ T (1) S (1) . (6.9)
Therefore, the genus 2 modular transformations M1 and S2 indeed induce ordinary genus
1 modular transformations on the Prym period. As a result, all the modular transformations
in Hε induce modular transformations on τε , which belong to the genus 1 modular group,
according to the above correspondence.
6.3. Modular orbits of the twists under Hε
First, a parametrization of the twists is needed,

0 0 0 0 0 1 0 1
2ε1 = , 2ε2 = , 2ε3 = , 2ε4 = ,
0 0 0 1 00 01

0 0 0 1 1 0 1 0
2ε5 = , 2ε6 = , 2ε7 = , 2ε8 = ,
1 0 1 0 00 01

1 0 1 1 0 0 1 1
2ε9 = , 2ε10 = , 2ε11 = , 2ε12 = ,
1 0 1 1 11 00

0 1 1 1 1 0 1 1
2ε13 = , 2ε14 = , 2ε15 = , 2ε16 = .
1 1 0 1 11 10
(6.10)
Notice that the standard twist adopted previously is given by ε = ε2 , and that ε1 is no
twist at all. The modular transformations under the subgroup Hε are then easily computed
from the transformation formula for the twists.
Simple inspection of Table 5 reveals 4 orbits,
O0 [ε] = {ε1 },
Oε [ε] = {ε2 },
O+ [ε] = {ε3 , ε4 , ε7 , ε8 , ε12 , ε14 },
O− [ε] = {ε5 , ε6 , ε9 , ε10, ε11 , ε13 , ε15 , ε16 }. (6.11)
Table 5
The modular group Hε acting on twists
M ε1 ε2 ε3 ε4 ε5 ε6 ε7 ε8 ε9 ε10 ε11 ε12 ε13 ε14 ε15 ε16
M1 ε1 ε2 ε3 ε4 ε5 ε6 ε12 ε14 ε16 ε15 ε11 ε7 ε13 ε8 ε10 ε9
T2 ε1 ε2 ε4 ε3 ε9 ε10 ε7 ε8 ε5 ε6 ε15 ε14 ε16 ε12 ε11 ε13
S2 ε1 ε2 ε12 ε14 ε5 ε16 ε3 ε4 ε6 ε15 ε11 ε7 ε10 ε8 ε13 ε9
The non-trivial orbits O± [ε] may be characterized in a simple modular covariant manner,
υ ∈ O± [ε] ⇔ ε|υ = ±1 and υ = ε1 , ε2 . (6.12)
6.4. Subgroups leaving two twists invariant
The first twist may be denoted ε and chosen as in (6.1). The non-trivial cases arise when
the second twist α belongs to either orbit O± [ε]. In each orbit, any representative may be
taken for the second twist α. The stabilizer groups of two twists ε and α will be denoted
by Hε,α ; their generators may be deduced from inspection of Table 5, and given by
Hε,ε3 ∼ generators {M1 , M2 , M3 },
Hε,ε5 ∼ generators {M1 , S2 , (M3 S2 T2 )}. (6.13)
The group Hε,ε3 is Abelian, but the group Hε,ε5 is non-Abelian.
7. Asymptotic behavior of the chiral measure
Given the choice of homology basis adopted throughout this paper, the period matrix
may be parametrized as follows,

τ τ
Ω= 1 , (7.1)
τ τ2
where τ1,2 , τ ∈ C, subject to the constraint Im Ω > 0. Degenerations fall into two
classes according to whether the degeneration separates the surface into two disconnected
components or leaves the surface connected. Separating (respectively non-separating)
degenerations result from the shrinking of a homologically trivial (respectively non-trivial)
1-cycle. The limit τ → 0 is separating, while the limits τ1 → +i∞ or τ2 → +i∞ are
non-separating.
The flat space–time chiral measure depends on the spin structure and the limiting
behavior of the measure therefore depends upon the inter-relation between the spin
structure the homology of the degenerating cycle. These degenerations were worked out
explicitly—to leading order—in [4, Section 8].
The Z2 -twisted chiral measure (1.10) depends on the spin structure δ, AND on the
twist ε. Its limiting behavior will therefore depend upon the inter-relations between not
only the spin structure and the homology of the degenerating cycle, but also upon the
twist. Here, we shall derive the limit of only two representative cases, one separating, the
other non-separating; the other cases are similar, but their number is simply too large to
discuss usefully here. Finally, only the case where n = 4 dimensions are Z2 -twisted will
be discussed, since this will be the case of greatest physical interest, as discussed in the
next subsection.
7.1. Separating degeneration
This limit corresponds to letting τ → 0, while keeping τ1,2 fixed. The leading behavior
was given in [4]; the leading and subleading behaviors are as follows,

µi
ϑ (0, Ω) = ϑi (0, τ1 )ϑj (0, τ2 ) 1 + 2τ 2 ∂ ln ϑi (0, τ1 )∂ ln ϑj (0, τ2 ) ,

µj

µ
Ξ6 i (Ω) = 28 µi |ν0 µj |ν0 η(τ1 )12 η(τ2 )12
µj

τ2
× −1 + 3∂ ln ϑi4 (0, τ1 )∂ ln ϑj4 (0, τ2 )
2

− ∂ ln η(τ1 ) ∂ ln ϑj (0, τ2 ) − ∂ ln η(τ2 ) ∂ ln ϑi (0, τ1 ) ,
12 4 12 4
Ψ10 (Ω) = 212 (2πτ )2 η(τ1 )24 η(τ2 )24 1 + 48τ 2 ∂ ln η(τ1 )∂ ln η(τ2 ) . (7.2)
Here, µi and µj represent the even spin structures on each genus 1 component; the even
spin structure which restricts to the odd spin structure on each genus 1 component will not
be needed here.
First, the limiting behavior is needed for the Prym period, since it enters into the form
of the chiral measure. Using the Schottky relation (1.6) as well as the above limiting
behaviors, we find

τε = τ1 + τ 2 ∂ ln ϑ3 (0, τ2 )ϑ4 (0, τ2 ) + O τ 4 (7.3)
up to shifts τε → τε + 4. Second, we need the bosonic twisted partition function factor
ϑ[δj+ ](0, Ω)2ϑ[δj− ](0, Ω)2

= ϑ3 (0, τ2 )2 ϑ4 (0, τ2 )2 + O τ 4 (7.4)
ϑj (0, τε )4
which is clearly independent of j to this order. Third, the limit of the terms involving
Γ [δ; ε] is needed. When n = 4, as is being assumed here, it is advantageous to consider
directly the quantity
ZC [δiσ , ε] σ iσ ϑi (0, τε )4
Γ δi , ε = ν0 |µi , i = 2, 3, 4, σ = ±. (7.5)
ZM [δi ]
σ (2π)7 η(τε )12
In view of the relation between τ1 and τε , the limit as τ → 0 of this quantity is regular and
is obtained simply from the above formula by replacing τε by τ1 . Putting all together, we
find
2 ν0 |µi ϑi (0, τ1 )
4
dµC δiσ , ε pεµ = eiπτε pε 6 8
2 π η(τ1 )12

−σ 1 π2 2
× 2 − 2πi σ + ∂ ln ϑ i (0, τ 1 ) − σ p .
τ ϑ2 (0, τ2 )4 2 2 ε
(7.6)
The physical analysis of this limit is as follows.
The 1/τ 2 singularity, which is familiar from the flat space–time chiral measure,
produces tachyon and massless scalar singularities in the channel connecting both genus 1
components. This is as expected, since the six spin structures δiσ restrict to even spin
structures on each genus 1 component. The summation over σ corresponds to the GSO
projection imposed on states in the NS sector which are traversing the A2 cycle. This part
of the GSO projection eliminates the corresponding tachyon intermediate state, and indeed
the partially summed measure is regular, as may be seen from

dµC [δi+ , ε] pεµ + dµC [δi− , ε] pεµ
2 ν0 |µi ϑi (0, τ1 )4
= −ieiπτε pε ∂ ln ϑi (0, τ1 ). (7.7)
25 π 7 η(τ1 )12
Notice also that a partial GSO resummation over the spin structures of the first genus 1
component vanishes to this order,

dµC δiσ , ε pεµ = 0 (7.8)
i=2,3,4
in view of the Riemann identities for genus 1,

ν0 |µi ϑi (0, τ1 )4 = ν0 |µi ϑi (0, τ1 )4 ∂ ln ϑi (0, τ1 ) = 0. (7.9)
i=2,3,4 i=2,3,4
Again, this is as expected in view of space–time supersymmetry.
7.2. Non-separating degenerations
As mentioned in the opening paragraph to this subsection, many cases need to be

distinguished based on the inter-relation of the spin structure, the twist and the shrinking
cycle. The case considered here corresponds to letting the cycle B2 grow to infinite length;
more specifically, letting τ2 → +i∞ while keeping τ and τ1 fixed. As usual, we introduce
the variable q ≡ exp{iπτ2 }, in terms of which the limit is given by q → 0. We recall from
[4] the following limits,

µi
ϑ (0, Ω) = ϑi (0, τ1 ) + O(q), µj = (00), 0 12 . (7.10)
µj
As a result, the Schottky relation giving τε in terms of the period matrix Ω + I J simply
yield τε = τ1 up to shifts in 4Z. The ratio ZC /ZM → 1 in this limit. The following
combination,
ZC [δiσ , ε] σ iσ ϑi (0, τ1 )4
Γ δ i , ε → ν0 |µ i (7.11)
ZM [δiσ ] (2π)7 η(τ1 )12
has a smooth limit. Putting all together, we obtain the following limits for the full measure,
2 σ ν0 |µi ϑi (τ/2, τ1 ) + ϑ1 (τ/2, τ1 )
4 4
dµC δiσ ; ε (pε ) = eiπτε pε + O(1). (7.12)
q 28 π 6 η(τ1 )6 ϑ1 (τ, τ1 )2
This result coincides with the untwisted case, as expected.
8. Applications to physical theories
In this section, specific physical superstring theories are considered that are constructed
from flat Minkowski space–time orbifolded by groups acting by reflections and shift.
When the action of the orbifold group is the same (respectively different) on left and
right movers, the orbifold is referred to as symmetric (respectively asymmetric). The
models of [9], for example, are asymmetric orbifolds of type II superstring theory. For
definiteness, the present study will concentrate on type II theories with six or fewer
compactified dimensions, but the methods may be extended to heterotic orbifolds and/or
more compactified dimensions.
The construction of symmetric orbifolds may be carried out directly from the functional
integral by inserting projection operators and including twisted sectors; for completeness,
it will be briefly reviewed below. The construction of asymmetric orbifolds, on the other
hand cannot, in general, be carried out directly from the functional integral. Therefore,
properties that are known to hold generally for symmetric orbifolds (such as their behavior
under modular transformations) may or may not hold for asymmetric orbifolds. Various
methods have been proposed to circumvent these obstacles. The method of [40] uses a
doubling of the left- and right-moving degrees of freedom, while the results of [41] rather
suggest the use of operator methods. (For reviews and further references, see, e.g., [42].)
The approach taken here to the construction of asymmetric orbifold superstring theories
will be based on chiral splitting. The starting point will be the construction of the chiral
blocks, carried out separately for the left-movers and for the right-movers. Once the chiral
blocks are in hand, string amplitudes are obtained by assembling these left and right chiral
blocks in a manner consistent with the definition of the asymmetric orbifold model as well
as with modular invariance. This approach will be discussed in enough detail here so that
the cosmological constant in the KKS models can be investigated. Further study will appear
in a forthcoming publication [24].
8.1. Chiral splitting of symmetric orbifolds
Let G be an orbifold group acting on Rn , with n 6, i.e., with at least four

uncompactified space–time dimensions. An element g ∈ G is an Euclidean transformation
on Rn , consisting of a rotation Rg ∈ O(n) and a shift vg ∈ Rn , so that g = (Rg , vg ),
possibly supplemented by an action on internal quantum numbers. The action on the fields
is given by gx = Rg x + vg and gψ± = Rg ψ± , and thus symmetric on left- and right-

movers. Clearly, the non-trivial action of G on the fermion fields and on the momentum
and oscillator modes of the bosonic fields is only by the subgroup of rotations (Rg , 0)
(obtained by simply omitting the shift vg for each g ∈ G), which form the point group PG .
The subgroup of elements of the form (I, v) on the other hand only acts on the zero modes
of x and forms a lattice ΛG ; the coset TG = Rn /ΛG is a torus. The symmetric orbifold
may be constructed in two equivalent ways;
Rn /G = TG /P̄G , (8.1)
i.e., as the coset of Rn by the full orbifold group G, or as the coset of the torus TG by the
group P̄G = G/ΛG , which is isomorphic to PG . The construction via TG /P̄G is generally
more convenient (because the groups P̄G are usually finite, while G is always infinite)
and is better suited for dealing with asymmetric orbifolds; therefore it will be adopted
throughout. The orbifold is Abelian (respectively non-Abelian) if the point group PG ∼ P̄G
is Abelian (respectively non-Abelian).
The orbifold string theory is constructed as follows. Let AI , BI , I = 1, 2 be a canonical
µ
homology basis and define the fermion fields ψ± with a spin structure δ. To the basis cycles
(AI , BI )we associate a sector (aI , bI ) of elements of P̄G which satisfies the homotopy
relation I =1,2 aI bI aI−1 bI−1 = I . (For Abelian orbifolds, this condition is automatically
fulfilled.) The sector is untwisted if aI and bI equal the identity in P̄G , and is twisted
otherwise. In a given sector λ = (aI , bI ), the fields obey the following monodromy
conditions,11

(x, ψ± )(z + AI ) = aI x(z), (−1)2δI aI ψ± (z) ,

(x, ψ± )(z + BI ) = bI x(z), (−1)2δI bI ψ± (z) . (8.2)
The string amplitude in the sector λ is given by the familiar functional integral but where
the fields now obey the monodromy conditions (8.2),

AC [δ; λ] = DEM DΩM δ(T )
A
DXµ e−Im . (8.3)
(aI ,bI )
The same procedure of gauge-fixing and chiral splitting as was carried out in (3.12) may
be applied here, and results in a set of chiral blocks.
The chiral blocks are indexed by the sector labels λ = (aI , bI ) and by the internal mo-
menta associated respectively with uncompactified (pu ) and compactified (pλ ) dimensions.
The effects due to the uncompactified internal momenta pu are the same as in flat space–
time and yield the familiar measure factor (det Im Ω)−5+n/2 ; henceforth, the dependence
on pu will suppressed. The internal momenta pλ associated with the compactified direc-
tions are discrete and will take values in the lattice Λ∗G dual to ΛG . The chiral blocks
will be denoted by dµC [δ; λ](pλ ) = dµC [δ; aI , bI ](pλ ) and the amplitude AC take on the
11 In the fermion monodromy, an extra sign may arise depending on the reference spin structure chosen in the
given homology basis. For simplicity, this factor has been omitted here.
form,

AC [δ; λ] = (det Im Ω)−5+n/2 dµC [δ; λ](pλ)2 . (8.4)
M2 pλ ∈Λ∗G
The physical problem with this amplitude is the presence of the tachyon, which in
particular causes the integration over moduli space to diverge.
In the GSO projection, left and right fermions are treated independently (via chiral
splitting, see [20,21]), and assigned spin structures δL and δR which are summed over
independently. A consistent GSO projection will systematically eliminate the tachyon in
all superstring loops. Isolating the chiral blocks introduces a phase arbitrariness. Thus, the
GSO projection requires a consistent set of phases ηL [δL ; λ] and ηR [δR ; λ], which may
depend upon the sector λ. The GSO projected amplitude for the symmetric orbifold is,

ZG = (det Im Ω)−5+n/2 dµL [λ](pλ ) ∧ dµR [λ](pλ ), (8.5)
M2 λ pλ ∈Λ∗G
where the GSO resummed chiral blocks are given by

dµL [λ](pλ ) = ηL [δL ; λ] dµC [δL ; λ](pλ ),
δL

dµR [λ](pλ ) = ηR [δR ; λ] dµC [δR ; λ](pλ ). (8.6)
δR
As is familiar from flat space–time, the phase assignments on the left and right GSO
summation may be taken to be the same (as in type IIA) or different (as in type IIB) from
one another, even though for symmetric orbifolds the chiral blocks for fixed spin structure
for right movers are simply the complex conjugate of those for left movers.
8.2. Asymmetric orbifolds
An asymmetric orbifold results from taking the coset of Minkowski space–time by a

group G that acts asymmetrically on left and right moving degrees of freedom. More
precisely, an asymmetric orbifold group G consists of pairs of Euclidean transformations
of Rn , g = (RgL , vgL ; RgR , vgR ). On the genuine conformal fields ∂x and ψ± , propagating
say on a cylinder, the action of the asymmetric orbifold group elements is clearly given by
the point group PG
g(∂z x, ψ+ ) = (RgL ∂z x, RgL ψ+ ),
g(∂z̄ x, ψ− ) = (RgR ∂z̄ x, RgR ψ− ), (8.7)
where x is viewed as a field taking values in a torus TG . The action by the shifts on the zero
mode part of x, however, is more subtle; we shall not make use of it here and postpone a
more detailed discussion to [24].
For a higher genus surface, the group elements aI , bI assigned to the canonical
homology cycles AI , BI now also each come in a pair of left and right rotations, aI =
(aLI ; aRI ) and bI = (bLI ; bRI ). One would naturally be led to consider fields obeying the
monodromy conditions

(∂z x, ψ+ )(z + AI ) = aLI ∂z x(z), (−1)2δI aLI ψ+ (z) ,

(∂z x, ψ+ )(z + BI ) = bLI ∂z x(z), (−1)2δI bLI ψ+ (z) ,

(∂z̄ x, ψ− )(z + AI ) = aRI ∂z̄ x(z), (−1)2δ̄I aRI ψ− (z) ,

(∂z̄ x, ψ− )(z + BI ) = bRI ∂z̄ x(z), (−1)2δ̄I bRI ψ− (z) . (8.8)
These monodromy conditions are, however, only formal, as no real field x in the functional
integral will exist that satisfies conditions with aLI = aRI or bLI = bRI .
The strategy adopted here is to obtain the amplitudes to higher loop order for the
asymmetric theory from the chiral blocks of symmetric orbifold theories. Specifically, each
element f ∈ PG is written as a pair f = (fL ; fR ) of rotations of Rn . The set of all elements
fL forms a group, which will be denoted PL , while the set of all elements fR forms a group
PR . The group PG is thus viewed as a subgroup of PL ⊗ PR subject to the pairing relation
f = (fL ; fR ). The chiral blocks for the left movers are the holomorphic blocks of the
symmetric theory with orbifold group PL , while the chiral blocks for the right movers are
the antiholomorphic blocks of the symmetric theory with orbifold group PR . The internal
momenta of the left and right movers do not need to match, as is familiar from toroidal and
orbifold compactifications of the heterotic string. They will be denoted by pL and pR ; the
pair (pL , pR ) belongs to a self-dual even lattice associated with G. When the elements of
P̄G have non-trivial shifts, the self-dual lattice will, in general, depend on the sector label
λ = (aI , bI ).
Thus, the block associated with the spin structures δL and δR and sector λ = (aI , bI ),
where aI = (aLI ; aRI ) and bI = (bLI ; bRI ), and chiral sector labels λL = (aLI , bLI ),
λR = (aRI , bRI ), is given by
dµC [δL ; λL ](pL ) ∧ dµC [δR ; λR ](pR ).
Just as in the case of flat Minkowski space–time (or, as explained in the previous
subsection, for symmetric orbifold compactifications) a GSO projection must be performed
to eliminate the tachyon. For most interesting asymmetric orbifolds, this projection is also
carried out in a chiral manner, i.e., summing independently over the spin structures of left
and right fermions. The relevant GSO resummed blocks are then precisely those of (8.6)
but now with different sector labels for left and right,

dµL [λL ](pL ) = ηL [δL ; λL ] dµC [δL ; λL ](pL ),
δL

dµR [λR ](pR ) = ηR [δR ; λR ], dµC [δR ; λR ](pR ). (8.9)
δR
In the next subsection, constraints will be established on the phases ηL and ηR arising from
modular symmetry.
The proposal made here for the partition function ZG of the asymmetric orbifold theory
with asymmetric orbifold group G is given by the expression

ZG = (det Im Ω)−5+n/2 K(λL , λR ; pL , pR )
pL ,pR λL ,λR
M2
× dµL [λL ](pL ) ∧ dµR [λR ](pR ). (8.10)
Here, the coefficients K(λL , λR , pL , pR ) may depend on both sets of left and right group
elements and internal momenta. Non-trivial internal momentum dependence in K will
arise in particular when the group elements of G involve non-trivial shifts. Clearly, much
of the structure of the original asymmetric group G is encoded in these coefficients. Their
evaluation can be a subtle issue, even to one-loop, and we shall return to this in [24]. In this
paper, we shall restrict attention to the pointwise vanishing of the cosmological constant,
and examine only the vanishing of each GSO resummed chiral block dµL [λL ](pL ) and
dµR [λR ](pR ).
8.3. GSO projection phases and Z2 -twisting
In the models to be studied below, the point group PG will be Abelian and generated by
chiral Z2 reflections (possibly together with an action on internal quantum numbers, such
as by the operators (−)F ). In all such cases, the chiral blocks will be given (up to phases)
by the Z2 -twisted blocks dµC [δ; ε](pε ) which were derived earlier. The sector labels will
be abbreviated by λ ≡ (aI , bI ). The twist ε, which enters the chiral blocks dµC [δ; ε](pε )
will be determined by the sector labels λ, whence the notation ε = ε(λ).
It is assumed that the GSO summation over δ is carried out independently on left-
and right-movers. For the models of greatest physical interest, the number of Z2 -twisted
dimensions is 4. This is the number of twisted dimensions in the KKS model, as well as in
the orbifolding of the type II theories by a pure Z2 twist which yield again the same type II
theory. Henceforth, the number of twisted dimensions will be assumed to be 4.
In view of the above assumptions, the relevant summation is a chiral summation over
the measure calculated in (5.11) with n = 4. The result is given by (we use the abbreviation
δ = δL , λ = λL and p = pL ),

dµL [λ](p) = ηL [δ; λ] dµC δ, ε(λ) (p) = dµ(1) (2)
L [λ](p) + dµL [λ](p), (8.11)
δ
where the partial measures are defined by (still using the notation ε = ε(λ))
2 ZC [δ, ε]
dµ(1)
L [λ](p) = ηL [δ; λ]eiπτε p iπp2 − 4∂τε ln ϑi (0, τε ) Γ [δ; ε],
ZM [δ]
δ
(2)
2 ZC [δ, ε] Ξ6 [δ]ϑ[δ]4
dµL [λ](p) = ηL [δ; λ]eiπτε p . (8.12)
ZM [δ] 16π 6 Ψ10
δ
Here, ηL [δ; λ] are the left chiral GSO projection phases, which remain to be determined.
The assumption n = 4 actually leads to considerable simplifications. A key ingredient in
the chiral blocks is the ratio of chiral partition functions for the twisted space dimensions.
For n = 4, this quantity simplifies,

+ 2 − 2
ZC [δ; ε] ϑ[δj ] ϑ[δj ] ϑ[δ +
]
2
= (8.13)
ZM [δ] ϑj4 ϑ[δ]2
since now only squares of ϑ-constants are involved.
Next, the expression for Γ [δ; ε] was calculated in (5.44). Finally, the Prym period τε
is determined by the Schottky relations in terms of the super-period matrix, which is now
simply denoted by ΩI J . In our formalism, ΩI J is interpreted as the period matrix of a
bosonic spin structure independent Riemann surface. Therefore, the Prym period τε must
be viewed as independent of the spin structures δ, and the Gaussian involving τε may be
factored out of the sum over δ. The following simplified expressions are obtained,

i ∂ iπτε p 2
dµ(1)
L [λ](p) = e αη [δ
L i
α
; λ]ν |µ ϑ
0 i i ,
4
(2π)7 η(τε )12 ∂τε α=±
i=2,3,4
eiπτε p2 ϑ[δj+ ]2 ϑ[δj− ]2

dµ(2)
L [λ](p) = ηL [δ; λ] Ξ6 [δ] ϑ[δ]2 ϑ[δ + ε]2 .
16π 6 Ψ10 ϑj4 δ
(8.14)
It remains to put constraints on the GSO phases and determine all possible solutions to
these constraints.
8.4. Modular covariance constraints on GSO phases
Constraints on the GSO phases arise from the requirement of modular invariance of
the full string measure. Under a general modular transformation M ∈ Sp(4, Z), the period
matrix Ω, the internal momenta pL , the spin structure δ, the sector label λ = (aI , bI ) and
the twist ε are transformed to Ω̃, p̃L , δ̃, λ̃ and ε̃, according to familiar rules, which are
summarized in Appendix B. Modular invariance requires the following transformation law
for the GSO resummed measure,
dµL [λ̃](p̃L , Ω̃) = ϕ(λ, M)(cτε + d)−2 dµL [λ](pL , Ω). (8.15)
The factor (cτε + d)−2 represents the effect of the modular transformation induced by M
on the Prym period τε , and ϕ is a set of phases.
The constraints on the GSO phases η are most easily established by restricting M to
(1)
the subgroup Hε of the modular group which leaves ε invariant. The two terms dµL
(2)
and dµL are functionally independent, and therefore must each result from a modular
(2)
covariant GSO summation over δ. The modular transformations of dµL are obtained by
combining the relation
Ξ6 [δ̃](Ω̃)ϑ[δ̃](0, Ω̃)4 = det(CΩ + D)8 Ξ6 [δ](Ω)ϑ[δ](0, Ω)4 (8.16)

familiar from [4] with the transformation law of the ϑ-constants,
ϑ[δ̃ + ε](0, Ω̃)
(δ + ε, M) ϑ[δ + ε](0, Ω)
= . (8.17)
ϑ[δ̃](0, Ω̃)
(δ, M) ϑ[δ](0, Ω)
Therefore, modular covariance requires that

(δ + ε, M)2
ηL [δ̃; λ̃] = ϕ(λ, M)ηL [δ; λ] for all M ∈ Hε . (8.18)

(δ, M)2
The ratios of
2 may be computed by inspecting Table 3. The symmetric pairs (δ, δ + ε)
take the values (δ1 , δ2 ), (δ3 , δ4 ) and (δ7 , δ8 ). It is manifest from Table 3 that the
2 values
for both members of each pair coincide for all M ∈ Hε . Thus, we have the following
condition for each given λ,
ηL [δ̃; λ̃] = ϕ(λ, M)ηL [δ; λ] for all M ∈ Hε . (8.19)
It is straightforward to show that the second part of the chiral measure dµ(1) L is
then automatically modular covariant. To solve these constraints requires more detailed
information on the structure of the orbifold group G. Two specific examples will be
discussed in the remainder of this section.
From Table 3, it is clear that the modular subgroup, as it acts on the even spin structures,
has two distinct orbits. Actually, these orbits are distinguished by the signatures ε|δ in
the following manner,
ε|δ = +1 orbit {δ1 , δ2 , δ3 , δ4 , δ7 , δ8 },
ε|δ = −1 orbit {δ5 , δ6 , δ9 , δ0 }. (8.20)
The orbit {δ1 , δ2 , δ3 , δ4 , δ7 , δ8 } is precisely the one that contains all even spin structures δ
for which δ + ε is also even; these are the only spin structures that enter here.
8.5. Z2 orbifolds
We now consider the model where 4 directions are compactified by an orbifold group G
whose point PG group consists of single Z2 chiral reflection rL of the fields x and ψ. Since
rL2 = 1, all sectors may be labeled precisely by a single twist ε, i.e., λ = ε. As a result, the
modular covariance condition (8.19) further simplifies and becomes,
ηL [δ̃; ε] = ϕ(ε, M)ηL [δ; ε] for all M ∈ Hε . (8.21)
By a modular transformation, ε may be brought to the standard form (6.1). The spin
structures transform under Hε in the two orbits listed in (8.20), and (8.21) is to be solved
separately in each orbit.
Observe that if a modular transformation M ∈ Hε leaves any one of the spin structures
δ invariant, one has ηL [δ; ε] = ϕ(ε, M)ηL [δ; ε], and thus ϕ(ε, M) = 1. (The alternative,
ηL [δ; ε] = 0 would lead to zero GSO phases throughout the orbit, and thus a vanishing
partition function, which is excluded.) It also follows that any spin structures related by M
will have the same GSO phase ηL [δ; ε] = ηL [δ̃; ε].
Applying these observations to the case of orbit {δ1 , δ2 , δ3 , δ4 , δ7 , δ8 }, simple inspection
of Table 3 reveals that M1 , M3 and T2 have fixed points, so that ϕ(ε, M1 ) = ϕ(ε, M3 ) =
ϕ(ε, T2 ) = 1, and ηL [δ1 ; ε] = ηL [δ2 ; ε], ηL [δ3 ; ε] = ηL [δ4 ; ε], and ηL [δ7 ; ε] = ηL [δ8 ; ε].
From the first two relations, it follows that ϕ(ε, S2 ) = 1, while from the last that ϕ(ε, M2 ) =
1. Using the action of S2 and M2 it follows that ηL [δ; ε] is independent of δ within this
entire orbit, and may be set equal to ηL [δ; ε] = 1. (The analysis in the case of the orbit
{δ5 , δ6 , δ9 , δ0 } leads to the conclusion that ηL [δ; ε] is constant throughout also this orbit.)
Since the chiral measure for Z2 reflections is non-zero for only the first orbit above, the
GSO resummed measure becomes simply,

dµL [ε](pL ) = dµL [δ, ε](pL ). (8.22)
δ
Notice that with a single twist, the assignment of GSO phases consistent with modular
invariance (in the even spin structure sector) is unique, a situation that is familiar from flat
space–time [4].
(2)
For the above GSO phases, the chirally summed measure dµC vanishes pointwise on
moduli space. To prove this, the starting point is a relation, derived first in [4],
z

Ξ6 [δ]ϑ[δ] Sδ (z, w) =
4 2
Ξ6 [δ]ϑ[δ] (0)ϑ[δ]
2 2
ωI =0 (8.23)
δ δ w
for any pair of points z, w on the surface. It suffices to choose the pair to be branch points,
canonically associated with the twist via ε = −νa + νb and z = ∆ + νb and w = ∆ + νa ,
where ∆ is the Riemann class. Using (B.9) and the specific form of the standard twist ε,
the desired relation is readily obtained,

Ξ6 [δ]ϑ[δ]2 ϑ[δ + ε]2 = 0 (8.24)
δ
which proves that dµ(2) (1)

L [ε](pL ) = 0. The vanishing of the measure dµL [ε](pL ) is even
simpler, as it results directly from the Riemann identities for the Prym variety,

ν0 |µi ϑi (0, τε )4 = 0. (8.25)
i=2,3,4
(1,2)
The vanishing of dµL might have been expected, on the grounds that the model obtained
by orbifolding by a single Z2 reflection group produces an orbifold superstring theory that
coincides with the original type II theory.12
8.6. Z2 × Z2 orbifolds: KKS models
The models introduced by Kachru–Kumar–Silverstein (KKS) in [9] are constructed on

a square torus TG at self-dual radius and an asymmetric orbifold group G whose point
group P̄G is generated by two elements,
5
f = (rL , sR )1−4 , 1, sR2 , (sL , sR )6 ; (−)FR ,
6
g = (sL , rR )1−4 , (sL , sR )5 , sL2 , 1 ; (−)FL . (8.26)
12 We are grateful to Eva Silverstein and Shamit Kachru for detailed explanations of this point.
The superscripts indicate the dimension labels on which the corresponding generator acts.
The operators (−)FL and (−)FR denote the parity of left and right chirality worldsheet
fermion number respectively, with value + on NS states and − on R states. The operations
rL , rR , sL and sR are chiral reflections and shifts, defined below.
8.6.1. Chiral reflections and shifts

The chiral reflections rL and rR act on x µ and ψ µ by reflecting respectively the left and
right oscillators of both fields and the momenta of x µ . Since rL and rR are reflections, their
order is 2. The product r = rL rR is the usual reflection on the non-chiral fields, given by
rx µ = −x µ and rψ µ = −ψ µ . The operators sL and sR leave the full fields ψ µ , as well as
the oscillators of x µ invariant. Their action on x µ is solely through chiral shifts of the zero
mode. The operators may be represented in terms of the left and right momentum operators
µ µ
pL and pR ,
µR µR
sL = exp 2πi pL , sR = exp 2πi pR . (8.27)
µ
2 µ
2
The product
√ s = sL sR is the usual shift by half a circumference. At the self-dual radius
R = 1/ 2, both operators sL and sR are of order 4. Compactification to a radius R restricts
the momentum parameters (pL , pR ) to a discrete Lorentzian lattice ΓR . To simplify the
notation, we shall not indicate this explicitly.
8.6.2. GSO phases

Of relevance here is the action of each generator on the left chiral blocks, including
their spin structure dependence. The action of the chiral shift operator sL is diagonal on
the chiral blocks and will enter only when assembling left and right chiral blocks together.
Thus, when carrying out the GSO projection by summation over spin structures in each
chirality sector separately, the action of the chiral shift operators is immaterial. Effectively,
the action of the orbifold group G on left chiral blocks reduces to the action of the point
group PG , whose generators are
f = (fL ; fR ), fL = rL1−4 , fR = (−)FR ,

g = (gL ; gR ), gL = (−)FL , gR = rR1−4 . (8.28)
Clearly, each generator is of order 2, whence the fact that on the left chiral blocks, the
action of the orbifold group is by the Abelian point group PG = Z2 × Z2 .
All possible sectors of the Z2 × Z2 theory on the genus 2 surface may be parametrized
by two half-characteristics ε and α. To establish this, denote the assignments of elements
of Z2 × Z2 around each homology cycle as before, (AI , BI ) → (aI , bI ) for I = 1, 2 and
aI , bI ∈ Z2 × Z2 . One has the following correspondence between the twist values ε, α and
the group element assignments on the homology cycles,
2ε 2α 2ε 2α 2ε 2α 2ε 2α
a1 = fL 1 gL 1 ; fR 1 gR 1 , b1 = fL 1 gL 1 ; fR 1 gR 1 ,
2ε 2α 2ε 2α 2ε 2α 2ε 2α
a2 = fL 2 gL 2 ; fR 2 gR 2 , b2 = fL 2 gL 2 ; fR 2 gR 2 . (8.29)
Given a spin structure δ, the effect of gL is to assign a sign factor depending on whether the
state traversing the corresponding cycle is NS or R. This assignment takes a simple form
in terms of the signature,
NS state traversing cycle α α|δ = +1,
R state traversing cycle α α|δ = −1. (8.30)
To check this, it suffices to work out a specific twist assignment, e.g., the standard twist
α = ε of (6.1), whose only non-zero entry is ε2 = 1/2. If δ2 = 0, the state traversing the
A2 cycle is indeed NS and the signature assignment comes out to be +1, while if δ2 = 1/2,
the state is R and we get −1.
The net effect in the sector (aI , bI ), parametrized by two twists ε and α, is to produce a
Z2 reflection of the fields x and ψ+ on the cycle ε, and an insertion of the phase α|δ. As
(aI , bI ) will run over all sectors, λ ≡ (ε, α) will run over all pairs of half-characteristics. It
follows that the set of chiral blocks for the Z2 × Z2 theory is given by
dµL [δ; ε, α](pL) = α|δ dµL [δ; ε](pL). (8.31)

Assembling all ingredients, the GSO summed chiral measure for given twists ε and α is
obtained as follows,

dµL [ε, α](pL ) = dµL [δ; ε, α](pL ) = α|δ dµL [δ, ε](pL ). (8.32)
δ δ
Under modular transformations that leave the twist ε invariant, the measure transforms
covariantly,
dµL [ε, α̃](p̃L , Ω̃) ≡ ϕ(ε, α, M)(cτε + d)−2 dµL [ε, α](pL , Ω). (8.33)
Here, the phase factor ϕ(ε, α, M) = α|(M)0 depends only on M and α.13 The presence
of the phase factor α|δ represents a generalization of discrete torsion [43] to the chiral
case.
By analyzing the action of the modular subgroup that leaves two twists invariant,
as done in Section 6.4, one may show that the above phase assignment is the unique
non-trivial choice consistent with modular invariance. The proof is similar to the one
given to determine the unique phase assignment for the Z2 case in Section 8.5. This
will be proven here for the case α ∈ O+ [ε] only; the other case is analogous. Upon
making the choice α = ε3 , the stabilizer of both twists, Hε,α , is generated by M1 , M2
and M3 . Acting on the spin structures {δ1 , δ2 , δ3 , δ4 , δ7 , δ8 }, M1 and M3 have fixed
points. Therefore, ϕ(ε, α; M1 ) = ϕ(ε, α; M3 ) = 1 and thus ηL [δ1 ; ε, α] = ηL [δ3 ; ε, α],
ηL [δ2 ; ε, α] = ηL [δ4 ; ε, α], and ηL [δ7 ; ε, α] = ηL [δ8 ; ε, α]. In view of the last equality,
we must have ϕ(ε, α; M2 ) = 1, and we therefore conclude that
ηL [δ1 ; ε, α] = ηL [δ2 ; ε, α] = ηL [δ3 ; ε, α] = ηL [δ4 ; ε, α],

ηL [δ7 ; ε, α] = ηL [δ8 ; ε, α]. (8.34)
13 To be precise, (M) is the inhomogeneous contribution in the modular transformation of spin structures
0
given in (B.12), which is common to all spin structures.
If all η are equal, we recover the case of twisting by a single Z2 . The only other linearly
independent case is when
ηL [δi ; ε, α] = +1, i = 1, 2, 3, 4,
ηL [δ7 ; ε, α] = ηL [δ8 ; ε, α] = −1 (8.35)
but this assignment is readily seen to coincide with ηL [δ; ε, α] = α|δ, which is what we
set out to prove.
8.7. Analysis of the GSO resummed measure for KKS models
For each sector λ = (ε, α), the chiral block dµC [ε, α](pL , Ω) may be analyzed by
exploiting its modular properties. Using a first modular transformation, ε may be mapped
into the standard form, familiar from (6.1),

0 0
ε= . (8.36)
0 12
Under the modular subgroup Hε (which leaves ε invariant), twists transform in 4 orbits,
as was shown in Section 6.3. The contribution to the GSO resummed measure must be
analyzed orbit-by-orbit.
(1) α ∈ O0 [ε]. This orbit corresponds to the untwisted sector and has only one element,
α = 0. Using the results of Section 7.2, one establishes dµL [ε, 0](pL, Ω) = 0.
(2) α ∈ Oε [ε]. This orbit also has only one element, α = ε. Using the results (8.20), and
(7.2), one establishes dµL [ε, ε](pL , Ω) = 0.
(3) α ∈ O− [ε]. The twists in this orbit are such that α|ε = −1. One may choose the
representative

0 0
α = 1 .
2 0
The signature with the spin structures becomes explicit: δi± |α = ±1. Therefore, the
summation over i = 2, 3, 4 vanishes and we have dµ(1) L [ε](pL , Ω) = 0. The measure
(2)
dµL [ε](pL , Ω) involves the sum

Ξ6 [δi+ ] − Ξ6 [δi− ] ϑ[δi+ ]2 ϑ[δi− ]2 .
i=2,3,4
In the separating degeneration, this sum vanishes to order τ 2 included. This may be seen
from the following limit, derived from (7.2) for each j = 3, 4,
2 2
µi µi µi
Ξ6 ϑ ϑ = O τ4 (8.37)
µj 00 0 12
i=2,3,4
using the Riemann identities of (7.9) for the summation over i = 2, 3, 4. It may be shown
that the order τ 4 term, on the other hand is non-vanishing.
(4) α ∈ O+ [ε]. The twists in this orbit are such that α|ε = +1. One may choose the
representative
1
0 2
α=
0 0
(1)
so that δ|α = (−)2δ1 . The measure dµL [ε](pL , Ω) manifestly vanishes, as the contri-
(2)
butions from α = + and α = − cancel one another out. The measure dµL [ε](pL , Ω) on
the other hand vanishes only to leading order as τ → 0, but is non-vanishing to order τ 2 .
To see this, the following limit is taken

α|δΞ6 [δ]ϑ[δ]2 ϑ[δ + ε]2
δ

= −256π 2τ 2 η(τ1 )12 η(τ2 )12 ϑ38 − ϑ48 (τ1 )ϑ24 ϑ32 ϑ42 (τ2 ) + O τ 4 . (8.38)
Thus, the chiral measure dµL [ε, α](pε ) is non-vanishing on both orbits α ∈ O± [ε].
8.8. Conclusions on the KKS model
The non-vanishing of the contribution from the orbit O− [ε] would appear to contradict
the conclusions of [9], where, instead, it is claimed to be vanishing. The arguments made
in [9] are based on the fact that, when viewed as elements of the full orbifold group G, the
generators f and g do not commute with one another. Since their commutator is a pure
translation, the elements f and g do commute, however, when viewed as elements of the
point group P̄G .
Standard arguments for symmetric orbifolds, based on functional integral methods,
guarantee the equivalence between coseting Rn by G and coseting TG by P̄G . They also
guarantee
that no contributions will arise from any sector in which the homotopy relation
−1 −1
I =1,2 I I aI bI = 1 fails to be satisfied, whether aI , bI take values in G or in P̄G .
a b
For asymmetric orbifolds, however, no direct functional integral formulation is
available. Therefore, the arguments made in the case of symmetric orbifolds may or
may not hold for asymmetric orbifolds. In particular, it is unclear how the coset theory
Rn /G should be constructed, since it is unclear how the zero mode of the field x should
be handled. Only the construction of the coset TG /P̄G is well-defined. Its construction
guarantees
that no contributions will arise from a sector in which the homotopy relation
I =1,2 a I b I aI−1 bI−1 = 1 fails to be satisfied, only when aI , bI take values in the space
group PG (but not necessarily when f and g take values in the full group G).
To summarize the situation for orbit O− [ε]: the GSO resummed chiral blocks in orbit
O− [ε] are non-vanishing. Yet, it is conceivable that assembling left and right blocks of the
asymmetric orbifold will effectively enforce the vanishing of contributions arising from
sectors in which the homotopy relation I =1,2 aI bI aI−1 bI−1 = 1 fails to be satisfied, when
aI , bI take values in the full group G. No evidence in favor of this possibility is available,
however, at this time.
In the orbit O+ [ε], the chiral GSO resummed blocks are also non-vanishing. As their
contribution is O(τ 2 ) in the separating degeneration limit, they dominate in this limit
(over the orbit O− [ε], which is O(τ 4 )). This result contradicts the claim of [9], where
the corresponding chiral contribution was found to vanish. Noticing that the sectors for
this block all satisfy the homotopy relation, no cancellation should be expected when
assembling left and right blocks.
To summarize, the non-vanishing of the contribution of orbit O+ [ε] shows that the
cosmological constant cannot vanish pointwise on moduli space and gives strong evidence
that the full cosmological constant in the KKS models is non-vanishing. The net effect
of the correct treatment of the superstring measure given here is that the cosmological
constant in KKS models involves the sum over the new modular object Ξ6 [δ], discovered
in [4], which, in contrast to the measure derived from the BRST picture changing operator
formalism does not lead to a cancellation in the chiral blocks.
Acknowledgements
We are happy to acknowledge conversations with Michael Gutperle, Per Kraus, and
Edward Witten. We are especially grateful to Shamit Kachru and Eva Silverstein for
several helpful discussions of various fine points in the theory of asymmetric orbifold
compactifications and for detailed explanations of their specific models.
Appendix A. Genus 1 ϑ-function identities
Let ω and ω be the two half period of the torus and τ = ω /ω its modulus. The
Weierstrass function obeys

℘ (z)2 = 4 ℘ (z) − e1 ℘ (z) − e2 ℘ (z) − e3 . (A.1)
Recall the prime form E and the Szegö kernel Sµ for even spin structure µ at genus 1,
ϑ1 (z − w|τ ) ϑ[µ](z − w|τ )
E(z, w) = , Sµ (z, w) = . (A.2)
ϑ1 (0|τ ) ϑ[µ](0|τ )E(z, w)
The correspondence between the classical Jacobi ϑi -functions and the spin structures is
given by ϑi (z|τ ) ≡ ϑ[µi ](z|τ ) with

µ1 = 12 12 , µ2 = 12 0 , µ3 = (0|0), µ4 = 0 12 . (A.3)
The Dedekind η-function is related to the product of all ϑ-constants,
ϑ1 (0|τ ) = −2πη(τ )3 = −πϑ2 (0|τ )ϑ3 (0|τ )ϑ4 (0|τ ). (A.4)
We have the following doubling formulas,
ϑ1 (2z|2τ ) = ϑ1 (z|τ )ϑ2 (z|τ )/ϑ4 (0|2τ ),

ϑ2 (2z|2τ ) = ϑ3 (z|τ )2 − ϑ4 (z|τ )2 /ϑ2 (0|2τ ),

ϑ3 (2z|2τ ) = ϑ3 (z|τ )2 + ϑ4 (z|τ )2 /ϑ3 (0|2τ ),
ϑ4 (2z|2τ ) = ϑ3 (z|τ )ϑ4 (z|τ )/ϑ4 (0|2τ ). (A.5)
The doubling formulas for the ϑ-constants are as follows,

2ϑ2 (0|2τ )2 = ϑ3 (0|τ )2 − ϑ4 (0|τ )2 ,
2ϑ3 (0|2τ )2 = ϑ3 (0|τ )2 + ϑ4 (0|τ )2 ,
ϑ4 (0|2τ )2 = ϑ3 (0|τ )ϑ4 (0|τ ). (A.6)
We have the following derivation formula,
∂ ϑ1 (z|τ ) ϑ[µ ](z|τ )ϑ[µ ](z|τ )
ln = −πϑ[µ]2 (0|τ ) . (A.7)
∂z ϑ[µ](z|τ ) ϑ1 (z|τ )ϑ[µ](z|τ )
This formula is valid for any even spin structure µ; the spin structures µ and µ above are
such that the triple (µ, µ , µ ) is a permutation of (µ2 , µ3 , µ4 ).
Appendix B. Genus 2 Riemann surfaces
In this appendix, we review from [4] some fundamental facts about genus 2 Riemann
surfaces, their spin structures, ϑ-functions modular properties, and relations between the
ϑ-function and the hyperelliptic formulations.
B.1. Spin structures
On a Riemann surface Σ of genus 2, there are 16 spin structures, of which 6 are odd
(usually denoted by ν) and 10 are even (usually denoted by δ). Each spin structure κ can be
identified with a ϑ-characteristic κ = (κ |κ ), where κ , κ ∈ {0, 1/2}2 , represented here
by column matrices. The parity of the spin structure κ is that of the integer 4κ · κ . The
signature assignment between (even or odd) spin structures κ and λ is defined by

κ|λ ≡ exp 4πi(κ λ − κ λ ) . (B.1)
• If κ1 and κ2 have the same parity (respectively opposite parity), we have

κ1 |κ2 = +1 ⇔ κ1 − κ2 even (respectively odd),
κ1 |κ2 = −1 ⇔ κ1 − κ2 odd (respectively even). (B.2)
• If ν1 , ν2 and ν3 are odd and all distinct, then
ν1 |ν2 ν2 |ν3 ν3 |ν1 = −1. (B.3)
It will be convenient to use a definite basis for the spin structures, in a given choice of
homology basis, as in [4]. The odd spin structures may be labeled as follows,

0 0 0 1 1 0
2ν1 = , 2ν3 = , 2ν5 = ,
11 11 11

1 1 1 1 1 1
2ν2 = , 2ν4 = , 2ν6 = . (B.4)
00 01 10
The pairs for which νi |νj = −1 are 14, 16, 23, 25, 35, 46; all others give νi |νj = +1.
The even spin structures may be labeled by,

0 0 0 0 0 1 0 1
2δ1 = , 2δ2 = , 2δ3 = , 2δ4 = ,
00 01 00 01

0 0 0 1 1 0 1 0
2δ5 = , 2δ6 = , 2δ7 = , 2δ8 = ,
10 10 00 01

1 0 1 1
2δ9 = , 2δ0 = . (B.5)
10 11
The pairs for which δi |δj = −1 are 25, 26, 29, 20, 37, 38, 39, 30, 45, 46, 47, 48, 58, 50,
67, 69, 70, 89; all others give +1.
For genus 2 there exist many relations between even and odd spin structures, some of
which will be needed here. First, the sum of all odd spin structures is a double period,
ν1 + ν2 + ν3 + ν4 + ν5 + ν6 = 4δ0 . (B.6)
Second, each even spin structure δ can be written as δ = νi1 + νi2 + νi3 (modulo integral
periods), where the νia , a = 1, 2, 3 are odd and pairwise distinct,
ν1 + ν2 + ν3 = δ7 + 2ν3 , ν1 + ν2 + ν4 = δ5 + 2ν4 ,
ν1 + ν2 + ν5 = δ3 + 2ν5 , ν1 + ν2 + ν6 = δ2 + 2ν6 ,
ν1 + ν3 + ν4 = δ8 + 2ν3 , ν1 + ν3 + ν5 = δ0 + 2ν1 , (B.7)
ν1 + ν3 + ν6 = δ9 + 2ν3 , ν1 + ν4 + ν5 = δ4 + 2ν5 ,
ν1 + ν4 + ν6 = δ1 + 2δ0 , ν1 + ν5 + ν6 = δ6 + 2ν5 .
Clearly, the mapping {νi1 , νi2 , νi3 } → δ is 2 to 1, with νi1 + νi2 + νi3 and its complement
4δ0 − (νi1 + νi2 + νi3 ) corresponding to the same even spin structure, in view of (B.6).
B.2. ϑ-functions
The ϑ-function is an entire function in the period matrix Ω and ζ ∈ C2 , defined by

ϑ[κ](ζ, Ω) ≡ exp πi(n + κ )Ω(n + κ ) + 2πi(n + κ )(ζ + κ ) . (B.8)

n∈Z2
Here, ϑ is even or odd in ζ depending on the parity of the spin structure. The following
useful periodicity relations hold, in which m , m ∈ Z2 and λ , λ ∈ C2 ,
ϑ[κ](ζ + m + Ωm , Ω) = ϑ[κ](ζ, Ω)

× exp −iπm Ωm − 2πim (ζ + κ ) + 2πiκ m ,

ϑ[κ + m , κ + m ](ζ, Ω) = ϑ[κ , κ ](ζ, Ω) exp{2πiκ m },
ϑ[κ + λ](ζ, Ω) = ϑ[κ](ζ + λ + Ωλ , Ω)

× exp iπλ Ωλ + 2πiλ (ζ + λ + κ ) . (B.9)

The ϑ-function with vanishing characteristic is often denoted by ϑ(ζ, Ω) = ϑ[0](ζ, Ω).
For each odd spin structure ν we have ϑ[ν](0, Ω) = 0. For each even spin structure δ
one defines the particularly important ϑ-constants, ϑ[δ] ≡ ϑ[δ](0, Ω). For every odd spin
structure, there exists a Riemann relation for ϑ-constants,

ν|δϑ[δ]4 = 0. (B.10)
δ
B.3. The action of modular transformations
Modular transformations M form the infinite discrete group Sp(4, Z), defined by
T
A B A B 0 I A B 0 I
M= , = , (B.11)
C D C D −I 0 C D −I 0
where A, B, C, D are integer valued 2 × 2 matrices and the superscript T denotes
transposition. To exhibit the action of the modular group on 1/2 characteristics, it is
convenient to assemble the 1/2 characteristics into a single column of 4 entries. In this
notation, the action of the modular group on spin structures κ and twists ε is then given by
[39]

κ̃ D −C κ 1 CD T
= + diag ,
κ̃ −B A κ 2 AB T

ε̃ D −C ε
= . (B.12)
ε̃ −B A ε
Here and below, diag(M) of a n × n matrix M is an 1 × n column vector whose entries
are the diagonal entries on M. On the period matrix, modular transformations act by
Ω̃ = (AΩ + B)(CΩ + D)−1 , while on the Jacobi ϑ-functions, we have

T
ϑ[κ̃] (CΩ + D)−1 ζ, Ω̃
−1 Cζ
=
(κ, M) det(CΩ + D)1/2 eiπζ(CΩ+D) ϑ[κ](ζ, Ω), (B.13)
where κ = (κ |κ ) and κ̃ = (κ̃ |κ̃ ). The phase factor
(κ, M) depends upon both κ and
the modular transformation M and obeys
(κ, M)8 = 1. Its expression was calculated in
[39], and is given by
(δ, M) =
0 (M) exp{2πiφδ (M)}, where

1

0 (M) = exp 2πi tr(M − I ) ,
2
8
1 T 1
φδ (M) = − δ D Bδ + δ B T Cδ − δ C T Aδ
2 2
1
+ δ D T − δ C T diag AB T . (B.14)
2
The modular group is generated by the following elements

I Bi 1 0 0 0 0 1
Mi = , B1 = , B2 = , B3 = ,
0 I 0 0 0 1 1 0

0 I
S= ,
−I 0
Table 6
Modular transformations of even spin structures
δ M1 M2 M3 S Σ T
2 (δ, M1 )
2 (δ, M2 )
2 (δ, M3 )
δ1 δ3 δ2 δ1 δ1 δ1 δ1 1 1 1
δ2 δ4 δ1 δ2 δ5 δ3 δ4 1 1 1
δ3 δ1 δ4 δ3 δ7 δ2 δ3 1 1 1
δ4 δ2 δ3 δ4 δ9 δ4 δ2 1 1 1
δ5 δ6 δ5 δ6 δ2 δ7 δ5 1 i 1
δ6 δ5 δ6 δ5 δ8 δ8 δ6 1 i 1
δ7 δ7 δ8 δ8 δ3 δ5 δ9 i 1 1
δ8 δ8 δ7 δ7 δ6 δ6 δ0 i 1 1
δ9 δ9 δ9 δ0 δ4 δ9 δ7 i i −1
δ0 δ0 δ0 δ9 δ0 δ0 δ8 i i −1

σ 0 0 1
Σ= , σ= ,
0 −σ −1 0

τ+ 0 1 1 1 0
T= , τ+ = , τ− = . (B.15)
0 τ− 0 1 −1 1
The transformation laws for even spin structures under these generators are given in
Table 6; those for odd spin structures will not be needed here and may be found in [4].
We shall be most interested in the modular transformations of ϑ-constants ϑ 2 [δ] and
thus in even spin structures δ and the squares of
, which are given by
 

(δ, M1 ) = exp{2πiδ1(1 − δ1 )},
 2

(δ, S) = −1,
2

(δ, M2 ) = exp{2πiδ2(1 − δ2 )},
2

(δ, Σ)2 = −1, (B.16)

 

(δ, M3 )2 = exp{−4πiδ1 δ2 },
(δ, T )2 = +1.
A convenient way of establishing these values is by first analyzing the case of the shifts
Mi , whose action may be read off from the definition of the ϑ-function. The values for the
transformations S, Σ and T may be obtained by letting the surface undergo a separating
degeneration Ω12 → 0, and using the sign assignments of genus 1 ϑ-functions. The non-
trivial entries for
2 are listed in Table 6.
B.4. The hyperelliptic representation
Every genus 2 surface admits a hyperelliptic representation, given by a double cover of

the complex plane with three quadratic branch cuts supported by 6 branch points, which
we shall denote ui , i = 1, . . . , 6. The full surface Σ is obtained by gluing together two
copies of C along, for example, the cuts from u2j −1 to u2j , 1 j 3. The surface is then
parametrized by14

6
s2 = (x − pa ). (B.17)
a=1
In the hyperelliptic representation, there is another convenient way of identifying spin
structures. Each spin structure can be viewed then as a partition of the set of branch points
pa , a = 1, . . . , 6 into two disjoint subsets, in the following way.
ν odd ⇔ branch point pa ,
δ even ⇔ partition A ∪ B, A = {pa , pb , pc }, B = {pd , pe , pf }, (B.18)
where (a, b, c, d, e, f ) is a permutation of (1, 2, 3, 4, 5, 6).
B.5. Thomae-type formulas via Mab
As shown in [4], Thomae relations may be derived most easily in terms of the bilinear
ϑ-constants Mab ≡ Mνa νb which were defined in [4],
Mab ≡ ∂1 ϑ[νa ](0, Ω)∂2ϑ[νb ](0, Ω) − ∂2 ϑ[νa ](0, Ω)∂1 ϑ[νb ](0, Ω). (B.19)
The abbreviation ∂I ϑ[ν] ≡ ∂I ϑ[ν](0, Ω) will be convenient. The holomorphic 1-form
ωνa (z), defined for any odd spin structure νa by ωνa (z) ≡ ωI (z)∂I ϑ[νa ], has a unique
double zero at pa , which is the branch point that may be canonically associated with νa .
The following fundamental results may be derived from a comparison of holomorphic
forms in the ϑ-functions and hyperelliptic formulations,
Nνb (pa − pb ) ωνb (pa ) Mνa νb
= = , (B.20)
Nνc (pa − pc ) ωνc (pa ) Mνa νc
where Nνa depends only on a and not on b. Taking the cross ratio of four branch points
(with a, d = b, c), the normalization factors N cancel out and we get the desired identity
pa − pb pc − pd M νa νb M νc νd
· = . (B.21)
pa − pc pb − pd M νa νc M νb νd
This is a Thomae-type formula, relating ϑ-constants to rational expressions of branch
points.
B.6. Mab in terms of ϑ-constants
In [4], it was shown that the following relation holds between Mab and the ϑ-constants
for even spin structure,

Mab = mab π 2 ϑ[νa + νb + νc ], (B.22)
c=a,b
14 It is customary to introduce a local coordinate system z(x) = (x, s(x)), which is well-defined also at the
branch points. Throughout, the formulas in the hyperelliptic representation will be understood in this way.
However, to simplify notation, the local coordinate z(x) will not be exhibited explicitly.
where m2ab = 1. The sign factor mab is not intrinsic, for two reasons. First, Mab is odd
under a ↔ b, while the product of ϑ-constants on the rhs is even. Second, the ϑ-constants
on the lhs and on the rhs arise to the first power, so that Mab is actually sensitive to
shifts in the ν’s by single periods (although they are invariant under shifts by any double
period). Here, the convention will be taken that the odd spin structures ν are normalized
as given in (B.4), while the even spin structures νa + νb + νc are truncated to the standard
normalization given in (B.5),
νa + νb + νc ≡ δi , i = 0, 1, . . . , 9. (B.23)
In the calculation of Γ [δ; ε] above, the signs are in fact needed. The sign factors are then
given as follows,
mab = −mba ,
mab = +1, ab = 15, 23, 25, 26, 35, 45, 46,
mab = −1, ab = 12, 13, 14, 16, 24, 34, 36, 56. (B.24)
The derivation was given in [4]. To obtain the results in the above form, one proceeds as
follows. Let ν0 be the unique genus 1 odd spin structure and let µ1 , µ3 and µ5 (and denote
an independent set by µ2 , µ4 and µ6 ) denote the three distinct genus 1 even spin structures,
obeying µ1 + µ3 + µ5 = ν0 , then any pair of genus 2 odd spin structures may be expressed
in the following basis,

µ1 ν µ3
ν1 = , ν2 = 0 , ν3 = . (B.25)
ν0 µ2 ν0
The result of [4] is then,

µ3 µ5 µ1 µ1
Mν1 ν2 = −π 2 ϑ ϑ ϑ ϑ ,
µ2 µ2 µ4 µ6

ν µ5 µ5 µ5
Mν1 ν3 = −π 2 σ (µ1 , µ3 )ϑ 0 ϑ ϑ ϑ , (B.26)
ν0 µ2 µ4 µ6
with the sign factor σ (µ1 , µ3 ) = −σ (µ3 , µ1 ) given as follows,

σ (0|0), 0 1 = σ 1 0 , (0|0) = σ 1 0 , 0 1 = +1.
2 2 2 2 (B.27)
Notice that the ordering of µ1 and µ2 determines the sign on the right hand side. Going
through all cases, (B.24) is readily derived.
References
[1] E. D’Hoker, D.H. Phong, Two-loop superstrings I, main formulas, Phys. Lett. B 529 (2002) 241–255, hep-
th/0110247.
[2] E. D’Hoker, D.H. Phong, Two-loop superstrings II, the chiral measure on moduli space, Nucl. Phys. B 636
(2002) 3–60, hep-th/0110283.
[3] E. D’Hoker, D.H. Phong, Two-loop superstrings III, slice independence and absence of ambiguities, Nucl.
Phys. B 636 (2002) 61–79, hep-th/0111016.
[4] E. D’Hoker, D.H. Phong, Two-loop superstrings IV, the cosmological constant and modular forms, Nucl.
Phys. B 639 (2002) 129–181, hep-th/0111040.
[5] E. D’Hoker, D.H. Phong, Lectures on two-loop superstrings, Hangzhou, Beijing, 2002, hep-th/0211111.
[6] Z.J. Zheng, J.B. Wu, C.J. Zhu, Two-loop superstrings in hyperelliptic language. I: the main results, Phys.
Lett. B 559 (2003) 89, hep-th/0212191;
Z.J. Zheng, J.B. Wu, C.J. Zhu, Two-loop superstrings in hyperelliptic language. II: the vanishing of the
cosmological constant and the non-renormalization theorem, Nucl. Phys. B 663 (2003) 79, hep-th/0212198;
Z.J. Zheng, J.B. Wu, C.J. Zhu, Two-loop superstrings in hyperelliptic language. III: the four-particle
amplitude, Nucl. Phys. B 663 (2003) 95, hep-th/0212219.
[7] C.J. Zhu, Two-loop computation in superstring theory, hep-th/0301018;
W.J. Bao, C.J. Zhu, Comments on two-loop four-particle amplitude in superstring theory, JHEP 0305 (2003)
056, hep-th/0303152.
[8] E. Martinec, Non-renormalization theorems and fermionic string finiteness, Phys. Lett. B 171 (1986) 189;
R. Iengo, C.J. Zhu, Notes on non-renormalization theorem in superstring theories, Phys. Lett. B 212 (1988)
309;
O. Lechtenfeld, W. Lerche, On non-renormalization theorems for four-dimensional superstrings, Phys. Lett.
B 227 (1989) 373.
[9] S. Kachru, J. Kumar, E. Silverstein, Vacuum energy cancellation in a nonsupersymmetric string, Phys. Rev.
D 59 (1999) 106004, hep-th/9807076;
S. Kachru, E. Silverstein, Self-dual nonsupersymmetric type II string compactifications, JHEP 9811 (1998)
001, hep-th/9808056;
S. Kachru, E. Silverstein, On vanishing two loop cosmological constants in nonsupersymmetric strings,
JHEP 9901 (1999) 004, hep-th/9810129.
[10] J. Harvey, String duality and non-supersymmetric strings, hep-th/9807213;
Z. Kakushadze, S.H.H. Tye, Brane world, Nucl. Phys. B 548 (1999) 180, hep-th/9809147;
G. Shiu, S.H.H. Tye, Bose–Fermi degeneracy and duality in non-supersymmetric strings, Nucl. Phys. B 542
(1999) 45, hep-th/9808095.
[11] S. Deser, B. Zumino, Broken Supersymmetry and Supergravity, Phys. Rev. Lett. 38 (1977) 1433;
E. Cremmer, B. Julia, J. Scherk, S. Ferrara, L. Girardello, P. van Nieuwenhuizen, Spontaneous symmetry
breaking and Higgs effect in supergravity without cosmological constant, Nucl. Phys. B 147 (1979) 105;
E. Cremmer, S. Ferrara, C. Kounnas, D.V. Nanopoulos, Naturally vanishing cosmological constant in N = 1
supergravity, Phys. Lett. B 133 (1983) 61;
A.B. Lahanas, D.V. Nanopoulos, The road to no-scale supergravity, Phys. Rep. 145 (1987) 1;
L. Andrianopoli, R. D’Auria, S. Ferrara, M.A. Lledo, Gauging of flat groups in four-dimensional
supergravity, JHEP 0207 (2002) 010, hep-th/0203206.
[12] D. Friedan, E. Martinec, S. Shenker, Conformal invariance, supersymmetry, and string theory, Nucl. Phys.
B 271 (1986) 93.
[13] E. Verlinde, H. Verlinde, Multiloop calculations in covariant superstring theory, Phys. Lett. B 192 (1987)
95–102;
H. Verlinde, A note on the integral over fermionic supermoduli, Utrecht Preprint No. THU-87/26, 1987,
unpublished.
[14] R. Iengo, C.J. Zhu, Evidence for non-vanishing cosmological constant in non-SUSY superstring models,
JHEP 0004 (2000) 028, hep-th/9912074.
[15] E. Gava, R. Iengo, Modular invariance and the two loop vanishing of the cosmological constant, Phys. Lett.
B 207 (1988) 283;
C.J. Zhu, Two loop computations in superstring theories, Int. J. Mod. Phys. A 4 (1989) 3877.
[16] J. Atick, J. Rabin, A. Sen, An ambiguity in fermionic string perturbation theory, Nucl. Phys. B 299 (1988)
279–294.
[17] G. Moore, A. Morozov, Some remarks on two-loop string calculations, Nucl. Phys. B 306 (1988) 387–404.
[18] H. La, P. Nelson, Unambiguous fermionic string amplitudes, Phys. Rev. Lett. 63 (1989) 24–27.
[19] J. Atick, G. Moore, A. Sen, Some global issues in string perturbation theory, Nucl. Phys. B 308 (1988) 1;
J. Atick, G. Moore, A. Sen, Catoptric tadpoles, Nucl. Phys. B 307 (1988) 221–273.
[20] E. D’Hoker, D.H. Phong, Superholomorphic anomalies and supermoduli space, Nucl. Phys. B 292 (1987)
317.
[21] E. D’Hoker, D.H. Phong, Conformal scalar fields and chiral splitting on super Riemann surfaces, Commun.
Math. Phys. 125 (1989) 469–513.
[22] R. Dijkgraaf, E. Verlinde, H. Verlinde, c = 1 conformal field theories on Riemann surfaces, Commun. Math.
Phys. 115 (1988) 649–690.
[23] E. D’Hoker, D.H. Phong, The geometry of string perturbation theory, Rev. Mod. Phys. 60 (1988) 917–1065.
[24] K. Aoki, E. D’Hoker, D.H. Phong, On the construction of some asymmetric orbifold models, in press.
[25] E. Verlinde, H. Verlinde, Chiral bosonization, determinants and the string partition function, Nucl. Phys.
B 288 (1987) 357;
E. Verlinde, H. Verlinde, Superstring Perturbation Theory, in: M. Green, et al. (Eds.), Superstrings, vol. 88,
World Scientific, Singapore, 1989, pp. 222–240.
[26] R. Russo, S. Sciuto, Twisted determinants on higher genus Riemann surfaces, Nucl. Phys. B 669 (2003) 207;
R. Russo, S. Sciuto, Twisted determinants and bosonic open strings in an electromagnetic field, hep-
th/0312205.
[27] J. Fay, Theta Functions on Riemann Surfaces, in: Springer Lecture Notes in Mathematics, vol. 352, Springer,
Berlin, 1973.
[28] P. Ginsparg, Applied Conformal Field Theory, in: E. Brézin, J. Zinn-Justin (Eds.), Les Houches 1988,
Session XLIX, Fields, Strings and Critical Phenomena, North-Holland, Amsterdam, 1990.
[29] L. Dixon, P. Ginsparg, J.A. Harvey, ĉ = 1 superconformal field theory, Nucl. Phys. B 306 (1988) 470.
[30] A. Erdelyi (Ed.), Higher Transcendental Functions, in: Bateman Manuscript Project, vol. II, Krieger,
Melbourne, FL, 1981.
[31] F. Gliozzi, J. Scherk, D. Olive, Supersymmetry, supergravity theories and the dual spinor model, Nucl. Phys.
B 122 (1977) 253;
E. Witten, in: W.A. Bardeen, A.R. White (Eds.), Symposium on Anomalies, Geometry, Topology, World
Scientific, Singapore, 1985, p. 61;
N. Seiberg, E. Witten, Spin structures in string theory, Nucl. Phys. B 276 (1986) 272.
[32] M.B. Green, J.H. Schwarz, Supersymmetrical string theories, Phys. Lett. B 109 (1982) 444.
[33] D.J. Gross, J.A. Harvey, E.J. Martinec, R. Rohm, Heterotic string theory, II. The interacting heterotic string,
Nucl. Phys. B 267 (1986) 75.
[34] E. Martinec, Superspace geometry of superstrings, Phys. Rev. D 28 (1983) 2604;
D. Friedan, Notes on string theory and two-dimensional conformal field theory, in: M.B. Green, D.J.
Gross (Eds.), Unified String Theories: Proceedings of the 1985 Santa Barbara Workshop, World Scientific,
Singapore, 1986;
S. Chaudhuri, H. Kawai, H. Tye, Path integral formulation of closed strings, Phys. Rev. D 36 (1987) 1148.
[35] E. D’Hoker, D.H. Phong, Loop amplitudes for the fermionic string, Nucl. Phys. B 278 (1986) 225;
G. Moore, P. Nelson, J. Polchinski, Strings and supermoduli, Phys. Lett. B 169 (1986) 47.
[36] A. Belavin, V. Knizhnik, Algebraic geometry and the theory of quantized strings, Phys. Lett. B 168 (1986)
201.
[37] E. D’Hoker, D.H. Phong, Superstrings, Super Riemann Surfaces, and Supermoduli Space, in: Symposia
Mathematica, in: String Theory, vol. XXXIII, Academic Press, London, 1990, Preprint UCLA-89-TEP-32.
[38] K. Aoki, E. D’Hoker, D.H. Phong, Unitarity of closed superstring perturbation theory, Nucl. Phys. B 342
(1990) 149.
[39] J.I. Igusa, Theta Functions, Springer-Verlag, Berlin, 1972;
J.I. Igusa, On Siegel modular forms of genus two, Am. J. Math. 84 (1962) 175;
J.I. Igusa, On the graded ring of theta-constants, Am. J. Math. 86 (1964) 219.
[40] K.S. Narain, M.H. Sarmidi, C. Vafa, Nucl. Phys. B 288 (1987) 551.
[41] R. Dijkgraaf, C. Vafa, E. Verlinde, H. Verlinde, Commun. Math. Phys. 123 (1989) 485.
[42] E. Kiritsis, Introduction to string theory, hep-th/9709062;
W. Lerche, A.N. Schellekens, N.P. Warner, Phys. Rep. 177 (1989) 1.
[43] C. Vafa, Nucl. Phys. B 273 (1986) 592.
Closed strings as imaginary D-branes

Davide Gaiotto, Nissan Itzhaki, Leonardo Rastelli
Physics Department, Princeton University, Princeton, NJ 08544, USA
Received 27 January 2004; accepted 16 March 2004
Abstract
Sen has recently drawn attention to an exact time-dependent boundary conformal field theory
with the space–time interpretation of brane creation and annihilation. An interesting limit of this
BCFT is formally equivalent to an array of D-branes located in imaginary time. This raises the
question: what is the meaning of D-branes in imaginary time? The answer we propose is that
D-branes in imaginary time define purely closed string backgrounds. In particular we prove that
the disk scattering amplitude of m closed strings off an arbitrary configuration of imaginary branes is
equivalent to a sphere amplitude with m + 1 closed string insertions. The extra puncture is a specific
closed string state, generically normalizable, that depends on the details of the brane configuration.
We study in some detail the special case of the array of imaginary D-branes related to Sen’s BCFT
and comment on its space–time interpretation. We point out that a certain limit of our set-up allows to
study classical black hole creation and suggests a relation between Choptuik’s critical behavior and
a phase-transition à la Gregory–Laflamme. We speculate that open string field theory on imaginary
D-branes is dual to string theory on the corresponding closed string background.
1. Introduction
Worldsheet duality between open and closed strings is one of the truly fundamental
ideas of string theory. Many modern developments have originated from the application
of this idea to supersymmetric configurations of D-branes. In this paper we show that the
study of non-supersymmetric, unstable D-branes gives some clues for a new intriguing
incarnation of open/closed string duality.
Our starting point is the real-time process of D-brane creation and annihilation, which
has recently attracted much attention [1–21]. In particular, Sen [2] introduced a simple
E-mail address: lrastell@princeton.edu (L. Rastelli).
doi:10.1016/j.nuclphysb.2004.03.017
D. Gaiotto et al. / Nuclear Physics B 688 (2004) 70–100 71
class of models in bosonic string theory obtained by perturbing the flat-space c = 26 CFT
with the exactly marginal deformation

λ dt cosh X0 (t) , (1.1)
where t is a coordinate on the worldsheet boundary, and λ is a free parameter in the range
0 λ 12 . This is a family of exact solutions of classical open string theory whose space–
time picture is that of an unstable brane being created at a time X0 ∼ −τ and decaying at
a time X0 ∼ +τ , with τ = − log(sin(πλ)). For λ = 12 the lifetime of the brane is zero, that
is, there is no brane to be found anywhere. Moreover, the corresponding boundary state
appears to vanish identically [2]. This is fascinating as it seems to suggest that for λ = 12
the BCFT (1.1) describes the stable closed string vacuum, where open string degrees of
freedom are absent. Somehow the boundary perturbation (1.1) with λ = 12 must get rid of
the worldsheet boundaries!
In the framework of (1.1), we have the opportunity to precisely test a scenario [22,23]
for how purely closed string amplitudes may be obtained at the tachyon vacuum. The basic
idea is that as the tachyon condenses, worldsheets with large holes are suppressed, and the
integration over moduli space should localize to the region where the holes shrink to points.
This heuristic picture was made somewhat more concrete in [23], where it was argued that
amplitudes for m external closed strings on the disk reduce at the tachyon vacuum to sphere
amplitudes with the same m closed string punctures plus an additional insertion of a zero-
momentum state (possibly a soft dilaton), a remnant of the shrunk boundary. This analysis
[23] was performed in the framework of a regulated version of vacuum string field theory
[24]. However, due to subtleties in the regulation procedure, it was difficult to make this
conclusion completely precise.
Since we wish to focus on the case λ = 12 , it is very useful to realize that at this critical
value the BCFT admits a simple description. By Wick rotation X0 → iX, one obtains the
well-known exactly marginal deformation λ cos(X) [25] (in fact this is how Sen arrived
at (1.1) in the first place), which for λ = 12 is equivalent to an infinite array of D-branes
located at X = 2π(n + 12 ) [25,26]. We could thus say that at the critical value λ = 12 ,
the time-dependent boundary deformation (1.1) becomes an array of D-branes located at
imaginary times

1
X0 = i2π n + , n ∈ Z. (1.2)
2
This is an empty statement if we do not define the meaning of D-branes in imaginary time.
Our approach to that issue is very simple: any quantity that one wishes to compute for a
configuration of D-branes in imaginary time should be obtained by Wick rotation of the
configuration in real space. This prescription gives consistent answers with an interesting
physical meaning.
A natural class of observables is given by scattering amplitudes of closed strings on the
disk in the background of the brane configuration (1.2). We find that these reduce to sphere
amplitudes with m + 1 punctures. The extra puncture is, however, not a soft dilaton as in
[23], but a non-trivial closed string state that involves the whole tower of massive modes.
Thus the BCFT (1.1) with λ = 12 describes a purely closed string background with no
72 D. Gaiotto et al. / Nuclear Physics B 688 (2004) 70–100
physical open string degrees of freedom. This background, however, not is the closed string
vacuum, but a specific time-dependent state with non-zero energy. The detailed features of
this background are very reminiscent of ‘tachyon matter’ [3,4].
While these results were first obtained in the special case (1.2), the basic conclusion is
much more general. A generic configuration of imaginary D-branes defines a closed string
background. The details of the background depend on the details of the configuration of
imaginary D-branes. Exactly marginal open string deformations, for example deformations
that move the positions of the imaginary branes, are naturally reinterpreted as deformations
of the closed string background. The case (1.2) is seen to be very special as the closed string
state has divergent norm [12], and the background admits an additional exactly marginal
deformation which is associated with the creation of an actual brane in real time.
An outline of the paper is as follows. In Section 2 we spell out the basic prescription for
how to deal with the array of D-branes in imaginary time. We consider a more general
case in which X0 = ia(n + 12 ), where a is an arbitrary parameter, and find a general
formula for disk amplitudes associated with such an array. In Section 3 we analyze in
detail disk amplitudes for scattering of m external closed strings from the array of D-
branes in imaginary time. By an exact computation, we show that they are equivalent
to sphere amplitudes with the same m closed strings insertions plus the insertion of an
extra closed string state |W . The details of this closed string state, in particular its space–
time interpretation, are studied in Section 4. The energy of the state is finite and of order
O(gs0 ) = O(1) for any a > 2π . The case a = 2π , corresponding to λ = 12 in the BCFT
(1.1), is seen to be special as the normalization and energy of the state diverge [12].
We suggest a heuristic mechanism to cutoff this divergence, namely we point out that
the gravitational back-reaction makes the effective distance aeff = 2π + γ gs , with γ > 0,
leading to a total energy of order 1/gs .
In Section 5 it is shown that although the imaginary array of branes does not have
propagating open string degrees of freedom, open strings still play an important role.
There still exists a discrete set of on-shell open string vertex operators corresponding to
exactly marginal deformations, for example deformations that move the branes around in
imaginary time. These open string moduli are re-interpreted as closed string deformations,
according to a precise dictionary. In Section 6 we show that, subject to certain reality
conditions, one can distribute D-branes quite freely in the complex X0 plane. The reduction
of disk amplitudes to sphere amplitudes still holds in this general case. We briefly comment
on extensions to the superstring. In Section 7 we briefly discuss some ideas about the open
string field theory associated with D-branes in imaginary time and speculate that a version
of vacuum string field theory may be obtained in the limit a → ∞.
Section 8 is devoted to the case a = 2π . In this case there is an additional exactly
marginal open string deformation (which we may label as ‘cosh(X0 )’) which is not dual
to a purely closed deformation, as it introduces instead an actual D-brane in real time. We
give a treatment of (1.1) for all λ 12 by representing the boundary state as an infinite array
of some specific smeared sources, and then performing the Wick rotation. We find that for
λ < 12 a time-delay of order τ is introduced between the incoming and the outgoing parts
of the closed string wave, while for |X0 | < τ there is an actual source for the closed string
fields.
Finally in Section 9 we describe a curious application of our set-up. We consider an

array of branes in imaginary time, and tune various parameters in such a way that the closed
string state |W becomes simply a classical, spherically symmetric dilaton wave with
barely enough energy to form a black hole. Fascinating critical behavior was discovered
by Choptuik [27] in such a system. We observe that in the corresponding Euclidean theory
this critical point corresponds to a phase transition reminiscent of the Gregory–Laflamme
[28] phase transition. We speculate on a possible realization of this phase transition in large
N open string field theory.
While this paper was in preparation we received [12] that has some overlapping results.
2. Preliminaries
We wish to give a meaning to the notion of an array of D-branes located at imaginary

times X0 = i(n + 12 )a. The distance a = 2π (in units α = 1) corresponds to the λ = 12
case of the BCFT (1.1), but it is interesting and not more difficult to keep a arbitrary. In
this section we define our basic prescription to compute disk amplitudes of external closed
strings in the background of this ‘imaginary array’. We first give a naive argument why
all such scattering amplitudes vanish. Then we illustrate, via a simple example, a natural
analytic continuation prescription that actually yields non-zero answers. Finally we derive
a simple general formula that expresses the scattering amplitude S in the background of the
imaginary array in terms of the scattering amplitude Ã for a single D-brane in real time.
2.1. Naive argument
Let us denote the scattering amplitude of some closed strings off a single D-brane
located at X = 0 by Ã. Here X is a spatial coordinate. To find the scattering amplitude
S for an array of D-branes located at imaginary time we could proceed in two steps:
(i) Find the scattering amplitude S̃ for an array of D-branes located at the real positions
X = (n + 12 )a, n ∈ Z. This is given by
∞
1
S̃(P , . . .) = Ã(P , . . .)ei(n+ 2 )aP
n=−∞
∞

= Ã(P , . . .) (−1)n 2πδ(P a − 2πn), (2.1)
n=−∞
where P is the total momentum in the X direction and the dots denote other
variables that the amplitude may depend on. This is a precise equality in the sense
of distributions. Simply put, the momentum has to be quantized due to the periodicity
in X.
(ii) Apply (inverse) Wick rotation
X → −iX0 , P → iE, (2.2)

where now X0 and E are real. This has the effect of turning the spatial coordinate X
into a temporal coordinate X0 and of rotating the array of D-branes from the real axis
to the imaginary axis, X0 = i(n + 12 )a. So it takes us exactly to the set-up that we wish
to study. Wick rotating (2.1) we get the formal expression
∞

S(E, . . .) ≡ S̃(iE, . . .) = Ã(iE, . . .) (−1)n 2πδ(iEa − n). (2.3)
n=−∞
Naively this implies that S(E, . . .) is identically zero since for any real E the delta
functions vanish. However one has to be more careful, as Ã(iE, . . .) may blow up for
some real values of E yielding a non-zero S(E). Clearly the discussion so far has been
quite formal, for example the summation (2.1) does not commute with the Wick rotation
(2.2) in the sense that if we first Wick rotate and then sum over the array, the sum does not
converge for any real E. We need to specify an unambiguous prescription for the analytic
continuation (2.2). Let us illustrate how a natural prescription comes about in a simple
example.
2.2. Example
The example we wish to study is

1
Ã(P , . . .) = , (2.4)
P 2 + c2
where c is a real number (we take for definiteness c > 0) that can depend on the other
variables but not on P . This is the one-dimensional Euclidean propagator. In position space

π
G̃(X) = dP Ã(P )eiP X = e−c|X| , (2.5)
c
2
2 − c )G̃(X) = −δ(X). All the amplitudes we study in this paper can be
d 2
which obeys ( dX
expanded in terms of (2.4). Hence this example is of special importance.
The advantage of working in position space is that now we can simply sum over all the
contributions of the array of D-branes explicitly:
∞
π −c|a(n+ 1 )+X|
G̃array (X) = e 2 . (2.6)
c n=−∞
This is a periodic function with period a (see Fig. 1), given in a neighborhood of X = 0 by
π cosh(cX) a
G̃array (X) = , |X| . (2.7)
c sinh( ca
2 ) 2
Now we wish to Wick rotate. Of course, G̃array (X) is not an analytic function, precisely
because of the δ-function sources located at X = (n + 12 )a. So we need to specify what we
exactly mean by the analytic continuation X → −iX0 . A natural prescription is to focus
Fig. 1. A graph of G̃array (X), which has the interpretation of the field produced by an infinite array of δ-function
sources (‘D-branes’) located at X = n(a + 12 ). The dashed line represents the analytic continuation to |X| > a2 of
the branch around the origin.
on the branch around the origin to find

π cos(cX0 )
Garray X0 = . (2.8)
c sinh( ca
2 )
Fourier transforming (2.8) back to momentum space we find
π
S(E, . . .) = ca δ(E − c) + δ(E + c) . (2.9)
2c sinh( 2 )
As anticipated by the heuristic discussion in the previous subsection, while Ã(P , . . .) is
non-zero for any real P , we find that S(E, . . .) has support only for those values of E
corresponding to singularities of Ã(P , . . .), namely E = −iP = ±c. As we shall see in the
context of string theory this will have the interpretation of a change in the dimension of the
moduli space.
2.3. General prescription
In principle, all amplitudes considered in the paper could be expanded in terms of the
example studied above. It would be nice however to have a general formula that gives
S(E, . . .) in terms of Ã(P , . . .). To this end consider, for generic Ã(P , . . .),
∞ ∞

G̃array (X) = dP eiP X (−1)n 2πδ(aP − 2πn)Ã(P , . . .). (2.10)
−∞ n=−∞
Fig. 2. Integration contours in the complex P plane. The zeros of sin(aP /2) are denoted by the symbols ‘x’
along the real P axis. The black dots represent possible poles of Ã and the thick line represents a possible cut.
We assume that the only singularities of Ã are on the imaginary P axis.
In all cases that we study in the present paper Ã(P , . . .) is an analytic function with poles
or cuts only along the imaginary P axis. Moreover Ã(P , . . .) goes to zero for |P | → 0
sufficiently fast to validate to following argument.
With the help of the residues theorem we can write

1 Ã(P , . . .)
G̃array (X) = dP eiP X , (2.11)
2i sin( aP
2 )
C
where the contour C is depicted in Fig. 2. By our analyticity assumptions on Ã(P , . . .), the
curve C can be deformed to C˜ without crossing any singularities (see Fig. 2) and without
picking up any contributions from the two semi-circles at infinity (that are not shown in
Fig. 2). So we conclude, after Fourier transforming back to momentum space, that

S(E) = F (E) DiscE Ã(iE) , (2.12)
where
1
F (E) = . (2.13)
2 sinh( aE
2 )
Here by DiscE we mean the discontinuity with respect to E, namely

f (E + i) − f (E − i)
DiscE f (E) = , (2.14)
i
so for example DiscE (1/E) = −2πδ(E). Let us check our master formula (2.12), (2.13)
in the example studied in the previous subsection. From (2.4), Ã(iE) = 1/(−E 2 + c2 );
applying (2.12), (2.13) we immediately reproduce the result (2.9).
3. Disk amplitudes
To gain some insight into the meaning of the D-brane array at imaginary times (1.2), we
now turn to a detailed analysis of disk scattering amplitudes of external closed strings. We
start by considering the simplest possible case, namely the amplitude of two closed string
tachyons. We first study this concrete example using standard methods, and then describe
a more abstract point of view, that can be generalized easily to m-point functions involving
arbitrary on-shell closed strings. A clear physical interpretation will emerge.
3.1. Tachyon two-point amplitude
The simplest non-trivial case that we consider is the disk amplitude with the insertion
of two closed string tachyons. This example, which we are going to work out in full detail,
contains already much of the essential physics.
To apply the prescription derived in the previous section, we need to evaluate the
disk two-point function Ã(p1 , p2 ) for a standard D-brane. We consider a D(p − 1)-brane
with Dirichlet boundary conditions for X̃M , M = 0, . . . , 25 − p, and Neumann boundary
conditions for X̃m , m = 26 − p, . . . , 25. We are using a notation that will be natural for
the theory after the Wick rotation: X̃0 is a spatial coordinate that will become timelike
after Wick rotation. In order to have a standard D-brane with Neumann conditions in time,
we take one of the X̃m to be timelike. So the Wick rotation is actually a double Wick
rotation: it transforms X̃0 from spacelike to timelike and one of the parallel directions X̃m
from timelike to spacelike.1 In practice, the amplitude Ã(p1 , p2 ) will be written in terms
of kinematic invariants, so these distinctions are of no much consequence.
It is convenient to break up the momenta into parallel and perpendicular directions,
M m
pµ = p⊥ , p , (3.1)
and define the kinematic invariants
s = p1
2
= p2
2
,
t = (p1 + p2 )2 = (p1⊥ + p2⊥ )2 . (3.2)
1 The case p = 0 is special, as we must choose one of the transverse directions X̃M to be timelike.
In our conventions2 α = 1, so that the on-shell condition for closed string tachyons is
p2 = 4.
The two-point disk amplitude has one modulus (four real coordinates minus three
conformal Killing vectors). We work with a slightly unconventional parametrization of this
moduli space that will be easy to generalize later when we turn to higher-point functions.
Namely we shall fix the positions of the two vertex operators and integrate over the radius
of the disk. We represent the disk as the complex domain Hρ = {z ∈ C, |z| ρ}, that is we
cut out a hole of radius ρ from the complex plane. The measure for the modulus ρ is
dρ
(b0 + b̄0 ) . (3.3)
ρ
Using the standard doubling trick, a closed string vertex operator V (p, z, z̄) is replaced by
the two chiral insertions VL (p, z) and VL (p , ρ 2 /z̄), where p = (−p⊥ , p ). For closed
string tachyons, V (p, z, z̄) = c(z)c̄(z̄) exp(ipX(z, z̄)) and VL (p, z) = c(z) exp(ipXL (z)).
Fixing the tachyon vertex operators at z = 1 and z = ∞, we have
1
1 dρ
Ã(p1 , p2 ) = dz
2πi ρ
0 |z|=ρ

× VL (p1 ; ∞)VL(p2 ; 1)b(z)VL p2 ; ρ 2 VL (p1 ; 0) , (3.4)
where the symbol
, denotes a CFT correlator on the plane. A short calculation gives
1
s−2 (t/4 − 1) (s − 1)
Ã(p1 , p2 ) = dρ ρ t /2−3 1 − ρ 2 = . (3.5)
2 (t/4 + s − 2)
0
This is of course a standard result, with a familiar interpretation [29,30]. The scattering
amplitude of a closed string off a D-brane shows the usual ‘dual’ structure with poles both
in the open and in the closed string channel. In the open string channel, the poles are located
at s = 1, 0, −1, . . . and arise from expanding around ρ = 1 where the vertex operators
approach the boundary. In the closed strings channel, we see poles at t = 4, 0, −4, . . .,
arising from expanding around ρ = 0 where the boundary shrinks to zero size.
We are now in a position to apply the prescription (2.12).3 The discontinuity with
respect to E = E1 + E2 = −i(p10 + p20 ) comes from the poles in the variable t, and so
we find that
∞

1
S(p1 , p2 ) = fk (s)δ(t/4 − 1 + k), (3.6)
2 sinh( a|E|
2 ) k=0
2 Moreover in writing a string amplitude Ã(p , . . . , p ), we treat all momenta as incoming. Finally our
1 m
convention for the Minkowski metric is ‘mostly plus’.
3 Notice that Ã(p , p ) obeys our analyticity assumptions, indeed it is an analytic function of the total
1 2
momentum P 0 with singularities (poles) only for imaginary P 0 , and behaves as 1/|P 0 |2s for large |P 0 |.
where
(−1)k (s − 1) (2 − s)(3 − s) · · · (1 + k − s)
fk (s) = = . (3.7)
2k! (s − k − 1) 2k!
Several remarks are in order. First, all the contributions to S come from the ρ → 0
region of the moduli space. The sharpest way to see this is to introduce a cut-off ρ 1.
Then the amplitude Ã becomes analytic in t since the poles in the closed string channel
disappear. Applying (2.12) yields S = 0 for any . This vindicates the original intuition that
the boundary state for the array of imaginary D-branes should correspond to a ‘hole of zero
size’ (an extra puncture) in the worldsheet. This intuition will be made very precise below.
Second, S has no poles in s, the open string channel! The external closed strings do not
couple to any on-shell open string degrees of freedom. Since there are no D-brane sources
in real time, this is as expected. We are describing a purely closed string background.
More precisely, (3.6) describes a sphere amplitude with the two tachyons insertions and
an additional on-shell closed string state that involves excitation at all levels of the tower of
massive modes. Indeed, the prefactor fk (s) is a polynomial of degree 2k in p , consistently
with the fact that a closed string mode at level k has up to 2k Lorentz indices.
3.2. Boundary state computation
In order to secure this result, and to prepare for the generalization to higher-point
amplitudes, we now repeat the computation in a more abstract language. We still write
the amplitude in the domain Hρ , but instead of using the doubling trick, we represent the
effect of the boundary at |z| = ρ by the insertion of a boundary state |B̃ p−1|z|=ρ . This is
the full boundary state defining the D(p − 1)-brane located at X̃M = 0.
A boundary state is a ghost number three state in the closed string Hilbert space, obeying
among other things the conditions

(QB + Q̄B )B̃ p−1 = 0, (b0 − b̄0 )B̃ p−1 = 0. (3.8)
In radial quantization, a state at radius ρ can be obtained from a state at radius ρ = 1 by
propagation in the closed string channel,
p−1
B̃ = ρ L0 +L̄0 p−1
B̃ . (3.9)
|z|=ρ |z|=1
We can then write
1
dρ
Ã(p1 , p2 ) =
|V (p1 ; ∞, ∞)V (p2 ; 1, 1)(b0 + b̄0 )ρ L0 +L̄0 B̃ p−1 |z|=1 . (3.10)
ρ
0
To proceed, we insert a complete set of states {|k, i},
dρ
1
k2
Ã(p1 , p2 ) = dk
|V (p1 ; ∞, ∞)V (p2 ; 1, 1)|k, iρ ( 2 +2li )
ρ
i 0

×
k, i|(b0 + b̄0 )B̃ p−1 |z|=1 , (3.11)
where li is the level of the state |k, i, and we integrate over the modulus ρ to get
1
Ã(p1 , p2 ) = dk
|V (p1 ; ∞, ∞)V (p2 ; 1, 1)|k, i 2
2 + 2li
k
i
p−1
×
k, i|(b0 + b̄0 )B̃ |z|=1
. (3.12)
This expression exhibits the decomposition of the amplitude into a source term from the
boundary state, a closed string propagator, and the 3-point interaction vertex. If we now
apply the master formula (2.12) we find
S(p1 , p2 ) =
|V (p1 ; ∞, ∞)V (p2 ; 1, 1)|W , (3.13)
where4
δ(k 2 /2 + 2l(i))
|W ≡ dk |k, i
k, i|(b0 + b̄0 )B p−1 |z|=1
i 2 sinh( a|E|
2 )
δ(L0 + L̄0 )
= (b0 + b̄0 )B p−1 |z|=1 . (3.14)
2 sinh( a|E|
2 )
We see that the interaction of the two tachyons with the imaginary array is captured by a
ghost number two closed string state |W . Using (3.8) and {b0 , QB } = L0 , we have
(QB + Q̄B )|W = 0, (b0 − b̄0 )|W . (3.15)

These are precisely the physical-state conditions for closed strings. Moreover, we have the
freedom to add to |W BRST trivial states, that would decouple in the computation of the
correlator (3.13). We recognize |W as an element of the closed string cohomology. A more
detailed discussion of this state and of its space–time interpretation will be carried out in
Section 4.
Finally we can trade the state |W with a vertex operator insertion at the origin, and
re-write the amplitude (3.13) as a CFT correlator in the plane,

S(p1 , p2 ) = V (p1 ; ∞, ∞)V (p2 ; 1, 1)W(0, 0) . (3.16)
This is manifestly the scattering amplitude for three closed strings on the sphere! Having
explicitly performed the integral over ρ there are no remaining moduli, as it should be for
a sphere three-point function.
Clearly this conclusion does not depend on the vertex operators V (pi , z, z̄) being closed
string tachyons. The analysis immediately generalizes to arbitrary physical closed string
vertex operators: in the background of the imaginary array, a two-point function on the
disk is equal to a three-point function on the sphere, with the on-shell vertex operator W
as the extra insertion.
4 Notice that in the formula below we drop the ‘tilde’ on the symbol for the boundary state: |B p−1 denotes
the boundary state after double Wick rotation.
3.3. Higher-point disk amplitudes
The computation of higher-point amplitudes is a straightforward generalization. We still

represent the disk as the complex domain Hρ and describe the moduli space of m closed
strings on the disk by fixing the positions of two vertex operators, say V1 at z = ∞ and
V2 at z = 1, and varying the positions of the other m − 2 insertions and the radius ρ
of the hole. More precisely, we vary the m − 2 coordinates over the full complex plane,
{zi ∈ C, i = 3, . . . , m}, and for a given choice of the {zi } we vary the radius ρ between 0
and the distance of the closest insertion, 0 ρ ρ0 = min[|zi |, i = 1, . . . , m] (see Fig. 3).
This way we cover moduli space exactly once, as can be easily checked for example by
mapping the above configuration to the interior of the unit disk. We then have
ρ0
dρ
Ã(p1 , . . . , pn ) = d z3 · · · d zm
2 2
ρ
0

×
|R V1 (p1 ; ∞, ∞)V2(p2 ; 1, 1) · · · Vm (pm ; zm , z̄m )

× (b0 + b̄0 )ρ L0 +L̄0 B̃ p−1 |z|=1 , (3.17)
where R{·} denotes radial ordering. Inserting as before an intermediate complete set of
states, and performing the ρ integral, we find

Ã(p1 , . . . , pn ) = dk d 2 z3 · · · d 2 zm
i

×
|R V1 (p1 ; ∞, ∞)V2(p2 ; 1, 1) · · ·Vm (pm ; zm , z̄m ) |k, i
k2
+2li
ρ02
×
k, i|(b0 + b̄0 )B̃ p−1 |z|=1 . (3.18)
k2
2 + 2li
Now we extract the discontinuity with respect to the energy E. Clearly we get
contributions from the poles in the propagators ∼ k 2 /2+2l
1
. A priori, the integrals over
i
the coordinates zi may generate additional singularities. However it is not difficult to show
that extra singularities can only arise when a (proper) subset of the vertex operators have
an on-shell total momentum. We can define the amplitude by analytic continuation away
from these singular points, and then disregard these singularities. This is reminiscent of the
celebrated canceled propagator argument [31]. We can then write5 (see Fig. 3)

S(p1 , . . . , pm ) = d 2 z3 · · · d 2 zm

× V1 (p1 ; ∞, ∞)V2(p2 ; 1, 1) · · · Vm (pm ; zm , z̄m ) W(0, 0) .
(3.19)
5 Notice that the dependence on ρ drops because of the delta function δ(L + L̄ ). In other terms, the ρ
0 0 0
integral localizes to ρ → 0 and the upper limit of integration ρ0 is immaterial.
Fig. 3. Before the double Wick rotation we have a standard disk amplitude. The disk can be viewed as the region
Hρ , which is the complex plane with a hole of radius ρ. There are contributions to the scattering amplitude from
all values of ρ ρ0 , where ρ0 is the distance of the closest puncture. After the double Wick rotation the only
contribution is coming from ρ = 0. The hole shrinks to a point leaving behind an extra puncture W inserted at
the origin.
The coordinates zi , i = 3, . . . , m, are integrated over the full complex plane, so this is
the string theory amplitude for m + 1 insertions on the sphere. We can check that the
counting of moduli is consistent with this result. We started with 2m − 3 moduli for a
disk with m closed punctures and performed an explicit integration over ρ. This gives
2m − 4 = 2(m + 1) − 6, which is the number of moduli for a sphere with m + 1 punctures.
3.4. Interpretation
Let us summarize what we have learned. Disk amplitudes for m external closed strings
off the D-brane array at the imaginary times X0 = i(n + 1/2)a are completely equivalent
to (m + 1)-point amplitudes on the sphere, with the extra closed string insertion

gs d 2 z W(z, z̄),
where W is given by (3.14). The explicit factor of gs , which was omitted in the previous
formulae, comes from the relative normalization of disk amplitudes to standard sphere
amplitudes. Unlike the case of an ordinary brane, we are finding that for the imaginary array
disk amplitudes give a purely closed correction (of order gs ) to the background. There are
no D-brane sources in real time, only closed string radiation satisfying the homogeneous
wave equation.
The absence of sources in real time can also be deduced from the fact that all one-
point functions on the disk are trivially zero in our prescription. This is consistent with
the computations in [3], where the stress tensor (related to the graviton one-point function)
was found to vanish in the BCFT (1.1) for λ = 12 . There are two equivalent ways to see
this in our language. The one-point function on the disk is a smooth function of the total
energy E, so there is no discontinuity in E. Alternatively, the discussion above implies that
a disk one-point function is equal to a sphere two-point function with the extra insertion
of W; but sphere two-point functions are zero because of the infinite volume of the unfixed
moduli.
Since disk amplitudes provide the first order correction to the background, it is very
natural to expect that amplitudes with multiple boundaries give the necessary higher-order
corrections. A genus zero amplitude with m external closed strings and b boundaries,
is expected to become after Wick rotation a sphere amplitude with m + b punctures,
with b insertions of gs d 2 z W. Notice that since the boundaries are indistinguishable,
the sum over boundaries exponentiates to the insertion of exp( d 2 z gs W), which has
the interpretation of a coherent state of closed string radiation. One may investigate
this issue rigorously by representing the effect of the boundaries using boundary states,
and decomposing the moduli space of a surface with b boundaries and m closed string
punctures in terms of closed string vertices and propagators [32]. The moduli space
integration should localize to the region where each boundary shrinks to an extra puncture.
A priori, one may also expect that regions of moduli space where zero-size boundaries
collide could provide additional contributions. For example one may expect that an annulus
amplitude would reduce to a sphere amplitude with two insertions of gs W, plus a sphere
amplitude with a single insertion of a new operator gs2 W1 coming from the shrinking of
both boundaries to the same point. This issue should be investigated further.
4. The closed string state
In this section we study the physical properties of |W in more detail and make contact
with [11,12]. Our starting point is the boundary state |B p−1, which is obtained from
double Wick rotation of |B̃ p−1 . So |B p−1 is associated with a D(p − 1)-brane located at
XM = 0, in a space with metric (ηMN , δmn ). One has
∞
p−1 M
B = N δ X exp − µ†
a Sµν ã ν†
|0; k = 0. (4.1)
n n
n=1
N is a normalization constant that can be found for example in [33], and
SMN = −ηMN , Smn = δmn . (4.2)

With the help of (3.14) we can expand |W in terms of closed string physical states,
25−p
dk⊥ 1 1
|W = N c1 c̄1 |0gh ⊗
(2π) 25−p 2|E| 2 sinh( a|E| )
2
2
× k = 4 − Sµν ∂Xµ ∂X ¯ ν k 2 = 0 + · · · ,
where the dots indicate higher massive modes. The first term is the closed string tachyon,
which is an artifact of the bosonic string and is standard practice to ignore. Below we first
discuss the massless modes and then consider the massive modes.
4.1. Massless sector
In order to read off the dilaton and gravity wave profile from the second term we have
to undo their mixing. The unmixed dilaton and graviton (in the Einstein frame) take the
form (see, e.g., [33])
S · (φ)
hµν = Sµν − ηµν , φ = S · (φ) , (4.3)
η · (φ)
where
1
(φ) = (ηµν − kµ lν − kν lµ ), k · l = 1, l 2 = 0. (4.4)
2
Let us first look at the case p = 0 (an array of imaginary D(−1)-branes). One finds
that the Einstein metric is completely flat as all the expectation values of hµν vanish.
This means that the part of the leading term in the ADM mass, which scales like 1/gs ,
vanishes (see also [12]).6 The dilaton, on the other hand, does not vanish. So for p = 0, the
massless fields in |W consist of a spherically symmetric dilaton wave φ(r, X0 ) in 25 + 1
dimensions, whose energy is of order O(gs0 ) = O(1).
For p > 0 the 26-dimensional metric in the Einstein frame is non-trivial. However the
fields profiles are translationally invariant in the p longitudinal directions Xm , and to read
off the ADM mass we are instructed to dimensionally reduce to the 26 − p transverse
dimensions. One finds that in the (26 − p)-dimensional Einstein frame the metric is zero,
and we have again only a spherically-symmetric dilaton wave. We conclude that to order
1/gs the mass vanishes for general p.
6 This is consistent with Sen’s observation [3] that the stress tensor vanishes for the λ = 1 BCFT (1.1).
2
The 7
√ space–time profile of this dilaton wave is quite interesting [11]. For fixed radius
r = X X , the field decays exponentially fast as X → ±∞. For fixed X0 , the field
M M 0
decays as 1/r 23−p as r → ∞, just like the fields produced by an ordinary Dp-brane, but
with a different numerical coefficient. (These two asymptotic behaviors match in a region
of thickness of the order of a around the light-cone X0 = ±r.) To be precise, up to an
overall numerical constant that will not be relevant for the discussion below, the leading
asymptotic behavior as r → ∞ of the various fields is, in the 26-dimensional Einstein
frame,
(1 + K)(d − 3) − 2p 1
h00 → ,
d −2 r 23−p

δmn 2d − 2p − 5 − K 1
hmn → ,
2 d −2 r 23−p

δMN 2p + 1 + K 1
hMN → ,
2 d −2 r 23−p
d − 2p − 3 − K 1
φ→ . (4.5)
4 r 23−p
Here d = 26 and the parameter K is set to 1 for the standard Dp-brane and is set to −1 for
the background |W .
An exercise one can now do is to compute the force acting on a probe D0-brane in this
dilaton wave background [11]. One has to be a bit careful here with the exact meaning of
the ‘force’ between the brane and the wave since this is not a static set-up. The ‘force’
that can be computed using (4.5) for K = −1 is acting during a finite time interval
−T X0 T at a distance r in the limit Tr → 0. In the Einstein frame in 26 dimensions
the DBI action for the D0-brane takes the form

11 √
dX0 e− 12 φ g00 . (4.6)
From this and (4.5) with p = 0 we can deduce that the ratio between the force acted upon
the D0 by a standard brane and the ‘force’ acted on it by the dilatonic wave in |W is
Fstandard 11
12 φ(K = 1) + 12 h00 (K = 1) 12
= = . (4.7)
FW 11
12 φ(K = −1) + 1
2 h00 (K = −1) 11
This is in agreement with [11] where this ratio was calculated in a different way.
4.2. Massive modes
The massive closed string states in |W obey the dispersion relation E 2 = k⊥

2 + m2 , with
m2 = 4(n − 1). For n > 1, their field profile (proportional to 1/ sinh(a|E|/2) is strongly
peaked at energies |E| − m 1/a. This means that for any a 2π , the states are with good
approximation non-relativistic already at level n = 2, with the approximation improving at
7 From the discussion in Section 6 of [11] it is not obvious that Eq. (6.27) in that section describes just a
dilaton wave. However, this can be verified with the help of Eq. (4.3).
higher levels. Their fields at X0 = 0 are Gaussians of the form

ma ak 2
δ(k ) exp − exp − ⊥ , (4.8)
2 4m
that is, the closed string modes occupy
√ the directions Xm and are localized at XM = 0
(M = 0) with a width of order a/m in the transverse directions. The time evolution
of these √field profiles follows non-relativistic Schrödinger equation, so their width scales
as |X0 |/ am for large times. Interestingly, the massive modes behave as non-relativistic
matter located at the would-be position of the brane. This conclusion was also reached
in [12].
Let us now compute the normalization of |W and its the space–time energy, following
[12].8 The normalization has the interpretation of (the expectation value of) the total
number of particles n̄ in the background. One has
∞
25−p
dk⊥ 1 1
n̄ =
W|c0 c̄0 |W = N 2
dn ,
(2π)25−p 2|E| 4 sinh2 ( aE
2 )
n=0
∞
25−p
dk⊥ 1 |E|
Ē = N 2 dn , (4.9)
(2π)25−p 2|E| 4 sinh2 ( aE
2 )
n=0
where dn can be computed from the generating function

∞
∞

−24

dn w = f (w)
n
, f (w) = 1 − wn . (4.10)
n=0 m=1
√
The asymptotic behavior of dn for large n is [34] dn ∼ n−27/4 e4π n . It is easy to see that
a > 2π the exponential suppression from 1/ sinh2 ( a|E| 2 ) wins over the exponential growth
of states, and both n̄ and Ē are perfectly finite. On the other hand, in the limit a → 2π we
get
∞ −2(a−2π) √
e n
Ē ∼ ∼ (a − 2π)p/2−1. (4.11)
np/4
n=0
For a = 2π the expectation value of the energy diverges for p = 0, 1, 2. Naively, for
p > 2 the energy is finite. However, for any p the expectation values of powers
E k
will eventually diverge for sufficiently high k [12], and hence for a = 2π the uncertainty
in the energy9 is infinite for any p.
This divergence has a very natural physical interpretation. As reviewed in the
introduction and further discussed in Section 8, for a = 2π the background of imaginary
branes admits an infinitesimal deformation that introduces a D-brane source in real time,
and we should then expect that this background also has an energy of order 1/gs .
8 Eqs. (4.9) and (4.11) were not obtained independently of [12].

9 A similar behavior is found for n̄, which is logarithmically divergent for p = 0 and finite for p > 0, but with
higher moments diverging for any p.
Sincewe are computing this energy in the limit gs → 0 in a perturbative expansion

Ē = n=0 (gs )n E (n) , it is natural to find that leading term E (0) = ∞. At finite gs this
divergence should be regulated in such a way that Ē ∼ 1/gs .
Before discussing a heuristic mechanism that supports this expectation, we would like
to point out that for a = 2π the state |W has all the features to be identified with ‘tachyon
matter’ [3,4]. Indeed for a = 2π the energy is stored in very massive closed strings modes
that behave like non-relativistic matter strongly localized along the directions Xm . This is
the closed string dual of Sen’s discussion on ‘tachyon matter’ in the context of open string
field theory; for large times |X0 | → ∞ the classical open string solution corresponding to
the BCFT (1.1) approaches, for all λ, an open string configuration with zero pressure and
with all the energy localized along the Xm directions.10
It is very tempting to suspect that at finite gs the distance a is renormalized to an
effective value
aeff = 2π + ε, where ε = γ gsb , (4.12)

where both b and γ are positive numbers of order one. This would make the energy finite.
Restricting in the following to p = 0,11 we find from (4.11)
1
Ē ∼ . (4.13)
ε
So to obtain the expected scaling ∼ 1/gs the parameter b has to be equal to one. An
argument why this is plausible was given in [12] from the point of view of open string field
theory. Here we provide an alternative heuristic argument from the closed string channel,
that justifies why b = 1 and also why γ > 0.
Let us think about this issue from the point of view of the array before Wick rotation.
When gs = 0 the distance between the branes is 2π . When gs is turned on the branes will
slightly curve space–time due to their mass so that the distance between them is no longer
2π . Actually since their mass is ∼ 1/gs and the Newton constant scales like gs2 the back-
reaction is of order gs and so the proper distance between the branes is indeed as in (4.12).
γ is positive simply because the metric components in the transverse direction to the brane
scale like
γ̃
g⊥ ∼ 1 + gs , (4.14)
r 23−p
with γ̃ > 0.
It is interesting to check if this argument generalizes when we put N rather than just
one brane at each site. First, the back-reaction of the metric scales like gs N which means
that now ε ∼ gs N . Combining this with (4.13) and with the fact that |W gets multiplied
10 More precisely, as shown in [12] and further elaborated in Section 6 of this paper, the limit X0 → ∞ of
(1.1) corresponds for all λ (up to a trivial time translation) to the outgoing (X0 > 0) part of our state |W , and
symmetrically X0 → −∞ of (1.1) corresponds to the incoming (X0 < 0) part of |W .
11 For p > 0 (4.11) cannot be trusted as the higher moments of the energy (
E k −
Ek )1/k have a worse
degree of divergence than the mean value Ē =
E.
by a factor of N we find for the total energy

N2 N
ĒN ∼ ∼ , (4.15)
ε gs
which is the correct scaling for the tension of N D-branes.
Clearly this heuristic reasoning is not powerful enough to fix γ in (4.12). One obvious
reason is that the distance between the branes is of the order of the string scale and hence
the gravity approximation should not be trusted. A closely related point is the following.
Suppose that we could somehow fix γ and that we found the correct D-brane mass. If γ
was just a fixed number then it is easy to see that the uncertainty in the D-brane mass would
be of the same order as the mass itself, which is of course not the case for D-branes at weak
coupling. This seems to suggest that γ should not be viewed as a constant but rather as a
fluctuating field.
5. On open and closed string moduli
In Section 3 we showed that scattering amplitudes off the imaginary array do not have
any open string poles. This means that there are no propagating open strings degrees of
freedom, consistently with the fact that there are no branes in real time. However there still
is a discrete set of on-shell open string vertex operators, which demand an interpretation.
In this section we argue that they are dual to deformations of the closed string state |W .
For simplicity let us start by considering the case p = 0, the array of D(−1)-branes at
imaginary times X0 = i(n + 12 )a. The open string spectrum should be read off from the
theory before double Wick rotation, where we have an array of D(−1)-branes at the spatial
locations X̃0 = (n + 12 )a. For generic12 distance a, the only matter primaries of dimension
(n) (n)
one are Ṽµ = ∂ X̃µ , where µ are space–time indices and n ∈ Z labels the position of the
D(−1)-brane. In the double-Wick rotated theory the physical open string states are just the
same, up to trivial relabeling. So the most general state in the open string cohomology of
the theory of D(−1)-branes at imaginary times can be written as

25 ∞
µ
|f = f(n) ∂Xµ(n) (0)c1 |0. (5.1)
µ=0 n=−∞
Clearly these are the exactly marginal open string deformations that correspond to moving
the positions of the D(−1)-branes.
Next we can consider disk amplitudes S(p1 , . . . , pm ; f ) for m closed strings scattering
off the imaginary array, with one additional insertion of (5.1) on the boundary of the disk.
Without any open strings puncture, S(p1 , . . . , pm ) was shown in Section 3 to be a sphere
amplitude with an extra insertion of the closed string state |W . How does the addition of
|f change this conclusion?
µ
To get some insight, consider first the case f(n) = f µ for all n. This deformation
is simply a Goldstone mode associated with a rigid translation Xµ → Xµ + fµ of
12 The case a = 2π is of course special and will be discussed in Section 8.

the whole array. It is clear that in this simple case S(p1 , . . . , pm ; f ) is obtained from
S(p1 , . . . , pm ; f ) by replacing |W with its infinitesimal translation, |W → f µ ∂µ |W .
In this special example the open string insertion |f corresponds to a symmetry of the
vacuum broken by the closed string background |W .
In the more general case (5.1), it is not difficult to show13 that S(p1 , . . . , pm ; f ) will
still be a sphere amplitude with m + 1 punctures, where the extra closed puncture is now
δf W, the infinitesimal deformation of W obtained by displacing the branes according
to |f . Generically this deformation will change the total energy of the closed string state.
The intuitive picture is that the branes act as sources for the closed string fields. Since these
sources are not in real time, the resulting closed string fields are homogeneous solutions of
the wave equation, and we have a purely closed background. Changing the positions of the
branes changes the details of the closed string field profiles. Thus the open string moduli
(5.1) are re-interpreted as deformations of the closed string background. In the next section
this picture is made precise.
There is one important restriction on the allowed motions of the branes: the resulting
µ
closed string fields must be real. This imposes certain constraints on the moduli f(n) .
Starting from the array at X0 = in(a + 12 ), reality of the closed string fields demands
that the D(−1)-branes be moved in pairs, that is, the kth D(−1)-brane at X0 = ia(k + 12 ),
k > 0, together with its mirror partner at X0 = ia(−k + 12 ). The reality condition is
µ µ ∗
f(k) = f(−k) ∀k ∈ N, (5.2)
where ∗ denotes complex conjugation. As long as we refrain from considering branes with
complex spatial coordinates Xi , i = 1, . . . , 25, we should keep the spatial moduli f(k)
i real,
and then (5.2) implies f(k) = f(−k) . For the time coordinate X on the other hand, there
i i 0
are two possible motions for a given pair of branes, as illustrated in Fig. 4.14
The reality condition can be simply rephrased by saying that the D-brane configuration
must be symmetric under reflection with respect to real time axis. A simple way to see
this constraint is to focus for example on the massless closed string fields. The field
produced by some δ-function sources located at the origin in space and at times X0 = Yk0
is schematically [11]
1
φ(Xµ ) ∼ , (5.3)
k
[−(X − Yk )2 + (Xi )2 ]s
0 0
where the power s depends on the number of transverse directions. We are interested in
measuring the field φ(Xµ ) for real values of its arguments Xµ . It follows that for φ(Xµ )
to be real, the locations of the sources {Yk0 } must either be real or come in complex
conjugate pairs. In the next section we write down the prescription to obtain the closed
string background associated with an arbitrary configuration for imaginary branes, and it
will be apparent that the same reality condition is valid in full generality.
13 We cannot directly apply the prescription of Section 2. However, the relevant disk amplitude is simple
enough that can be evaluated directly in the theory of the spatial array X̃0 = a(n + 12 ), and then Wick rotated.
14 Interestingly, for a pair of branes before the double Wick rotation there are also two allowed exactly marginal
motions along X = −iX0 , but of course they are the independent translations of each brane along the real X axis.
Fig. 4. The two possible motion modes of a pair of D-branes in the complex X0 plane.
For p > 0 there are additional on-shell open strings since the open string tachyon can
be put on-shell provided pm pm = 1, where pm is the momentum along the Neumann
directions. These are precisely the exactly marginal deformations studied in the original
work by Callan et al. [25]. For example the deformation λ cos(Xm ), where Xm is one of
the directions along the brane, continuously interpolates between Neumann and Dirichlet
boundary conditions for the coordinate Xm . At the critical value λ = 12 , we get an array of
D(p − 2)-branes localized at Xm = 2π(n + 12 )i. The deformed boundary state is known for
all values of λ [25] and we can easily apply our prescription (3.14) to compute |W(λ ).
As λ varies, these states are a family of purely closed string backgrounds. An open
string deformation, in this case cos(Xm ), is then again re-interpreted as a closed string
deformation.
6. A more general set-up
One important conclusion from the discussion in the previous section is that there is
nothing fundamentally special about the array of D-branes at X0 = i(n + 12 )a. Exactly
marginal open string deformations allow to move the positions of the branes in the complex
X0 plane, with the only constraints coming from the reality condition. By sending off most
of the branes to large imaginary time, we can also consider configurations with a finite
number of branes.
The simplest configuration consists of only one pair of branes at X0 = ±iβ (with β real,
β > 0). As in Section 2.2, we start with the pair in Euclidean space at X = ±β. The analog
of (2.7) is
2π −βc
G̃pair (X) = e cosh(cX), |X| β. (6.1)
c
Notice that the only change relative to (2.7) is in the prefactor. This was foreordained since
in a finite neighborhood of X = 0 both G̃pair (X) and G̃array (X) obey homogeneous wave
2 − c )G̃ = 0 and are even under X → −X. Wick rotation then gives the usual
d 2
equation ( dX
sourceless solution ∼ cos(cX0 ). In Fourier transform
π
Spair (E, . . .) = e−βc δ(E − c) + δ(E + c) , (6.2)
2c
which we can write
Spair (E, . . .) = Fpair (E) DiscE [A], (6.3)
where
Fpair (E) = sign(E)e−β|E| . (6.4)
This result should be compared with (2.9), (2.12), (2.13).
This formula can be generalized further by displacing the D-brane pair along real time,
that is at X0 = α ±iβ. In this case Fpair (E) → sign(E)e−β|E|+iαE . By linear superposition,
an arbitrary configuration of M D-brane pairs at positions X0 = αk ± iβk , k = 1, . . . , M,
leads again to an amplitude of the form (2.12), where now the prefactor takes the general
form

M
F (E) = sign(E)e−βk |E|+iαk E . (6.5)
k=1
The usual array is a special case of this formula: with = ∞, αk = 0, βk = a(k − 12 ) we

immediately recover (2.13).
All the conclusions of Section 3 are valid in this more general case. The basic step is
the extraction of the discontinuity in E, which localizes the moduli space integration to
ρ → 0. A general configuration of imaginary D-branes leads to a sphere amplitude with an
extra closed string insertion W, only the details of this insertion change. All we have to do
2 ) in (3.14) with the general F (E) given in (6.5). The state |W is
is to replace 1/ sinh( Ea
normalizable and has finite energy as long as all D-branes are at distances βk > π from the
real axis, generalizing the condition a > 2π that we had for the array.
For a generic configuration of a finite number of imaginary D-branes, the space–time
dependence of W is, however, quite different than the case of the infinite array. This
difference is sharpest for the massless fields outside the lightcone, r |X0 |. For the array,
they have the same dependence ∼ 1/r 23−p as for a static Dp-brane, whereas for a finite
configuration they decay one power faster ∼ 1/r 24−p . This nicely dovetails with the fact
that the infinite array with a = 2π admits the exactly marginal deformation that creates an
actual Dp-brane in real time. The incoming and outgoing radiation that makes up Warray
are precisely tuned to admit the deformation that reconstructs the brane, whereas for a
generic finite configuration, brane creation would require an abrupt change of the field
asymptotics.
It is worth pointing out that our prescription to compute the closed string fields
associated with imaginary D-branes is equivalent to the second-quantized point of view
taken in Section 3 of [12]. They propose to obtain the wavefunction for the closed string
fields at X0 = 0 by cutting open the Euclidean path integral in the presence of D-brane
sources located at Euclidean time X̃0 < 0. To evaluate the expectation value of the closed
string fields in such a wave function, one needs to construct the full configuration of sources
symmetric under X̃0 → −X̃0 (obtained by simply reflecting the sources located at X̃0 < 0
to X̃0 > 0), and read off the solution at X̃0 = 0. This is the second-quantized version of
what we do here.
An interesting open question is whether as we vary the configuration of imaginary
D-branes, the state |W spans the full closed string cohomology. It is clear that to have
any chance of success we must introduce more general boundary states than the ones
considered in this paper (for example, D-branes with magnetic and electric fields on their
worldvolume), and allow for an infinite number of imaginary branes. If the answer to
this question is positive then the open/closed string duality takes a new form since at the
fundamental level closed strings and D-branes are unified. Each closed string puncture in
a string theory amplitude could be effectively represented by a hole in the worldsheet with
appropriate boundary conditions.
6.1. Superstring
This general set-up can be generalized to the superstring in a straightforward way.

There is now more variety of D-brane sources, since we can consider both stable BPS Dp-
branes and unstable non-BPS Dp-branes. For non-BPS branes the discussion is completely
analogous to the bosonic case. One can distribute non-BPS brane pairs freely in the
complex X0 plane, subject to exactly the same reality condition as we discussed in
Section 5.
BPS Dp-branes introduce on the other hand an important novelty: they are sources of
Ramond–Ramond fields. The reality condition for the RR fields forces us to consider pairs
composed of a BPS brane at X0 = α + iβ and of its anti-brane partner at X0 = α − iβ.
It can be checked that the RR fields produced by a generic configuration of such pairs are
non-vanishing only inside the lightcone, r |X0 | (see also [11]); this is consistent with
the fact that the configuration has zero total RR charge, so we do not see long range RR
fields.
Like in the bosonic string, infinite ‘critical’ arrays of branes at X0 = iacrit(n + 12 )
correspond to special limits
√ of BCFTs related to real √time processes of brane creation and
annihilation. Here acrit = 2π (this is the familiar 2 in translating between the bosonic
string and the superstring). There are two interesting classes of examples. One can consider
a critical array of non-BPS D(p − 1)-branes in imaginary time, which is dual to the closed
string background related to the decay of a Dp–D̄p pair; or a critical array of alternating
BPS D(p − 1)/D̄(p − 1) in imaginary time, related to the decay of non-BPS Dp. (Needless
to say, given p, any of these examples makes sense in either type IIA or IIB, but not in
both.) The subject is potentially quite rich.
7. Open string field theory
We have found an intriguing relation between D-branes in imaginary time and purely
closed backgrounds. Since D-branes admit an open string description, this suggests that one
may be able to obtain a dual description of closed string theories in terms of open strings.
We have already seen in Section 5 that the exactly marginal open string deformations have
a natural reinterpretation as deformations of the closed string background. However since
there are no propagating on-shell open string degrees of freedom on imaginary D-branes,
it is clear that if such a complete open/closed duality exists, it must involve the off-shell
open strings. In this section we offer some very brief and incomplete speculations in this
direction.
We would like to propose that the open string field theory (OSFT) on a configuration
of imaginary D-branes is dual to the corresponding closed string theory. To make sense
of this speculation we must define what is the OSFT for imaginary branes. Applying our
usual strategy, we start with OSFT on standard D-branes, and double Wick rotate. While
from a first quantized point of view it may be subtle to define the Wick-rotated open string
theory, in the second quantized approach we have the luxury of a space–time action, which
seems straightforward to analytically continue.
Let us sketch how this may come about in the example of the array of D(−1)-branes.
We start with an array of D(−1)-branes at X̃0 = (n + 12 )a. The open string field Ψj k has
Chan–Paton labels j, k ∈ Z running over the positions of the D-instantons, and the cubic
OSFT action [35] takes the form

1 1
jk 1
S[Ψ ] = − 2 Ψj k , QB Ψj k +
Ψj k , Ψj l ∗ Ψlk +
Ψjj , C . (7.1)
g0 2 jk
3
j kl j
Here we have also included the gauge-invariant open/closed vertex [23,36–39]

Ψ, C that
couples external on-shell closed strings C to the open string field. Notice that the string
fields Ψij do not depend at all on the zero modes of the space–time coordinates. Double
Wick rotation is then immediate. There is little to do in all the purely open terms of the
action.15 Only the open/closed vertex is affected. We conjecture that the resulting action
describes the non-trivial closed string state |W .
In principle it should be possible to recover the results of Section 3 from the point of
view of the second quantized Feynman rules for the action (7.1). In OSFT, amplitudes
of external closed string on the disk are obtained by computing expectation values of
the gauge-invariant operators discussed in [23,36–39], or in other terms by the use of
the open/closed vertex. We expect that for (7.1) such amplitudes collapse to the region
of moduli space where the open string propagators have zero length, reducing to sphere
amplitudes with an extra insertion of |W . The mechanism for such a collapse must be of a
somewhat different nature than in [23] or in [41], since here we are using the conventional
BRST operator but a highly unconventional state-space.
15 Note however that one has to be careful here with the reality conditions discussed in the previous sections
[40].
Finally, the OSFT (7.1) may provide us with new clues about the string field theory
around the tachyon vacuum. As a → ∞, the energy of |W goes to zero, and we approach
the tachyon vacuum. Interestingly, in the same limit |W does not vanish completely, but it
becomes purely a zero-momentum dilaton. Thus we recover precisely the scenario of [23].
Since for any a such an open string field theory should make sense this may provide us
with a consistent regularization of vacuum string field theory, which should be related to
the one considered in [23] by some non-trivial field redefinition.
8. a = 2π and reconstruction of the brane
As discussed in Section 4, for a = 2π the normalization and the energy associated with
the closed string state |W diverge [12]. This singularity signals the appearance of new
open string degrees of freedom. Open strings stretched between neighboring branes in
imaginary time have conformal dimension L0 = (a/2π)2 , which equals one for a = 2π .
The infinite periodic array has the special property that a specific linear combination of
these marginal operators (the one which is invariant under translations X0 → X0 + 2π and
is even under X0 → −X0 ) is in fact exactly marginal.16 A new branch of moduli space
opens up for a = 2π . Indeed, the theory for a = 2π is equivalent to the λ = 12 critical point
of the BCFT (1.1). Turning on the exactly marginal open string deformation (‘cosh(X0 )’)
we can reduce the value of λ, from λ = 12 (the purely closed string background |W ) to
λ = 0 (the usual brane with Neumann boundary condition in time). This is the sense in
which the array for a = 2π is very special: it admits an exactly marginal deformation that
cannot be interpreted purely as a deformation of the closed background.
8.1. Smearing and brane creation
One can obtain many insights into the physics for λ < 1/2 by a simple extension of the
methods in Section 2. The basic idea is that in the Euclidean BCFT with X = −iX0 , taking
λ < 12 amounts to regulating the delta-function sources located at X = 2π(n + 1/2) with
smooth lumps. The precise way to make this regulation can be gleaned from the boundary
state [25] for general λ. Focusing on the oscillator free part of the boundary state,
∞

n −nτ
|B0 ∼ 1 + 2 (−1) e cos nX(0) |0, τ ≡ − log sin(πλ) . (8.1)
n=1
By Poisson resummation, one finds

∞

˜ 1 τ
|B0 ∼ jτ X + 2π n + |0, j˜τ (X) = . (8.2)
−∞
2 π(X2 + τ 2)
16 In principle also the odd operator, ‘sinh(X0 )’, is exactly marginal. However in the bosonic string this
deformation would bring us to the ‘wrong’ side of the tachyon potential for X0 < 0.
The interpretation of this formula is clear: for λ < 12 , the boundary state corresponds to an
infinite array of smeared sources j˜τ ; indeed j˜τ is a well-known representation for δ(X) in
the limit τ → 0.
If we sum the source j˜τ (X) over the infinite array and then Wick rotate X → −iX0 we
find
tanh((X0 + τ )/2) − tanh((X0 − τ )/2)
Jτ X0 = . (8.3)
4π
Already at this stage we see the crucial difference compared to λ = 12 (τ = 0). Now there
is a non-zero source localized at real time for |X0 | < τ , which is to say that an unstable D-
brane appears at X0 = −τ and disappears at X0 = τ . For later use let us record the Fourier
transform of this function,
sin(τ E)
ρτ (E) = . (8.4)
sinh(πE)
The next step is to repeat the exercise for the fields generated by this smeared array. The
Fourier transform of the Euclidean source j˜τ (X) is e−τ |P | . Therefore, in the notations of
Section 2.2, Eq. (2.4) should be replaced by
e−τ |P |
Ã(P , . . .) = . (8.5)
P 2 + c2
As in Section 2.3, we sum over the array and use the residue theorem to write the total field
as a contour integral,

1 e−τ |P |
G̃τ (X) = dP eiP X 2 . (8.6)
2i (P + c2 ) sin(πP )
C
If we now move the contour over the imaginary P axis as in Fig. 2, and Wick rotate, we
find

1 eiτ E e−iτ E
Sτ (E) = − , (8.7)
2 sinh(πE) (E + i)2 − c2 (E − i)2 − c2
which we recognize as
1
Sτ (E) = Gret (E)eiτ E − Gadv (E)e−iτ E . (8.8)
2 sinh(πE)
This way of writing the answer makes the space–time interpretation manifest (see also
[12]). For X0 < −(τ + r), we have purely incoming radiation. For X0 > τ + r, we have
purely outgoing radiation. Outside of these two cones (for (X0 )2 − r 2 < τ 2 ), the fields are
the same as the ones produces by a static source. The thickness of the transition regions
is of the order of the string length. Notice that the outgoing radiation is produced by the
rapid change of the source for X0 ∼ τ , and similarly the incoming radiation is correlated
with the change at X0 ∼ −τ . This process can be described some finely tuned incoming
closed strings that create an unstable D-brane at X0 = −τ , which then decays at X0 = τ
into outgoing closed strings.
We can also write

π
Sτ (E, . . .) = GF (E)ρτ (E) + cos(τ E) ca δ(E − c) + δ(E + c) , (8.9)
2c sinh( 2 )
where GF (E) is the Feynman propagator and ρ(E) the Fourier transform of the source
in real time Jτ (X0 ), see (8.4). The two terms in (8.9) have very different interpretations.
While the first term contains a propagator, and is non-zero for any finite E, the second term
has support only for E = ±c. The second term in (8.9) is proportional to (2.9); in Section 3
it was seen to corresponds to the extra insertion |W in a sphere amplitude. The first term
is instead a real D-brane source, associated to an earnest disk amplitude.
In the limit λ → 12 (τ → 0) only the second term in (8.9) contributes, since the source
ρτ (E) → 0. It is a bit less obvious to see why only the first term contributes as λ → 0.
In this limit, τ → ∞ and since the oscillations of both terms are controlled by the
dimensionless parameter τ E, to have a non-zero contribution E must go to zero, hence for
any fixed finite c the second term vanishes. The fact the E → 0 is consistent with the fact
that the for λ = 0 the unstable D-brane exists for an infinite amount of time; X0 becomes
an ordinary Neumann direction and the total energy (with all insertions as incoming) must
vanish.
We see that as we vary λ, not only does the D-brane source vary, but the closed
string background (captured by the second term in (8.9)) also changes in a very non-
trivial way. This phenomenon requires some comments. Although the BCFT (1.1) may
be naively thought of as a deformation of the open string background only, in fact we need
to simultaneously change the closed string background to cancel tadpoles. Indeed, while
the operator cosh(X0 ) is exactly marginal in the open string sense, one point functions of
generic on-shell closed strings have a non-trivial dependence on λ. Of course, tadpoles
for closed string operators are also generated in the familiar case of time-independent
boundary deformations. However in the time-independent case only one-point functions of
zero-energy on-shell bulk operators can change, whereas for a time-dependent perturbation
like (1.1) there are tadpoles with a non-trivial space–time dependence. As we turn on the
cosh deformation, the closed string background needs to be corrected introducing closed
string matter. Happily, as (8.9) shows, this seems to be automatically taken care of by the
analytic continuation procedure.
The discussion in this section has been at the level of the discussion of Section 2.
It would be very interesting to repeat the analysis of Section 3 and compute scattering
amplitudes of closed strings for general λ. It would also be of interest to find the behavior
of the open string spectrum for 0 < λ < 12 .
9. From Choptuik to Gregory–Laflamme
As was explained in Section 4 the massless sector of |W is a spherically symmetric

dilaton wave in 26 − p dimensions. In the limit that the Newton constant GN goes to zero
the linearized solution is an exact solution to the gravity-dilaton equation of motion. When
the coupling constant is turned on the non-linearity of the equation of motion becomes
more and more important. With spherical symmetry much is known about the non-linear
aspects of the system. In particular, in [27] Choptuik showed, via numerical analysis, that
a universal behavior occurs at the critical point where black hole formation first occurs.
A crude summary of his results is the following. Consider a spherically symmetric wave
of a massless scalar field

ηφ r, X0 . (9.1)
The strength of the non-linear effects of the gravitational back-reaction of the wave grows
with the overall coefficient η. When η → 0 the linearized approximation is exact while
for η → ∞ a large black hole will be formed, with an exponentially small amount of
energy escaping the black hole formation as an outgoing radiation.17 Therefore, for a given
field profile φ(r, X0 ) there is a special value η∗ where the black hole formation first takes
place. While η∗ certainly depends on the details of φ, the time evolution of the system
for η → η∗ admits scaling behavior that is fixed by a certain constant, ∆. This constant
is universal in the sense that it does not depend on φ, but it can depend for example on
the number of spatial directions. Another critical exponent δ is related to the scaling of the
mass of the black hole near the transition. These fascinating results have been explored
quite extensively in the last decade.
In this section we take advantage of the fact that a spherically symmetric dilaton wave
can be viewed as a configuration of D-branes located in imaginary time to relate Choptuik’s
findings to a phase transition somewhat reminiscent of the Gregory–Laflamme instability
[28]. For simplicity we will mostly phrase the discussion in the context of the bosonic
string, ignoring as usual the closed string tachyon. The extension to the superstring of the
scaling arguments given below is straightforward.
The first step is to understand how to get from our |W a background containing only
a classical dilaton wave. It is clear that to suppress massive closed string modes we need
to take las → ∞. (Here we have restored the string length ls , that was set to one in the rest
of the paper.) However the total energy of the wave will then go to zero as 1/a 25−p in
this limit (see (4.9), and a black hole will not be formed this way. A simple way to obtain
a configuration with enough energy without exciting the massive modes is to increase the
number of branes at each point X0 = ia(n + 12 ), as in the discussion at the end of Section 4.
With N D-brane at each point the particles density, n̄/V , and the energy density = Ē/V
(V is the volume along the wave) of the wave for large a are (from (4.9)
24−2p 24−2p
n̄ N 2 ls N 2 ls
∼ , ∼ . (9.2)
V a 24−p a 25−p
Let us now state the exact conditions that must be satisfied in the limit we want to take:
(1) The string coupling gs should go to zero to suppress quantum effects, and the total
number of particles n̄ should go to infinity so that |W is well described by a classical
wave.
(2) las → 0 to ensure that the massive closed strings decouple.
17 The system is classical so this radiation is classical and should not be confused with Hawking radiation.
(3) The gravitational radius associated with the total energy of the wave
1
rG = (GN ) 23−p , (9.3)
where should be comparable to the wavelength, a (here GN is the Newton constant in
26 dimensions). If we define the dimensionless parameter
1
48−2p
rG (gs2 N 2 ls /a 25−p ) 23−p
ζ= = , (9.4)
a a
then in our set-up ζ plays the role of η in (9.1) That is, for small ζ the linearized
approximation is valid and a black hole will not be formed while for large ζ non-linear
effects are important and lead to black hole formation. So the critical value ζ ∗ is a
number of order one.
It is easy to verify that these conditions are satisfied in the following limit
24+β−p
a fixed, ls → 0, gs ∝ lsβ , N ∝ 1/ ls , β > 0. (9.5)
So far what we have done is to take a certain ‘decoupling’ limit that leaves us with
the classical picture of black-hole formation. The obvious question to ask is what does
this mean from the D-brane point of view. In particular, does anything special happens at
ζ = ζ ∗ in the array of D(p − 1)-branes located at X̃0 = (n + 12 )a, i.e., in the array before
the double Wick rotation?
The gravitational radius associated with N D-branes located at each site is
1
lG ∼ (gs N) 24−p ls . (9.6)
When lG is much smaller than the distance between the D-branes, a, the gravitational
interactions between them is small and they can be considered as separated points.
However, when lG is larger than a the gravitational interaction between them is so strong
that effectively they form a black hole along X̃0 . It is easy to see from (9.6), (9.4) that
this transition occurs at the same point as the wave to black hole transition. Since the two
processes are related by Wick rotation it is very tempting to suspect that there is a precise
numerical relation between the exponents of [27] and the exponents, yet to be found, in the
phase transition just described.
It should be stressed that we although the transition considered here is somewhat
reminiscent of the Gregory–Laflamme instability, there are also obvious differences. In the
Gregory–Laflamme case one starts with a black p-brane and finds, at the level of linearized
equations of motion, an instable mode; condensation of this tachyon is then conjectured to
lead to an array of black (p − 1)-branes. (This expectation however may even be false, see,
e.g., [42–44] for recent discussions.) In our case we start from the array and we turn on
the coupling constant to eventually form a (possibly non-uniform) black p-brane, so we
are going in the opposite direction. More crucially, we do not start with an array of black
p-branes. The D-branes are extremal to begin with in the sense that they have no finite
area horizon. In the process of turning on the coupling constant the radius of their zero
area horizon grows until they meet. At that point a combined finite area horizon will be
formed.
Let us illustrate this in the context of the superstring. In this case there are also no closed
string tachyon so the whole discussion makes more sense. We can consider either an array
of BPS D(p − 1)-branes, which must then be alternating as D/D̄, or an array of non-BPS
D(p − 1)-branes. Consider for example an array of D3/D̄3 in type IIB. For small ζ the D3-
branes interact weakly and so they can be treated as separated stable objects. As we turn
on ζ their gravitational radius start to touch. At that stage the fact that we have alternating
branes anti-branes is crucial since the charges can annihilated to form a black 4-brane. This
black 4-brane does not carry any 6-form charge (we are in type IIB).
It would be extremely interesting to understand Choptuik’s critical behavior from the
point of view of the dual open string field theory discussed very briefly in Section 7. Very
schematically, the way this could work is as follows. From the open string field theory
point of view the classical non-linear gravity dynamics is obtained in the loop expansion.
We expect that the lightest open string mode stretched between the branes, which at the
classical level has large positive m2 , becomes tachyonic for ζ > ζ ∗ due to quantum open
string effects. It may be possible to see this effect in a one-loop computation, in the spirit
of [45]. If this scenario is correct, the universal physics at the transition point would be
completely captured by a massless field, and we would have an explanation of gravitational
critical behavior in terms of a second order phase transition in the dual OSFT.
Acknowledgements
We thank C. Callan, A. Hashimoto, I. Klebanov, H. Liu, J. Maldacena, J. McGreevy,

H. Ooguri, A. Sen, C. Vafa and H. Verlinde for discussions. This material is based upon
work supported by the National Science Foundation under Grant No. PHY 9802484.
Any opinions, findings, and conclusions or recommendations expressed in this material
are those of the author and do not necessarily reflect the views of the National Science
Foundation.
References
[1] M. Gutperle, A. Strominger, Spacelike branes, JHEP 0204 (2002) 018, hep-th/0202210.
[2] A. Sen, Rolling tachyon, JHEP 0204 (2002) 048, hep-th/0203211.
[3] A. Sen, Tachyon matter, JHEP 0207 (2002) 065, hep-th/0203265.
[4] A. Sen, Field theory of tachyon matter, Mod. Phys. Lett. A 17 (2002) 1797, hep-th/0204143.
[5] A. Sen, Time evolution in open string theory, JHEP 0210 (2002) 003, hep-th/0207105.
[6] P. Mukhopadhyay, A. Sen, Decay of unstable D-branes with electric field, JHEP 0211 (2002) 047, hep-
th/0208142.
[7] A. Strominger, Open string creation by S-branes, hep-th/0209090.
[8] A. Sen, Time and tachyon, hep-th/0209122.
[9] F. Larsen, A. Naqvi, S. Terashima, Rolling tachyons and decaying branes, JHEP 0302 (2003) 039, hep-
th/0212248.
[10] M. Gutperle, A. Strominger, Timelike boundary Liouville theory, hep-th/0301038.
[11] A. Maloney, A. Strominger, X. Yin, S-brane thermodynamics, hep-th/0302146.
[12] N. Lambert, H. Liu, J. Maldacena, Closed strings from decaying D-branes, hep-th/0303139.
[13] B. Chen, M. Li, F.L. Lin, Gravitational radiation of rolling tachyon, JHEP 0211 (2002) 050, hep-th/0209222.
[14] S.J. Rey, S. Sugimoto, Rolling tachyon with electric and magnetic fields: T-duality approach, hep-
th/0301049.
[15] N. Moeller, B. Zwiebach, Dynamics with infinitely many time derivatives and rolling tachyons, JHEP 0210
(2002) 034, hep-th/0207107.
[16] S. Sugimoto, S. Terashima, Tachyon matter in boundary string field theory, JHEP 0207 (2002) 025, hep-
th/0205085.
[17] J.A. Minahan, Rolling the tachyon in super-BSFT, JHEP 0207 (2002) 030, hep-th/0205098.
[18] T. Okuda, S. Sugimoto, Coupling of rolling tachyon to closed strings, Nucl. Phys. B 647 (2002) 101, hep-
th/0208196.
[19] J. Kluson, Exact solutions in open bosonic string field theory and marginal deformation in CFT, hep-
th/0209255.
[20] I.Y. Aref’eva, L.V. Joukovskaya, A.S. Koshelev, Time evolution in superstring field theory on non-BPS
brane, I: rolling tachyon and energy–momentum conservation, hep-th/0301137.
[21] A. Ishida, S. Uehara, Rolling down to D-brane and tachyon matter, JHEP 0302 (2003) 050, hep-th/0301179.
[22] S.L. Shatashvili, On field theory of open strings, tachyon condensation and closed strings, hep-th/0105076.
[23] D. Gaiotto, L. Rastelli, A. Sen, B. Zwiebach, Ghost structure and closed strings in vacuum string field theory,
hep-th/0111129.
[24] L. Rastelli, A. Sen, B. Zwiebach, String field theory around the tachyon vacuum, Adv. Theor. Math. Phys. 5
(2002) 353, hep-th/0012251.
[25] C.G. Callan, I.R. Klebanov, A.W. Ludwig, J.M. Maldacena, Exact solution of a boundary conformal field
theory, Nucl. Phys. B 422 (1994) 417, hep-th/9402113.
[26] J. Polchinski, L. Thorlacius, Free fermion representation of a boundary conformal field theory, Phys. Rev.
D 50 (1994) 622, hep-th/9404008.
[27] M.W. Choptuik, Universality and scaling in gravitational collapse of a massless scalar field, Phys. Rev.
Lett. 70 (1993) 9.
[28] R. Gregory, R. Laflamme, Black strings and p-branes are unstable, Phys. Rev. Lett. 70 (1993) 2837, hep-
th/9301052.
[29] I.R. Klebanov, L. Thorlacius, The size of p-branes, Phys. Lett. B 371 (1996) 51, hep-th/9510200.
[30] A. Hashimoto, I.R. Klebanov, Decay of excited D-branes, Phys. Lett. B 381 (1996) 437, hep-th/9604065.
[31] J. Polchinski, String Theory, vol. 1, Cambridge Univ. Press, 1995.
[32] B. Zwiebach, Closed string field theory: quantum action and the B–V master equation, Nucl. Phys. B 390
(1993) 33, hep-th/9206084.
[33] P. Di Vecchia, M. Frau, I. Pesando, S. Sciuto, A. Lerda, R. Russo, Classical p-branes from boundary state,
Nucl. Phys. B 507 (1997) 259, hep-th/9707068.
[34] M. Green, J. Schwarz, E. Witten, Superstring Theory, vol. 1, Cambridge Univ. Press, 1987.
[35] E. Witten, Noncommutative geometry and string field theory, Nucl. Phys. B 268 (1986) 253.
[36] J.A. Shapiro, C.B. Thorn, BRST invariant transitions between closed and open strings, Phys. Rev. D 36
(1987) 432.
[37] J.A. Shapiro, C.B. Thorn, Closed string–open string transitions and Witten’s string field theory, Phys. Lett.
B 194 (1987) 43.
[38] B. Zwiebach, Interpolating string field theories, Mod. Phys. Lett. A 7 (1992) 1079, hep-th/9202015.
[39] A. Hashimoto, N. Itzhaki, Observables of string field theory, JHEP 0201 (2002) 028, hep-th/0111092.
[40] Work in progress.
[41] N. Drukker, Closed string amplitudes from gauge fixed string field theory, hep-th/0207266.
[42] G.T. Horowitz, K. Maeda, Fate of the black string instability, Phys. Rev. Lett. 87 (2001) 131301, hep-
th/0105111.
[43] S.S. Gubser, On non-uniform black branes, Class. Quantum Grav. 19 (2002) 4825, hep-th/0110193.
[44] B. Kol, Topology change in general relativity and the black-hole black-string transition, hep-th/0206220.
[45] S.R. Coleman, E. Weinberg, Radiative corrections as the origin of spontaneous symmetry breaking, Phys.
Rev. D 7 (1973) 1888.
The three-loop splitting functions in QCD:

the non-singlet case
S. Moch a , J.A.M. Vermaseren b , A. Vogt b
a Deutsches Elektronensynchrotron DESY, Platanenallee 6, D-15735 Zeuthen, Germany
b NIKHEF Theory Group, Kruislaan 409, 1098 SJ Amsterdam, The Netherlands
Received 23 March 2004; accepted 31 March 2004
Abstract
We compute the next-to-next-to-leading order (NNLO) contributions to the three splitting
functions governing the evolution of unpolarized non-singlet combinations of quark densities in
perturbative QCD. Our results agree with all partial results available in the literature. We find that
the correct leading logarithmic (LL) predictions for small momentum fractions x do not provide
a good estimate of the respective complete results. A new, unpredicted LL contribution is found
for the colour factor d abc dabc entering at three loops for the first time. We investigate the size
of the corrections and the stability of the NNLO evolution under variation of the renormalization
scale. Except for very small x the corrections are found to be rather small even for large values
of the strong coupling constant, in principle facilitating a perturbative evolution into the sub-GeV
regime.
 2004 Published by Elsevier B.V.
PACS: 12.38.-t; 12.38.Bx; 13.60.Hb
1. Introduction
Parton distributions form indispensable ingredients for the analysis of all hard-scattering
processes involving initial-state hadrons. The dependence of these quantities on the fraction
x of the hadron momentum carried by the quark or gluon cannot be calculated in
perturbation theory. However, the scale-dependence (evolution) of the parton distributions
can be derived from first principles in terms of an expansion in powers of the strong
E-mail address: sven-olaf.moch@desy.de (S. Moch).
0550-3213/$ – see front matter  2004 Published by Elsevier B.V.

doi:10.1016/j.nuclphysb.2004.03.030
102 S. Moch et al. / Nuclear Physics B 688 (2004) 101–134
coupling constant αs . The corresponding nth-order coefficients governing the evolution

are referred to as the n-loop anomalous dimensions or splitting functions. Parton densities
evolved by including the terms up to order αsn+1 in this expansion constitute, together
with the corresponding results for the partonic cross sections for the observable under
consideration, the Nn LO (leading-order, next-to-leading-order, next-to-next-to-leading-
order, etc.) approximation of perturbative QCD.
Presently the next-to-leading-order is the standard approximation for most important
processes. The corresponding one- and two-loop splitting functions have been known for
a long time [1–11]. The NNLO corrections need to be included, however, in order to
arrive at quantitatively reliable predictions for hard processes at present and future high-
energy colliders. These corrections are so far known only for structure functions in deep-
inelastic scattering [12–15] and for Drell–Yan lepton-pair and gauge-boson production in
proton–(anti)proton collisions [16–19] and the related cross sections for Higgs production
in the heavy-top-quark approximation [17,20–22]. Work on NNLO cross sections for jet
production is under way and expected to yield results in the near future, see Ref. [23] and
references therein. For the corresponding three-loop splitting functions, on the other hand,
only partial results have been obtained up to now, most notably the lowest six/seven (even
or odd) integer-N Mellin moments [24–26].
These Mellin moments already provide a rather accurate description of the splitting
functions at large momentum fractions x [25,27–29]. Their much-debated behaviour at
small values of x, on the other hand, can only be determined by a full calculation. As
we will demonstrate below for the non-singlet cases, this statement holds despite the
existence of resummation predictions for the leading small-x logarithms [30,31], since
(a) the correctly predicted logarithms do not dominate the three-loop splitting functions
at any practically relevant value of x and (b) a term of the same size occurs with a new
colour factor at third order which could not have been predicted from lower-order results,
analogous to the situation for the four-loop β-function of QCD [32].
In this article we present the (unpolarized) flavour non-singlet (ns) splitting functions
at the third order in perturbative QCD. The corresponding flavour singlet results will
appear in a forthcoming publication [33]. The present article is organized as follows:
in Section 2 we set up our notations for the three independent third-order splitting
functions and briefly discuss the method of our calculation. The Mellin-N space results
are written down in Section 3 together with their explicit large-N limit which is relevant
for the soft-gluon threshold resummation [34–36] at next-to-next-to leading logarithmic
accuracy [37]. A surprising relation is found between the leading large-N term at two
loops and the subleading (ln N)/N contribution at third order. In Section 4 we present the
exact results as well as compact parametrizations for the x-space splitting functions and
study their behaviour at small x. The numerical implications of these results for the scale
dependence of the non-singlet quark distributions are illustrated in Section 5. Except for
very small values of x, the perturbation series appears to be well-behaved even down to
sub-GeV scales where the initial distributions have been studied using non-perturbative
methods for example in Refs. [38–43]. Finally we briefly summarize our findings in
Section 6.
S. Moch et al. / Nuclear Physics B 688 (2004) 101–134 103
2. Notations and method
We start by setting up our notations for the non-singlet combinations of parton

distributions and the splitting functions governing their evolution. The number distributions
of quarks and antiquarks in a hadron are denoted by qi (x, µ2f ) and q̄i (x, µ2f ), respectively,
where x represents the fraction of the hadron momentum carried by the parton and µf
stand for the factorization scale. There is no need to introduce a renormalization scale µr
different from µf at this point. The subscript i indicates the flavour of the (anti)quark, with
i = 1, . . . , nf for nf flavours of light quarks.
The general structure of the (anti)quark–(anti)quark splitting functions, constrained by
charge conjugation invariance and flavour symmetry, is given by
Pqi qk = Pq̄i q̄k = δik Pqq

v
+ Pqq
s
,
Pqi q̄k = Pq̄i qk = δik Pqvq̄ + Pqsq̄ . (2.1)
In the expansion in powers of αs the flavour-diagonal (‘valence’) quantity Pqq v starts at
first order, while Pq q̄ and the flavour-independent (‘sea’) contributions Pqq and Pqsq̄ are of
v s
order αs2 . A non-vanishing difference Pqq s − P s occurs for the first time at the third order.
q q̄
This general structure leads to three independently evolving types of non-singlet
distributions: the evolution of the flavour asymmetries
±
qns,ik = qi ± q̄i − (qk ± q̄k ) (2.2)
± , is governed by
and of linear combinations thereof, hereafter generically denoted by qns
±
Pns = Pqq
v
± Pqvq̄ . (2.3)
The sum of the valence distributions of all flavours,
nf

v
qns = (qr − q̄r ), (2.4)
r=1
evolves with
s −
v
Pns = Pqq
v
− Pqvq̄ + nf Pqq − Pqsq̄ ≡ Pns + Pns
s
. (2.5)
The first moments of Pns − and P v vanish, since the first moments of the distributions q −
ns ns
v
and qns reflect conserved additive quantum numbers.
We expand the splitting functions in powers of as ≡ αs /(4π), i.e., the evolution
i (x, µ2 ), i = ±, v, are written as
equations for qns f

d αs (µ2f ) n+1 (n)i
2
qns x, µf =
i 2
Pns (x) ⊗ qns
i
x, µ2f , (2.6)
d ln µf 4π
n=0
where ⊗ represents the standard Mellin convolution.

Our calculation is performed in Mellin-N space, i.e., we compute the non-singlet

(n)i
anomalous dimensions γns (N) which are related by the Mellin transformation
1
(n)i
γns (N) = − dx x N−1 Pns
(n)i
(x) (2.7)
0
to the splitting functions discussed above. The relative sign is the standard convention.
Note that in the older literature an additional factor of two is often included in Eq. (2.7).
The calculation follows the methods of Refs. [24–26,44,45]. We employ the optical
theorem and the operator product expansion to calculate Mellin moments of the deep-
inelastic structure functions. Since we treat the Mellin moment N as an analytical
parameter, we cannot apply the techniques of Refs. [24–26], where the M INCER program
[46,47] was used as the tool to solve the integrals. Instead, the introduction of new
techniques was necessary, and various aspects of those have already been discussed in
Refs. [45,48–50]. Here we briefly summarize our approach, focusing on some parts which
have not been presented yet. It should be emphasized that we have at our disposal a very
powerful check on all our derivations and calculations by letting, at any point, N be some
positive integer value. Then we can resort to the approach of Refs. [24–26] and, with the
help of the M INCER program, the checking of all programs greatly simplifies.
We start by constructing the diagrams for the forward Compton reactions
quark(P ) + vector(Q) → quark(P ) + vector(Q), (2.8)

which contribute to the non-singlet structure functions F2 , FL and F3 of deep-inelastic
scattering. The N th Mellin moment is given by the N th derivative with respect to the
quark momentum P at P = 0. The diagrams are generated automatically with the diagram
generator Q GRAF [51] and for all symbolic manipulations we use the latest version of
F ORM [52,53]. The calculation is performed in dimensional regularization [54–57] with
D = 4 − 2. The unrenormalized results in Mellin space are formulae in terms of the
invariants determined by the colour group [58], harmonic sums [6,7,59–61] and the values
ζ3 , ζ4 , ζ5 of the Riemann ζ -function. In physics results the terms with ζ4 cancel in N -
space. With the help of an inverse Mellin transformation the results can be transformed
to harmonic polylogarithms [62–64] in Bjorken-x space. Details have been discussed in
Refs. [45,65]. The renormalization is carried out in the MS-scheme [66,67] as described in
Refs. [24–26].
The complete non-singlet contributions to the structure functions can be obtained from
three Lorentz projections of the amplitude for the process (2.8), that is with g µν , P µ P ν and
with P Qµν ≡ αβµν Pα Qβ . For the projection with g µν and P µ P ν one has two vector-like
couplings, whereas for the projection with P Qµν one has the product of a vector and an
axial-vector coupling. The axial nature leads to the need for additional renormalizations
with ZA , the axial renormalization, and with Z5 , the finite renormalization due to the
treatment of the γ5 . This is all described in the literature [68]. For the anomalous
dimensions we need only the divergent parts of the g µν and P Qµν projections, but just as
for the fixed moments we can also obtain the finite pieces which lead to the coefficient
functions in N3 LO. The determination of the latter for F2 and FL requires also the
computation of the P µ P ν projection which is still in progress. The results for the three-
loop coefficient functions will thus be presented in a future publication [69].
To solve the integrals we apply the following strategy [45,49]. We set up a hierarchy
of classes among all diagrams depending on the topology, for instance ladder, Benz or
non-planar. Within a certain topology, we define a sub-hierarchy depending on the number
of P -dependent propagators. We define basic building blocks (BBBs) as diagrams of a
given topology in which the quark momentum P flows only through a single line in the
diagram, while composite building blocks (CBBs) denote all diagrams with more than one
P -dependent propagator. We determine reduction schemes that map the CBBs of a given
topology class to the BBBs of the same topology class or to simpler CBB topologies.
Subsequently, we use reduction identities that express the BBBs of a given topology class
in terms of BBBs of simpler topologies.
This procedure has been discussed to some extent in Refs. [45,49]. It exploits various
categories of relations between the integrals which can be derived as follows. For a generic
loop integral depending on external momenta P and Q, the first category are integration-
by-parts identities [54,70],

∂ µ
d D pn µ pj × (· · ·) = 0. (2.9)
n
∂p i
These give a number of non-trivial relations by making various choices for the pi and pj
from the loop momenta. Additionally pj can be equal to P or Q. The second category
is based on scaling arguments [45] in Mellin space. They involve applying one of the
operators
∂ ∂ ∂
Qµ , Pµ , Pµ (2.10)
∂Qµ ∂Qµ ∂P µ
both inside the integral and to the integrated result. The scaling in Mellin space tells us
the effect of these operators on the integrated result, while inside the integral we just work
out the derivative. These relations naturally involve polynomials linear in N . The fourth
operator of this kind,
∂
Qµ , (2.11)
∂P µ
cannot be used naively in this context, because it does not commute with the limit
P · P → 0. More care is needed in this case and we will come back to this shortly.
A third category of relations is obtained along the lines of the Passarino–Veltman
decomposition into form factors [71]. In Mellin space we write

µ
d D pn pi × (· · ·) = Qµ IQ + P µ IP , (2.12)
n
where IQ and IP are the two form factors. By contracting Eq. (2.12) either with Qµ
or Pµ , the IQ and IP are determined in terms of a number of integrals. Next, by taking
the derivative with respect to Qµ , the relevant identities can be obtained. Because the
momentum pi can be any of the loop momenta, Eq. (2.12) gives us as many relations as
there are loops. Again, in Mellin space, these relations contain polynomials linear in N .
The fourth and the fifth category of relations are new. Together with the form factor
relations from Eq. (2.12) they were crucial in setting up the reduction scheme for the three-
loop topologies. They are based on operators that do not commute with the limit P · P → 0.
In the fourth category, one considers the dimensionless operators
P ·Q µ ∂
O1 = Q , (2.13)
Q·Q ∂P µ
∂ ∂
O2 = P · Q µ , (2.14)
∂P ∂Qµ
(P · Q)2 ∂ ∂
O3 = . (2.15)
Q · Q ∂P ∂P µ
µ
Each individual operator Oi does not commute with the limit P · P → 0, but certain
linear combinations of the Oi do. However, one has to extend the ansatz based on scaling
arguments in N -space. Specifically, one has for the N th moment of an integral I (N)

2P · Q N (0)
I (N) = (Q · Q)α CN
Q·Q

2P · Q N−2 P · P (2)
+ (Q · Q)α CN + · · · , (2.16)
Q·Q Q·Q
(0) (2)
where the CN and CN are dimensionless functions of N , and α adjusts the mass
(2)
dimensions. The novel feature is here the term CN proportional to P · P , which one
may call higher twist. In contrast, for the relations based on Eq. (2.10) it was sufficient to
(0)
restrict the ansatz to CN .
Applying the differential operators Oi in Eqs. (2.13)–(2.15) to the ansatz (2.16), one
finds that the combinations
2(α + 1 − N)O1 − O2 , (2N − 4 + D)O1 − O3 (2.17)

do commute with the limit P · P → 0. That is to say, any dependence on the higher twist
(2) (0)
term CN vanishes in this limit and one is left with only contributions from CN . Eq. (2.17)
adds two more relations, which in Mellin space contain quadratic polynomials in N due
to the differential operators of second order. We have checked that differential operators of
yet a higher order in P and Q do not add any new information.
Finally, the fifth category of relations again uses the form factor approach of Eq. (2.12).
However, now we do not take the derivative with respect to Qµ but with respect to Pµ .
Some extra book-keeping is needed here, since one has to take along terms proportional to
P · P . Let us write Eq. (2.12) as
µ
pi I = Qµ IQ + P µ IP . (2.18)
Taking the derivative of Eq. (2.18) with respect to Pµ in N -space one finds
∂ µ ∂
µ
pi I = Qµ µ IQ + (D + N − 1)IP . (2.19)
∂P ∂P
Solving Eq. (2.18) for IQ and IP as above, however keeping all terms P · P , substituting
into Eq. (2.19) and finally taking the limit P · P → 0, we find
∂ µ P · pi µ ∂ Q · pi P · Q − Q · QP · pi
p I= Q I + (D + N − 2) I. (2.20)
∂P µ i P ·Q ∂P µ (P · Q)2
Again, as the momentum pi can be any of the loop momenta, Eq. (2.20) gives us as many
relations with polynomials linear in N as there are loops.
Taken together, the reductions of category one to five suffice to obtain a complete
reduction scheme. In particular, the reduction equations of category two to five involve
explicitly the parameter N of the Mellin moment. They give rise to difference equations
in N for an integral I (N),
a0 (N)I (N) + a1 (N)I (N − 1) + · · · + am (N)I (N − m) = G(N), (2.21)
in which the function G refers to a combination of integrals of simpler topologies. Zeroth

order equations are of course trivial, although sometimes the function G can contain
thousands of terms. First order difference equations can be solved analytically in a closed
form, introducing one sum. Higher order difference equations on the other hand can be
solved constructively, sometimes with considerable effort, by making an ansatz for the
solution in terms of harmonic sums. For the present calculation we had to go up to fourth
order for certain types of integrals.
Due to the difference equations, which have to be solved in a successive way, a strict
hierarchy for topology classes is introduced in the reduction scheme. For a given integral I ,
a difference equation as in Eq. (2.21), with some (often lengthy) function G expressed in
terms of harmonic sums, can be solved in terms of harmonic sums again. Subsequently,
the result for I can be part of the inhomogeneous term in a difference equation for another,
more complicated integral. This requires the tabulation of a large number CBB and BBB
integrals, because each integral is typically used many times, thus it saves computer time
and disk space. Only this tabulation, which required the addition of features to F ORM [53],
renders the calculation feasible with current computing resources. For the complete project,
including Refs. [33,69], we have collected tablebases with more than 100 000 integrals and
a total size of tables of more than 3 GByte.
3. Results in Mellin space
±,s
Here we present the anomalous dimensions γns (N) in the MS-scheme up to the
third order in the running coupling constant αs , expanded in powers of αs /(4π). These
quantities can be expressed in terms of harmonic sums [6,7,59,60]. Following the notation
of Ref. [59], these sums are recursively defined by

M
(±1)i
S±m (M) = (3.1)
im
i=1
and

M
(±1)i
S±m1 ,m2 ,...,mk (M) = Sm2 ,...,mk (i). (3.2)
i m1
i=1
The sum of the absolute values of the indices mk defines the weight of the harmonic sum.
In the n-loop anomalous dimensions written down below one encounters sums up to weight
2n − 1.
In order to arrive at a reasonably compact representation of our results, we employ the
abbreviation Sm ≡ Sm (N) in what follows, together with the notation
N± Sm = Sm (N ± 1), N±i Sm = Sm (N ± i) (3.3)
for arguments shifted by ±1 or a larger integer i. In this notation the well-known one-loop
(LO) anomalous dimension [1,2] reads

(0)
γns (N) = CF 2(N− + N+ )S1 − 3 , (3.4)
and the corresponding two second-order (NLO) non-singlet quantities [4,6] are given by

17 28
γns (N) = 4CA CF 2N+ S3 −
(1)+
− 2S−3 − S1
24 3

151 11
+ (N− + N+ ) S1 + 2S1,−2 − S2
18 6

1 4 11 1
+ 4CF nf + S1 − (N− + N+ ) S1 − S2
12 3 9 3

3
+ 4CF 2 4S−3 + 2S1 + 2S2 − + N− [S2 + 2S3 ]
8

− (N− + N+ )[S1 + 4S1,−2 + 2S1,2 + 2S2,1 + S3 ] , (3.5)

CA
(1)−
γns (N) = γns
(1)+
(N) + 16CF CF −
2

× (N− − N+ )[S2 − S3 ] − 2(N− + N+ − 2)S1 . (3.6)
+ (N)
The three-loop (NNLO, N2 LO) contribution to the anomalous dimension γns
corresponding to the upper sign in Eq. (2.3) reads

3 5 10 10 4 2
(2)+
γns (N) = 16CA CF nf ζ3 − + S−3 − S3 + S1,−2 − S−4 + 2S1,1
2 4 9 9 3 3

25 257 2 2 2
− S2 + S1 − S−3,1 − N+ S2,1 − S3,1 − S4
9 27 3 3 3

23
− (N+ − 1) S3 − S2
18

1237 11 317 16
− (N− + N+ ) S1,1 + S1 + S3 − S2 + S1,−2
216 18 108 9

2 1 1 1 1 1
− S1,−2,1 − S1,−3 − S1,3 − S2,1 − S2,−2 + S1 ζ3 + S3,1
3 3 2 2 3 2

1657 15 31 67
+ 16CF CA2 − ζ3 + 2S−5 + S−4 − 4S−4,1 − S−3
576 4 6 9
11 3
+ 2S−3,−2 + S−3,1 + S−2 − 6S−2 ζ3 − 2S−2,−3 + 3S−2,−2
3 2
1883 16
− 4S−2,−2,1 + 8S−2,1,−2 − S1 − 10S1,−3 − S1,−2 + 12S1,−2,1
54 3
5 1 176 13
+ 4S1,3 − 4S2,−2 − S4 + S5 + S2 + S3
2 2 9 3
+ (N− + N+ − 2)[3S1 ζ3 + 11S1,1 − 4S1,1,−2 ]

9737 19 91
+ (N− + N+ ) S1 − 3S1,−4 + S1,−3 + 8S1,−3,1 + S1,−2
432 6 9
29 19
− 6S1,−2,−2 − S1,−2,1 + 8S1,1,−3 − 16S1,1,−2,1 − 4S1,1,3 − S1,3
3 4
11
+ 4S1,3,1 + 3S1,4 + 8S2,−2,1 + 2S2,3 − S3,−2 + S3,1 − S4,1
12

1 1967 121
− 4S2,−3 + S2,−2 − S2 + S3
6 216 72

1
− (N− − N+ ) 3S2 ζ3 + 7S2,1 − 3S2,1,−2 + 2S2,−2,1 − S2,3
4

3 29 11 1
− S3,−2 − S3,1 + S4,1 + S2,−3 − S2,−2
2 6 4 2

28 2376 8 5
+ N+ S3 − S2 − S4 − S5
9 216 3 2

17 13 2
+ 16CF n2f − S1 + S2
144 27 9

2 11 1
+ (N− + N+ ) S1 − S2 + S3
9 54 18

45 151 89 134
+ 16CF 2 CA ζ3 − − 10S−5 − S−4 + 20S−4,1 + S−3
4 64 6 9
31 9
− 2S−3,−2 − S−3,1 + 2S−3,2 − S−2 + 18S−2 ζ3 + 10S−2,−3
3 2
26
− 6S−2,−2 + 8S−2,−2,1 − 28S−2,1,−2 + 46S1,−3 + S1,−2
3
28 185
− 48S1,−2,1 + S1,2 − S3 − 8S1,3 + 2S3,−2 − 4S5
3 6

133 209 242
− (N− + N+ − 2) 9S1 ζ3 − S1 + S1,1 − 14S1,1,−2 − S2
36 6 18

33 14
+ 9S2,−2 + S4 − 3S3,1 + S2,1
4 3

107 173
+ (N− + N+ ) 17S1,−4 − S1,−3 − 32S1,−3,1 − S1,−2
6 9
103
+ 16S1,−2,−2 + S1,−2,1 − 2S1,−2,2 − 36S1,1,−3 + 56S1,1,−2,1
3
109 43 11
+ 8S1,1,3 − S1,2 − 4S1,2,−2 + S1,3 − 8S1,3,1 − 11S1,4 + S2,2
9 3 3
+ 21S2,−3 − 30S2,−2,1 − 4S2,1,−2 − 5S2,3 − S4,1

31 67
+ S2,−2 − S2,1
6 9

+ (N− − N+ ) 9S2 ζ3 + 2S2,−3 + 4S2,−2,1 − 12S2,1,−2 − 2S2,3

1 11 33 59 127 1153
+ 13S4,1 + S2,−2 + S4 − S3,1 + S3 + S2,1 − S2
2 2 2 9 6 72

4 23 73 151
+ N+ 8S3,−2 + S3,1 − 2S3,2 + 14S5 + S4 + S3 + S2
3 6 3 24

23 3 4 59 4
+ 16CF 2 nf − ζ3 + S−3,1 − S2 + S−4
16 2 3 36 3
20 20 8 8 4
− S−3 + S1 − S1,−2 − S1,1 − S1,2
9 9 3 3 3

25 4 1 67 4 4
+ N+ S3 − S3,1 − S4 − (N+ − 1) S2 − S2,1 + S3
9 3 3 36 3 3

325 2 32 4
+ (N− + N+ ) S1 ζ3 − S1 − S1,−3 + S1,−2 − S1,−2,1
144 3 9 3
4 16 4 11 2 10 1
+ S1,1 + S1,2 − S1,3 + S2 − S2,−2 + S2,1 + S4
3 9 3 18 3 9 2

2 8
− S2,2 − S3
3 9

29 15
+ 16CF 3 12S−5 − − ζ3 + 9S−4 − 24S−4,1 − 4S−3,−2 + 6S−3,1
32 2
− 4S−3,2 + 3S−2 + 25S3 − 12S−2 ζ3 − 12S−2,−3 + 24S−2,1,−2
67
− 52S1,−3 + 4S1,−2 + 48S1,−2,1 − 4S3,−2 + S2 − 17S4
2

31
+ (N− + N+ − 2) 6S1 ζ3 − S1 + 35S1,1 − 12S1,1,−2 + S1,2
8

+ 10S2,−2 + S2,1 + 2S2,2 − 2S3,1 − 3S5

+ (N− + N+ ) 23S1,−3 − 22S1,−4 + 32S1,−3,1 − 2S1,−2 − 8S1,−2,−2
− 30S1,−2,1 − 6S1,3 + 4S1,−2,2 + 40S1,1,−3 − 48S1,1,−2,1 + 8S1,2,−2

+ 4S1,2,2 + 8S1,3,1 + 4S1,4 + 28S2,−2,1 + 4S2,1,2 + 4S2,2,1 + 4S3,1,1
− 4S3,2 + 8S2,1,−2 − 26S2,−3 − 2S2,3 − 4S3,−2 − 3S2,−2

3
− 3S2,2 + S4
2

81
+ (N− − N+ ) 12S2,1,−2 − 6S2 ζ3 − 2S2,−3 + 3S2,3 + 2S3,−2 − S2,1
4

1 15 1
+ 14S3,1 − 5S2,−2 − S2,2 + S2 + S3 − 13S4,1 + 4S5
2 8 2

265 87
+ N+ 14S4 − S2 − S3 − 4S4,1 − 4S5 . (3.7)
8 4
− (N) corresponding to the lower sign
The third-order result for the anomalous dimension γns
in Eq. (2.3) is given by

CA 367
(2)−
γns (N) = γns(2)+
(N) + 16CA CF CF − (N− + N+ − 2) S1 + 12S1 ζ3
2 18
140
+ 2S1,−2 + 4S1,−3 + 8S1,−2,1 + S1,1 − 16S1,1,−2
3
− S5 − 8S3,1 − S4

70
+ (N− − N+ ) 4S5 − 12S2 ζ3 − 4S2,−3 − 8S2,−2,1 − S2,1
3
70 13 41
+ 16S2,1,−2 + 4S3,−2 − 8S4,1 + S3,1 + S4 − S2
3 3 18
152
+ 2S2,−2 − S3
9

+ 4(N+ − 1) 4S2,−2 − 8S2 − S3

CA 61 8
+ 16CF nf CF − (N− + N+ − 2) S1 − S1,1
2 9 3

4 41 38 4 4
+ (N− − N+ ) S2,1 − S2 + S3 − S3,1 − S4
3 9 9 3 3

CA
+ 16CF 2 CF − (N− + N+ − 2)[8S1,−2 − 15S1 − 12S1 ζ3
2
− 12S1,−3 − 60S1,1 + 24S1,1,−2 + 8S1,2 + 40S2 − 12S2,−2 + 8S2,1
+ 7S3 + 12S3,1 + 6S5 ]
+ (N− − N+ )[12S2 ζ3 − 24S2 + 12S2,−3 + 8S2,−2 + 30S2,1
− 24S2,1,−2 − 4S2,2 − 15S3 − 38S3,1 + 4S3,2 + 24S4,1 − 12S5 ]

− (N+ − 1)[8S3,−2 + 26S4 ] . (3.8)
s (N) corresponding to the last term in Eq. (2.5) starts at three loops
Finally the quantity γns
with

d abc dabc 25 11 5 1
γns (N) = 16nf
(2)s
(N− + N+ ) S1 + S1,−3 − S1,−2,1 − S1,1,−2
nc 3 12 3 6

13 91 3 1 91
+ (N− + N+ − 2) S1,−2 + S1,1 − S1,3 − S2,−2 − S2
12 24 8 4 48

3 5
+ S3 + S3,1
16 8
2
+ (N+ − N+2 )[S4 + S2,−2 − S3,1 ]
3
2
− (N−2 + N+2 )[S1,−3 − S1,−2,1 − S1,1,−2 ]
3
1 1
+ (N− − 1) S4 + S5
4 2

1 1 109 41 67
+ (N− − N+ ) S2,−3 + S2,−2 − S2 − S2,1 + S3
2 2 48 24 48

1 1 1 3
− S3,1 − S2,1,−2 + S2,3 + S3,−2 − S4,1
2 4 2 4

50 1
− S1 − S1,−3 + 2S1,−2,1 − S1,1,−2 . (3.9)
3 2
Eqs. (3.7)–(3.9) represent new results of this article, with the exception of the (identical)
n2f parts of Eqs. (3.7) and (3.8) which have been obtained by Gracey in Ref. [72] and of the
contribution linear in nf in Eq. (3.7) which we have published before [49]. All our results
agree with the fixed moments determined before using the M INCER program [46,47], i.e.,
Eq. (3.7) reproduces the even moments N = 2, . . . , 14 computed in Refs. [24–26], while
Eqs. (3.8) and (3.9) reproduce the odd moments N = 1, . . . , 13 also obtained in Ref. [26].
The results (3.4), (3.5) and (3.7) for γns+ (N) are assembled in Fig. 1 for four active
flavours and a typical value αs = 0.2 for the strong coupling constant (recall that the terms
up to order αsn+1 are included at Nn LO). Numerically, the colour factors take the values
CF = 4/3, CA = 3 and d abc dabc /nc = 40/9. Note that the latter normalization is different
from that employed in Ref. [58].
The NNLO corrections are rather small under these circumstances, amounting to less
than 2% for N 2. At large N the anomalous dimensions behave as

ln N 1
γns(n−1)±,v
(N) = An (ln N + γe ) − Bn − Cn +O , (3.10)
N N
where γe is the Euler–Mascheroni constant and the coefficients are specified in the next
+ = γ + / ln N , also shown in Fig. 1, approaches a constant for N → ∞.
paragraph. Thus γ̃ns ns
(n)+ (n)+
The asymptotic results are indicated by replacing γ̃ns (N = 15) by γ̃ns (N → ∞) for
+
Fig. 1. The perturbative expansion of the anomalous dimension γns (N ) for four flavours at αs = 0.2. In the right
part the leading N -dependence for large N has been divided out, and the corresponding asymptotic limits are
indicated as discussed in the text.
the respective highest term included in the curves (e.g., for n = 2 at NNLO). Obviously
the approach to the asymptotic limit is very slow. Yet the results at N → ∞, which can be
derived much easier than the full N -dependence [73], do provide a reasonable first estimate
of the corrections.
The leading large-N coefficients An , which are also relevant for the soft-gluon
(threshold) resummation [34–37], are given by
A1 = 4CF ,

67 5
A2 = 8CF − ζ2 CA − nf ,
18 9

245 67 11 11 55
A3 = 16CF CA2 − ζ2 + ζ3 + ζ22 + 16CF2 nf − + 2ζ3
24 9 6 5 24

209 10 7 1
+ 16CF CA nf − + ζ2 − ζ3 + 16CF nf −
2
. (3.11)
108 9 3 27
The nf -independent contribution to the three-loop coefficient A3 is also a new result of
the present article. Inserting the numerical values of the ζ -function and the QCD colour
factors it reads A3 |nf =0 ∼
= 1174.898, in agreement with the previous numerical estimate
of Ref. [37]. The constants Bn can be read off directly from the terms with δ(1 − x) in
Eqs. (4.5), (4.6) and (4.9). Surprisingly, the coefficients Cn in Eq. (3.10), which are also
best determined using those x-space results, turn out to be related to the An by
C1 = 0, C2 = 4CF A1 , C3 = 8CF A2 . (3.12)

Especially the relation for C3 is very suggestive and seems to call for a structural
explanation.
4. Results in x-space
(n)±,s
The splitting functions Pns (x) are obtained from the N -space results of the previous
section by an inverse Mellin transformation, which expresses these functions in terms
of harmonic polylogarithms [62–64]. The inverse Mellin transformation exploits an
isomorphism between the set of harmonic sums for even or odd N and the set of harmonic
polylogarithms. Hence it can be performed by a completely algebraic procedure [45,64],
based on the fact that harmonic sums occur as coefficients of the Taylor expansion of
harmonic polylogarithms.
Our notation for the harmonic polylogarithms Hm1 ,...,mw (x), mj = 0, ±1 follows
Ref. [64] to which the reader is referred for a detailed discussion. The lowest-weight
(w = 1) functions Hm (x) are given by
H0 (x) = ln x, H±1 (x) = ∓ ln(1 ∓ x). (4.1)

The higher-weight (w 2) functions are recursively defined as

1
lnw x, if m1 , . . . , mw = 0, . . . , 0,
Hm1 ,...,mw (x) = w!x (4.2)
0 dz fm1 (z)Hm2 ,...,mw (z), else
with
1 1
f0 (x) = , f±1 (x) = . (4.3)
x 1∓x
A useful short-hand notation is
H0,...,0,±1,0,...,0,±1,... (x) = H±(m+1),±(n+1),... (x). (4.4)

m n
For w 3 the harmonic polylogarithms can be expressed in terms of standard polyloga-

rithms; a complete list can be found in Appendix A of Ref. [45]. All harmonic polyloga-
rithms of weight w = 4 in this article can be expressed in terms of standard polylogarithms,
Nielsen functions [74] or, by means of the defining relation (4.2), as one-dimensional inte-
grals over these functions. A F ORTRAN program for the functions up to weight w = 4 has
been provided in Ref. [75].
For completeness we recall the one- and two-loop non-singlet splitting functions [3,8]

Pns(0)
(x) = CF 2pqq (x) + 3δ(1 − x) (4.5)
and

67 11
(1)+
Pns (x) = 4CA CF pqq (x) − ζ2 + H0 + H0,0
18 6
14
+ pqq (−x)[ζ2 + 2H−1,0 − H0,0] + (1 − x)
3

17 11
+ δ(1 − x) + ζ2 − 3ζ3
24 3

5 1 2 1 2
− 4CF nf pqq (x) + H0 + (1 − x) + δ(1 − x) + ζ2
9 3 3 12 3

3
+ 4CF 2 2pqq (x) H1,0 − H0 + H2
4
− 2pqq(−x)[ζ2 + 2H−1,0 − H0,0 ]

3
− (1 − x) 1 − H0 − H0 − (1 + x)H0,0
2

3
+ δ(1 − x) − 3ζ2 + 6ζ3 , (4.6)
8

CA
(1)−
Pns (x) = Pns
(1)+
(x) + 16CF CF − pqq (−x)[ζ2 + 2H−1,0 − H0,0 ]
2

− 2(1 − x) − (1 + x)H0 . (4.7)
Here and in Eqs. (4.9)–(4.11) we suppress the argument x of the polylogarithms and use
pqq (x) = 2(1 − x)−1 − 1 − x. (4.8)

All divergences for x → 1 are understood in the sense of ‘+’-distributions.
The three-loop splitting function for the evolution of the ‘plus’ combinations of quark
densities in Eq. (2.2), corresponding to the anomalous dimension (3.8) reads

1 10 209 167
(2)+
Pns (x) = 16CA CF nf pqq (x) ζ2 − − 9ζ3 − H0 + 2H0 ζ2 − 7H0,0
6 3 36 18

− 2H0,0,0 + 3H1,0,0 − H3

1 3 5 10
+ pqq (−x) ζ3 − ζ2 − H−2,0 − 2H−1 ζ2 − H−1,0 − H−1,0,0
3 2 3 3

1 5
+ 2H−1,2 + H0 ζ2 + H0,0 + H0,0,0 − H3
2 3

1 257 43 1
+ (1 − x) ζ2 − − H0 − H0,0 − H1
6 54 18 6

2 1 1 1
− (1 + x) H−1,0 + H2 + ζ2 + H0 + H0,0
3 2 3 6

5 167 1 25
+ δ(1 − x) − ζ2 + ζ22 + ζ3
4 54 20 18

5 69
+ 16CA CF 2 pqq (x) ζ3 − ζ22 − H−3,0 − 3H−2 ζ2 − 14H−2,−1,0
6 20
151 41 17
+ 3H−2,0 + 5H−2,0,0 − 4H−2,2 − H 0 + H 0 ζ2 − H 0 ζ3
48 12 2
13 23 2
− H0,0 − 4H0,0 ζ2 − H0,0,0 + 5H0,0,0,0 + H3 − 24H1ζ3
4 12 3
67 31
− 16H1,−2,0 + H1,0 − 2H1,0ζ2 + H1,0,0 + 11H1,0,0,0 + 8H1,1,0,0
9 3

67 11
− 8H1,3 + H4 + H2 − 2H2 ζ2 + H2,0 + 5H2,0,0 + H3,0
9 3

1 67 31
+ pqq (−x) ζ22 − ζ2 + ζ3 + 5H−3,0 − 32H−2 ζ2 − 4H−2,−1,0
4 9 4
31 31 9
− H−2,0 + 21H−2,0,0 + 30H−2,2 − H−1 ζ2 − 42H−1ζ3 + H0
6 3 4
− 4H−1,−2,0 + 56H−1,−1ζ2 − 36H−1,−1,0,0 − 56H−1,−1,2
134 31
− H−1,0 − 42H−1,0ζ2 − H3,0 + 32H−1,3 − H−1,0,0
9 6
31 13 29 67
+ 17H−1,0,0,0 + H−1,2 + 2H−1,2,0 + H0 ζ2 + H0 ζ3 + H0,0
3 12 2 9

89 31
+ 13H0,0ζ2 + H0,0,0 − 5H0,0,0,0 − 7H2 ζ2 − H3 − 10H4
12 6

133 167
+ (1 − x) + 4H0,0,0,0 − ζ3 − 2H0ζ3 − 2H−3,0 + H−2 ζ2
36 4
77 209
+ 2H−2,−1,0 − 3H−2,0,0 + H0,0,0 − H1 − 7H1 ζ2
4 6

14
+ 4H1,0,0 + H1,0
3

43 25
+ (1 + x) ζ2 − 3ζ22 + H−2,0 − 31H−1ζ2 − 14H−1,−1,0
2 2
13 55 1457
− H−1,0 + 24H−1,2 + 23H−1,0,0 + H0 ζ2 + 5H0,0ζ2 + H0
3 2 48

1025 155
− H0,0 − H2 + H2 ζ2 − 15H3 + 2H2,0,0 − 3H4
36 6
1 2 37 242
− 5ζ2 − ζ2 + 50ζ3 − 2H−3,0 − 7H−2,0 − H0 ζ3 − H0 ζ2 − H0
2 2 9
185 28
− 2H0,0ζ2 + H0,0 − 22H0,0,0 − 4H0,0,0,0 + H2 + 6H3
6 3

151 205 247 2 211 15
+ δ(1 − x) + ζ2 ζ3 − ζ2 − ζ2 + ζ3 + ζ5
64 24 60 12 2

245 67 12 1 1043
+ 16CA2 CF pqq (x) − ζ2 + ζ22 + ζ3 + H0 + H−3,0
48 18 5 2 216
3 31
+ 4H−2,−1,0 − H−2,0 − H−2,0,0 + 2H−2,2 − H0 ζ2 + 4H0ζ3
2 12
389
+ H0,0 − 2H2,0,0 − H0,0,0,0 + 9H1 ζ3 + 6H1,−2,0 − H1,0 ζ2
72
11 31
− H1,0,0 − 3H1,0,0,0 − 4H1,1,0,0 + 4H1,3 + H0,0,0
4 12

11
+ H3 + H4
12

67 11 11
+ pqq (−x) ζ2 − ζ22 − ζ3 − H−3,0 + 8H−2ζ2 + H−2,0
18 4 6
11
− 4H−2,0,0 − 3H−1,0,0,0 + H−1 ζ2 + 12H−1ζ3 − 16H−1,−1ζ2
3
67
+ 8H−1,−1,0,0 + 16H−1,−1,2 + H−1,0 − 8H−2,2 + 11H−1,0ζ2
9
11 11 3 1
+ H−1,0,0 − H−1,2 − 8H−1,3 − H0 − H0 ζ2 − 4H0 ζ3
6 3 4 6

67 31 11
− H0,0 − 3H0,0 ζ2 − H0,0,0 + H0,0,0,0 + 2H2ζ2 + H3 + 2H4
18 12 6

1883 1 1
+ (1 − x) − H0,0,0,0 + 11H1 − H−2,−1,0 + H−3,0
108 2 2
1 1 523 13 5
− H−2 ζ2 + H−2,0,0 + H0 + H0 ζ3 − H0,0 − H0,0,0
2 2 36 3 2

+ 2H1 ζ2 − 2H1,0,0

8
+ (1 + x) 8H−1 ζ2 + 4H−1,−1,0 + H−1,0 − 5H−1,0,0 − 6H−1,2
3
13 3 43 5 11 1
− ζ2 + ζ22 − ζ3 − H−2,0 − H0 ζ2 − H2 ζ2
3 8 4 2 2 2

5 1 3
− H0,0 ζ2 + 7H2 − H2,0,0 + 3H3 + H4
4 4 4
1 1 2 8 17 19 5
+ H0,0 ζ2 + ζ2 − ζ2 + ζ3 + H−2,0 − H0 + H0 ζ2 − H0 ζ3
2 4 3 2 2 2
13 5 1
+ H0,0 + H0,0,0 + H0,0,0,0
3 2 2

1657 281 1 97 5
− δ(1 − x) − ζ2 + ζ22 + ζ3 − ζ5
576 27 8 9 2

1 1 5
+ 16CF n2f pqq (x) H0,0 − + H0
18 3 3

13 1 17 5 1
+ (1 − x) + H0 − δ(1 − x) − ζ2 + ζ3
54 9 144 27 9

1 55 5
+ 16CF 2 nf pqq (x) 5ζ3 − 4H1,0,0 − + H 0 + H 0 ζ2
3 16 8

3 10 10
+ H0,0 − H0,0,0 − H1,0 − H2 − 2H2,0 − 2H3
2 3 3

2 5 3 10
+ pqq (−x) ζ2 − ζ3 + H−2,0 + 2H−1 ζ2 + H−1,0 + H−1,0,0
3 3 2 3

1 5
− 2H−1,2 − H0 ζ2 − H0,0 − H0,0,0 + H3
2 3

10 19 4 2 4
− (1 − x) + H0,0 − H1 + H1,0 + H2
9 18 3 3 3

4 25 1 2 7
+ (1 + x) H−1,0 − H0 + H0,0,0 + H0 + H0,0
3 24 2 9 9

4 23 5 29 17
+ H2 − δ(1 − x) − ζ2 − ζ22 + ζ3
3 16 12 30 6

9 2
+ 16CF 3 pqq (x) ζ − 2H−3,0 + 6H−2 ζ2 + 12H−2,−1,0
10 2
3 3 13
− 6H−2,0,0 − H0 − H0 ζ2 + H0 ζ3 + H0,0 − 2H0,0,0,0 + 8H1,3
16 2 8
+ 12H1ζ3 + 8H1,−2,0 − 6H1,0,0 − 4H1,0,0,0 + 4H1,2,0 − 3H2,0

+ 2H2,0,0 + 4H2,1,0 + 4H2,2 + 4H3,0 + 4H3,1 + 2H4

7 9
+ pqq (−x) ζ22 − ζ3 − 6H−3,0 + 32H−2ζ2 + 8H−2,−1,0 + 3H−2,0
2 2
− 26H−2,0,0 − 28H−2,2 + 6H−1 ζ2 + 36H−1 ζ3 + 8H−1,−2,0
− 48H−1,−1ζ2 + 40H−1,−1,0,0 + 48H−1,−1,2 + 40H−1,0ζ2 + 3H−1,0,0
3 3
− 22H−1,0,0,0 − 6H−1,2 − 4H−1,2,0 − 32H−1,3 − H0 − H0 ζ2
2 2
9
− 13H0ζ3 − 14H0,0ζ2 − H0,0,0 + 6H0,0,0,0 + 6H2 ζ2 + 3H3
2
+ 2H3,0 + 12H4

31
+ (1 − x) 2H−3,0 − + 4H−2,0,0 + H0,0 ζ2 − 3H0,0,0,0 + 35H1
8

5
+ 6H1 ζ2 − H1,0 + H2,0
2

37 2 93 81
+ (1 + x) ζ2 − ζ2 − ζ3 − 15H−2,0 + 30H−1ζ2 + 12H−1,−1,0
10 4 2
539 191
− 2H−1,0 − 26H−1,0,0 − 24H−1,2 − H0 − 28H0ζ2 + H0,0
16 8
85
+ 20H0,0,0 + H2 − 3H2,0,0 − 2H3,0 + 13H3 − H4
4
67
+ 4ζ2 + 33ζ3 + 4H−3,0 + 10H−2,0 + H0 + 6H0ζ3 + 19H0ζ2
2
− 25H0,0 − 17H0,0,0 − 2H2 − H2,0 − 4H3

29 9 18 2 17
+ δ(1 − x) − 2ζ2 ζ3 + ζ2 + ζ2 + ζ3 − 15ζ5 . (4.9)
32 8 5 4
The x-space counterpart of Eq. (3.8) for the evolution of the ‘minus’ combinations (2.2) is
given by

CA 134
(2)−
Pns (x) = Pns
(2)+
(x) + 16CA CF CF − pqq (−x) ζ2 − 4ζ22 − 11ζ3
2 9
22 44
− 4H−3,0 + 32H−2 ζ2 + H−2,0 − 16H−2,0,0 − 32H−2,2 + H−1 ζ2
3 3
+ 48H−1ζ3 − 64H−1,−1ζ2 + 32H−1,−1,0,0 + 64H−1,−1,2
268 22 44
+ H−1,0 + 44H−1,0ζ2 + H−1,0,0 − 12H−1,0,0,0 − H−1,2
9 3 3
2 134
− 32H−1,3 − 3H0 − H0 ζ2 − 16H0ζ3 − H0,0 − 12H0,0ζ2
3 9
31 22
− H0,0,0 + 4H0,0,0,0 + 8H2 ζ2 + H3 + 8H4
3 3

367 1 2
+ (1 − x) + ζ2 + 2H−3,0 − 2H−2 ζ2 − 4H−2,−1,0 − 10H−2,0
18 2

140
− 2H0,0 + 2H−2,0,0 + 2H0 ζ3 + H0,0 ζ2 − H0,0,0,0 + 8H1ζ2 + H1
3

26
+ (1 + x) 32H−1ζ2 − 18ζ2 − 23ζ3 + H−1,0 − 16H−1,0,0
3

481 70
− 32H−1,2 − H0 − 29H0 ζ2 + 5H0,0,0 + 24H3 + H2
18 3

− 2ζ2 − 2ζ3 + 32H0 + 14H0ζ2 + 2H0,0,0 − 16H3

CA 20 4 8
+ 16CF nf CF − pqq (−x) 2ζ3 − ζ2 − H−2,0 − H−1 ζ2
2 9 3 3
40 4 8 2 20
− H−1,0 − H−1,0,0 + H−1,2 + H0 ζ2 + H0,0
9 3 3 3 9
4 4
+ H0,0,0 − H3
3 3

61 8 8 41 4
+ (1 − x) − H1 + (1 + x) 2H0,0 − H−1,0 + H0 − H2
9 3 3 9 3

C A
+ 16CF 2 CF − pqq (−x)[9ζ3 − 7ζ22 + 12H−3,0 − 64H−2ζ2
2
− 16H−2,−1,0 − 6H−2,0 + 52H−2,0,0 + 56H−2,2 − 12H−1 ζ2
− 72H−1ζ3 − 16H−1,−2,0 + 96H−1,−1ζ2 − 80H−1,−1,0,0
− 96H−1,−1,2 − 80H−1,0ζ2 − 6H−1,0,0 + 44H−1,0,0,0 + 12H−1,2
+ 8H−1,2,0 + 64H−1,3 + 3H0 + 3H0 ζ2 + 26H0ζ3 + 28H0,0ζ2
+ 9H0,0,0 − 12H0,0,0,0 − 12H2ζ2 − 6H3 − 4H3,0 − 24H4]
− (1 − x)[15 + 8H−3,0 + 8H−2,0,0 + 61H0 + 6H0 ζ3 + 2H0,0ζ2
− 6H0,0,0,0 + 12H1ζ2 + 60H1 + 8H1,0]
+ (1 + x)[24ζ2 + 57ζ3 + 10H−2,0 − 48H−1 ζ2 − 4H−1,0 + 40H−1,0,0
+ 48H−1,2 + 59H0ζ2 − 22H0,0 − 35H0,0,0 − 22H2 − 4H2,0 − 44H3]
+ 8ζ2 − 42ζ3 − 4H−2,0 + 42H0

− 38H0ζ2 + 14H0,0 − 16H2 + 26H0,0,0 + 24H3 . (4.10)
(2)s
Finally the Mellin inversion of γns (N) in Eq. (3.9) leads to the following result for the
(2)s
leading (third-order) difference Pns (x) of the ‘valence’ and ‘minus’ splitting functions:

d abc dabc 1 50 41 5
(2)s
Pns (x) = 16nf (1 − x) + ζ2 − ζ22 − H−3,0 + H−2 ζ2
nc 2 3 12 4
9 3 1
− H−2,0,0 + H3 + 2H−2,−1,0 + H0,0ζ2 − H1 ζ2
4 2 2
3 91
− H1,0,0 + H1
4 12

1 3 3 13 1
+ (1 + x) H−1,−1,0 − H−1 ζ2 + H0 − H−1,0 + H−1,0,0
2 2 4 6 2
3 9 29 41
+ 2H−1,2 − H−2,0 + H0 ζ2 + H0,0 + H2 − H2 ζ2
2 4 12 12
1 3
− H2,0,0 + H4
2 2

1 1
− + x 2 [3H−1 ζ2 + 2H−1,−1,0 − 2H−1,0,0 − 2H−1,2 + H1 ζ2 ]
3 x
1
+ x 2 [5ζ3 − 2H3 + 2H−2,0 + 4H0ζ2 − 2H0,0,0 + 2H1 ζ2 ]
3
91 9
+ H0 + ζ3 − ζ2 + ζ22 − H0 ζ3 − H0 ζ2 − 2H0,0ζ2
24 2
3 1 1
+ H0,0 − H0,0,0 + H0,0,0,0 + H−2,0 − H3 . (4.11)
8 4 2
Of particular interest is the end-point behaviour of the harmonic polylogarithms at

x → 0 or x → 1, where logarithmic singularities occur. In the limit x → 0, the factors
ln x are related to trailing zeroes in the index field, whereas in the limit x → 1 factors of
ln(1 − x) emerge from leading indices of value 1. In both limits, the logarithms can be
factored out by repeated use of the product identity for harmonic polylogarithms,

Hm w (x)Hnv (x) = Hlw+v (x). (4.12)
lw+v =m
w
nv
Here m w nv represents all mergers of m w and nv in which the relative orders of the
elements of m w and n v are preserved. All algorithms for this algebraic procedure have
been coded in F ORM, some explicit examples are given in Refs. [64,76].
(n)±,v
The large-x behaviour of splitting functions Pns (x) reflects the large-N behaviour
of the corresponding anomalous dimensions in Eq. (3.10). Specifically, the (identical)
(2)±,v
large-x behaviour of Pns (x) is given by
(2)±,v A3
Px→1 (x) = + B3 δ(1 − x) + C3 ln(1 − x) + O(1). (4.13)
(1 − x)+
The constants A3 and C3 have been specified in Eqs. (3.11) and (3.12), respectively, while
the coefficients of δ(1 − x) are explicit in Eq. (4.9). At small x the splitting functions can
(n)±,s
be expanded in powers of ln x. For the three-loop non-singlet splitting functions Pns (x)
one finds
(2)i
Px→0 (x) = D0i ln4 x + D1i ln3 x + D2i ln2 x + D3i ln x + O(1). (4.14)
Generally, terms up to ln2k x occur at order αsk+1 . Keeping only the highest n + 1 of
these, one arrives at the Nn Lx small-x approximation. Like the large-x coefficients, these
(2)+
contributions can be readily extracted from our full results using Eq. (4.12). For Pns
we obtain
2
D0+ = CF3 ,
3
22 4
D1+ = CF2 CA − 4CF3 − CF2 nf ,
3 3

+ 121 472
D2 = − 30ζ2 CF CA +
2
+ 96ζ2 CF2 CA + [4 − 104ζ2]CF3
9 9
44 64 4
− CF CA nf − CF2 nf + CF n2f ,
9 9 9

3934 370
D3+ = − 92ζ2 CF CA2 + + 216ζ2 + 48ζ3 CF2 CA
27 9

1268
− [30 + 192ζ2 + 96ζ3]CF3 − − 8ζ2 CF CA nf
27
88 2 88
− CF nf + CF n2f , (4.15)
9 27
or, after inserting CA = 3 and CF = 4/3 and the numerical values of ζ2 and ζ3 ,
D0+ ∼
= 1.58025,
+∼
D = 29.6296 − 2.37037nf ,
1
D2+ ∼
= 295.042 − 32.1975nf + 0.592592n2f ,
D3+ ∼
= 1261.11 − 152.597nf + 4.345679n2f . (4.16)
(2)−
The corresponding coefficients for Pns are given by
10 3
D0− = −CF CA2 + 4CF2 CA − C ,
3 F
40 14 20 16
D1− = CF CA2 − CF2 CA − 4CF3 + CF2 nf − CF CA nf ,
9 9 9 9
152
D2− = [81 + 14ζ2 ]CF CA2 − + 96ζ2 CF2 CA − [60 − 104ζ2]CF3
3
196 80 4
− CF CA nf + CF2 nf + CF n2f ,
9 3 9
3442 100 1850 680
D3− = + ζ2 + 112ζ3 CF CA2 + − ζ2 − 336ζ3 CF2 CA
27 3 9 3

88 568 32 2
+ CF n2f − [286 − 192ζ2 − 224ζ3]CF3 + + CF nf
27 9 3

2252 8
− − ζ2 CF CA nf , (4.17)
27 3
and
D0− ∼
= 1.43210,
−∼
D = 35.5556 − 3.16049nf ,
1
D2− ∼
= 399.205 − 39.7037nf + 0.592592n2f ,
∼ 1465.93 − 172.693nf + 4.345679n2 .
D3− = (4.18)
f
The coefficients D0± of the leading logarithms in Eqs. (4.15) and (4.17) agree with the
predictions in Ref. [31] based of the resummation of Ref. [30]. Finally the small-x
(2)s
expansion of Pns (x) reads
d abc dabc 1
D0s = nf ,
nc 3
abc d
d abc 2
D1 =
s
nf − ,
nc 3
d abc dabc
D2s = nf (18 − 10ζ2),
nc
d abc dabc
D3s = nf (56 + 2ζ2 − 16ζ3), (4.19)
nc
(2) +
Fig. 2. The nf -independent three-loop contribution P+,0 (x) to the splitting function Pns (x), multiplied by (1−x)
for display purposes. Also shown in the left part is the uncertainty band derived in Ref. [29] from the lowest six
even-integer moments [24–26]. In the right part our exact result is compared to the small-x approximations
defined in Eq. (4.14) and the text below it.
or, inserting the QCD value of 40/9 for the group factor d abc dabc /nc ,
D0s ∼
= +1.48148nf , D1s ∼
= −2.96296nf ,
s ∼ s ∼
D = +6.89182nf ,
2 D = +178.030nf .
3 (4.20)
(2)±
The n0f and n1f parts of the functions Pns (x) in Eqs. (4.9) and (4.10) are separately
shown in Figs. 2–4 together with the approximate expressions derived in Ref. [29]
mainly from the integer-N results of Refs. [24–26]. Also shown for the non-fermionic
contributions in Figs. 2 and 3 are the successive approximations by small-x logarithms as
defined in Eq. (4.14) and the text below it. As can be seen from Eqs. (4.16) and (4.18),
(2)±
the coefficients of lnk x for Pns increase sharply with decreasing power k. Consequently
the shapes of the full results in Figs. 2 and 3 are reproduced only after all logarithmically
enhanced terms have been included. Even then the small-x approximations underestimate
(2)+ (2)−
the complete results by factors as large as 2.7 and 2.0, respectively, for Pns and Pns at
−4
x = 10 , rendering the small-x expansion (4.14) ineffective for any practically relevant
value of x. Keeping only the Lx (ln4 x) or NLx (ln4 x and ln3 x) contributions leads
to a reasonable description only at extremely small values of x. Therefore, meaningful
estimates of higher-order effects based on resumming leading (and subleading) logarithms
in the small-x limit appear to be difficult.
(2)s
The new three-loop n1f contribution Pns with the colour structure d abc dabc /nc is
graphically displayed in Fig. 5 for nf = 1. Rather unexpectedly, also this function behaves
−
Fig. 3. As Fig. 2, but for the splitting function Pns (x). The first seven odd moments underlying the previous
approximations [29] also shown in the left part have been computed in Ref. [26].
(2) ±
Fig. 4. The n1f three-loop contributions P±,1 (x) to the splitting functions Pns (x), compared to the uncertainty
bands of Ref. [29] based on the integer moments calculated in Refs. [24–26].
(2) s (x), compared to the

Fig. 5. The first non-vanishing contribution Ps,1 (x) to the splitting functions Pns
approximations of Ref. [29] (where, assuming the completeness of the resummation [30,31], the possibility of a
ln4 x term was disregarded) and to the small-x expansion in powers of ln x.
like ln4 x for x → 0, and here the leading small-x terms do indeed provide a reasonable
approximation. In fact, this function dominates the small-x behaviour of the non-singlet
(2)±
splitting functions, for nf = 4 being, for example, about 7 times larger than Pns (x) at
x = 10−4 . The presence of a (dominant) leading small-x logarithm in a term unpredictable
from lower-order structures appears to call into question the very concept of the small-x
resummation of the double logarithms αsk+1 ln2k x.
(2)i
In view of the length and complexity of the exact expressions for the functions Pns (x),
it is useful to have at ones disposal also compact approximate representations involving,
besides powers of x, only simple functions like the ‘+’-distribution and the end-point
logarithms
D0 = 1/(1 − x)+ , L1 = ln(1 − x), L0 = ln x. (4.21)

(2)+
Inserting the numerical values of the QCD colour factors, Pns in Eq. (4.9) can be
represented by
(2)+
Pns (x) ∼
= +1174.898D0 + 1295.384δ(1 − x) + 714.1L1 + 1641.1
− 3135x + 243.6x 2 − 522.1x 3 + L0 L1 [563.9 + 256.8L0]
+ 1258L0 + 294.9L20 + 800/27L30 + 128/81L40

+ nf −183.187D0 − 173.927δ(1 − x) − 5120/81L1 − 197.0
+ 381.1x + 72.94x 2 + 44.79x 3 − 1.497xL30 − 56.66L0L1 − 152.6L0

− 2608/81L20 − 64/27L30

+ n2f −D0 − (51/16 + 3ζ3 − 5ζ2 )δ(1 − x)
+ x(1 − x)−1 L0 (3/2L0 + 5) + 1

+ (1 − x) 6 + 11/2L0 + 3/4L20 64/81. (4.22)
(2)−
A corresponding parametrization of Pns in Eq. (4.10) is given by
(2)−
Pns (x) ∼
= +1174.898D0 + 1295.470δ(1 − x) + 714.1L1 + 1860.2 − 3505x
+ 297.0x 2 − 433.2x 3 + L0 L1 [684 + 251.2L0] + 1465.2L0 + 399.2L20
+ 320/9L30 + 116/81L40

+ nf −183.187D0 − 173.933δ(1 − x) − 5120/81L1 − 216.62
+ 406.5x + 77.89x 2 + 34.76x 3 − 1.136xL30 − 65.43L0L1

− 172.69L0 − 3216/81L20 − 256/81L30

+ n2f −D0 − (51/16 + 3ζ3 − 5ζ2 )δ(1 − x)
+ x(1 − x)−1 L0 (3/2L0 + 5) + 1

+ (1 − x) 6 + 11/2L0 + 3/4L20 64/81. (4.23)
(2)s
Finally the splitting function Pns in Eq. (4.11) can be approximated by

(2)s
Pns (x) ∼
= nf L1 −163.9x −1 − 7.208x + 151.49 + 44.51x − 43.12x 2

+ 4.82x 3 [1 − x]
+ L0 L1 [−173.1 + 46.18L0] + 178.04L0 + 6.892L20

+ 40/27 L40 − 2L30 . (4.24)
(2)±
The identical n2f parts of Pns , the ‘+’-distribution contributions (up to a numerical
truncation of the coefficients involving ζi ), and the rational coefficients of the (sub-)leading
regular end-point terms are exact in Eqs. (4.22)–(4.24). The remaining coefficients have
been determined by fits to the exact results, for which we have used the F ORTRAN package
(2)i
of Ref. [75]. Except for x values very close to zeros of Pns (x), the above parametrizations
deviate from the exact expressions by less than one part in thousand, which should be
sufficiently accurate for foreseeable numerical applications. For a maximal accuracy for the
convolutions with the quark densities, also the coefficients of δ(1 − x) have been slightly
adjusted, by 0.02% or less, using low integer moments. Also the complex-N moments
of the splitting functions can be readily obtained to a perfectly sufficient accuracy using
Eqs. (4.22)–(4.24). The Mellin transform of these parametrizations involve only simple
harmonic sums Sm>0 (N) (see, e.g., the appendix of Ref. [60]) of which the analytic
continuations in terms of logarithmic derivatives of Euler’s
-function are well known.
5. Numerical implications
In this section we illustrate the effect of our new three-loop splitting functions
(2)±,v ±,v
Pns (x) on the evolution (2.6) of the non-singlet combinations qns (x, µ2f ) of the
quark and antiquark distributions. For all figures we employ the same schematic, but
characteristic model distribution,
±,v

xqns x, µ20 = x a (1 − x)b (5.1)
with
a = 0.5, b = 3, (5.2)
facilitating a direct comparison of the various splitting functions contributing to Eq. (2.6).
For the same reason the reference scale is specified by an order-independent value for the
strong coupling constant usually chosen as

αs µ20 = 0.2. (5.3)
This value corresponds to µ20 25–50 GeV2 for αs (MZ2 ) = 0.114–0.120 beyond the
leading order, a scale region relevant for deep-inelastic scattering both at fixed-target
experiments and, for much smaller x, at the ep collider HERA. Our default for the number
of effectively massless flavours is nf = 4. The normalization of qns i is irrelevant for our
purposes, as we consider only the logarithmic derivatives q̇ns i ≡ d ln q i /d ln µ2 .

ns f
The scale derivatives of the three non-singlet distributions are graphically displayed in
Figs. 6 and 7 over a wide region of x. At large x the NNLO corrections are very similar in
all cases, amounting to 2% or less for x 0.2, thus being smaller than the NLO corrections
by a factor of about eight. The same suppression factor is also found for qns − (x) in the region
10−5 x 10−2 . The NNLO effects are even smaller for qns + at small x, but considerably
larger for qns v at x < 10−3 . For example, at x 10−4 , where P (2)v (x) exceeds P (2)− (x)
ns ns
by a factor of about 8 as discussed in the paragraph above Eq. (4.21), the ratio of the
corresponding corrections in Fig. 7 amounts to 2.5. Recall that the scale derivatives (2.6) do
not probe the splitting functions locally in x due to the presence of the Mellin convolution.
The numerical values for q̇ns v (x, µ2 ) are presented in Table 1 for four characteristic
0
values of x. Also illustrated in this table is the dependence of the results on the shape of
the initial distribution, the number of flavours and the value of the strong coupling constant.
The relative corrections are rather weakly dependent of the large-x power b in Eq. (5.1).
They increase at small-x with increasing small-x power a, i.e., with decreasing size of qns v.
abc s
At large x, where the nf d dabc /nc contribution Pns is negligible, the NNLO corrections
decrease with increasing nf . At small-x this decrease is overcompensated in q̇ns v by the
−3
effect of Pns . Except for very small momentum fractions x 10 (where the non-singlet
s
quark densities play a minor role for most important observables) the NNLO corrections
amount to 15% or less even for a strong coupling constant as large as αs = 0.5. Hence
the non-singlet evolution at intermediate and large x appears to remain perturbative down
to very low scales as used in the phenomenological analyses of Refs. [77,78] and in non-
perturbative studies of the initial distributions like those of Refs. [38–43].
+
Fig. 6. The perturbative expansion of the logarithmic scale derivative d ln qns /d ln µ2f for a characteristic
+
non-singlet quark distribution xqns = x 0.5 (1 − x)3 at the standard scale µr = µf .
−,v
Fig. 7. As Fig. 6, but for the scale derivatives of the two other non-singlet combinations qns .
Table 1
The LO, NLO and NNLO logarithmic derivatives q̇ns v ≡ d ln q v /d ln µ2 at four representative values of x,
ns f
together with the ratios rn = Nn LO/Nn−1 LO − 1 for the default input parameters specified in the first paragraph
of this section and some variations thereof
x LO NLO NNLO r1 r2 r2 /r1
default (Fig. 7)
10−4 6.546 × 10−2 8.424 × 10−2 9.163 × 10−2 0.287 0.088 0.31
0.002 5.632 × 10−2 6.875 × 10−2 7.041 × 10−2 0.221 0.024 0.11
0.25 −5.402 × 10−2 −6.331 × 10−2 −6.457 × 10−2 0.172 0.020 0.12
0.75 −1.949 × 10−1 −2.189 × 10−1 −2.222 × 10−1 0.123 0.015 0.12
a = 0.8
10−4 1.660 × 10−1 2.351 × 10−1 2.818 × 10−1 0.417 0.198 0.48
0.002 1.249 × 10−1 1.583 × 10−1 1.650 × 10−1 0.268 0.042 0.16
0.25 −4.352 × 10−2 −5.171 × 10−2 −5.283 × 10−2 0.188 0.022 0.12
0.75 −1.930 × 10−1 −2.168 × 10−1 −2.200 × 10−1 0.123 0.015 0.12
b=5
10−4 6.474 × 10−2 8.278 × 10−2 8.917 × 10−2 0.279 0.077 0.28
0.002 5.324 × 10−2 6.432 × 10−2 6.546 × 10−2 0.208 0.018 0.09
0.25 −7.835 × 10−2 −9.022 × 10−2 −9.180 × 10−2 0.151 0.018 0.12
0.75 −2.300 × 10−1 −2.580 × 10−1 −2.619 × 10−1 0.122 0.015 0.12
nf = 3
10−4 6.546 × 10−2 8.480 × 10−2 9.187 × 10−2 0.295 0.083 0.28
0.002 5.632 × 10−2 6.942 × 10−2 7.174 × 10−2 0.233 0.033 0.14
0.25 −5.402 × 10−2 −6.406 × 10−2 −6.588 × 10−2 0.186 0.028 0.15
0.75 −1.949 × 10−1 −2.219 × 10−1 −2.269 × 10−1 0.139 0.023 0.16
nf = 3 and αs = 0.5
10−4 1.636 × 10−1 2.845 × 10−1 3.949 × 10−1 0.739 0.388 0.53
0.002 1.408 × 10−1 2.227 × 10−1 2.589 × 10−1 0.581 0.163 0.28
0.25 −1.350 × 10−1 −1.978 × 10−1 −2.262 × 10−1 0.465 0.144 0.31
0.75 −4.871 × 10−1 −6.563 × 10−1 −7.346 × 10−1 0.347 0.119 0.34
Another conventional way to assess the reliability of perturbative calculations is to

investigate the stability of the results under variations of the renormalization scale µr .
For µr
= µf the expansion in Eq. (2.6) has to be replaced by

(0) (1),i µ2f
i
Pns (µf , µr ) = as Pns + as µr Pns − β0 Pns ln 2
µ2r 2 2 (0)
µr

(2),i µ2f
+ as3 µ2r Pns − β1 Pns(0)
+ 2β0 Pns
(1),i
ln 2
µr
2
µ
(0) 2 f
+ β02 Pns ln 2 + · · · , (5.4)
µr
where βk represent the MS expansion coefficients of the β-function of QCD [79–82].
In Fig. 8 the consequences of varying µr over the rather wide range 18 µ2f µ2r 8µ2f
+ at six representative values of x. The scale dependence is considerably
are displayed for q̇ns
+ +
Fig. 8. The dependence of the NLO and NNLO predictions for q̇ns ≡ d ln qns /d ln µ2f on the renormalization
scale µr for six typical values of x. The initial conditions are as in Fig. 5.
reduced by including the third-order corrections over the full x-range. At NNLO both
the points of fastest apparent convergence and the points of minimal µr -sensitivity,
+ /∂µ = 0, are rather close to the ‘natural’ choice µ = µ for the renormalization
∂ q̇ns r r f
scale.
The relative scale uncertainties of the average results, conventionally estimated by
i (x, µ2 = 1 µ2 · · · 4µ2 )] − min[q̇ i (x, µ2 = 1 µ2 · · · 4µ2 )]
max[q̇ns r 4 f f ns r 4 f f
i
q̇ns ≡ (5.5)
i (x, µ2 = 1 µ2 · · · 4µ2 )]|
2|average[q̇ns r 4 f f
is shown in Fig. 9 for all three cases i = ±, v. These uncertainty estimates amount to 2%
or less except for x 10−3 , an improvement by more than a factor of three with respect
to the corresponding NLO results. Taking into account also the apparent convergence of
Fig. 9. The renormalization scale uncertainty of the NLO and NNLO predictions for the scale derivative of qns i ,
i = ±, V , as obtained from the quantity q̇ns defined in Eq. (5.5). Here and in Figs. 6 and 7 the spikes close to
i
x = 0.1 reflect the sign-change of q̇ns
i and do not constitute appreciable absolute corrections and uncertainties.
the series in Figs. 6 and 7, it is not unreasonable to expect that the effect of the four-
loop non-singlet splitting functions—which most likely will remain uncalculated for quite
some time—will be less than 1% for x > 10−3 . This expectation is consistent with the
(3)i
Padé estimates of Pns employed in Ref. [83] for the N3 LO large-x evolution of the
deep-inelastic structure functions F2 and F3 . At very small values of x the higher-order
corrections will presumably be considerably larger.
6. Summary
We have calculated the complete third-order contributions to the splitting functions

governing the evolution of unpolarized non-singlet parton distribution in perturbative
QCD. Our calculation is performed in Mellin-N space and follows the previous fixed-N
computations [24–26] inasmuch as we compute the partonic structure functions in deep-
inelastic scattering at even or odd N using the optical theorem and a dispersion relation as
discussed in [25]. Our calculation, however, is not restricted to low fixed values of N but
provides the complete N -dependence from which the x-space splitting functions can be
obtained by a (by now) standard Mellin inversion. This progress has been made possible
by an improved understanding of the mathematics of harmonic sums, difference equations
and harmonic polylogarithms [45,59,64], and the implementation of corresponding tools,
together with other new features [53], in the symbolic manipulation program F ORM [52]
which we have employed to handle the almost prohibitively large intermediate expressions.
Our results have been presented in both Mellin-N and Bjorken-x space, in the latter case
we have also provided easy-to-use accurate parametrizations. Our results agree with all
partial results available in the literature, in particular we reproduce the lowest seven even-
or odd-integer moments computed before [24–26]. We also agree with the resummation
predictions [30,31] for the leading small-x logarithms ln4 x of the splitting functions
+ (x) and P − (x) governing the evolution of flavour differences of quark–antiquark sums
Pns ns
and differences. However, an unpredicted term of the same size is found also for the new
d abc dabc /nc contributions Pns s to the splitting function for the total valence distribution.
At large x we find that the coefficient of the leading integrable term ln(1 − x) at order n
is proportional to the coefficient of the (only) ‘+’-distribution 1/(1 − x)+ at order n − 1,
a result that seems to point to a yet unexplored structure.
We have investigated the numerical impact of the three-loop (NNLO) contributions on
the evolution of the various non-singlet densities. The effect of the new contribution Pns s (x)
is very small at√large x but rises sharply towards x → 0, reaching 10% for a standard
Regge-inspired x initial distributions at x 10−5 . At x > 10−3 , on the other hand, the
perturbative expansions for the scale dependences d ln qns (x, µ2f )/d ln µ2f appear to be
very well convergent. For αs = 0.2, for example, the NNLO corrections amount to 2% or
less for four flavours, a factor of about 8 less than the NLO contributions. Also the variation
of the renormalization scale leads to effects of about ±2% at NNLO in this region of x.
Corrections of this size are comparable to the dependence of the predictions on the number
of quark flavours, rendering a proper treatment of charm effects rather important even for
large-x non-singlet quantities, see Refs. [84,85] and references therein.
F ORM files of our results, and F ORTRAN subroutines of our exact and approximate
x-space splitting functions can be obtained from the preprint server http://arXiv.org by
downloading the source. Furthermore they are available from the authors upon request.
Note added in proof

(2)−
The colour structure of the anomalous dimension γns in Eq. (3.8) is in agreement with
(2)+ (2)−
the prediction of Ref. [86], i.e., the difference γns − γns is proportional to CF − CA /2.
The absence of higher powers of ln N in Eq. (3.10) and the maximally non-Abelian
colour structure of the coefficient A3 in Eq. (3.11) have been predicted in Ref. [87]. The
interpretation of the coefficient A2 in Eq. (3.11) in the context of the threshold resummation
has first been given in Ref. [88]. We thank A.L. Kataev, G.P. Korchemsky and L. Trentadue
for pointing out the respective references to us.
Acknowledgements
The preparations for this calculation have been started by J.V. following a suggestion
by S.A. Larin. For stimulating discussions during various stages of this project we
would like to thank, in chronological order, S.A. Larin, F.J. Yndurain, E. Remiddi,
E. Laenen, W.L. van Neerven, P. Uwer, S. Weinzierl and J. Blümlein. M. Zhou

has contributed some F ORM routines during an early stage of the calculation. The
work of S.M. has been supported in part by Deutsche Forschungsgemeinschaft in
Sonderforschungsbereich/Transregio 9. The work of J.V. and A.V. has been part of the
research program of the Dutch Foundation for Fundamental Research of Matter (FOM).
References
[1] D.J. Gross, F. Wilczek, Phys. Rev. D 8 (1973) 3633.

[2] H. Georgi, H.D. Politzer, Phys. Rev. D 9 (1974) 416.
[3] G. Altarelli, G. Parisi, Nucl. Phys. B 126 (1977) 298.
[4] E.G. Floratos, D.A. Ross, C.T. Sachrajda, Nucl. Phys. B 129 (1977) 66.
[5] E.G. Floratos, D.A. Ross, C.T. Sachrajda, Nucl. Phys. B 152 (1979) 493.
[6] A. Gonzalez-Arroyo, C. Lopez, F.J. Yndurain, Nucl. Phys. B 153 (1979) 161.
[7] A. Gonzalez-Arroyo, C. Lopez, Nucl. Phys. B 166 (1980) 429.
[8] G. Curci, W. Furmanski, R. Petronzio, Nucl. Phys. B 175 (1980) 27.
[9] W. Furmanski, R. Petronzio, Phys. Lett. B 97 (1980) 437.
[10] E.G. Floratos, C. Kounnas, R. Lacaze, Nucl. Phys. B 192 (1981) 417.
[11] R. Hamberg, W.L. van Neerven, Nucl. Phys. B 379 (1992) 143.
[12] W.L. van Neerven, E.B. Zijlstra, Phys. Lett. B 272 (1991) 127.
[13] E.B. Zijlstra, W.L. van Neerven, Phys. Lett. B 273 (1991) 476.
[14] E.B. Zijlstra, W.L. van Neerven, Phys. Lett. B 297 (1992) 377.
[15] E.B. Zijlstra, W.L. van Neerven, Nucl. Phys. B 383 (1992) 525.
[16] R. Hamberg, W.L. van Neerven, T. Matsuura, Nucl. Phys. B 359 (1991) 343;
R. Hamberg, W.L. van Neerven, T. Matsuura, Nucl. Phys. B 644 (2002) 403, Erratum.
[17] R.V. Harlander, W.B. Kilgore, Phys. Rev. Lett. 88 (2002) 201801, hep-ph/0201206.
[18] C. Anastasiou, L.J. Dixon, K. Melnikov, F. Petriello, Phys. Rev. Lett. 91 (2003) 182002, hep-ph/0306192.
[19] C. Anastasiou, L. Dixon, K. Melnikov, F. Petriello, hep-ph/0312266.
[20] C. Anastasiou, K. Melnikov, Nucl. Phys. B 646 (2002) 220, hep-ph/0207004.
[21] V. Ravindran, J. Smith, W.L. van Neerven, Nucl. Phys. B 665 (2003) 325, hep-ph/0302135.
[22] R.V. Harlander, W.B. Kilgore, Phys. Rev. D 68 (2003) 013001, hep-ph/0304035.
[23] E.W.N. Glover, Nucl. Phys. B (Proc. Suppl.) 116 (2003) 3, hep-ph/0211412.
[24] S.A. Larin, T. van Ritbergen, J.A.M. Vermaseren, Nucl. Phys. B 427 (1994) 41.
[25] S.A. Larin, P. Nogueira, T. van Ritbergen, J.A.M. Vermaseren, Nucl. Phys. B 492 (1997) 338, hep-
ph/9605317.
[26] A. Retey, J.A.M. Vermaseren, Nucl. Phys. B 604 (2001) 281, hep-ph/0007294.
[27] W.L. van Neerven, A. Vogt, Nucl. Phys. B 568 (2000) 263, hep-ph/9907472.
[29] W.L. van Neerven, A. Vogt, Phys. Lett. B 490 (2000) 111, hep-ph/0007362.
[30] R. Kirschner, L.N. Lipatov, Nucl. Phys. B 213 (1983) 122.
[31] J. Blümlein, A. Vogt, Phys. Lett. B 370 (1996) 149, hep-ph/9510410.
[32] T. van Ritbergen, J.A.M. Vermaseren, S.A. Larin, Phys. Lett. B 400 (1997) 379, hep-ph/9701390.
[33] A. Vogt, S. Moch, J.A.M. Vermaseren, hep-ph/0404111.
[34] G. Sterman, Nucl. Phys. B 281 (1987) 310.
[35] S. Catani, L. Trentadue, Nucl. Phys. B 327 (1989) 323.
[36] S. Catani, L. Trentadue, Nucl. Phys. B 353 (1991) 183.
[37] A. Vogt, Phys. Lett. B 497 (2001) 228, hep-ph/0010146.
[38] H. Weigel, L.P. Gamberg, H. Reinhardt, Phys. Lett. B 399 (1997) 287, hep-ph/9604295.
[39] D. Diakonov, et al., Nucl. Phys. B 480 (1996) 341, hep-ph/9606314.
[40] D. Diakonov, et al., Phys. Rev. D 56 (1997) 4069, hep-ph/9703420.
[41] P.V. Pobylitsa, et al., Phys. Rev. D 59 (1999) 034024, hep-ph/9804436.
[42] O. Schroeder, H. Reinhardt, H. Weigel, Nucl. Phys. A 651 (1999) 174, hep-ph/9902322.
[43] H. Weigel, E. Ruiz Arriola, L.P. Gamberg, Nucl. Phys. B 560 (1999) 383, hep-ph/9905329.
[44] D.I. Kazakov, A.V. Kotikov, Nucl. Phys. B 307 (1988) 721;
D.I. Kazakov, A.V. Kotikov, Nucl. Phys. B 345 (1990) 299, Erratum.
[45] S. Moch, J.A.M. Vermaseren, Nucl. Phys. B 573 (2000) 853, hep-ph/9912355.
[46] S.G. Gorishnii, et al., Comput. Phys. Commun. 55 (1989) 381.
[47] S.A. Larin, F.V. Tkachev, J.A.M. Vermaseren, NIKHEF-H-91-18.
[48] J.A.M. Vermaseren, S. Moch, Nucl. Phys. B (Proc. Suppl.) 89 (2000) 131, hep-ph/0004235.
[49] S. Moch, J.A.M. Vermaseren, A. Vogt, Nucl. Phys. B 646 (2002) 181, hep-ph/0209100.
[50] J.A.M. Vermaseren, S. Moch, A. Vogt, Nucl. Phys. B (Proc. Suppl.) 116 (2003) 100, hep-ph/0211296.
[51] P. Nogueira, J. Comput. Phys. 105 (1993) 279.
[52] J.A.M. Vermaseren, math-ph/0010025.
[53] J.A.M. Vermaseren, Nucl. Phys. B (Proc. Suppl.) 116 (2003) 343, hep-ph/0211297.
[54] G. ’t Hooft, M. Veltman, Nucl. Phys. B 44 (1972) 189.
[55] C.G. Bollini, J.J. Giambiagi, Nuovo Cimento B 12 (1972) 20.
[56] J.F. Ashmore, Lett. Nuovo Cimento 4 (1972) 289.
[57] G.M. Cicuta, E. Montaldi, Lett. Nuovo Cimento 4 (1972) 329.
[58] T. van Ritbergen, A.N. Schellekens, J.A.M. Vermaseren, Int. J. Mod. Phys. A 14 (1999) 41, hep-ph/9802376.
[59] J.A.M. Vermaseren, Int. J. Mod. Phys. A 14 (1999) 2037, hep-ph/9806280.
[60] J. Blümlein, S. Kurth, Phys. Rev. D 60 (1999) 014018, hep-ph/9810241.
[61] S. Moch, P. Uwer, S. Weinzierl, J. Math. Phys. 43 (2002) 3363, hep-ph/0110083.
[62] A.B. Goncharov, Math. Res. Lett. 5 (1998) 497, available at http://www.math.uiuc.edu/K-theory/0297.
[63] J.M. Borwein, et al., math.CA/9910045.
[64] E. Remiddi, J.A.M. Vermaseren, Int. J. Mod. Phys. A 15 (2000) 725, hep-ph/9905237.
[65] S. Moch, J.A.M. Vermaseren, Nucl. Phys. B (Proc. Suppl.) 89 (2000) 137, hep-ph/0006053.
[66] G. ’t Hooft, Nucl. Phys. B 61 (1973) 455.
[67] W.A. Bardeen, et al., Phys. Rev. D 18 (1978) 3998.
[68] S.A. Larin, J.A.M. Vermaseren, Phys. Lett. B 259 (1991) 345.
[69] J.A.M. Vermaseren, A. Vogt, S. Moch, in preparation.
[70] K.G. Chetyrkin, F.V. Tkachev, Nucl. Phys. B 192 (1981) 159.
[71] G. Passarino, M. Veltman, Nucl. Phys. B 160 (1979) 151.
[72] J.A. Gracey, Phys. Lett. B 322 (1994) 141, hep-ph/9401214.
[73] C.F. Berger, Phys. Rev. D 66 (2002) 116002, hep-ph/0209107.
[74] L. Lewin, Polylogarithms and Associated Functions, North-Holland, New York, 1981.
[75] T. Gehrmann, E. Remiddi, Comput. Phys. Commun. 141 (2001) 296, hep-ph/0107173.
[76] J. Blümlein, hep-ph/0311046.
[77] M. Glück, E. Reya, A. Vogt, Z. Phys. C 67 (1995) 433.
[78] M. Glück, E. Reya, A. Vogt, Eur. Phys. J. C 5 (1998) 461, hep-ph/9806404.
[79] W.E. Caswell, Phys. Rev. Lett. 33 (1974) 244.
[80] D.R.T. Jones, Nucl. Phys. B 75 (1974) 531.
[81] O.V. Tarasov, A.A. Vladimirov, A.Y. Zharkov, Phys. Lett. B 93 (1980) 429.
[82] S.A. Larin, J.A.M. Vermaseren, Phys. Lett. B 303 (1993) 334, hep-ph/9302208.
[84] E. Laenen, S. Riemersma, J. Smith, W.L. van Neerven, Nucl. Phys. B 392 (1993) 162.
[85] A. Chuvakin, J. Smith, W.L. van Neerven, Phys. Rev. D 61 (2000) 096004, hep-ph/9910250.
[86] D.J. Broadhurst, A.L. Kataev, C.J. Maxwell, hep-ph/0403037.
[87] G.P. Korchemsky, Mod. Phys. Lett. A 4 (1989) 1257.
[88] J. Kodaira, L. Trentadue, Phys. Lett. B 112 (1982) 66.
Vacuum polarization and hadronic contribution

to muon g − 2 from lattice QCD
QCDSF Collaboration
M. Göckeler a,b , R. Horsley c , W. Kürzinger d,e , D. Pleiter d ,
P.E.L. Rakow f , G. Schierholz d,g
a Institut für Theoretische Physik, Universität Leipzig, D-04109 Leipzig, Germany
b Institut für Theoretische Physik, Universität Regensburg, D-93040 Regensburg, Germany
c School of Physics, University of Edinburgh, Edinburgh EH9 3JZ, UK
d John von Neumann-Institut für Computing NIC, Deutsches Elektronen-Synchrotron DESY,
D-15735 Zeuthen, Germany
e Institut für Theoretische Physik, Freie Universität Berlin, D-14195 Berlin, Germany
f Theoretical Physics Division, Department of Mathematical Sciences, University of Liverpool,
Liverpool L69 3BX, UK
g Deutsches Elektronen-Synchrotron DESY, D-22603 Hamburg, Germany
Received 22 December 2003; accepted 22 March 2004
Abstract
We compute the vacuum polarization on the lattice in quenched QCD using non-perturbatively
improved Wilson fermions. Above Q2 of about 2 GeV2 the results are very close to the predictions
of perturbative QCD. Below this scale we see signs of non-perturbative effects which we can describe
by the use of dispersion relations. We use our results to estimate the light quark contribution to the
muon’s anomalous magnetic moment. We find the result 446(23) × 10−10 , where the error only
includes statistical uncertainties. Finally we make some comments on the applicability of the operator
product expansion to our data.
PACS: 11.15.Ha; 12.38.-t; 12.38.Bx; 12.38.Gc; 14.60.Ef; 16.65.+i
Keywords: QCD; Lattice; e+ e− annihilation; Muon anomalous magnetic moment
E-mail address: rakow@amtp.liv.ac.uk (P.E.L. Rakow).
doi:10.1016/j.nuclphysb.2004.03.026
136 QCDSF Collaboration / Nuclear Physics B 688 (2004) 135–164
1. Introduction
The vacuum polarization Π(Q2 ) provides valuable information on the interface

between perturbative and non-perturbative physics. It has been the subject of intensive
discussions in the literature.
The vacuum polarization tensor is responsible for the running of αem , which must be
known very accurately for high-precision electro-magnetic calculations. To calculate the
hadronic contribution to the anomalous magnetic moment of the muon we need to know
the vacuum polarization at scales from ∼ 100 MeV to many GeV. Perturbative QCD will
be unreliable at the low end of this scale, so a non-perturbative calculation on the lattice
would be useful.
The vacuum polarization Π(Q2 ) is defined by

Πµν (q) = i d 4 x eiqx 0|T Jµ (x)Jν (0)|0 ≡ qµ qν − q 2 gµν Π Q2 , (1)
where Jµ is the hadronic electro-magnetic current

2 1
Jµ (x) = ef ψ̄f (x)γµ ψf (x) = ū(x)γµ u(x) − d̄(x)γµ d(x) + · · · (2)
3 3
f
and Q2 ≡ −q 2 (so that Q2 > 0 for spacelike momenta, Q2 < 0 for timelike). Π can be
computed on the lattice for spacelike momenta Q2 > 0.
Π can also be calculated in perturbation theory. Π has to be additively renormalized,
even the one-loop diagram (with no gluons involved) is logarithmically divergent. This
renormalization implies that the value of Π can be shifted up and down by a constant
depending on scheme and scale without any physical effects. However, the Q2 dependence
of Π(Q2 ) is physically meaningful, and it must be independent of renormalization scheme
or regularization.
Experimentally Π can be calculated from data for the total cross section of e+ e− →
hadrons with theoretical predictions of QCD by means of the dispersion relations [1]
n n

∞
2 (−1) d R(s)
2
12π Q Π Q 2
=Q
2
ds , (3)
n! d Q2 (s + Q2 )n+1
4m2π
where

σe+ e− →hadrons(s) 3s
R(s) = = 2
σe+ e− →hadrons(s). (4)
σe+ e− →µ+ µ− (s) 4παem
The first derivative of the vacuum polarization (the n = 1 term in Eq. (3)) is referred to as
the Adler D-function [2]:
dΠ(Q2 )
D Q2 = −12π 2 Q2 . (5)
dQ2
The anomalous magnetic moment of the muon (g − 2)µ can be calculated to very high
order in QED (5 loop), and measured very precisely. (g − 2)µ is more sensitive to high-
energy physics than (g − 2)e , by a factor m2µ /m2e , so it is a more promising place to look
QCDSF Collaboration / Nuclear Physics B 688 (2004) 135–164 137
Fig. 1. A typical Feynman diagram contributing to the CΠ part of the vacuum polarization Π is shown as
diagram (a). A diagram contributing to the AΠ part of Π is shown in (b). In this paper we only consider diagrams
of the first type, with both photons attached to the same quark line.
for signs of new physics, but to identify new physics we need to know the conventional
contributions very accurately. QED perturbative calculations take good account of muon
and electron loops, but at the two-loop level quarks can be produced, which in turn will
produce gluons. The dominant contribution comes from photons with virtualities ∼ m2µ ,
which is a region where QCD perturbation theory will not work well.
Π(Q2 ) can be split into two contributions with a different dependency on the quark
charges, as shown in Fig. 1,
2
−12π 2 Π Q2 = ef CΠ µ2 , Q2 , mf + ef ef AΠ µ2 , Q2 , mf , mf . (6)
f f,f
The CΠ term begins with a tree-level term which is O(αs0 ), while the first contribution to
the AΠ term is O(αs3 ). Furthermore, if flavour SU (3) is a good symmetry, the contribution
to AΠ from the three light flavours (u, d, s) cancels because eu + ed + es = 0, and the only
surviving contributions to AΠ come when both f and f are heavy quarks (c, b, t). In this
paper we will concentrate on the CΠ term, both because it is larger, and because it is much
easier to measure on the lattice.
For large (spacelike) momenta CΠ (µ2 , Q2 , mf ) can be expressed by means of the
operator product expansion, OPE

CΠ µ2 , Q2 , mf = c0 µ2 , Q2 , mf
c4F (µ2 , Q2 ) c4G (µ2 , Q2 ) αs 2
+ m f ψ̄f ψ f + G
(Q2 )2 (Q2 )2 π µν

c F (µ2 , Q2 ) 1
+ 4 m f ψ̄
f fψ + O , (7)
(Q2 )2
(Q2 )3
f
where c0 , c4F , c4 F and c4G are the Wilson coefficients and µ is the renormalization scale
parameter in some renormalization scheme such as MS. The f sum extends over the
flavours of the sea quarks (internal quark loops not directly connected to the photon lines).
The Wilson coefficients can be computed in perturbation theory, while the non-perturbative
physics is encoded in the condensates mf ψ̄f ψf (µ), αs G2µν (µ), etc.
Perturbatively the functions D, Eq. (5), and R, Eq. (4), are known to four loops for
massless quarks [3,4], while the coefficient c0 is known to three loops in the massive
case [5]. The coefficients c4F , c4 F and c4G are known up to O(αs2 ) [6–8] for massless quarks.
In this paper we shall compute Π(Q2 ) on the lattice and compare the result with current
phenomenology [9]. Preliminary results of this calculation were presented in [10].
The structure of this paper is as follows. After this Introduction we discuss the lattice
setup in Sections 2 and Appendix A. The results are presented in Section 3 and Appendix
C. In Section 4 and Appendix B we compare with perturbation theory. In Sections 5 and 6
we present a simple model which describes our lattice data well. In Section 7 we use this
model to give a lattice estimate of the hadronic contribution to the muon’s anomalous
magnetic moment. Finally in Section 8 we make some comments on the applicability of
the operator product expansion, and give our conclusions in Section 9.
2. Lattice setup
We work with non-perturbatively improved Wilson fermions. For the action and
computational details, as well as for results of hadron masses and decay constants, see [11–
13]. Full details of the vacuum polarization calculation can be found in [14].
To discretize the hadronic electro-magnetic current

Jµem (x) = ef Jµf (x) (8)
f
we use the conserved vector current

1
Jµf (x) = ψ̄f (x + a µ̂)(1 + γµ )Uµ† (x)ψf (x)
2
− ψ̄f (x)(1 − γµ )Uµ (x)ψf (x + a µ̂) , (9)
where a is the lattice spacing. From now on we work with a Euclidean metric and write

Jµf (q̂) = eiq(x+a µ̂/2)Jµf (x) (10)
x
with

2 aqµ
q̂µ = sin . (11)
a 2
f f
On the lattice the photon self-energy Π is not simply given by Jµ (x)Jν (0). This is
because the lattice Feynman rules include vertices where any number of photons couple to
a quark line (not just the single-photon vertex of the continuum) [15]. We are therefore led
to define
Πµν (q̂) = Πµν

(1)
(q̂) + Πµν
(2)
(q̂), (12)
with

(1)
Πµν (q̂) = a 4 eiq(x+a µ̂/2−a ν̂/2) Jµf (x)Jνf (0) , (13)
x
and

(2)
Πµν (q̂) = −aδµν Jµ(2) (0) , (14)
where
1
Jµ(2) (x) = ψ̄(x + a µ̂)(1 + γµ )Uµ† (x)ψ(x)
2
+ ψ̄(x)(1 − γµ )Uµ (x)ψ(x + a µ̂) . (15)
It then follows that (12) fulfils the Ward identity
q̂µ Πµν (q̂) = q̂ν Πµν (q̂) = 0. (16)
The conserved vector current (9) is on-shell as well as off-shell improved [16] in forward
matrix elements. In non-forward matrix elements, such as the vacuum polarization, it needs
to be further improved,
imp 1
Jµ (x) = Jµ (x) + ia cCVC ∂ν ψ̄(x)σµν ψ(x) , (17)
2
where we have some freedom in choosing ∂ν on the lattice. Any choice must preserve (16),
and ideally it should keep O(a 2 ) corrections small. Considering the fact that the conserved
vector current Jµ (x) ‘lives’ at x + a µ̂/2, the natural (or naive) choice for the improved
operator would be
imp 1
Jµ (q̂) = Jµ (q̂) + iacCVC eiq(x+a µ̂/2)
2 x
1
× Tµν (x + a ν̂) + Tµν (x + a µ̂ + a ν̂)
4a
− Tµν (x − a ν̂) − Tµν (x + a µ̂ − a ν̂)

1 aqµ
= Jµ (q̂) + cCVC cos sin(aqν ) eiqx Tµν (x), (18)
2 2 x
where Tµν (x) ≡ ψ̄(x)σµν ψ(x). Unfortunately it turns out that this choice introduces very
large O(a 2 Q2 ) errors into Π . We found that

imp aqν iqx
Jµ (q̂) ≡ Jµ (q̂) + cCVC sin e Tµν (x) (19)
2 x
was a better choice of improved current, because it makes the O(a 2 Q2 ) terms much
smaller, and so we will adopt this definition. The difference between the definitions (18)
and (19) is O(a 2 Q2 ), so we are free to choose whichever definition leads to the largest
scaling region in Q2 (both agree at small Q2 ).
In Fig. 2 we show the unimproved polarization tensor along with the two different
choices of improvement term in the case of free fermions. The best agreement with
Fig. 2. The effects of improving the current operator. We compare various curves, all calculated in the free-field
case at the one-loop level with q ∝ (0, 0, 0, 1) and κ = 0.123 (corresponding to am = 0.065). The dotted curves
show the continuum one-loop result, shifted vertically to match with lattice results. The dashed line shows
the one-loop lattice result without any improvement of the current. The dot-dashed curve shows the result of
improving with (18) and the solid curve the result of improving with (19). The upper two curves differ by
O(a), the difference ∼ am ln a 2 q̂ 2 . The lower lattice curves are both O(a) improved, but we see that the naive
improvement (dot-dashed curve) has very large O(a 2 ) discretization errors. Improving with (19) produces a
lattice result much closer to the continuum result.
continuum physics comes from using (19), which is the prescription we will use in the
rest of this paper.
(1) (2)
In Appendix A we give explicit expressions for Πµν (q̂) and Πµν (q̂) in terms of the
link variables and the quark propagators.
3. Lattice calculation
To facilitate the extrapolation to the chiral and continuum limits, we have made
simulations at three different values of β with three or more different κ values at each β.
The parameters are listed in Table 1. The lattice data for the vacuum polarization for the
individual momenta and β and κ values are given in Appendix C.
First, let us discuss the value we use for cCVC . From the fermion-loop contribution to the
gluon propagator, computed in [18] to O(m) in lowest order of lattice perturbation theory,
Table 1
Parameters of the lattice simulation. The improve-
ment coefficient in the fermionic action was taken
to be cSW = 1.769 for β = 6.0, cSW = 1.614 for
β = 6.2 and cSW = 1.526 for β = 6.4 [17]
β κ V # Conf.
6.0 0.1333 164 97
6.0 0.1339 164 44
6.0 0.1342 164 44
6.0 0.1345 164 44
6.0 0.1345 324 16
6.2 0.1344 244 51
6.2 0.1349 244 38
6.2 0.1352 244 51
6.4 0.1346 324 50
6.4 0.1350 324 29
6.4 0.1352 324 50
Fig. 3. Chiral extrapolation of CΠ (q̂ 2 , m) at β = 6.0. The points shown range from a 2 q̂ 2 ≈ 3.0 to ≈ 6.5. All are
calculated with cCVC = 1.
Fig. 4. The slope M(q̂ 2 ) for the improved (solid points, cCVC = 1) and unimproved (open points, cCVC = 0)
vector current as a function of a 2 q̂ 2 at β = 6.0. The dashed line shows the height of the plateau in one-loop
lattice perturbation theory, setting cCVC = 1 in Eq. (20).
we obtain
(0)
c0 q̂ 2 , am = 3 ln a 2 q̂ 2 − 3am(1 − cCVC ) ln a 2 q̂ 2 − 3.25275141(5)
+ 1.19541770(1)cCVC − 7.06903716(4)cCVC2

+ am 6.46270704(30) − 5.29413266(6)cCVC

+ 1.67389761(2)cCVC
2
+ O a 2 , m2 , (20)
where am = 1/2κ − 1/2κc . We use c0(0) to refer to the lowest order, g 0 , perturbative
contribution to c0 . As said earlier, the q̂ 2 dependence of Π and c0 is physical. Therefore
in an O(a)-improved calculation there should be no O(a) terms which depend on q̂ 2 . On
the other hand a constant added to c0 has no physical effect, so there is no objection to
constant terms of O(a) in Eq. (20). We see that there is an unphysical am ln(a 2 q̂ 2 ) term in
(20) unless cCVC = 1 + O(g 2 ). In the following we take cCVC = 1 and make the ansatz

CΠ q̂ 2 , am = CΠ q̂ 2 , am = 0 + am M q̂ 2 . (21)
In Fig. 3 we show the quark mass dependence of CΠ (q̂ 2 , am) for several momenta, which
justifies assuming a linear am dependence. In Fig. 4 we show the slope M(q̂ 2 ) with and
without improvement of the conserved vector current. The derivative M should tend to a
constant when Q2
m2 . With no improvement term (open points) we see that M has
a logarithmic dependence on q̂ 2 , corresponding to an unphysical am ln(a 2 q̂ 2 ) term in
c0 . We see that for cCVC = 1 the slope is, within error bars, independent of q̂ 2 down to
small momenta. This shows that the choice cCVC = 1 has eliminated or greatly reduced
Fig. 5. A check for finite volume effects. At β = 6.0, κ = 0.1345 we have measured the polarization tensor both
on a 164 lattice (white points) and on a 324 lattice (black points). The agreement is excellent, and we conclude
that finite volume effects are negligible.
the unphysical logarithm in c0 . We conclude that the tree-level value cCVC = 1 is a good
choice.1
The lattice sizes in Table 1 were chosen such that the physical volume is approximately
equal for all three β values. As we have done simulations at β = 6.0, κ = 0.1345 on
two different lattice volumes we check for finite volume effects. We have not found any,
see Fig. 5. This figure also shows that on the larger lattice we can measure the vacuum
polarization at much lower values of Q2 , which is an important advantage. The quark
boundary conditions are antiperiodic in all four directions, while gluon and photon fields
have periodic boundary conditions in all directions.
4. Comparison with perturbation theory
The first thing to do is to compare lattice results with continuum perturbation

pert
theory, which we do in Fig. 6. For the perturbative contribution c0 (q̂ 2 , m) we use the
renormalization-group improved result given in Eqs. (B.14) and (B.16) of Appendix B. We
use ΛMS = 243 MeV [18,19] and µ = 1/a, and we identify Q2 with q̂ 2 . The r0 parameter
is used to fix the scale [20], with r0 = 0.5 fm.
CΠ calculated on the lattice and CΠ in the continuum can differ by an integration
constant which can depend on µ and a. In lowest-order perturbation theory this constant
is found by comparing c0(0) in MS, Eq. (B.2), with c0(0) calculated in lattice perturbation
theory, Eq. (20). Setting cCVC equal to 1 and µ = 1/a we find that c0 = c0lat − c0MS =
1 If c
CVC should ever be computed non-perturbatively, or in perturbation theory, it should be kept in mind that
we have used Eq. (19) as the definition of the improvement term.
Fig. 6. The deviation of lattice data from continuum perturbation theory. The data are at β = 6.0, κ = 0.1345 on
a 164 lattice (white points) and on a 324 lattice (black points). Below Q2 ∼ 2 GeV2 there is a visible deviation
from perturbation theory. The dotted line shows the effect expected from a gluon condensate as expected in the
OPE.
−22.379. In next to leading order this becomes

c0 = −22.379 − 12/11 ln α MS /α lat + O α lat . (22)
In view of this result, the value c0 ≈ −28.5 seen in Fig. 6 is reasonable.
Fig. 6 shows that the lattice results deviate from perturbation theory at large Q2 . This
deviation, which sets in at a 2 q̂ 2 ∼ 5, is probably a sign of O(a 2 ) lattice artefacts.
More interesting are the deviations from perturbation theory at low Q2 . These are
especially large at low quark mass. The OPE, Eq. (7), would suggest that the effects of
a gluon condensate should show up at small Q2 . In Fig. 6 we show the OPE prediction for
a gluon condensate (α/π)GG = 0.012 GeV4 . The deviations which we see in the data
do not look like the effects expected from a gluon condensate. The sign is the opposite of
what the OPE predicts, and the deviation is probably not growing as quickly as 1/Q4 .
5. A simple model of the vacuum polarization at low Q2
We have seen that perturbation theory, even when supplemented with higher twist terms
from the operator product expansion, has difficulty in explaining the low-Q2 region of the
data. Can we understand this region in some other way?
The cross section ratio R(s) is given by the cut in the vacuum polarization, or in other
words there are dispersion relations which give the vacuum polarization if we know R(s).
One way of modelling Π is to make a model for R(s), and then calculate Π from R,
using the dispersion relation Eq. (3). This should be quite robust, since Π gets contributions
Table 2
Lattice data on the vector meson used as input for the dispersion
relation fits [11,13]. Numbers in italics have been interpolated
from nearby κ values
β κ amPS amV 1/fV
6.0 0.1333 0.4122(9) 0.5503(20) 0.2055(14)
6.0 0.1339 0.3381(15) 0.5017(40) 0.2217(15)
6.0 0.1342 0.3017(13) 0.4904(40) 0.2293(14)
6.0 0.1345 0.2561(15) 0.4701(90) 0.2387(90)
6.2 0.1344 0.3034(6) 0.4015(17) 0.2210(20)
6.2 0.1349 0.2431(6) 0.3663(27) 0.2403(24)
6.2 0.1352 0.2005(9) 0.3431(60) 0.2474(47)
6.4 0.1346 0.2402(8) 0.3107(16) 0.2252(21)
6.4 0.1350 0.1933(7) 0.2800(20) 0.2423(19)
6.4 0.1352 0.1661(10) 0.2613(40) 0.2448(35)
from a large range in s, little inaccuracies in the model R get washed out, and we can hope
that even a crude representation of R will give a good result for Π . We will keep the model
simple so that we can do all the integrals analytically.
Perturbatively

αs
R(s) = ef Nc 1 +
2
+ ··· , (23)
π
f
where Nc is the number of colours (3 in our case). We know that really the low s behaviour
of R is more complicated than that, it is dominated by the ρ(770), ω(782) and φ(1020)
mesons. Following [7], let us make the following model for R. We ignore the splitting
between the ρ and ω, which comes from the AΠ -type diagrams which we have dropped.
We also treat these mesons as narrow resonances, each contributing a δ function to R. The
continuum part of R takes a while to climb up to the value in Eq. (23). We will represent
this rise by a step function at some value s0 . So, our model is

R(s) = ef2 Aδ s − m2V + BΘ(s − s0 ) , (24)
f
where we would expect B to be slightly above 3 in order to match Eq. (23).
Using the dispersion relation, this R(s) translates into a vacuum polarization
A
CΠ Q2 = B ln a 2 Q2 + a 2 s0 − + K, (25)
Q2 + m2V
where K is a constant which is not determined from the dispersion relation, and which
never appears in any physical quantity. Again we identify the continuum quantity Q2 with
the lattice quantity q̂ 2 .
The constant A can be expressed in terms of the decay constant fV , which has been
measured on the lattice. The cross section for the production of a narrow vector resonance,
V , is [7]
ΓV →e+ e−
σe+ e− →V (s) = 12π 2 δ s − m2V . (26)
mV
The partial width ΓV →e+ e− is related to a meson decay constant gV [7,21] by

2 m
4παem V
ΓV →e+ e− = , (27)
3 gV2
where
m2V
0|Jµem |V , ε = εµ . (28)
gV
Here εµ is the polarization vector of the meson. On the lattice it is more natural to define
decay constants fV in terms of currents with definite isospin [11]
m2V
0|Vµ |V , ε = εµ (29)
fV
with
1
VµI =1 = √ (ūγµ u − d̄γµ d) (30)
2
for the ρ 0 and
1
VµI =0 = √ (ūγµ u + d̄γµ d) (31)
2
for the ω (ignoring any s̄s admixture in the ω). The relationship between the two definitions
is
1 eu − ed 1 1 1
= √ =√ , (32)
gρ 2 fρ 2 fρ
1 eu + ed 1 1 1
= √ = √ . (33)
gω 2 fω 3 2 fω
In terms of these decay constants
m2
R(s) = 12π 2 δ s − m2V 2V + continuum (34)
V
gV
m2ρ (eu − ed )2
= 12π 2 δ s − m2ρ 2
fρ 2
m2 (eu + ed )2
+ 12π 2 δ s − m2ω ω2 + continuum. (35)
fω 2
Neglecting annihilation diagrams implies that mω = mρ and fω = fρ . Experimentally
both relations are fairly accurate. The mass ratio mω /mρ is 1.02. fω = fρ implies
Γω→e+ e− = (1/9)Γρ 0 →e+ e− , while the experimental ratio is 0.089(5) [22]. Equating the
mass and decay constant of the ω and ρ in Eq. (35) gives
m2ρ
R(s) = 12π 2 δ s − m2ρ 2 eu2 + ed2 + continuum (36)
fρ
Table 3
The result of fit I, Eq. (25), including only data with a 2 q̂ 2 < 5
β κ V a 2 s0 B K χ2
6.0 0.1333 164 0.395(32) 3.13(6) −32.98(11) 4.3
6.0 0.1339 164 0.397(43) 3.13(8) −33.14(14) 2.2
6.0 0.1342 164 0.391(39) 3.10(8) −33.18(13) 3.0
6.0 0.1345 164 0.409(61) 3.11(9) −33.30(16) 2.5
6.0 0.1345 324 0.403(79) 3.12(10) −33.30(18) 0.7
6.2 0.1344 244 0.282(21) 3.10(5) −32.94(7) 1.9
6.2 0.1349 244 0.278(24) 3.07(5) −33.06(8) 1.9
6.2 0.1352 244 0.253(25) 3.06(5) −33.10(7) 2.2
6.4 0.1346 324 0.164(13) 3.06(3) −32.68(5) 0.8
6.4 0.1350 324 0.150(14) 3.03(4) −32.76(6) 0.6
6.4 0.1352 324 0.124(12) 3.02(3) −32.78(5) 1.2
so
m2V
A = 12π 2 . (37)
fV2
6. Dispersion relation fits
We first try making fits to the lattice data for the vacuum polarization CΠ using Eq. (25)
with B, K and s0 as free parameters. To avoid problems from lattice artefacts of O(a 2 q̂ 2 )
we have only used data with a 2 q̂ 2 < 5. We will call this simple ansatz fit I. A, the weight
of the vector meson contribution, is determined by Eq. (37). The vector meson masses
and decay constants which we use are shown in Table 2. They have been taken from [11,
13]. We can compare the values for B and K with the one-loop lattice perturbation theory
result, Eq. (20), which gives B = 3 and K = −27.38 + am8.53.
A typical result is shown in the top panel of Fig. 7. As can be seen from Table 3, the
χ 2 values for the fits are low in every case. Nevertheless we see that the data deviates
quite strongly from the fit when a 2 q̂ 2 is large. The small deviations from the fit at smaller
q̂ 2 are not random—they occur in the same place in every data set, and depend on the
direction of q. Points where q is near the diagonal direction (1, 1, 1, 1) lie lower than
points measured for momenta away from this diagonal direction.
To correct for this behaviour we add terms to our fit to parameterize O(a 2 ) lattice errors.
There are two possible terms at O(a 2 ):
4
2 2 2 µ qµ
a q and a . (38)
q2
The second term has only cubic symmetry, and can give rise to a dependence of CΠ on the
direction of q, which would not occur in the continuum. To include these terms we fit with
Fig. 7. A comparison between fits I and II. The data are at β = 6.0, κ = 0.1333 on a 164 lattice. Fit I only uses
data with a 2 q̂ 2 < 5 (vertical dashed line). The second fit describes the data better.
the ansatz (fit II)

A
CΠ q̂ 2 = B ln a 2 q̂ 2 + a 2 s0 − + K + U1 a 2 q̂ 2 + U2 h(q) (39)
q̂ 2 + m2V
where
2

µ sin
4
aqµ − (1/4) µ sin
2
aqµ 4
µ qµ 1
h(q) = ≈a 2
− q2 . (40)
a 2 q̂ 2 q2 4
The function h(q) has been chosen so that it vanishes when q ∝ (1, 1, 1, 1), as most of
our momenta are near this direction. Looking at the lower panel of Fig. 7 we see that the
U1 term does a good job of describing the high q̂ 2 data, while the U2 term successfully
Fig. 8. The deviation of lattice data from the fit of Table 4. The data are at β = 6.0, κ = 0.1345 on a 164 lattice
(white points) and on a 324 lattice (black points). The fit describes the data extremely well.
Table 4
The result of fit II, including all data
β κ V a 2 s0 B K χ2
6.0 0.1333 164 0.357(50) 3.00(12) −32.97(17) 1.52
6.0 0.1339 164 0.357(68) 2.99(16) −33.13(23) 0.69
6.0 0.1342 164 0.354(62) 2.97(15) −33.18(21) 0.83
6.0 0.1345 164 0.37(10) 2.97(17) −33.30(25) 0.70
6.0 0.1345 324 0.35(11) 2.92(33) −33.22(28) 0.06
6.0 0.1345 164 &324 0.358(77) 2.93(11) −33.22(15) 1.49
6.2 0.1344 244 0.262(35) 3.00(12) −32.94(12) 0.14
6.2 0.1349 244 0.252(40) 2.94(13) −33.05(14) 0.12
6.2 0.1352 244 0.232(38) 2.95(12) −33.11(12) 0.13
6.4 0.1346 324 0.156(21) 3.01(11) −32.69(7) 0.08
6.4 0.1350 324 0.144(23) 2.99(12) −32.78(9) 0.10
6.4 0.1352 324 0.117(18) 2.97(10) −32.81(7) 0.11
describes the direction dependence of the data. Fits with the ansatz (39) show practically
no deviation from the data. This can also be seen in Fig. 8, where we show the deviation in
the case where we have data on two lattice sizes.
In Fig. 9 we show the results converted into physical units. The agreement between
the different β values is fair, and the value of s0 in agreement with phenomenology (for
example, [23] finds s0 = 1.66(22) GeV2 in the I = 1 channel).
Fig. 9. The threshold s0 (white points) compared with m2ρ (black points). Squares are for β = 6.0, triangles for
6.2, and circles for 6.4.
Fig. 10. The parameter B from Table 4. The β values are shown by the same symbols as in Fig. 9.
In Fig. 10 we show the values for the fit parameter B. The value we find is always very
close to the tree-level value 3.
7. The anomalous magnetic moment of the muon
With a good description of the vacuum polarization tensor in the low-Q2 region we can
make statements about phenomenologically interesting quantities such as the contribution
of the u, d and s-quarks to the muon anomalous magnetic moment [24].
Traditionally the hadronic contribution to the muon’s anomalous magnetic moment is
found from R(s) via a dispersion relation [25] (see [26] for a recent review)
2 ∞
αem ds s
aµhad = K R(s), (41)
3π 2 s m2µ
4m2π
where
1
s x 2 (1 − x)
K = . (42)
m2µ x 2 + (1 − x)s/m2µ
0
The muon mass, mµ , is 105.7 MeV. For us it is more useful to deform the contour and find
this same number from the vacuum polarization at spacelike momenta [27]
∞ 2
α2 dQ2 Q
aµhad = em2 2
F 2
12π 2 Π(0) − Π Q2
3π Q mµ
0
2 ∞ 2
αem dQ2 Q 2
= 2
ef F CΠ Q , mf − CΠ (0, mf )
3π 2 Q2 m2µ
f 0
+ annihilation, (43)
where the kernel F is
2
Q 16m4µ
F = . (44)
m2µ (Q2 )2 (1 + 1 + 4m2µ /Q2 )4 1 + 4m2µ /Q2
To calculate the value aµhad we evaluate the integral

∞ 2
dQ2 Q 2
I (mf ) =
f
2
F 2
CΠ Q , mf − CΠ (0, mf ) (45)
Q mµ
0
for each of our data sets. We then need to extrapolate (or interpolate) to the chiral limit (for
the u and d quarks) and to the strange quark mass. As can be seen in Fig. 11 the integral is
dominated by Q2 ∼ 3m2µ , so we need to extrapolate in Q2 , as our lowest measured value
Fig. 11. The product F (Q2 /m2µ )[CΠ (Q2 , mf ) − CΠ (0, mf )] for the case β = 6.0, κ = 0.1345. The shaded
region shows the uncertainty.
is at Q2 = 0.17 GeV2 . We do our extrapolation of CΠ by using the results of fit II from the
previous section.
The results are shown in Fig. 12, plotted against m2PS (the square of the pseudoscalar
mass). We see that I depends strongly on the quark mass, with heavier quarks making
a smaller contribution (as one would expect). There is also some dependence on β. To
extrapolate to the continuum we fit the data with an ansatz of the form

I f = A1 + A2 a 2 + B1 + B2 a 2 m2π . (46)
This describes the data well (χ 2 /dof = 0.53), and gives the continuum limit shown by
the dashed line in Fig. 12. The physically relevant values of I f are at the physical pion
mass (M 2 = 0.019 GeV2 ) for the u and d quarks, and at the mass of a hypothetical s̄s
pseudoscalar meson with M 2 = 2m2K − m2π = 0.468 GeV2 for I s . The values at these
points are
I u = I d = 0.0389(21), I s = 0.0287(9). (47)

This gives as our final lattice estimate of aµhad

α2 4 u 1 d 1 s
aµhad = em2 I + I + I = 446(23) × 10−10 . (48)
3π 9 9 9
This error reflects the statistical errors of the lattice calculation and the extrapolations
to the physical points. It does not include any estimate of the error due to the quenched
approximation used in the calculation. This value is somewhat lower than the experimental
value 692.4 ± 5.9exp ± 2.4radiative × 10−10 , found by applying Eq. (41) to experimental
R measurements [28,29]. Our value is similar to another lattice measurement [24], which
finds 460(78) × 10−10 .
Fig. 12. The results of the integral (45). The β values are shown by the same symbols as in Fig. 9. The dashed
line shows our extrapolation to the continuum limit, and the points shown with stars are our results for the light
(u, d) and strange quarks.
The shortfall in the value of aµhad can probably be attributed to quenching. In particular
a quenched calculation omits the process e+ e− → ππ , which is the only contribution
allowed at very low s.
8. The applicability of the operator product expansion
What does the success of our fit function Eq. (24) tell us about the applicability and
usefulness of the OPE?
To answer this question let us look at the large Q2 behaviour of the formula (25):
Bs0 − A Am2ρ − (B/2)s02

CΠ Q2 = B ln Q2 + K + + + ···
Q2 (Q2 )2
∞

n−1 B n 2 −n
= B ln Q2 + K + (−1)n A m2ρ − s0 Q . (49)
n
n=1
The expansion has the same form as the OPE, at least at leading order when there are no
logarithmic corrections to the Wilson coefficients. We can now look at the higher twist
terms in the expansion, and estimate how important they are in comparison with the gluon
condensate contribution.
By looking at the first few terms we can relate our parameters s0 and A to the
condensates in the OPE. In the chiral limit there is no 1/Q2 term, so
A = Bs0 . (50)
Fig. 13. Our simple model for the contribution of the u and d quarks to R(s).
This says that the area under the meson δ-function is the same as the grey area in Fig. 13,
a typical sum-rule style result. Substituting this into the 1/Q4 term we get

1 αs
Bs0 m2ρ − s0 = −π 2 GG. (51)
2 π
Again, a typical sum-rule result—if there were no gluon condensate, the meson would lie
at s0 /2, exactly in the middle of the gap. Putting m2ρ = 0.6 GeV2 (the physical value) and
(αs /π)GG = 0.012 GeV4 gives s0 = 1.26 GeV2 .
Next, in Fig. 14, we plot a comparison of the threshold model, Eq. (25), perturbation
theory (just 3 ln a 2 Q2 at this level) and perturbation theory plus the gluon condensate
contribution, which scales like 1/(Q2 )2 . The physically irrelevant constant K has been
set to 0 in all cases.
The curve from the threshold model looks physically sensible, it goes to a finite value
at Q2 = 0, which is what must happen if R(s) does not extend all the way down to s = 0.
However, even though we chose our parameters A and s0 so that the threshold model
would match the gluon condensate prediction at high Q2 , the two curves do not resemble
each other closely.
Let us now subtract out the perturbative piece, to see things more clearly, Fig. 15. We
can only find a region where the gluon condensate region is important when we concentrate
on the large Q2 region. If Fig. 15 is what the real world looks like, the prospects for getting
at the gluon condensate look poor. The gluon condensate term is overwhelmed by higher
order terms in the OPE before it has a chance to get significant.
What is the conclusion of this exercise? In the region where we can easily see deviations
from perturbation theory the OPE is not very useful, because many operators of very high
dimension are all contributing, not just the leading 1/(Q2 )2 contribution coming from the
gluon condensate. To see the gluon condensate uncontaminated by higher order operators
we would need to look in the region Q2 ∼ 10 GeV2 with very accurate data (error bars at
least 2 orders of magnitude smaller than in this paper). This is unfortunately not a realistic
prospect.
Fig. 14. A comparison of the threshold model, Eq. (25) (solid line), perturbation theory (dotted line) and
perturbation theory plus the gluon condensate contribution (dashed line). In this plot we have chosen the
physically irrelevant constant K to be 0 when all quantities are expressed in GeV.
Fig. 15. The same as Fig. 14, but with the perturbative piece subtracted. Note that in this figure we are
concentrating on the behaviour at rather large Q2 where non-perturbative effects are small.
9. Conclusions
We have computed the vacuum polarization in the limit of two light flavours in quenched
QCD for three β values (6.0, 6.2 and 6.4) and on lattices as large as 324 . It was important
to improve the action and the vector current. We found good agreement with three-
loop perturbation theory in the interval 2 Q2 20 GeV2 . The lattice data show some
indication of non-perturbative effects at the lower end of the Q2 range. We can describe
these very well with a model of R(s) which includes vector mesons and threshold effects.
In order to make firm predictions, we need more data at small Q2 and at smaller lattice
spacings with high statistics.
From the low-Q2 region of the vacuum polarization we can extract a lattice value for
aµhad , the hadronic contribution to the muon’s anomalous magnetic moment. We find the
value 446(23) × 10−10 which is of the right order of magnitude, though lower than the
physical value. Our estimate could be improved by using a larger lattice size, enabling us to
reach lower Q2 which would reduce uncertainties from extrapolation. Naturally, dynamical
calculations would be very interesting.
Acknowledgements
The numerical calculations were performed on the Quadrics computers at DESY

Zeuthen. We thank the operating staff for their support. This work was supported in part
by the European Community’s Human Potential Program under Contract HPRN-CT-2000-
00145, Hadrons/Lattice QCD, as well as by the DFG (Forschergruppe Gitter-Hadronen-
Phänomenologie) and BMBF.
We thank C. Michael and T. Teubner for useful comments.
Appendix A
(1)
We denote the quark propagator from lattice points x to y by G(x, y). For Πµν (q̂) we
then obtain
a 4 iq(x+a µ̂/2−a ν̂/2)
(1)
Πµν (q̂) = e
4 x

× Tr (1 + γν )Uν† (0)γ5 G† (x + a µ̂, 0)γ5 (1 + γµ )Uµ† (x)G(x, a ν̂)
− (1 − γν )Uν (0)γ5G† (x + a µ̂, a ν̂)γ5 (1 + γµ )Uµ† (x)G(x, 0)
− (1 + γν )Uν† (0)γ5G† (x, 0)γ5(1 − γµ )Uµ (x)G(x + a µ̂, a ν̂)

+ (1 − γν )Uν (0)γ5G† (x, a ν̂)γ5 (1 − γµ )Uµ (x)G(x + a µ̂, 0)
a5
− cCVC eiq(x+a µ̂/2)q̂λ
4 x

× Tr (1 + γµ )Uµ† (x)G(x, 0)σνλγ5 G† (x + a µ̂, 0)γ5

− (1 − γµ )Uµ (x)G(x + a µ̂, 0)σνλ γ5 G† (x, 0)γ5
a5
+ cCVC eiq(x−a ν̂/2) q̂σ
4 x

× Tr (1 + γν )Uν† (0)γ5 G† (x, 0)γ5σµσ G(x, a ν̂)

− (1 − γν )Uν (0)γ5G† (x, a ν̂)γ5 σµσ G(x, 0)
a 6 2 iqx
− cCVC e q̂λ q̂σ
4 x

× Tr σµσ G(x, 0)σνλ γ5 G† (x, 0)γ5 , (A.1)
(2)
where the improvement term has been integrated by parts. For Πµν (q̂) we obtain
a
(2)
Πµν (q̂) = δµν Tr (1 + γν )Uν† (0)G(0, a ν̂) + (1 − γν )Uν (0)G† (0, a ν̂) . (A.2)
2
(1) (2)
To compute Πµν (q̂) and Πµν (q̂) we have to do a minimum of five inversions for each
gauge field configuration (and each κ value), which makes the calculation computationally
quite expensive, in particular on 324 lattices.
Appendix B
Perturbative results
Before we describe the lattice calculation in detail, we present here the perturbative
Wilson coefficients c0 , c4F and c4G . We will work in the quenched approximation, in which
contributions from sea quarks are neglected, and which corresponds to c4 F = 0.
We write
αs (µ2 )
c0 µ2 , Q2 , m = c0(0) µ2 , Q2 , m + CF c0(1) µ2 , Q2 , m
π

αs (µ2 ) 2 2 (2) 2 2 (2)
+ CF c0 µ , Q , m + CF CA c0 µ2 , Q2 , m
π
+ ···, (B.1)
with CF = 4/3 and CA = 3 for SU(3). These coefficients can be found in Eqs. (27)–(30)
of [5] (recall that Q2 ≡ −q 2 ). In the MS scheme the coefficients c0(0) , c0(1) , c0(2) and c0(2)
read

2 2
(0) 2 9 20 4 Q2 m̄2 4m̄ 1 1 Q2
c0 µ , Q , m̄ = −
2
− ln 2 − 8 2 + + ln ,
4 9 3 µ Q Q2 4 2 m̄2
(B.2)
(1) 2 2

c0 µ , Q , m̄

9 55 Q2 4m̄2 Q2
=− − 4ζ (3) − ln 2 − 2 4 − 3 ln 2
4 12 µ Q µ
2 2 2
4m̄ 1 11 Q 3 2 Q2 3 Q2 Q2
+ + ζ(3) + ln 2 + ln − ln ln ,
Q2 24 8 m̄ 4 m̄2 2 m̄2 µ2
(B.3)
(2) 2
c0 µ , Q2 , m̄

9 143 37 1 Q2
=− − − ζ (3) + 10ζ(5) + ln 2
4 72 6 8 µ
2
4m̄ 1667 5 35 51 Q2 9 2 Q2
− 2 − ζ(3) − ζ(5) − ln 2 + ln 2 , (B.4)
Q 96 12 6 8 µ 4 µ

c0(2) µ2 , Q2 , m̄

9 44215 227 5 41 Q2 11 2 Q2 11 Q2
=− − ζ (3) − ζ(5) − ln 2 + ln 2 + ζ(3) ln 2
4 2592 18 3 8 µ 24 µ 3 µ
2 2 2
4m̄ 1447 4 85 185 Q 11 2 Q
− 2 + ζ(3) − ζ(5) − ln 2 + ln 2 , (B.5)
Q 96 3 12 24 µ 8 µ
where ζ (3) = 1.20206 · · · and ζ (5) = 1.03693 · · ·, and m̄ refers to the quark mass at the
scale µ in the MS scheme. In (B.2), (B.3) terms of O((m̄2 /Q2 )3 ) modulo logarithms have
been neglected, while in (B.4), (B.5) terms O((m̄2 /Q2 )2 ) are dropped.
The Wilson coefficients multiplying the quark and gluon condensate are [6,8]

2 αs (µ2 )
c4F (µ, q) = −12π 2 2+ + ··· ,
3 π

11 αs (µ2 )
c4G (µ, q) = −π 2 1 − + ··· . (B.6)
18 π
Note that both mq̄q and (αs /π)G2µν are renormalization group invariants, which means
that they do not depend on µ and the renormalization scheme. The vacuum polarization
Π(Q2 ) itself is not an observable, but its derivatives are. Therefore the result (7) can only
depend on µ and the scheme in terms of an integration constant (independent of Q2 ).
Renormalization group improvement
We can use renormalization group improvement to re-sum the logarithms in higher-

order terms. This should lead to a significant improvement, since the fact that we are
interested in measurements over a large Q2 range means that these logarithms are large.
If we calculate the Adler function in the chiral limit from Eqs. (B.2)–(B.5) we find the
result

αs (µ2 ) αs (µ2 ) 2 365 11 Q2
D Q 2
=3 1+ + − 11ζ(3) − ln 2 . (B.7)
π π 24 4 µ
We should be able to do better than this because the massless Adler function is known to
four loops [3]. We can use this perturbative result for the Adler function to improve the
result for c0 in the chiral limit. In quenched SU(3) we have

αs (Q2 ) αs (Q2 ) 2 365
R(s) = 3 ef2 1 + + − 11ζ(3)
π π 24
f

αs (Q2 ) 3 87029 1103 275 121 2
+ − ζ(3) + ζ(5) − π
π 288 4 6 48
2

αs (Q2 ) 3 55 5
+ ef − ζ(3) + O αs4 (B.8)
π 72 3
f
and

αs (Q2 )

αs (Q2 ) 2 365

D Q2 = 3 ef2 1 + + − 11ζ(3)
π π 24
f

αs (Q2 ) 3 87029 1103 275
+ − ζ(3) + ζ(5)
π 288 4 6
2

αs (Q2 ) 3
55 5
+ ef − ζ(3) + O αs4 . (B.9)
π 72 3
f
Note that although the first terms of R and D coincide, the αs3 term is different. The extra
“π 2 ” term in R arises from analytic continuation,
2 ln3 (Q2 /µ2 ) → [ln(s/µ2 )±iπ]3 , see [3].
The first part of (B.9), proportional to f ef , is the derivative of the CΠ term in (6), while

the second term, proportional to ( f ef )2 , comes from the derivative of the AΠ term.
From (B.9) we have a differential equation for c0 :
∂
Q2 2
c0 µ2 , Q2 , m = 0
∂Q

αs (Q2 ) αs (Q2 ) 2 365
=3 1+ + − 11ζ(3)
π π 24
3

2
αs (Q ) 87029 1103 275 4
+ − ζ(3) + ζ(5) + O αs . (B.10)
π 288 4 6
We can solve this by using the known β-function [30]
αs (Q2 ) i+2
∂ αs (Q2 )
Q2
=− βi (B.11)
∂Q2 π π
i
with
11 51 2857 149753 891
β0 = , β1 = , β2 = , β3 = + ζ(3) (B.12)
4 8 128 1536 64
in the case of quenched SU(3).
Table 5
The vacuum polarization CΠ (q̂ 2 , am) on the 164 lattice at β = 6.0
n κ = 0.1333 κ = 0.1339 κ = 0.1342 κ = 0.1345
(1, 0, 0, 0) −38.15(11) −38.60(15) −38.84(14) −39.03(16)
(1, 1, 0, 0) −36.59(10) −36.89(13) −37.05(12) −37.17(13)
(1, 1, 1, 0) −35.52(09) −35.75(12) −35.89(11) −35.99(12)
(1, 1, 1, 1) −34.72(09) −34.91(12) −35.04(11) −35.12(12)
(2, 1, 1, 1) −32.89(08) −33.06(10) −33.15(09) −33.24(10)
(2, 2, 1, 1) −31.82(07) −31.98(10) −32.07(09) −32.15(10)
(2, 2, 2, 1) −31.07(07) −31.22(09) −31.30(08) −31.40(09)
(2, 2, 2, 2) −30.47(07) −30.62(09) −30.71(08) −30.81(09)
(3, 2, 2, 2) −29.61(06) −29.76(09) −29.85(08) −29.95(09)
(3, 3, 2, 2) −28.98(06) −29.13(08) −29.22(08) −29.32(08)
(3, 3, 3, 2) −28.47(06) −28.63(08) −28.72(07) −28.82(08)
(3, 3, 3, 3) −28.05(06) −28.20(08) −28.29(07) −28.39(08)
(4, 3, 3, 3) −27.54(05) −27.70(08) −27.79(07) −27.89(08)
(4, 4, 3, 3) −27.12(05) −27.28(07) −27.37(07) −27.47(08)
(4, 4, 4, 3) −26.76(05) −26.91(07) −27.00(07) −27.11(07)
(4, 4, 4, 4) −26.42(05) −26.58(07) −26.67(07) −26.78(07)
(5, 4, 4, 4) −26.11(05) −26.26(07) −26.36(06) −26.46(07)
(5, 5, 4, 4) −25.82(05) −25.97(07) −26.07(06) −26.17(07)
(5, 5, 5, 4) −25.55(05) −25.70(07) −25.80(06) −25.90(07)
(5, 5, 5, 5) −25.29(05) −25.44(07) −25.54(06) −25.64(07)
(6, 5, 5, 5) −25.09(05) −25.25(06) −25.34(06) −25.45(07)
(6, 6, 5, 5) −24.90(04) −25.05(06) −25.15(06) −25.25(06)
(6, 6, 6, 5) −24.71(04) −24.86(06) −24.96(06) −25.06(06)
(6, 6, 6, 6) −24.52(04) −24.67(06) −24.77(06) −24.87(06)
(7, 6, 6, 6) −24.42(04) −24.57(06) −24.67(05) −24.77(06)
(7, 7, 6, 6) −24.31(04) −24.46(06) −24.56(05) −24.66(06)
(7, 7, 7, 6) −24.19(04) −24.34(06) −24.44(05) −24.54(06)
(7, 7, 7, 7) −24.07(04) −24.22(06) −24.32(05) −24.42(06)
The solution is

c0 µ2 , Q2 , 0

12 2 3403 αs (Q2 )
= 3 ln Q2 − ln αs Q + − + 12ζ(3)
11 242 π

2
2301587 273 2
αs (Q )
+ − + ζ (3) − 25ζ(5) + O αs3 + const. (B.13)
15972 2 π
The constant of integration can be fixed by comparing with (B.1)–(B.5), giving the final
renormalization group improved result2

2 2 Q2 12 αs (Q2 ) 3403 αs (Q2 )
c0 µ , Q , 0 = 3 ln 2 − ln + − + 12ζ(3)
µ 11 αs (µ2 ) 242 π
2 Note the interesting result that due to a cancellation there is no ζ (3) term in the coefficient of α (µ2 ) and no
s
ζ (5) in the coefficient of αs (µ2 )2 . This surprised us.
Table 6 Table 7
The vacuum polarization CΠ (q̂ 2 , am) The vacuum polarization CΠ (q̂ 2 , am) on the 244 lattice at
on the 324 lattice at β = 6.0 β = 6.2
n κ = 0.1345 n κ = 0.1344 κ = 0.1349 κ = 0.1352
(1, 0, 0, 0) −41.64(22) (1, 0, 0, 0) −40.24(14) −40.87(17) −41.16(16)
(1, 1, 0, 0) −40.67(19) (1, 1, 0, 0) −38.78(12) −39.17(14) −39.35(12)
(1, 1, 1, 0) −39.85(17) (1, 1, 1, 0) −37.74(11) −38.03(13) −38.17(11)
(1, 1, 1, 1) −39.16(16) (1, 1, 1, 1) −36.96(10) −37.19(12) −37.30(11)
(2, 1, 1, 1) −37.58(14) (2, 1, 1, 1) −35.23(09) −35.40(10) −35.47(09)
(2, 2, 1, 1) −36.51(13) (2, 2, 1, 1) −34.16(08) −34.31(09) −34.38(08)
(2, 2, 2, 1) −35.72(13) (2, 2, 2, 1) −33.39(07) −33.54(08) −33.60(07)
(2, 2, 2, 2) −35.09(12) (2, 2, 2, 2) −32.79(07) −32.94(08) −33.00(07)
(3, 2, 2, 2) −34.20(12) (3, 2, 2, 2) −31.93(06) −32.08(07) −32.13(07)
(3, 3, 2, 2) −33.52(11) (3, 3, 2, 2) −31.29(06) −31.44(07) −31.49(06)
(3, 3, 3, 2) −32.99(11) (3, 3, 3, 2) −30.78(06) −30.94(07) −30.99(06)
(3, 3, 3, 3) −32.54(11) (3, 3, 3, 3) −30.35(06) −30.52(06) −30.57(06)
(4, 3, 3, 3) −31.97(11) (4, 3, 3, 3) −29.81(06) −29.98(06) −30.03(06)
(4, 4, 3, 3) −31.50(10) (4, 4, 3, 3) −29.37(06) −29.54(06) −29.59(06)
(4, 4, 4, 3) −31.11(10) (4, 4, 4, 3) −28.99(05) −29.17(06) −29.22(06)
(4, 4, 4, 4) −30.77(10) (4, 4, 4, 4) −28.67(05) −28.84(06) −28.90(05)
(5, 4, 4, 4) −30.36(10) (5, 4, 4, 4) −28.29(05) −28.46(06) −28.52(05)
(5, 5, 4, 4) −30.01(10) (5, 5, 4, 4) −27.96(05) −28.14(05) −28.20(05)
(5, 5, 5, 4) −29.70(10) (5, 5, 5, 4) −27.67(05) −27.85(05) −27.91(05)
(5, 5, 5, 5) −29.43(9) (5, 5, 5, 5) −27.41(05) −27.58(05) −27.64(05)
(6, 5, 5, 5) −29.11(9) (6, 5, 5, 5) −27.12(05) −27.30(05) −27.37(05)
(6, 6, 5, 5) −28.84(9) (6, 6, 5, 5) −26.87(05) −27.05(05) −27.11(05)
(6, 6, 6, 5) −28.59(9) (6, 6, 6, 5) −26.64(05) −26.82(05) −26.89(05)
(6, 6, 6, 6) −28.36(9) (6, 6, 6, 6) −26.42(05) −26.60(05) −26.66(05)
(7, 6, 6, 6) −28.11(9) (7, 6, 6, 6) −26.21(05) −26.39(05) −26.45(05)
(7, 7, 6, 6) −27.88(9) (7, 7, 6, 6) −26.01(05) −26.19(05) −26.26(05)
(7, 7, 7, 6) −27.68(9) (7, 7, 7, 6) −25.82(04) −26.00(05) −26.07(04)
(7, 7, 7, 7) −27.48(9) (7, 7, 7, 7) −25.64(04) −25.82(05) −25.89(04)

2301587 273 αs (Q2 ) 2
+ − + ζ(3) − 25ζ(5)
15972 2 π

2
151 αs (µ ) 566749 5 αs (µ2 ) 2
−5+ + − + ζ(3)
484 π 383328 3 π
3 2 3 2
+ O αs Q , αs µ (B.14)
where µ is the MS scale.
One can check that the differences between (B.14) and (B.1) are indeed O(αs3 ) by
making the substitution

αs (Q2 ) αs (µ2 ) αs (µ2 ) 2 Q2 αs (µ2 ) 3 Q2
→ − β0 ln 2 − β1 ln 2
π π π µ π µ
3
2
αs (µ ) Q 2
+ β02 ln2 2 + O αs4 (B.15)
π µ
in (B.14).
Table 8
The vacuum polarization CΠ (q̂ 2 , am) on the 324 lattice at
β = 6.4
n κ = 0.1346 κ = 0.1350 κ = 0.1352
(1, 0, 0, 0) −41.88(16) −42.52(22) −42.84(18)
(1, 1, 0, 0) −40.38(15) −40.77(19) −40.97(15)
(1, 1, 1, 0) −39.33(14) −39.61(17) −39.75(14)
(1, 1, 1, 1) −38.53(13) −38.74(17) −38.86(14)
(2, 1, 1, 1) −36.82(11) −36.96(13) −37.02(11)
(2, 2, 1, 1) −35.73(10) −35.84(11) −35.89(10)
(2, 2, 2, 1) −34.93(09) −35.04(10) −35.09(09)
(2, 2, 2, 2) −34.31(08) −34.42(10) −34.46(08)
(3, 2, 2, 2) −33.44(08) −33.55(08) −33.58(08)
(3, 3, 2, 2) −32.79(07) −32.89(08) −32.93(07)
(3, 3, 3, 2) −32.27(07) −32.37(07) −32.41(07)
(3, 3, 3, 3) −31.83(06) −31.94(07) −31.98(06)
(4, 3, 3, 3) −31.28(06) −31.39(06) −31.42(06)
(4, 4, 3, 3) −30.83(06) −30.94(06) −30.98(06)
(4, 4, 4, 3) −30.45(05) −30.56(06) −30.60(05)
(4, 4, 4, 4) −30.12(05) −30.23(06) −30.28(05)
(5, 4, 4, 4) −29.73(05) −29.84(05) −29.88(05)
(5, 5, 4, 4) −29.39(05) −29.51(05) −29.55(05)
(5, 5, 5, 4) −29.09(05) −29.21(05) −29.25(05)
(5, 5, 5, 5) −28.83(04) −28.95(05) −28.99(05)
(6, 5, 5, 5) −28.53(04) −28.65(05) −28.69(04)
(6, 6, 5, 5) −28.26(04) −28.39(05) −28.43(04)
(6, 6, 6, 5) −28.02(04) −28.15(04) −28.19(04)
(6, 6, 6, 6) −27.80(04) −27.93(04) −27.97(04)
(7, 6, 6, 6) −27.56(04) −27.70(04) −27.74(04)
(7, 7, 6, 6) −27.35(04) −27.48(04) −27.52(04)
(7, 7, 7, 6) −27.15(04) −27.28(04) −27.32(04)
(7, 7, 7, 7) −26.96(04) −27.10(04) −27.14(04)
Renormalization group improvement of the mass terms in (B.2)–(B.5) is simpler

because there is no constant of integration involved. The result is

m2 (Q2 ) αs (Q2 )
c0 µ2 , Q2 , m̄ = c0 µ2 , Q2 , 0 + MS 2 18 + 48
Q π

19691 124 1045 αs (Q2 ) 2
+ + ζ(3) − ζ(5)
24 3 3 π
m2 (Q2 ) 2
Q2
+ MS
−9 − 18 ln
Q2 m2 (Q2 )
MS

2
Q Q2 αs (Q2 )
− 2 + 48ζ(3) + 66 ln 2 + 36 ln 2
2
.
m (Q2 ) m (Q2 ) π
MS MS
(B.16)
What have we gained by using the renormalization group? Most importantly, the Q2
derivative of (B.14) is correct to four loops, one order better than (B.1). The uncertainties
in (B.14) are O(αs3 (Q2 ), αs3 (µ2 )) while in the original formula (B.1) terms of order
O(αs3 (µ2 ) ln3 (Q2 /µ2 )) have been neglected. So, if we are interested in describing a large
range of Q2 , Eq. (B.14) ought to be the better formula. In addition, (B.14) exhibits correct
physics, because it is the sum of a piece depending only on Q2 and a piece depending only
on µ2 , which is only approximately the case for (B.1).
Appendix C
In Tables 5–8 we give the results for CΠ (q̂, am) at the various β and κ values. We label
the lattice momenta by the vector n, q̂µ = (2/a) sin(πnµ /L). The momenta were chosen
close to the diagonal of the Brillouin zone to avoid large O(a 2 ) effects.
References
[1] F.J. Reinders, H. Rubinstein, S. Yazaki, Phys. Rep. 127 (1985) 1;

S. Narison, QCD Spectral Sum Rules, World Scientific, 1989.
[2] S.L. Adler, Phys. Rev. D 10 (1974) 3714.
[3] S.G. Gorishny, A.L. Kataev, S.A. Larin, Phys. Lett. B 259 (1991) 144;
L.R. Surguladze, M.A. Samuel, Phys. Rev. Lett. 66 (1991) 560;
L.R. Surguladze, M.A. Samuel, Phys. Rev. Lett. 66 (1991) 2416, Erratum.
[4] K.G. Chetyrkin, Phys. Lett. B 391 (1997) 402, hep-ph/9608480.
[5] K.G. Chetyrkin, J.H. Kühn, M. Steinhauser, Nucl. Phys. B 482 (1996) 213.
[6] M.A. Shifman, A.I. Vainshtein, V.I. Zakharov, Nucl. Phys. B 147 (1979) 385.
[7] M.A. Shifman, A.I. Vainshtein, V.I. Zakharov, Nucl. Phys. B 147 (1979) 448.
[8] S.C. Generalis, J. Phys. G 15 (1989) L225;
K.G. Chetyrkin, S.G. Gorishny, V.P. Spiridonov, Phys. Lett. B 160 (1985) 149;
V.P. Spiridonov, K.G. Chetyrkin, Yad. Fiz. 47 (1988) 818;
K.G. Chetyrkin, Sov. J. Nucl. Phys. 47 (1988) 522.
[9] S. Eidelman, F. Jegerlehner, A.L. Kataev, O. Veretin, Phys. Lett. B 454 (1999) 369, hep-ph/9812521.
[10] M. Göckeler, R. Horsley, W. Kürzinger, V. Linke, D. Pleiter, P.E.L. Rakow, G. Schierholz, Nucl. Phys. B
(Proc. Suppl.) 94 (2001) 571, hep-lat/0012010;
M. Göckeler, R. Horsley, W. Kürzinger, V. Linke, D. Pleiter, P.E.L. Rakow, G. Schierholz, hep-lat/0310027.
[11] M. Göckeler, R. Horsley, H. Perlt, P. Rakow, G. Schierholz, A. Schiller, P. Stephenson, Phys. Rev. D 57
(1998) 5562, hep-lat/9707021.
[12] M. Göckeler, P.E.L. Rakow, R. Horsley, D. Petters, D. Pleiter, G. Schierholz, in: V. Mitrjushkin,
G. Schierholz (Eds.), Lattice Fermions and Structure of the Vacuum, Kluwer, Dordrecht, 2000, p. 201;
M. Göckeler, R. Horsley, D. Petters, D. Pleiter, P.E.L. Rakow, G. Schierholz, P. Stephenson, Nucl. Phys. B
(Proc. Suppl.) 83 (2000) 203, hep-lat/9909160.
[13] M. Göckeler, et al., in preparation.
[14] W. Kürzinger, Ph.D. Thesis (in German), Freie Universität Berlin, 2001, http://www.diss.fu-berlin.de/
2001/220/.
[15] I. Montvay, G. Münster, Quantum Fields on a Lattice, Cambridge Univ. Press, Cambridge, 1994.
[16] S. Capitani, M. Göckeler, R. Horsley, H. Perlt, P.E.L. Rakow, G. Schierholz, A. Schiller, Nucl. Phys. B 593
(2001) 183.
[17] M. Lüscher, S. Sint, R. Sommer, P. Weisz, U. Wolff, Nucl. Phys. B 491 (1997) 323.
[18] S. Booth, M. Göckeler, R. Horsley, A.C. Irving, B. Joo, S. Pickles, D. Pleiter, P.E.L. Rakow, G. Schierholz,
Z. Sroczynski, H. Stüben, Phys. Lett. B 519 (2001) 229.
[19] S. Capitani, M. Lüscher, R. Sommer, H. Wittig, Nucl. Phys. B 544 (1999) 669;
P. Boucaud, G. Burgio, F. Di Renzo, J.P. Leroy, J. Micheli, C. Parrinello, O. Pène, C. Pittori, J. Rodriguez-
Quintero, C. Roiesnel, K. Sharkey, JHEP 0004 (2000) 006.
[20] M. Guagnelli, R. Sommer, H. Wittig, Nucl. Phys. B 535 (1998) 389.
[21] O. Dumbrajs, R. Koch, H. Pilkuhn, G.C. Oades, H. Behrens, J.J. De Swart, P. Kroll, Nucl. Phys. B 216
(1983) 277.
[22] Particle Data Group, K. Hagiwara, et al., Phys. Rev. D 66 (2002) 010001.
[23] R.A. Bertlmann, G. Launer, E. de Rafael, Nucl. Phys. B 250 (1985) 61.
[24] T. Blum, Phys. Rev. Lett. 91 (2003) 052001;
T. Blum, hep-lat/0310064.
[25] C. Bouchiat, L. Michel, J. Phys. Radium 22 (1961) 121.
[26] A. Nyffeler, hep-ph/0305135.
[27] B.E. Lautrup, A. Peterman, E. de Rafael, Phys. Rep. 3 (1972) 193.
[28] S. Eidelman, F. Jegerlehner, Z. Phys. C 67 (1995) 585.
[29] F. Jegerlehner, J. Phys. G 29 (2003) 101;
K. Hagiwara, A.D. Martin, D. Nomura, T. Teubner, Phys. Lett. B 557 (2003) 69;
K. Hagiwara, A.D. Martin, D. Nomura, T. Teubner, hep-ph/0312250;
M. Davier, S. Eidelman, A. Hocker, Z. Zhang, Eur. Phys. J. C 27 (2003) 497.
[30] T. van Ritbergen, J.A.M. Vermaseren, S.A. Larin, Phys. Lett. B 400 (1997) 379;
J.A.M. Vermaseren, S.A. Larin, T. van Ritbergen, Phys. Lett. B 405 (1997) 327.
The analytic value of a 4-loop sunrise graph

in a particular kinematical configuration
S. Laporta a,b , P. Mastrolia c,d , E. Remiddi e,c,a
a INFN, Sezione di Bologna, I-40126 Bologna, Italy
b Dipartimento di Fisica, Università di Parma, I-43100 Parma, Italy
c Dipartimento di Fisica, Università di Bologna, I-40126 Bologna, Italy
d Institut für Theoretische Teilchenphysik, Universität Karlsruhe, D-76128 Karlsruhe, Germany
e Theory Division, CERN, CH-1211 Geneva 23, Switzerland
Received 2 December 2003; received in revised form 27 February 2004; accepted 6 March 2004
Abstract
The 4-loop sunrise graph with two massless lines, two lines of equal mass M and a line of mass m,
for external invariant timelike and equal to m2 is considered. We write differential equations in
x = m/M for the Master Integrals of the problem, which we Laurent-expand in the regularizing
continuous dimension d around d = 4, and then solve exactly in x up to order (d − 4)3 included; the
result is expressed in terms of harmonic polylogarithms of argument x and maximum weight 7. As a
by product, we obtain the x = 1 value, expected to be relevant in QED 4-loop static quantities like the
electron (g − 2). The analytic results were checked by an independent precise numerical calculation.
 2004 Published by Elsevier B.V.
PACS: 11.15.Bt; 12.20.Ds
Keywords: Feynman diagrams; Multi-loop calculations
1. Introduction
This paper is devoted to the analytic evaluation of the master integrals (MIs) associated
to the 4-loop sunrise graph with two massless lines, two massive lines of equal mass M,
another massive line of mass m, with m = M, and the external invariant timelike and equal
to m2 , as depicted in Fig. 1.
E-mail addresses: stefano.laporta@bo.infn.it (S. Laporta), pierpaolo.mastrolia@bo.infn.it (P. Mastrolia),

ettore.remiddi@bo.infn.it (E. Remiddi).
0550-3213/$ – see front matter  2004 Published by Elsevier B.V.

doi:10.1016/j.nuclphysb.2004.03.029
166 S. Laporta et al. / Nuclear Physics B 688 (2004) 165–188
Fig. 1. The considered 4-loop sunrise graph.
We will follow for the analytic integration the differential equation method already
proposed in [1], further developed in [2] (and then used in [3,4] and in many subsequent
applications which would be too long to report here), as well as the finite difference
method [5,6] for an independent numerical check.
The differential equation method was already followed in two similar two- and three-
loop calculations [7,8]; the fact that its use could be extended without major changes to the
present four-loop calculation witnesses for its generality and power. Among the advantages
of the method, it allows a rather clear separation of the merely algebraic part of the work
(which is, not surprisingly, always very heavy in this kind of multiloop calculations, and
can be most conveniently processed by a computer algebra program, in our case FORM [9]),
from the really analytic issues of the problem, which can then be better investigated without
the disturbance of the algebraic complexity. In our case, indeed, the heart of the analytic
calculation was the study of a homogeneous fourth order differential equation, whose
solutions turned out to be, in a remarkably simple way, either a rational fraction or repeated
quadratures of rational fractions. The required four-loop integral could then be obtained
almost mechanically by repeated quadratures in terms of harmonic polylogarithms [10].
Several other different approaches to the analytic evaluation of multiloop integrals are
available in the literature, such as the powerful asymptotic expansion method (a fairly
complete account can be found in the recent book [11]), and it would be interesting to
compare the advantages and drawbacks of the various methods for the exacting, four-loop
integration which we consider; but as the results of the present paper are new, a meaningful
comparison cannot yet be carried out.
Following, as already said above, the approach already used in [7,8], we identify the
MIs of the current problem within the continuous d-dimensional regularization, write the
system of differential equations in x = m/M satisfied by the MIs, convert it into a higher
order differential equation for a single MI, Laurent-expand in (d − 4) around d = 4, solve
the associated homogeneous equation at d = 4 (as in previous cases, the solutions of the
homogeneous equation are surprisingly simple) and then use recursively Euler’s method of
the variation of the constants for obtaining the coefficient of the (d − 4) expansion in closed
analytic form. The result involves harmonic polylogarithms (HPLs) [10] of argument x and
weight increasing with the order in d − 4. We push the analytic calculation, which works
up to virtually any order in d − 4, up to (d − 4)3 included, involving HPLs of weight up
to w = 7 included. The integration constants are fixed at x = 0; as a by product we obtain
the values at x = 1, which are relevant in the evaluation of 4-loop static quantities such
as the electron (g − 2) in QED. The result was checked and confirmed by an independent
numerical calculation performed with the method of [5,6].
S. Laporta et al. / Nuclear Physics B 688 (2004) 165–188 167
The plan of the paper is the following. In Section 2 we define the master integrals and
write the differential equations; in Section 3 we study the x → 0 behaviour; in Section 4 we
work out the Laurent-expansion in d −4 and discuss the associated homogeneous equation;
in Section 5 we write down the solutions by Euler’s method of the variation of the constants
and give the x = 1 values; Section 6 deals with the independent numerical calculation by
which we check the results at x = 1 of the previous section. in Section 7 we carry out the
evaluation of the x → 0 values by direct integration in the parametric space.
2. The master integrals and the differential equations
We find that the problem has 5 MIs, which we choose to be
Fi (d, M 2 , m2 , P 2 = −m2 )

1 d d k1 d d k2 d d k3 d d k4 Ni
= ,
(2π)4(d−2) k12 k22 (k32 + M 2 )(k42 + M 2 )[(P − k1 − k2 − k3 − k4 )2 + m2 ]
(2.1)
where the 5 numerators Ni are M 2 , k1 · k3 , p · k3 , k1 · k2 , p · k2 . In terms of the
dimensionless variable x = m/M and putting P = Mp one can introduce 5 dimensionless
functions Φi (d, x) through
Fi (d, M 2 , m2 , P 2 = −m2 ) = M 4d−8 C 4 (d)Φi (d, x), (2.2)

4−d
where C(d) = (4π) 2 (3 − d/2) is an overall loop normalization factor, with the limiting
value C(4) = 1 at d = 4. Some of the formulae which will follow (in particular, the
differential equations) are slightly simpler when written in terms of x 2 rather than x; but
as x is the most convenient variable for expressing the final analytic results, we stick to x
from the very beginning.
As in [7,8], the derivation of the system of differential equations is straightforward; the
derivatives of the MIs, i.e., of the 5 functions Φi (d, x), with respect to x are easily carried
out in their representation as loop-integrals Eq. (2.1); when the result is in turn expressed
in terms of the same MIs, one obtains the following linear system of first order differential
equations in x

dΦ1 (d, x) (3d − 7) 3(d − 2) 3(d − 2)
= + − Φ1 (d, x)
dx x 2(1 − x) 2(1 + x)

3(d − 2) 3(d − 2) 3(d − 2)
− + −
x 2(1 − x) 2(1 + x)

× 3Φ2 (d, x) − 3Φ3 (d, x) + Φ4 (d, x) − 3Φ5 (d, x) , (2.3)
dΦ2 (d, x) (d − 2)
=− Φ2 (d, x) − 2Φ3 (d, x) + Φ4 (d, x) − 2Φ5 (d, x) , (2.4)
dx x

dΦ3 (d, x) 3(d − 2) 3(d − 2)
=− − Φ1 (d, x) − 3Φ2 (d, x) − Φ4 (d, x)
dx 2(1 − x) 2(1 + x)

3(d − 2) 9(d − 2) 9(d − 2)
− + − Φ3 (d, x) + Φ5 (d, x) , (2.5)
x 2(1 − x) 2(1 + x)
dΦ4 (d, x) 2(d − 2)
= Φ2 (d, x) + Φ4 (d, x) , (2.6)
dx x
dΦ5 (d, x) 2(d − 2)
= Φ3 (d, x) + Φ5 (d, x) . (2.7)
dx x
At variance with the cases discussed in [7,8], the system is homogeneous; indeed, quite
in general the non homogeneous terms are given by the MIs of the “subtopologies” of the
considered graph, obtained by shrinking to a point any of its propagator lines. When that
is done for the 5-propagator “topology” of Fig. 1, one obtains the product of 4 tadpoles;
but as the considered graph has two massless propagators, at least one massless tadpole
is always present in the product; as in the d-dimensional regularization massless tadpoles
vanish, the product of the 4 tadpoles is always equal to zero—and therefore the differential
equations are homogeneous.
By inspection, one sees that Φ3 (d, x), Φ5 (d, x) appear in the r.h.s. of Eqs. (2.3)–(2.7)
only in the combination
Ψ3 (d, x) = Φ3 (d, x) + Φ5 (d, x); (2.8)

the other linearly independent combination of the two MIs, say
Ψ5 (d, x) = Φ3 (d, x) − Φ5 (d, x), (2.9)

decouples and can be expressed in terms of the other integrals by means of the trivial 1st
order differential equation

dΨ5 (d, x) 3(d − 2) 3(d − 2)
=− − Φ1 (d, x) − 3Φ2 (d, x) − Φ4 (d, x)
dx 2(1 − x) 2(1 + x)

5(d − 2) 9(d − 2) 9(d − 2)
− + − Ψ3 (d, x). (2.10)
x 2(1 − x) 2(1 + x)
As Ψ5 (d, x) does not enter in the r.h.s. of Eqs. (2.3)–(2.7) the 4 linear equations for
Φ1 (d, x), Φ2 (d, x), Ψ3 (d, x) and Φ4 (d, x) can be written as a fourth order equation for
the first master integral Φ1 (x), which will be called simply Φ(d, x) from now on, and
which is therefore equal to

C −4 (d) d d k1 d d k2 d d k3 d d k4
Φ(d, x) = ,
(2π)4(d−2) k12 k22 (k32 + 1)(k42 + 1)[(p − k1 − k2 − k3 − k4 )2 + x 2 ]
p2 = −x 2 . (2.11)
One obtains for Φ(d, x) the following 4th order differential equation
d 4 Φ(d, x) 3
2 d Φ(d, x)
x 3 (1 − x 2 ) + x 2
1 + 5x 2
− 3(d − 4)(1 − 3x )
dx 4 dx 3
d 2 Φ(d, x)
− x 12 + 6x 2 + (d − 4)(13 + 32x 2) + (d − 4)2 (1 + 26x 2)
dx 2

+ 12 − 18x + (d − 4)(25 − 2x ) + 8(d − 4) (2 + 5x )
2 2 2 2
dΦ(d, x)
+ 3(d − 4)3 (1 + 8x 2 )
dx

+ 4x +12 + 29(d − 4) + 23(d − 4)2 + 6(d − 4)3 Φ(d, x) = 0. (2.12)
The expression of the other MIs in terms of Φ(d, x) and its first 3 x-derivatives reads

1 1 x2 2x 2 7x 2
Φ2 (d, x) = − + − + Φ(d, x)
2 6(d − 2) 5(d − 2)3 3(d − 2)2 15(d − 2)

x(1 − 7x 2 ) x(1 + 16x 2) x(11 + 28x 2 ) dΦ(d, x)
− + −
10(d − 2)3 12(d − 2)2 60(d − 2) dx
2 2
x (1 − 2x ) x (1 − 10x ) d Φ(d, x)
2 2 2
− −
5(d − 2)3 30(d − 2)2 dx 2
x 3 (1 − x 2 ) d 3 Φ(d, x)
− , (2.13)
20(d − 2)3 dx 3

x2 1x 2 3x 2
Φ3 (d, x) + Φ5 (d, x) = − − + Φ(d, x)
15(d − 2)3 3(d − 2)2 5(d − 2)

x(1 − 7x 2 ) x(1 − 8x 2 ) x(1 − 12x 2) dΦ(d, x)
+ − +
30(d − 2)3 12(d − 2)2 20(d − 2) dx
2 2
x (1 − 2x ) x (2 − 5x ) d Φ(d, x)
2 2 2
+ −
15(d − 2)3 30(d − 2)2 dx 2
x 3 (1 − x 2 ) d 3 Φ(d, x)
− , (2.14)
60(d − 2)3 dx 3

1 1 4x 2 3x 2 43x 2
Φ4 (d, x) = − − + − + Φ(d, x)
2 6(d − 2) 5(d − 2)3 (d − 2)2 15(d − 2)

2x x(14 − 30x 2) x(11 + 43x 2) dΦ(d, x)
+ − −
5(d − 2)3 5(d − 2)2 15(d − 2) dx
2 2
4x (1 − 2x ) 3x (1 − 5x ) d Φ(d, x)
2 2 2
+ −
5(d − 2)3 10(d − 2)2 dx 2
x 3 (1 − x 2 ) d 3 Φ(d, x)
+ . (2.15)
5(d − 2)3 dx 3
3. The x → 0 behaviour of Φ(d, x)
By inspection, one finds that the most general solution of Eq. (2.12) can be expanded
for x → 0 in the form

4
Φ(d, x) = x αi
A(i)
n (d)x
2n
, (3.1)
i=1 n=0
where the values of the 4 exponent αi are
α1 = 0,
α2 = (d − 2),
α3 = −(d − 2),
α4 = (3d − 7); (3.2)
the A(i) (i)
0 (d) are the 4 integration constants, and all the other coefficients An (d) for n > 0
are determined by the differential equation Eq. (2.12) once the integration constants are
fixed.
It is interesting to recall the leading exponents of the x → 0 expansions of the solutions
(2)
of the corresponding equations encountered in [7,8]. Calling αi the exponents for the 2-
(3)
loop graph of [7] and αi those of the 3-loop graph of [8], one finds that there are always
4 exponents. The explicit values in the case of [7] are
(2)
• α1 = 0,
(2)
• α2 = (d − 2),
α3(2) = −(d − 2),
α4(2) = (d − 3) (3.3)
and in the case of [8]
(3)
α1 = 0,
• α2(3) = (d − 2),
α3(3) = −(d − 2),
(3)
α4 = (2d − 5), (3.4)
where the exponents marked by a bullet (•) correspond to the behaviours forced by the
inhomogeneous terms. The similarity between the 3 sets of behaviours is impressive: the
first 3 exponents are identical, the fourth differ in steps of d − 2 for each additional loop.
In more details, 2 of the exponents of Eqs. (3.3) correspond to the 2 independent
solutions of the associated homogeneous differential equation, which is of 2nd order,
while the other 2 are forced by the 2 independent behaviours for x → 0 developed by
the inhomogeneous terms (both products of 2 tadpoles, the first product of two tadpoles
of mass M, the second of a tadpole of mass M and a tadpole of mass m = Mx). In the
x → 0 expansion of the most general solution, one is therefore left with 2 undetermined
integration constants, corresponding to the two homogeneous solutions, while all the
terms with the behaviours of the inhomogeneous terms are fully determined by the
inhomogeneous equation itself. Similarly, in Eqs. (3.4) 3 exponents correspond to the 3
solutions of the 3rd order homogeneous equation, with the 4th exponent forced by the
inhomogeneous term (product of 3 tadpoles, of masses M, M and m = Mx), so that

there are in principle 3 undetermined integration constants, the term corresponding to the
remaining behaviour being fixed by the equation. In the present case, finally, the equation
is homogeneous (as already observed, there is no inhomogeneous term, as all the possible
products of 4 tadpoles involve at least a vanishing zero-mass tadpole) and the 4 exponents
of Eqs. (3.2) correspond to the behaviour of the 4 homogeneous solutions, so that one is
left in principle with 4 undetermined integration constants.
A qualitative inspection of the integrals which one tries to evaluate by means of the
differential equations (Eq. (2.11) of the present paper, Eq. (1.3) of [7] and Eq. (2.3) of [8])
shows that they are all finite (just finite, not analytic!) for x → 0+ and (d − 2) > 0; that
is sufficient to rule out from their expression as solutions of the differential equation the
terms with the behaviour of the third and the fourth exponent (which is negative when d is
just above 2).
In the case of [7], that fixes completely the solution. In the case of [8], one integration
constant is left undetermined; to fix it, one has to provide some independent information,
such as the value of the required Feynman integral at x = 0 (which corresponds to a simpler
vacuum amplitude); that value can be provided by an explicit “conventional” calculation,
say in parameter space, which is in any case much easier than a calculation for non-zero
values of the variable x. (A closer analysis carried out in [8] shows however that the
regularity of the integral at x = 1 is sufficient to fully determine the solution, so that the
actual knowledge of the x = 0 value can be used as an independent check.)
In the current case, as the equation for Φ(d, x) is homogeneous, the only information
is that A0 (d) and A0 (d) are both equal to zero due to the finiteness for x → 0+ ; by
(3) (4)
substituting the ansatz Eq. (3.1) in Eq. (2.12) and dropping A(3) (4)
0 (d), A0 (d), one finds for
Φ(d, x) Eq. (2.11) the x → 0 expansion

(1) 2(2d − 5)(3d − 8) 2
Φ(d, x) = A0 (d) 1 − x + O(x ) 4
3d(d − 4)

(2) (d − 3)(d − 4)(3d − 8) 2
+ A0 (d)x d−2
1+ x + O(x ) .
4
(3.5)
2d(2d − 7)
(1)
The expansion depends on the two as yet unspecified integrations constants A0 (d),
A(2)
0 (d)—which are to be provided by an independent, explicit calculation. That is done in
Section 7, see Eq. (7.14), by direct integration in the parametric space.
Let us note here, for completeness, that in the present case the knowledge of the
regularity of the solution at x = 1 does not provide any additional information.
4. The expansion in d − 4 and the homogeneous equation at d = 4
The Laurent’s expansion in d − 4 of Φ(d, x) Eq. (2.11) is

∞

Φ(d, x) = (d − 4)n Φ (n) (x), (4.1)
n=−4
as it is known on general grounds that it develops at most a fourth order pole in d − 4. By

substituting in Eq. (2.12) one obtains a system of inhomogeneous, chained equations for
the coefficients Φ (n) (x) of the expansion in d − 4; the generic equation reads

d4 d3
x 3 (1 − x 2 ) 4 + x 2 (1 + 5x 2 ) 3
dx dx
2
d d
− 6x(2 + x 2 ) 2 + 6(2 − 3x 2 ) + 48x Φ (n) (x) = K (n) (x), (4.2)
dx dx
with

d
K (n)
(x) = 24x + (3 + 24x ) 2
Φ (n−3) (x)
dx

d d2
+ 92x + (16 + 40x 2 ) − (x + 26x 3) 2 Φ (n−2) (x)
dx dx

d d2
+ 116x + (25 − 2x 2 ) − (13x + 32x 3) 2
dx dx
3
d
− (3x 2 − 9x 4 ) 3 Φ (n−1) (x), (4.3)
dx
which shows that the equation at a given order n for Φ (n) (x) involves in the inhomogeneous
term the coefficients Φ (k) (x) (and their derivatives) with k < n—hence the “chained
equations” expression used above (obviously, Φ (k) (x) = 0 when k < −4). Such a structure
calls for an algorithm of solution bottom-up, i.e., starting from the lowest value of n (which
is n = −4) and proceeding recursively to the next n + 1 value up to the required order.
Eq. (4.2) have all the same associated homogeneous equation, independent of n,

d4 d3
x 3 (1 − x 2 ) 4 + x 2 (1 + 5x 2 ) 3
dx dx
2
2 d 2 d
− 6x(2 + x ) 2 + 6(2 − 3x ) + 48x φ(x) = 0; (4.4)
dx dx
once the solutions of Eq. (4.4) are known, all Eq. (4.2) can be solved by the method of the
variation of the Euler constants.
To our (pleasant) surprise, the solutions of Eq. (4.4) are almost elementary. By trial and
error, a first solution is found to be
φ1 (x) = x 2 . (4.5)
We then substitute φ(x) = φ1 (x)ξ(x) in Eq. (4.4), obtaining the following 3rd order
equation for the derivative of ξ(x)

d3 d2
x 3 (1 − x 2 ) 3 + 3x 2 (3 − x 2 ) 2
dx dx

2 d
+ 6x(1 + 2x ) − 6(5 + 2x ) ξ (x) = 0,
2
(4.6)
dx
and find that it admits the solution

1
ξ2 (x) = 3 (1 − x 2 + x 4 ). (4.7)
x
Substituting ξ (x) = ξ2 (x)χ(x) in Eq. (4.6) we obtain the following 2nd order equation for
χ (x)

d2 d
x 2 (1 − x 2 )(1 − x 2 + x 4 ) 2 + 6x 5 (2 − x 2 )
dx dx

− 6(2 − 2x 4 + x 6 ) χ (x) = 0, (4.8)
which admits as solution

2 4 5 − 2x + 5x
1 2 4
χ3 (x) = (1 − x ) . (4.9)
x3 (1 − x 2 + x 4 )2
Finally, substituting χ (x) = χ3 (x)τ (x) in Eq. (4.8), we obtain the equation

d
x(1 − x 2 )5 (1 − x 2 + x 4 )(5 − 2x 2 + 5x 4 )
dx

− 2(1 − x ) (15 − 12x + 11x + 30x − 24x + 20x ) τ (x) = 0,
2 4 2 4 6 8 10
(4.10)
which has the solution

x6 1 − x2 + x4
τ4 (x) = . (4.11)
(1 − x 2 )5 5 − 2x 2 + 5x 4
By repeated quadratures in x and multiplications by the previous solutions we obtain
the explicit analytic expressions of the 4 solutions of Eq. (4.4); the nasty denominators
(1 − x 2 + x 4 ) and (5 − 2x 2 + 5x 4 ) disappear in the final results, while the repeated
integrations of the terms with denominators x, 1 + x and 1 − x generate, almost by
definition, harmonic polylogarithms [10] of argument x and weight increasing up to 3.
Explicitly, one finds
1
φ2 (x) = − (1 − x 4 ) − H (0, x)x 2, (4.12)
2
(5 + 18x 2 + 14x 6 + 5x 8) 1
φ3 (x) = + (12 + x 2 − 12x 4)H (0; x)
8x 2 2
+ 12x H (0, 0; x),
2
(4.13)
(1 + x 2 )(15 + 182x 2 + 15x 4)
φ4 (x) =
65536x
3(1 − x 2 )2 (5 + x 2 )(1 + 5x 2)
+ H (−1; x) + H (1; x)
131072x 2
9(1 − x 4 )
− H (0, −1; x) + H (0, 1; x)
8192
9x 2
− H (0, 0, −1; x) + H (0, 0, 1; x) . (4.14)
4096
The corresponding Wronskian has the remarkably simple expression

φ1 (x) φ2 (x) φ3 (x) φ4 (x)

φ (x) φ2 (x) φ3 (x)
φ4 (x) (1 − x 2 )3
W (x) = 1 = , (4.15)
φ1 (x) φ2 (x) φ3 (x)
φ4 (x) x

φ1 (x) φ2 (x) φ3 (x) φ4 (x)
in agreement (of course) with the coefficients of the 4t h and 3rd x-derivative of φ(x) in
Eq. (4.4).
5. The solution of the differential equations for the coefficients of the expansion
in d − 4
With the results established in the previous section one can use Euler’s method of the
variation of the constants for solving Eq. (4.2) recursively in n, starting from n = −4. We
write Euler’s formula as
x

4
dx (n)
Φ (n)
(x) = φi (x) Φi(n) + Mi (x )K (x ) , (5.1)
W (x )
i=1 0
where the φi (x) are the solutions of the homogeneous equation given in Eqs. (4.5), (4.14),
(n)
the Φi are the as yet undetermined integration constants, the Wronskian W (x) can be
read from Eq. (4.15), the Mi (x) are the minors of the φi (x) in the determinant Eq. (4.15),
and the K (n) (x) are the inhomogeneous terms of Eq. (4.3). The constants Φi(n) are then
fixed by comparing the expansion in x for x → 0 of Eq. (5.1) with the expansion in d − 4
for d → 4 of Eq. (3.5). Explicitly, we find, for instance,
1 2
Φ (−4) (x) = − x , (5.2)
64
1 9 2 1 4 1
Φ (−3) (x) = − + x − x − x 2 H (0; x), (5.3)
384 256 192 48
41 143 2 37 4 5 6
Φ (−2) (x) = − x + x − x
4608
3072 2304 9216
3 1 1
+ − x 2 x 2 H (0; x) − x 2 H (0, 0; x), (5.4)
64 96 48
805 205 2 743 4 599 6
Φ (−1) (x) = − + x − x + x
55296
4096 27648 221184
43 11 2 5 4 2
− − x + x x H (0; x)
768 384 2304

3 1 1
+ − x 2 x 2 H (0, 0; x) − x 2 H (0, 0, 0; x), (5.5)
64 96 48
3173 1 2443 2 1
Φ (0) (x) = − π 4x2 − x − x 4 ζ(3)
663552 3840 49152 96
15649 4 34061 6 1
+ x − x + ζ(3)
331776
5308416 96
5 1 1 121 4 599 6
− − ζ (3)x + 2
x +
2
x − x H (0; x)
1536 48 768 2304 55296

1 5 1 7 2 1 5 6
+ + − x + x4 + x
96 1536 x 2 256 96 1536

× H (1, 0; x) − H (−1, 0; x)

43 2 11 4 5 6
− x − x + x H (0, 0; x)
768 384 2304
(1 − x 4 )
+ H (0, 1, 0; x) − H (0, −1, 0; x)
32
1 2
+ x H (0, 0, 1, 0; x) − H (0, 0, −1, 0; x)
16

3 1 2 2 1
+ − x x H (0, 0, 0; x) − x 2 H (0, 0, 0, 0; x). (5.6)
64 96 48
The full results become quickly too lengthy to be reported explicitly here, so we give only
(k)
the values of the integration constants up to order 3 included. We find Φ4 = 0, for (all)
−4 k 3; the other constants are:
24978775 32419 4181 7

Φ1(1) = + ζ(3) + π 4 − ζ(5); (5.7)
603979776 786432 5242880 64
(1) 19014089 34481 8507
Φ2 = − + ζ(3) − π 4; (5.8)
905969664 393216 7864320
264575 871 19
Φ3(1) = − ζ(3) − π 4; (5.9)
452984832 196608 3932160
(2) 1791217627 476887 37601
Φ1 = − ζ(3) + π4
7247757312 3145728 62914560
87801 1 13
+ ζ (5) + ζ 2 (3) − π 6; (5.10)
262144 32 32256
(2) 674733523 691111 79069 43165
Φ2 = − ζ(3) + π4 − ζ(5); (5.11)
10871635968 4718592 18874368 131072
191322157 27041 1787 133
Φ3(2) = − + ζ(3) − π4 − ζ(5); (5.12)
5435817984 2359296 9437184 65536
(3) 40220928899 6041939 775961 3855
Φ1 = − + ζ(3) − π4 + ζ(5)
21743271936 9437184 150994944 1048576
12543 2 54353 1 605
− ζ (3) + π6 + ζ(3)π 4 − ζ(7); (5.13)
131072 44040192 384 512
Fig. 2. The on-shell 4-loop sunrise graph.
(3) 336259373 1366211 7536877

Φ2 = + ζ(3) − π4
16307453952 14155776 1132462080
1939871 15281 2 69631
+ ζ (5) + ζ (3) − π 6; (5.14)
1572864 196608 66060288
3741658007 381061 275627
Φ3(3) = − ζ(3) + π4
16307453952 7077888 566231040
41641 19 2 247
− ζ (5) + ζ (3) − π 6. (5.15)
786432 32768 33030144
As a byproduct, using the results of [12], we obtain the values at x = 1, i.e., the on
mass-shell values at −p2 = m2 = M 2 of Eq. (2.1), depicted in Fig. 2,
1 7 17
Φ(d, x = 1) = − + −
64(d − 4)4 256(d − 4)3 768(d − 4)2
835 7379 1
+ − − π4
73728(d − 4) 1769472 1920

6766055 53 2 7
+ (d − 4) − + π + π4
42467328 1152 7680

7 2 29
− π ζ(3) + ζ(5)
256 256

2 1449210865 1913 2 53 2 1855
+ (d − 4) − π + π ln 2 − ζ(3)
1019215872 4608 128 2304
17 49 2 203
− π4 + π ζ(3) − ζ(5)
23040 1024 1024
63 2 9 3 2 4
− π ζ(3) ln 2 − a4 π 2 − π ln 2
256 32 256
3 4 2 245 2 661
+ π ln 2 + ζ (3) + π6
256 1024 322560

182188906799 734695 2 1913 2
+ (d − 4)3 − + π − π ln 2
24461180928 331776 512
66955 583 2 2 8707 4 265 4
+ ζ(3) + π ln 2 − π + ln 2
9216 384 147456 768
265 119 2 493
+ a4 − π ζ(3) + ζ(5)
32 3072 3072
441 2 21 2 4 21 4 2
+ π ζ(3) ln 2 + π ln 2 − π ln 2
1024 1024 1024
661 63 1715 2
− π6 + a4 π 2 − ζ (3)
184320 128 4096
627 2 123 2 5 9917 2
− π ζ(3) ln2 2 − π ln 2 + π ζ(5)
512 1280 4096
107 4 3 109 4 2205 2
+ π ln 2 − π ζ(3) + ζ (3) ln 2
1536 12288 1024
45 15 6 81
+ ζ(3) ln4 2 − π ln 2 − a4 π 2 ln 2
128 2048 32
45 81 45 1395
+ a4 ζ(3) − a5 π 2 − a5 ln2 2 + ζ(5) ln2 2
16 32 16 1024
45 45 135 45 45
− a6 ln 2 + b6 ln 2 − a7 + b7 − d7
4 16 8 8 32

15 29335
+ ln7 2 + ζ(7) + O (d − 4)4 . (5.16)
1792 2048
At variance with [12], we have expressed the results in terms of the constants listed in
Table 1; the first column is the name of the constant, the second column its value as
harmonic polylogarithm of suitable argument, the third as Nielsen polylogarithm (when
available), the last the numerical value.
Once the explicit analytic expressions of the coefficients of the Laurent-expansion of
Φ(d, x) in (d − 4) are known, one can use Eqs. (2.13)–(2.15) for obtaining the coefficients
of Φ2 (d, x), Ψ3 (d, x) and Φ4 (d, x). The coefficients of the expansion of Ψ5 (d, x) are then
Table 1
Constants up to weight 7 appearing in Eq. (5.16)
Constant HPL NPl Numerical value
ζ (3) H (0, 0, 1; 1) S2,1 (1) 1.2020569031595942854
a4 H (0, 0, 0, 1; 1/2) S3,1 (1/2) 0.51747906167389938633
ζ (5) H (0, 0, 0, 0, 1; 1) S4,1 (1) 1.0369277551433699263
a5 H (0, 0, 0, 0, 1; 1/2) S4,1 (1/2) 0.50840057924226870746
a6 H (0, 0, 0, 0, 0, 1; 1/2) S5,1 (1/2) 0.50409539780398855069
b6 H (0, 0, 0, 0, 1, 1; 1/2) S4,2 (1/2) 0.0087230030575968884272
ζ (7) H (0, 0, 0, 0, 0, 0, 1; 1) S6,1 (1) 1.0083492773819228268
a7 H (0, 0, 0, 0, 0, 0, 1; 1/2) S6,1 (1/2) 0.50201456332470849457
b7 H (0, 0, 0, 0, 0, 1, 1; 1/2) S5,2 (1/2) 0.0041965726953603256975
d7 H (0, 0, 0, 0, 1, −1, −1; 1) – 0.0022500546439578516764
recovered by integrating in x Eq. (2.10); the quadrature is trivial to carry out in terms of
HPLs, and the integration constants are fixed by the condition Ψ5 (d, 0) = 0, which follows
at once from Eqs. (2.9), (2.2), (2.1).
6. The independent numerical calculation
In this section we will calculate Φ(d, x = 1) with high numerical precision by suitably
using the difference equation method described in [5,6] and references therein. At variance
with [6], we will apply the method directly to the Master Integral with massless lines. This
will imply some additional work for establishing the initial conditions of the difference
equations.
We define

−2d d d k1 d d k2 d d k3 d d k4
I5 (d, n) = π ,
(k12 + 1)n ((k2 − k1 )2 + 1)((k3 − k2 )2 + 1)(k4 − k3 )2 (p − k4 )2
p2 = −1, (6.1)
so that I5 (d, 1) is equal to Φ(d, x) of Eq. (2.11) at x = 1 up to a known multiplicative
factor
4
I5 (d, 1) = 4(1 + ) Φ(d = 4 − 2, x = 1). (6.2)
By combining identities obtained by integration by parts one finds that I5 (d, n) satisfies
the third order difference equation
32(n − 1)(n − 2)(n − 3)(n − 3d + 5)I5 (d, n)

− 4(n − 2)(n − 3) 15n2 + (39 − 50d)n + 27d 2 + 5d − 54 I5 (d, n − 1)

+ 2(n − 3) 12n3 − (38d + 24)n2 + (23d 2 + 133d − 84)n

+ 9d 3 − 141d 2 + 134d − 24 I5 (d, n − 2)
+ (n − d − 1)(n − 2d + 1)(2n − 3d)(2n − 5d + 4)I5 (d, n − 3) = 0. (6.3)
We will solve this difference equation by using the Laplace ansatz
1
I5 (d, n) = t n−1 v5 (d, t) dt, (6.4)
0
giving for v5 (d, t) the fourth order differential equation
d4
4t 4 (8t + 1)(t − 1)2 v5 (d, t)
dt 4
d3
+ 4t 3 (t − 1) 24(d + 1)t 2 + (18 − 26d)t − 7d + 12 3 v5 (d, t)
dt

+ t 2 576(d − 1)t 3 + (−108d 2 − 420d + 648)t 2 + (46d 2 + 38d − 144)t
d2
+ 71d 2 − 284d + 288 v5 (d, t)
dt 2

+ t (576d − 960)t 3 + (−216d 2 + 360d)t 2 + (−18d 3 + 190d 2 − 496d + 384)t
d
+ 77d 3 − 533d 2 + 1236d − 960 v5 (d, t)
dt
+ (d − 3)(2d − 5)(3d − 8)(5d − 12)v5 (d, t) = 0. (6.5)
We will look for the solution of Eq. (6.5) in the form of a power series expansions;
inserted in Eq. (6.4) and integrated term by term it will provide very accurate values of
I5 (d, n). As the convergence is faster for larger n, we will consider “large enough” values
of the index n (see below); the repeated use “top–down” of Eq. (6.3) (i.e., using it for
expressing I5 (d, n − 3) in terms of the I5 (k) with k = n, n − 1, n − 2) will give the values
corresponding to smaller indices, till I5 (d, 1) is eventually obtained. To go on with this
program, initial conditions for v5 (d, t) are needed.
From the definition Eq. (6.1), and introducing spherical coordinates in d-dimension for
the loop momentum k1 , d d k1 = k1d−1 dk1 dΩ(d, k̂1) one has
∞
1 (k12 )d/2−1 dk12
I5 (d, n) = d f5 d, k12 , (6.6)
2
(k12 + 1)n
0

dΩ(d, k̂1)
f5 d, k12 = I4 d, 1, (p − k1 )2 , (6.7)
Ω(d)
where Ω(d) is the d-dimensional solid angle, and I4 (d, 1, (p − k1 )2 ) is the 3-loop (off
mass-shell) sunrise integral

I4 d, n, (p − k1 )2

d d k2 d d k3 d d k4
= π −3d/2 . (6.8)
((k2 − k1 ) + 1) ((k3 − k2 )2 + 1)(k4 − k3 )2 (p − k4 )2
2 n
By the change of variable 1/(k12 + 1) = t, k12 = (1 − t)/t, one finds
1
1 d d
I5 (d, n) = d t n−1 (1 − t) 2 −1 t − 2 f5 d, 1−t
t dt, (6.9)
2 0
from which one gets the relation between v5 (d, t) and f5 (d, (1 − t)/t)
1
d (1 − t) 2 −1 t − 2 f5 d, 1−t
d d
v5 (d, t) = t dt. (6.10)
2
From that relation we see that we can derive boundary conditions for v5 (d, t) in the
t → 1 limit from the expansion of f5 (d, k12 ) in the k1 → 0 limit, which is easy to obtain.
Indeed, only the first denominator of Eq. (6.8) depends on k1 ; expanding for small k1 and
performing the angular integration one gets

dΩ(d, k̂1) 1
Ω(d) (k2 − k1 )2 + 1

dΩ(d, k̂1) 1 k12 − 2k1 · k2 (k12 − 2k1 · k2 )2
= − + + · · ·
Ω(d) k22 + 1 (k22 + 1)2 (k22 + 1)3

1 1 4 k22
= 2 + k12 − 2 + + ···. (6.11)
k2 + 1 (k2 + 1)2 d (k22 + 1)3
The above result gives the expansion of f5 (d, k12 ) at k12 = 0:
(0) (1) (0)
f5 (d, k12 ) = f5 (d) + f5 (d)k12 + O k14 , f5 (d) = I4 (d, 1, p2 ),
4
f5(1) (d) = −I4 (d, 2, p2 ) + I4 (d, 2, p2 ) − I4 (d, 3, p2 ) . (6.12)
d
Note that f5 (d, k12 ) is regular in the origin.
By inspecting the differential equation (6.5) one finds that the behaviour at t = 1 of the
4 independent solution is ∼ (1 − t)αi , with α1 = d/2 − 1, α2 = d/2, α3 = 0, and α4 = 1;
for comparison with Eq. (6.10) the behaviours α3 = 0, and α4 = 1 are ruled out and the
expansion reads
(0)
v5 (d, t) = (1 − t) 2 −1 v5 (d) + v5 (d)(1 − t) + O(1 − t)2 ;
d (1)
(6.13)
by comparison with Eq. (6.12) (t = 1 corresponds to k12 = 0), one obtains

1 1 d
v5(0) (d) = d f5(0) (d), v5(1) (d) = d f5(0) (d) + f5(1) (d) . (6.14)
2 2 2
The values I4 (d, n) of I4 (d, n, p2 ) at p2 = −1 are therefore required

d d k2 d d k3 d d k4
I4 (d, n) ≡ I4 (d, n, p2 ) = π −3d/2 ,
(k2 + 1)n ((k3 − k2 )2 + 1)(k4 − k3 )2 (p − k4 )2
2
p2 = −1. (6.15)
The problem of evaluating the I4 (d, n) is similar to the original problem of evaluating
the I5 (d, n), but in fact it is much simpler, as the I4 (d, n) involve one less loop and one
less propagator. As above, by using integration-by-parts identities one finds that I4 (d, n)
satisfies the third order difference equation
6(n − 1)(n − 2)(n − 3)I4 (d, n) − (n − 2)(n − 3)(10n − 7d − 10)I4 (d, n − 1)

+ (n − 3) 2n2 + (2d − 18)n − 7d 2 + 29d − 8 I4 (d, n − 2)
+ (n − d − 1)(n − 2d + 1)(2n − 3d)I4 (d, n − 3) = 0. (6.16)
We solve the difference equation by using again the Laplace ansatz
1
I4 (d, n) = t n−1 v4 (d, t) dt, (6.17)
0
where v4 (d, t) satisfies the differential equation
d3
2t 3 (3t + 1)(t − 1)2 v4 (d, t)
dt 3
d2
+ t 2 (t − 1)(36t 2 + (6 − 7d)t − 9d + 18) v4 (d, t)
dt 2
d
+ t 36t 3 − 14dt 2 + (−7d 2 + 33d − 36)t + 13d 2 − 61d + 72 v4 (d, t)
dt
+ (d − 3)(2d − 5)(3d − 8)v4 (d, t) = 0. (6.18)
Following the procedure used above, we write
∞
1 (k22 )d/2−1 dk22
I4 (d, n) = d f4 k22 , (6.19)
2
(k22 + 1)n
0

dΩ(d, k̂2)
f4 k22 = I3 d, (p − k2 )2 (6.20)
Ω(d)

d d k3 d d k4
I3 d, (p − k2 )2 = π −d , (6.21)
((k3 − k2 ) + 1)(k4 − k3 )2 (p − k4 )2
2
1 d d
v4 (d, t) = d (1 − t) 2 −1 t − 2 f4 d, 1−t
t . (6.22)
2
At variance with the previous case, the function f4 (d, k22 ) is not regular for k2 → 0, as at
k2 = 0 the value of the external momentum squared (p − k2)2 becomes the threshold of the
2-loop sunrise graph associated to I3 (d, p2 ). But it is not difficult to evaluate analytically
I3 (d, q 2 ) for generic off-shell q 2 by using Feynman parameters:

2(5 − d) 3 − d2 2 d2 − 1
I3 (d, q ) =
2
d 2 F1 3 − d, 2 − 2 ; 2 ; −q ,
d d 2
(d − 4) (3 − d) 2
2
where 2 F1 is the Gauss hypergeometric function. The expansion of I3 (d, q 2 ) in q 2 = −1

consists of the sum of two series,

I3 (d, q 2 ) = a0 (d) 1 + O(q 2 + 1) + b0 (d)(q 2 + 1)2d−5 1 + O(q 2 + 1) ,

2(5 − d) 3 − d2 2 d2 − 1 (2d − 5)
a0 (d) = I3 (d, −1) = ,
(4 − d)2 (3 − d) 32 d − 3 (d − 2)

b0 (d) = 2 d2 − 1 (5 − 2d). (6.23)
Inserting Eq. (6.23) into Eq. (6.20), setting q = p − k2 and performing the angular
integration over k̂2 by means of the formula (see Eq. (88) of Ref. [5]) valid for k2 → 0

1 dΩ(d, k̂2) 2 − N2 d2 N2
≈ (k2 ) , k2 → 0; (6.24)
Ω(d) ((p − k2 )2 + 1)N 2(N) 12 (d − N)
with N = −(2d − 5), as in the term (q 2 + 1)2d−5 of Eq. (6.23), one obtains

f4 (d, k22 ) = a0 (d) 1 + O k22

d2 52 − d 1
+ 1 b0 (d)(k22 ) 2 (2d−5) 1 + O k22 . (6.25)
2(5 − 2d) 2 (3d − 5)
Using the variable 1/(k22 + 1) = t in Eq. (6.25) and inserting it in Eq. (6.22) one gets the
initial condition for v4 (d, t) at the singular point t = 1
a0 (d) d
v4 (d, t) = d (1 − t) 2 −1 1 + O(1 − t)
2

52 − d 1
+ 1 b0 (d)(1 − t) 2 (3d−7) 1 + O(1 − t) . (6.26)
2(5 − 2d) 2 (3d − 5)
By inspecting the equation (6.18) one gets that the behaviour at t = 0 of v4 (d, t) is
v4 (d, t → 0) ≈ c4 (d)t −d+3 + c4 (d)t −2d+5 + c4 (d)t 2 (−3d+8),

(1) (2) (3) 1
(6.27)
so that for d → 4 the integral (6.17) is convergent for n 4.
All the quantities depending on d are then systematically expanded in d − 4, the series
are truncated at some fixed number of terms, and the calculations with the truncated series
are performed by using the program SYS [5]; as the first 12 terms of the series are lost in
the intermediate steps of the calculations, in order to obtain the final results, from 1/ 4 up
to O( 6 ), 23 initial terms are needed. We solve finally the differential equation (6.18) with
the initial condition (6.26) by a first expansions in series at t = 1; due to the presence in
Eq. (6.18) of a singular point at t = −1/3, to have fast convergence till t = 0 we switch
to subsequent series expansions at the intermediate points 1/2, 1/4, 1/8 and 0; then we
calculate the integral (6.17) for n = 4, 5, 6, 7, 8 by integrating the series term by term
(about 300 terms are needed to reach a precision of 77 digits). By applying repeatedly
“top–down” the recurrence relation (6.16) to I4 (d, 8), I4 (d, 7), I4 (d, 6), we obtain I4 (d, 5)
and I4 (d, 4) (which are cross-checked with the values obtained by direct integration), then
I4 (d, 3), I4 (d, 2) and I4 (d, 1)

I4 (d, 1) = (1 + )3 0.3333333333333333333333333333333 −3 + −2
+ 1.3611111111111111111111111111111 −1
+ 0.4254662447518595926183272929396
− 31.041066530239171582155933528974
− 110.36756287612408836668594378878 2
− 611.01919192086881174365300012691 3
− 1734.0854215005636534710373849833 4

− 7316.7112322583252172396619274906 5 + O( 6 ) . (6.28)
Those values of I4 (d, n) are used to determine the initial condition for v5 (d, t), Eqs. (6.13),
(6.14), (6.12). We then solve the differential equation (6.5) by expansions in series centered
in the points t = 1, 1/2, 1/4, 1/8, 1/16 and 0 (as above, this subdivision is due to the
presence of a singular point at t = −1/8). By inspecting the equation (6.5) one gets that
the behaviour at t = 0 of v4 (d, t) is
v5 (d, t → 0) ≈ c5 (d)t −d+3 + c5 (d)t −2d+5 + c5 (d)t (−3d+8)/2

(1) (2) (3)
(4)
+ c5 (d)t (−5d+12)/2, (6.29)
so that for d → 4 the integral (6.4) is surely convergent for n 5; then we calculate the
integral (6.4) for n = 5, 6, 7, 8, 9 by integrating the series term by term. By using repeatedly
“top–down” the recurrence relation (6.16) starting from n = 9, we obtain I5 (d, 6), I5 (d, 5)
(used for cross-check), I5 (d, 4), . . . , I5 (d, 1). The result, up to the coefficient of order 5 in
(d − 4) included, is

I5 (d, 1)=(1 + )4 −0.25 −4 − 0.875 −3
− 1.416666666666666666666666666667 −2
− 1.449652777777777777777777777778 −1
− 14.055442461941065705599451765901
− 90.416062304531327135375791542063
− 1170.3684076603804614545918785105 2
− 5299.6462727245240600241060624284 3
− 37132.219579420859394978093377604 4

− 144381.92488313453838475109116166 5 + O( 6 ) (6.30)
Taking into account the normalization (6.2) one finds that the first 8 terms of Eq. (6.30)
agree with Eq. (5.16).
We want only to mention that we have also independently checked the numerical
result (6.30) by calculating the master integral with all masses equal to one by difference
equations, and then by using the value so obtained as initial condition for the integration
of a differential equation in the photon mass λ from λ = 1 to λ = 0.
7. The x → 0 values
We evaluate in this section the x → 0 values of Φ(d, x), Eq. (2.11).

By combining the familiar formulae
1 1
[(k − l)2 + a)]α ] (l 2 + b)β
1
(α + β) x α−1 (1 − x)β−1
= dx , (7.1)
(α)(β) [(l − xk)2 + x(1 − x)k 2 + ax + b(1 − x)]α+β
0
and

−1 dd l 1 1 n − d2 1
C (d) = , (7.2)
(2π)d−2 [(l − q)2 + c]n 4 3 − d2 (n) cn− d2
one gets

dd l 1 1
C −1 (d)
(2π) d−2 [(k − l) + a] (l + b)β
2 α 2
1 d d
1 α + β − d2 x 2 −β−1 (1 − x) 2 −α−1
= dx , (7.3)
4 (α)(β) 3 − d2 [k 2 + d(a, b, x)]α+β− 2
d
0
where
ax + b(1 − x)
d(a, b, x) = .
x(1 − x)
We rewrite Eq. (2.11) as

C −4 (d) dd q d d k2 d d k1
Φ(d, x) =
(2π) 4(d−2) [(p − q) + x ] (q − k2 )
2 2 2 (k2 − k1 )2

dd l
× , (7.4)
[(k1 − l)2 + 1](l 2 + 1)
and then use repeatedly Eq. (7.3), using in the order the parameters y, y1 , y2 , z for
integrating the loops l, k1 , k2 , q, obtaining
(1 − 2(d − 4))
Φ(d, x) =
1024(d − 3)(d − 4)(2d − 5)(2d − 7) 4 1 − 12 (d − 4)
1 1
d d d
−2
× dy y 2 −2 (1 − y) 2 −2 dy1 y12 (1 − y1 )d−3
0 0
1 d
−2 3
× dy2 y22 (1 − y2 ) 2 d−4 Ψ (d, x, y, y1, y2 ), (7.5)
0
where
1 3
z3− 2 d
Ψ (d, x, y, y1, y2 ) = dz , (7.6)
[(1 − z)2 x 2 + zD(y, y1 , y2 )]5−2d
0
1
D(y, y1 , y2 ) = . (7.7)
y(1 − y)(1 − y1 )(1 − y2 )
The above formulae are valid for any x; form now on we take 0 < x 1. For definiteness,
we take also d to be “just bigger” than 2 (i.e., d = 2 + η, with 0 < η 1). The z-integral
will be carried out first. To that aim, introduce an infinitesimal parameter Z, such that
0 < x 2 Z 1 (a possible choice might be Z = −x 2 ln x, but the exact value of Z will
be irrelevant), and split the integration interval as

1 Z 1
dz = dz + dz;
0 0 Z
correspondingly, we write the z-integral as
Ψ (d, x, y, y1, y2 ) = Ψ1 (d, x, y, y1, y2 ) + Ψ2 (d, x, y, y1, y2 ),
Z 3
z3− 2 d
Ψ1 (d, x, y, y1, y2 ) = dz ,
[(1 − z)2 x 2 + zD(y, y1 , y2 )]5−2d
0
1 3
z3− 2 d
Ψ2 (d, x, y, y1, y2 ) = dz . (7.8)
[(1 − z)2 x 2 + zD(y, y1 , y2 )]5−2d
Z
In the second term we can neglect x 2 in the denominator obtaining simply

1 3 1
z3− 2 d d
Ψ2 (d, x, y, y1, y2 ) dz = D(y, y1 , y2 )2d−5 dz z 2 −2
[zD(y, y1 , y2 )]5−2d
Z Z
2
= D(y, y1 , y2 )2d−5 , (7.9)
d −2
where we have neglected the contribution from the lower integration limit Z, which is
d
Z 2 −1 , as for d just bigger than 2 it vanishes with Z.
The first term is slightly more delicate. To start with, as 0 < z < Z 1 we can neglect
z with respect to 1 in (1 − z)2 x 2
Z 3
z3− 2 d
Ψ1 (d, x, y, y1, y2 ) dz 2 ;
[x + zD(y, y1 , y2 )]5−2d
0
for reasons which will be apparent in a moment, we rewrite it as
Z
z 2 −2
d
Ψ1 (d, x, y, y1, y2 ) = dz
x 2 5−2d
D(y, y1 , y2 ) + z
0
and integrate by parts the factor z 2 −2 ; the result is

d
Ψ1 (d, x, y, y1, y2 )
3 Z
2 z4− 2 d
=
d − 2 [x + zD(y, y1 , y2 )]
2 5−2d
0
Z 3

z3− 2 d
+ (2d − 5)x 2
dz 2
[x + zD(y, y1 , y2 )]6−2d
0
Z 3
2d − 5 2 z3− 2 d
=2 x dz ,
d −2 [x 2 + zD(y, y1 , y2 )]6−2d
0
as the end-point contributions vanish for d just bigger than 2. We can now modify the
integration interval for Ψ1 (d, x, y, y1, y2 ) into
Z ∞ ∞
dz = dz − dz,
0 0 Z
and write, correspondingly,
Ψ1 (d, x, y, y1, y2 ) = Ψ1∞ (d, x, y, y1, y2 ) − Ψ1Z (d, x, y, y1, y2 ), (7.10)
where, thanks to the previous integration by parts, the two resulting z-integrals (from 0 to
∞ for the first, from Z to ∞ for the second) are now both convergent for d just bigger
than 2.
We start again from the second term,
∞ 3
2d − 5 2 z3− 2 d
Ψ1Z (d, x, y, y1, y2 ) = 2 x dz ;
d −2 [x 2 + zD(y, y1 , y2 )]6−2d
Z
in the denominator we can neglect x2, the resulting integral is trivial and the result can be
written as

2d − 5 1 2 x 2 d −1 ∞
Ψ1Z (d, x, y, y1, y2 ) = 2 z2 ,
d − 2 [D(y, y1 , y2 )] 6−2d d −4 z Z
which vanishes with Z (recall x 2 Z), so that

Ψ1Z (d, x, y, y1, y2 ) = 0. (7.11)
In the first term of Eq. (7.10), Ψ1∞ (d, x, y, y1, y2 ), we substitute z = tx 2 /D(y, y1 , y2 ),
obtaining
∞ 3
2d − 5 (d−2) 3 d−4 t 3− 2 d
Ψ1∞ (d, x, y, y1, y2 ) = 2 x D(y, y1 , y2 ) 2 dt .
d −2 (1 + t)6−2d
0
By using the second representation of Euler’s Beta function
1 ∞
(α)(β) t α−1
B(α, β) = = du u α−1
(1 − u) β−1
= dt , (7.12)
(α + β) (1 + t)α+β
0 0
we finally obtain
Ψ1∞ (d, x, y, y1, y2 )
2d − 5 3 d−4
=2 B 4 − 32 d, 2 − 12 d x (d−2) D(y, y1 , y2 ) 2 . (7.13)
d −2
Fig. 3. The 4-loop watermelon graph.
By collecting the results Eqs. (7.13), (7.11), (7.10), (7.9), (7.8), one obtains the value of
the function Ψ (d, x, y, y1, y2 ) to be substituted in Eq. (7.5); recalling Eq. (7.7) one sees
that all the remaining integrations factorize and can be carried out in terms of Euler’s Beta-
functions Eq. (7.12).
It is clear that the final result for Eq. (2.11) for x → 0 consists of two terms, the first
constant (independent of x), the second proportional to x d−2 . In the notation of Eq. (3.5)
we have
(1) 3d − 11
A0 = −
8(d − 2)(d − 3)(d − 4)3 (2d − 5)(2d − 7)(3d − 8)(3d − 10)

(1 − (d − 4))(1 − 2(d − 4)) 2 1 + 12 (d − 4) 2 1 − 32 (d − 4)
× ,
4 1 − 12 (d − 4) (1 − 3(d − 4))
2(2d − 7)
A0 (2) = −
3(d − 2)2 (d − 3)(d − 4)4 (3d − 8)(3d − 10)

1 + 12 (d − 4) 1 − 32 (d − 4) 2 (1 − (d − 4))
× . (7.14)
2 1 − 12 (d − 4) (1 − 2(d − 4))
(1)
Let us observe that the term A0 is the value of the vacuum graph in Fig. 3, in agreement
with the result in Eq. (A.12) of [13] (up to a different normalization).
Acknowledgements
We are grateful to J. Vermaseren for his kind assistance in the use of the algebra
manipulating program FORM [9], by which all our calculations were carried out.
References
[1] A.V. Kotikov, Phys. Lett. B 254 (1991) 158.

[2] E. Remiddi, Nuovo Cimento A 110 (1997) 1435, hep-th/9711188.
[3] M. Caffo, H. Czyz, S. Laporta, E. Remiddi, Nuovo Cimento A 111 (1998) 365, hep-th/9805118.
[4] T. Gehrmann, E. Remiddi, Nucl. Phys. B 580 (2000) 485, hep-ph/9912329.
[5] S. Laporta, Int. J. Mod. Phys. A 15 (2000) 5087, hep-ph/0102033.
[6] S. Laporta, Phys. Lett. B 523 (2001) 95, hep-ph/0111123.
[7] M. Argeri, P. Mastrolia, E. Remiddi, Nucl. Phys. B 631 (2002) 388, hep-ph/0202123.
[8] P. Mastrolia, E. Remiddi, Nucl. Phys. B 657 (2003) 397, hep-ph/0211451.

[9] J.A.M. Vermaseren, Symbolic Manipulation with FORM, Version 2, CAN, Amsterdam, 1991;
J.A.M. Vermaseren, New features of FORM, math-ph/0010025.
[10] E. Remiddi, J.A.M. Vermaseren, Int. J. Mod. Phys. A 15 (2002) 725, hep-ph/9905237.
[11] V.A. Smirnov, Applied Asymptotic Expansions in Momenta and Masses, in: Springer Tracts in Modern
Physics, vol. 177, Springer, Berlin, 2002.
[12] J. Vermaseren, table available from the URL http://www.nikhef.nl/~form/FORMdistribution/packages/
harmpol/index.html.
[13] K. Kajantie, M. Laine, K. Rummukainen, Y. Schroder, JHEP 0304 (2003) 036, hep-ph/0304048.
Do neutrino flavor oscillations forbid large lepton

asymmetry of the universe?
A.D. Dolgov a,b,c , Fuminobu Takahashi d
a INFN, Sezione di Ferrara, Via Paradiso, 12-44100 Ferrara, Italy
b ITEP, Bol. Cheremushkinskaya 25, 113259 Moscow, Russia
c ICTP, 34014 Trieste, Italy
d Research Center for the Early Universe, Graduate School of Science,
University of Tokyo, Tokyo 113-0033, Japan
Received 6 February 2004; accepted 1 April 2004
Abstract
It is shown that hypothetical neutrino–majoron coupling can suppress neutrino flavor oscillations
in the early universe, in contrast to the usual weak interaction case. This reopens a window for a
noticeable cosmological lepton asymmetry which is forbidden for the large mixing angle solution in
the case of standard interactions of neutrinos.
1. Introduction
Cosmological lepton asymmetry is not directly measurable, in contrast to baryon asym-

metry, but may be observed or restricted through its impact on big bang nucleosynthesis
(BBN), large scale structure formation, and the angular spectrum of the cosmic microwave
background radiation (CMBR), for a review see, e.g., Ref. [1]. At the present time the best
bounds follow from the consideration of BBN. Primordial production of light elements is
especially sensitive to the value of asymmetry between electronic neutrinos and antineu-
trinos since they directly influence the neutron–proton transformations in weak reactions
νe n ↔ e− p and ν̄e p ↔ e+ n. The bound on the chemical potential of electronic neutrinos
obtained in Ref. [2] reads |ξe | ≡ |µe /T | < 0.1. The bound on muonic or tauonic asymme-
try is noticeably weaker because νµ or ντ produce an effect on BBN only through their
impact on the cooling rate of the universe and the degeneracy of these neutrino flavors is
E-mail address: fumi@resceu.s.u-tokyo.ac.jp (F. Takahashi).
doi:10.1016/j.nuclphysb.2004.04.002
190 A.D. Dolgov, F. Takahashi / Nuclear Physics B 688 (2004) 189–213
equivalent to an addition of
4 2
15 ξµ,τ ξµ,τ
Nν = +2 (1)
7 π π
massless neutrino species at the BBN epoch. If the latter is bounded by Nν < 1 (safe
bound) then |ξµ,τ | < 1.5. For less conservative bound, Nν < 0.4 [3], the dimensionless
chemical potentials should be below 0.9. Still, asymmetry of the order of unity would have
noticeable cosmological effects on large scale structure formation and CMBR. The limits
quoted above are valid if the chemical potential of only one kind of neutrino is non-zero.
If a conspiracy between different chemical potentials is allowed, such that a positive effect
of ξµ or ξτ is compensated by ξe or vice versa, the limits would be somewhat weaker.
According to Ref. [4] they are: |ξe | < 0.2 and |ξµ,τ | < 2.6.1
The bounds on chemical potentials of νµ and ντ can be significantly improved because
of the strong mixing between different neutrino flavors [6]. This mixing gives rise to the
fast transformation between νe , νµ , and ντ in the early universe and leads to equilibration
of asymmetries of all neutrino species. Thus the BBN bound on any chemical potential
becomes essentially that obtained for νe [7] (see also the papers [8]):
|ξe,µ,τ | < 0.07. (2)
In this case the cosmological impact of neutrino degeneracy would be negligible. However,
a compensation of ξe by other types of radiation is still possible, which was studied in
Ref. [9]. The authors of Ref. [9] put the constraints on ξe and Nν based on WMAP data
and BBN:
−0.1 ξe 0.3, −2 Nν 5. (3)
It is interesting to see if one could reasonably modify the standard model to allow large
muonic and/or tauonic charge asymmetries, together with a small electronic asymmetry,
to avoid conflict with BBN. This is the aim of this work. A natural generalization is to
introduce an additional interaction of neutrinos with massless or light (pseudo)Nambu–
Goldstone boson, majoron [10]. The idea to invoke majoron to modify neutrino oscillations
in primeval plasma was discussed in Refs. [11,12]. In these papers, two different
mechanisms which could block or suppress oscillations between active and hypothetical
sterile neutrinos have been considered. We have, however, some concerns about the
validity of their results which we will discuss in the next section. Let us note that in this
paper we consider an impact of neutrino majoron interactions on the oscillations between
active neutrinos and not on active–sterile oscillations, as is done in the mentioned above
papers.
1 We are not concerned in this paper with the origin of such large lepton asymmetry. Thus far, many
cosmological scenarios that explain both the small observed baryon asymmetry and a possible large lepton
asymmetry simultaneously, have been proposed [5].
A.D. Dolgov, F. Takahashi / Nuclear Physics B 688 (2004) 189–213 191
2. General discussion and approximate estimates
Neutrino oscillations may be suppressed in medium if the interaction of neutrinos with

the medium is sufficiently strong or in other words the refraction index, n (or effective
potential, V = E(n − 1)) of neutrinos is very large. Correspondingly the mixing angle in
medium θm becomes negligible:
s22
tan2 2θm = , (4)
(c22 + V /δE)2 + Γ 2 /(4δE 2 )
where s2 = sin 2θ , c2 = cos 2θ , θ is the vacuum mixing angle, δE = δm2 /2E, E is the
neutrino energy, and Γ is the rate of neutrino interaction with medium given by
80(1 + gL2 + gR2 )G2F ET 4
Γ = , (5)
3π 3
where T is the temperature of the primeval plasma, GF = 1.166 × 10−11 MeV−2 is the
Fermi coupling constant, gL = sin2 θW ± 1/2 and gR = sin2 θW where sin2 θW = 0.23 and
the sign “+” stands for νe and “−” stands for νµ,τ . The derivation of these equations can
be found, e.g., in lectures [13].
The diagonal components of the effective potential, created by the standard weak
interaction, for the active neutrino species are given by [14]:
G2F T 4 E
(w)
Vaa = ±C1 η(a) GF T 3 − C2a , (6)
α
where a = e, µ, τ labels the neutrino flavors, α = 1/137 is the fine structure constant, and
the signs “±” refer to neutrinos and antineutrinos respectively. The first term arises due
to a possible charge asymmetry of the primeval plasma, while the second one comes from
the non-locality of weak interactions associated with the exchange of W or Z bosons.
µ,τ
According to Ref. [14] the coefficients Cj are: C1 ≈ 0.95, C2e ≈ 0.62 and C2 ≈ 0.17
(for T < mµ ). These values are true in the limit of thermal equilibrium, otherwise these
coefficients are some integrals from the distribution functions over momenta. The charge
asymmetry of plasma is described by the coefficients η(a) which are equal to
η(e) = 2ηνe + ηνµ + ηντ + ηe − ηn /2 (for νe ), (7)
η(µ) = 2ηνµ + ηνe + ηντ − ηn /2 (for νµ ), (8)
and for ντ is obtained from Eq. (8) by the interchange µ ↔ τ . The individual charge
η(τ )
asymmetries, ηX , are defined as the ratio of the difference between particle–antiparticle
number densities to the number density of photons with the account of the 11/4-factor
emerging from the e+ e− -annihilation:
ηX = (NX − NX̄ )/Nγ . (9)
For “normal” values of charge asymmetry, i.e., |η| ∼ 10−9
the charge asymmetric term
in the potential (6) is subdominant but if the asymmetry is of the order of unity, then the
ratio Vaa /δE becomes huge:
−1
Vaa(w)
/δE ≈ 2 × 10−11ηET 3 δm2 (MeV)−2 ≈ 20η(T /MeV)4 eV2 /δm2 (10)
and, e.g., for δm2 = 10−2 eV2 and η ∼ 1 the mixing angle in matter would be suppressed
more than by three orders of magnitude in the BBN range of temperatures, T ∼ MeV.
However, this result is valid for mixing between active and sterile neutrinos and is not
true for active–active mixing. In the last case the effective potential has large off-diagonal
matrix elements [15], Vab , with a = b, which compensate the suppression induced by the
diagonal terms. This is the reason for non-suppressed oscillations between active flavors
and for the equilibration of all leptonic asymmetries [7].
Possible additional interactions of neutrinos with majoron can be described by the
Lagrangian:

Lint ∼ χ gab νaT Cνb + h.c. , (11)
where χ is the operator of the majoron field, C is the matrix of charge conjugation, and
gab are the coupling constants. We will assume for simplicity, though it is not necessary,
that only the flavor-diagonal coupling is non-vanishing, gab ∼ gaa δab . We will consider
the coupling matrix with more general form later. Let us also assume that the majoron is
so light that its decay and inverse decay are not essential at the BBN epoch. Neutrino–
(anti)neutrino scattering through majoron exchange gives rise to a contribution to neutrino
effective potential (proportional to forward scattering amplitude) which can be estimated
as
(χ)
Vab ∼ gaa gbb T . (12)
The constants gaa should satisfy several constraints to make the mechanism operative.
(χ)
Firstly, the diagonal part of the potential Vaa should be larger than the weak potential
V (w) given by Eq. (6), while its off-diagonal components must be much smaller than
(χ) (χ)
the diagonal ones, Vab Vaa , that is, the flavor symmetry in the neutrino–majoron
interactions should be strongly broken. These two conditions would ensure suppression
of flavor changing oscillations between active neutrinos. To realize these conditions the
following constraints should be imposed:
2
−11 (a) Tae
2
gaa > 10 η and gee gaa . (13)
1 MeV
Here a labels µ or τ and the coupling of majoron to νe is assumed to be much weaker than
those to νµ and/or ντ ; Tae is the temperature at which νa are effectively transformed into νe
in the standard theory. According to calculations of Ref. [7] it takes place around Tae = 1
MeV for the large mixing angle solution to solar neutrino deficit, while the transformation
between νµ and ντ in the standard theory takes place somewhat above 10 MeV. In fact, as
we see in what follows from numerical calculations, the more accurate bound is much less
restrictive than this simple estimate (see Eq. (49) or Figs. 6 and 7 below).
The second inequality (13) implies that the off-diagonal components of the effective
potential are suppressed in comparison with the diagonal ones and the oscillations
remain blocked. This is not so in the case of the standard weak interactions which are
flavor symmetric and the large value of the denominator due to large Vaa in Eq. (4) is
compensated by the same large off-diagonal components of the effective potential.
Secondly, the coupling constants gaa should not be too large, otherwise flavor non-
conserving reactions of the type νe νa ↔ ν̄e ν̄a (or similar) would lead to equilibration of all
leptonic charges. To avoid that the rate of these reactions, Γea ∼ σea T 3 , should be smaller
than the cosmological expansion rate H ∼ T 2 /mPl , where mPl = 1.221 × 1022 MeV is the
Planck mass. Thus, to suppress e − µ or e − τ transformation through direct reactions one
needs

T
2 2
gaa gee < 10−22 . (14)
1 MeV
This conditions should be satisfied for temperatures above the BBN range, i.e., T > 1
MeV. Similarly, if we require that νa νa ↔ ν̄a ν̄a should not occur efficiently, the coupling
constants must satisfy a similar inequality with gee replaced with gaa .
There are quite strong limits on possible coupling of majoron to neutrinos which follow
from astrophysics [16,17]; discussion and references to earlier works can be found e.g.
in the book [18]. Astrophysics allows either very small or quite large coupling constants.
The former is quite evident, while the latter appears because strongly interacting majorons,
though efficiently produced inside a star, cannot propagate out and carry away the energy,
thus opening a window for large values of the coupling. It is not so for the coupling
to νe because the latter is bounded from above by the data on double beta decay [19],
gee < 3 × 10−5 . Together with the supernova bounds, the upper limit is shifted down to
gee < 4 × 10−7 [17], with a small window around (2–3) × 10−5 . So we assume in the
following that gee < 4 × 10−7 . For µ or τ the allowed regions are: gaa < (3–5) × 10−6 or
gaa > (3–5) × 10−5 . Evidently the conditions specified above can be satisfied. As reference
values we may take gee = 10−7 and gaa = 5 × 10−6 which satisfy all constraints presented
above and would lead to a suppression of the transformation of νµ or ντ into νe , thus
permitting a large muonic or tauonic asymmetry combined with a small electronic one at
BBN. Note that it is not necessary to assume a large value of gaa 10−5 to accomplish this.
As shown later, such a large coupling constant would change the scenario into something
more complex.
If majoron is lighter than the mass difference of neutrinos then a heavier neutrino would
decay into a lighter one and majoron. This process might leave traces in the flavor ratios of
the high-energy neutrinos from distant astrophysical sources [20], which can be detected by
e.g., IceCube [21]. This could be potentially sensitive to very small values of the neutrino–
majoron coupling. Existing bound [22] on life-time/mass of decaying neutrinos based on
the solar neutrino data, τ/m 10−4 s/eV, is too weak to be essential for the mechanism
discussed in the present paper.
In the subsequent sections we will present more accurate calculations but before turning
to a closer examination of the scenario, let us make a few comments on earlier papers where
majoron suppression of neutrino oscillations in the early universe have been considered.
In Ref. [11] it was assumed that there exists a coupling of an active and sterile neutrinos
to majoron. Let us note that in the version of the theory that we consider here no sterile
neutrinos are introduced: the negative helicity states are identified with neutrinos, while
the positive helicity states are identified with antineutrinos. The coupling considered in
Ref. [11] has the form:
Las = igas χ ν̄a νs + h.c., (15)

where νa is an active neutrino flavor and νs is a sterile one. The authors argued that
the effective potential induced by the interactions of active and sterile neutrinos with
majoron strongly suppressed the oscillations. However they omitted off-diagonal Vas -term
in the effective potential which might invalidate their result. Possibly this mechanism of
oscillation suppression may operate in a more complicated version of the model.
In Ref. [12] a different mechanism of neutrino oscillation blocking by majoron has
been suggested. According to the equation of motion, the total leptonic current including
majoron and neutrino contributions is conserved:
f D 2 χ + Dµ J µ = 0, (16)
where Dµ is the covariant derivative in cosmological background, f is the vacuum
expectation value of the Higgs field responsible for the spontaneous breaking of leptonic
UL (1)-symmetry and J µ is the leptonic currents of fermions. In spatially homogeneous
case the solution to this equation is

ain 3
χ̇(t) = χ̇in − Jint /f + J t (t)/f, (17)
a(t)
where a(t) is the cosmological scale factor and the sub-index “in” indicates initial values.
We assume that χ̇in = 0. Hence the authors of Ref. [12] concluded that χ̇ = ηB nγ /f ,
where ηB = 6 × 10−10 is the baryon asymmetry of the universe. The contribution to the
neutrino effective potential from this term is δV χ = χ̇/f and, according to [12], with
f 2 = (5–9) GeV2 the potential would be strong enough to suppress the oscillations.
However, according to the usual scenarios, baryon asymmetry could be created at much
larger temperatures, much higher than this scale. Thus, after spontaneous symmetry
breaking, leading to creation of majoron, J t remains constant in comoving volume and
χ̇ = 0. A non-zero χ̇ might be created if lepton asymmetry was generated at spontaneous
breaking of leptonic UL (1) or by the neutrino oscillations themselves but both these cases
need further and more detailed investigation. So in our considerations we assume that the
mechanism of Ref. [12] is not operative.
3. Effective potential induced by neutrino–majoron interactions
In this section we derive the effective potential induced by the neutrino–majoron

interactions, which is one of essential ingredients of our paper. The relevant part of the
Lagrangian is given by
1
L = − ∂µ χ∂ µ χ − ν a γ µ ∂µ νa + Lint , (18)
2 a
1 ∗

Lint = χ gab ΦaT σ2 Φb + gab Φb† σ2 Φa∗
2
i ∗ †

= χ gab νaT Cνb + gab νb Cνa∗ , (19)
2
Fig. 1. s- and t-channel diagrams of neutrino–(anti)neutrino scattering through majoron exchange. Both diagrams
contribute to the off-diagonal component of the effective potential.
Fig. 2. Diagram of neutrino–majoron scattering through neutrino exchange. If majorons are not abundant in the
primeval plasma, this diagram gives negligible contribution to the effective potential.
where Φa and νa are two-component and four-component representations of neutrino

of flavor a, respectively. They are related to each other as νaT = (ΦaT , 0)T in the chiral
representation (see Appendix A for notations). Here and hereafter νa is taken to be the
left-handed field. In the following we neglect the effect of small masses of neutrinos, and
treat them as massless fields. In this limit, there are three diagrams2 that contribute to the
effective potential (see Figs. 1 and 2).
Let us first calculate the effective Hamiltonian corresponding to neutrino–(anti)neutrino
scattering processes. For that purpose we integrate out the majoron field using its
equation of motion, and substitute the solution into interaction Lagrangian (19). Then
we quantize neutrino fields perturbatively. The details of the derivation can be found in
the Appendix A. The effective Hamiltonian, which describes the neutrino–(anti)neutrino
scattering processes, is given by

(νν) (2π)3 δ (3)(p + q + r + s)e−iEtot t F (p, q)F (r, s)
Heff (t) = − dp dq dr ds (20)
16 p q Ep Eq (1 − p q cos θpq )
with
Ep ≡ |p|,
Etot ≡ p Ep + q Eq + r Er + s Es ,
p·q
cos θpq ≡ ,
|p||q|
∗
F (p, q) ≡ gab νa (p)T Cνb (q) + gab νb (−q)† Cνa∗ (−p), (21)
2 Note that there are other diagrams if small majorana masses are taken into account. However, the amplitudes
of such diagrams are suppressed by powers of mν /Eν 1 for relativistic neutrinos.
where dp ≡ d 3 p/(2π)3 . The momentum expansion of free neutrino field is

νi (x) = dp νi (p)e−ip t +ipx
0

= dp ai (p)up + bi (−p)† v−p e−ip t +ipx ,
0
(22)
with p0 ≡ p |p| = ±|p|. The sign of p is chosen to reproduce positive or negative

energy solution. Here ai (p) (bi † (p)) is the annihilation (creation) operator for negative
(positive)-helicity neutrinos with momentum p, up (vp ) represents left-handed Dirac spinor
of negative (positive)-helicity neutrinos. With thus obtained effective Hamiltonian, the
effective potential can be calculated by the technique of Sigl and Raffelt [23]. Although the
calculations are lengthy, the procedure is straightforward. The contribution to the effective
potential for neutrinos with momentum p is given by

(νν)
1 † T
Vp ab
= dq g ρq + ρ Tq g ab , (23)
4|p||q|
where the density matrices are defined as

aj† (p) ai (p ) = (2π)3 δ (3) (p − p )[ρp ]ij ,
†
bi (p) bj (p ) = (2π)3 δ (3) (p − p )[ρ p ]ij . (24)
If we take the coupling constant matrix gab to be diagonal, the effective potential becomes

(νν)
∗ 1 T
Vp ab
= gaa gbb dq ρq + ρ Tq ab (no summation)
4|p||q|
∗ T2
∼ gaa gbb .
|p|
This result confirms the rough estimate of the effective potential in the previous section.
The neutrino–majoron scattering shown in Fig. 2 also contributes to the effective
potential and can be calculated similarly. The result is

(νχ)
fχ (q) †
Vp ab
= dq g g ab , (25)
4|p||q|
where fχ (p) is the number density of majorons with momentum p. Note that this
expression vanishes if the abundance of majorons is negligible in comparison with
(eq)
their equilibrium value, i.e., fχ (p) fχ (p). Thus the complete effective potential for
neutrinos with momentum p induced by interactions with majorons is given by

(χ)
1 † T
Vp ab = dq g ρq + ρ Tq + fχ (q) · 1 g ab , (26)
4|p||q|
where 1 is the unit matrix in the flavor basis.
4. Possible role of neutrino–majoron reactions
As we have already mentioned reactions between neutrinos with an exchange of

majorons do not conserve leptonic charge and if they are efficient, lepton asymmetry of
the universe would be completely destroyed before BBN. One of the dangerous processes
is
2ν ↔ 2ν̄. (27)
We assume for simplicity that the diagonal coupling constants |gaa | are larger than non-
diagonal ones and that one flavor coupling dominates, e.g., |gτ τ | > |gµµ | > |gee |. In
this case only three diagrams presented in Fig. 3 contribute into the reaction (27). The
amplitude corresponding to any of these three diagram is equal to |g|2 and thus the
complete amplitude squared is simply |A(2ν ↔ 2ν̄)|2 = 9|g|4 (see Appendix B). Here and
in what follows we omit the sub-index aa at the coupling constant gaa . With the known
amplitude of the L-non-conserving reaction (L is the leptonic charge), the kinetic equation
can be written as
h 1 9|g|4
x∂x fν̄ (y 1 ) = ·
x2 2 29 π 5 xy1
3
d y2 d 3 y3 d 3 y4 (el)
× F (f )δ (4) (y1 + y2 − y3 − y4 ) + Icoll , (28)
y2 y3 y4
where f is the neutrino distribution function, (yi , yi ) = (Ei /T , pi /T ), x = (T /MeV)−1 ,
T is the temperature of the cosmic plasma, (H /MeV) = h/x 2 is the Hubble parameter
with h ≈ 4.5 × 10−22 , Icoll
(el)
is the contribution to the collision integral from reactions that
conserve leptonic charge, and

F (f ) = 1 − fν̄ (y1 ) 1 − fν̄ (y2 ) fν (y3 )fν (y4 )

− 1 − fν (y3 ) 1 − fν (y4 ) fν̄ (y1 )fν̄ (y2 ). (29)

We assumed that the temperature evolved according to Ṫ = −H T . The combinatorial
factor 1/2 in front of the r.h.s. comes from two factors 1/2! and one factor 2 because
two identical particle are produced (see, e.g., Ref. [24]).
The collision integral in Eq. (28) can be easily evaluated in the limit of Boltzmann
statistics, when f 1 and under assumption of kinetic equilibrium, so the distribution
functions have the form f = exp(−y + ξ ), where ξ = µ/T is the dimensionless chemical
(el)
potential. To avoid the contribution of Icoll to the equation governing the evolution of ξ we
take the difference of Eq. (28) and analogous equation for fν and integrate it over d 3 y1 .
Fig. 3. s-, t- and u-channel diagrams of 2ν ↔ 2ν̄ through majoron exchange.

(el)
The contribution of Icoll to this integrated difference vanishes. The resulting equation takes
the form:
9|g|4
∂x eξ − eξ̄ = − 7 3 e2ξ − e2ξ̄ , (30)
2 π h
where we have used the formula presented in Appendix C. It is also assumed above that
ξ = −ξ̄ . Otherwise the contribution from L-conserving ν̄ν-annihilation would not vanish
in the collision integral. If ξ 1 we may neglect exp(−ξ ) and solve Eq. (30) as

exp[ξ(0)]
exp ξ(x) = . (31)
1 + 9|g|4 x exp[ξ(0)]/(128π 3h)
The lepton asymmetry would not be destroyed by reaction (27) if

|g| < 2 × 10−5 T 1/4 exp −ξ(0)/4 . (32)

Here (and in what follows) T is given in MeV. Quantum statistics (Fermi) effects somewhat
weaken the bound. Their presence does not allow to solve kinetic equation analytically.
Numerical estimates give a similar constraint as shown in Fig. 5.
Muonic or tauonic charge asymmetry should be preserved until the annihilation
ν̄τ,µ ντ,µ into e+ e− -pairs is frozen. Otherwise we will return to the standard scenario
with zero lepton asymmetry. According to the estimates of Ref. [13] made in Boltzmann
approximation, the freezing temperature of the annihilation is Tf ≈ 5.3 MeV. Fermi
corrections would make its value slightly higher. If muonic and/or tauonic charge
asymmetry were erased by the oscillations or L-non-conserving reactions with exchange
of majorons below Tf then the total number density of ντ plus ν̄τ (or νµ + ν̄µ ) would be
conserved in the comoving volume and their distribution would be given by

−1
fνa = fν̄a = exp(E/T − ξa ) , (33)
where a = µ, τ . So effectively the energy density of these neutrinos would be the same
(up to a numerical factor of order of unity) as the energy density of the usual degenerate
neutrinos. If the mixing of ντ or νµ with νe would not change the distribution of the latter
then BBN would allow a large values of ξa and the degeneracy of νa may lead to noticeable
cosmological effects.
We make a simplifying assumption of absence of majorons in noticeable amount in the
primeval plasma. If majorons would be in equilibrium we should take into account their
(4/7)-contribution into the number of effective neutrino species at BBN and modification
of neutrino effective potential due to forward elastic scattering νχ → νχ . Even if majorons
are abundantly produced, the main conclusion of our paper remains unchanged but
numerical results would be somewhat different. The production of majorons could proceed
through the reaction ν + ν̄ → 2χ . The Feynman diagrams describing this process are
presented in Fig. 4. The amplitude squared of this process is

p1 · p3 p1 · p4
|Minv |2 = |g|4 + −2 ,
p1 · p4 p1 · p3
where p1(2) and p3,4 are four momenta for (anti)neutrino in the initial state and majorons in
the final state, and they satisfy the momentum conservation condition, p1 + p2 = p3 + p4 .
Fig. 4. t- and u-channel diagrams of ν + ν̄ ↔ 2χ through neutrino exchange.
Note that the collision integral with this amplitude squared involves an IR logarithmic
divergence, therefore we need to input a lower cutoff scale. Taking into account the
thermal effects, the dispersion relations for the majoron and the neutrinos change from
those at zero temperature. Especially, they obtain finite thermal masses, which provide the
desired infra-red cutoff. For |g| ∼ O(10−5 ), the thermal effect of the neutrino–majoron
interaction dominates over that due to the electroweak interaction, so the IR divergent part
is regularized as ∼ log(4E 2 /m2T ) with mT ∼ |g|T .
Kinetic equation governing the production of majorons is similar to Eq. (28), and has
the form:
h |g|4 log(4E 2 /m2T ) exp(−y)

2
x∂x fχ = + ···, (34)
x 27 π 3 x y
where we omitted elastic reactions. Integrating this equation we can find for the ratio of
the number density of χ to the equilibrium value:
nχ |g|4 log(4E 2 /m2T )x

(eq)
= . (35)
nχ 28 ζ (3)π 3 h
(eq)
Demanding that nχ < nχ we obtain
|g| < 2.1 × 10−5 T 1/4 , (36)
which gives a constraint similar to Eq. (32). However, if we take into account quantum
statistics, the bound becomes slightly weaker for large charge asymmetry of neutrinos, as
shown in Fig. 5.
So, to summarize, the muonic or tauonic lepton asymmetry would be erased if the
neutrino–majoron interaction is sufficiently strong, i.e., |g| 10−5 . If we restrict ourselves
to a scenario that the majoron–neutrino interactions suppress neutrino oscillations and
thereby keep the large lepton asymmetries intact, the largest coupling constant should
be smaller than ∼ 10−5 . However this does not necessarily mean that the conspiracy
between the speed-up effect of ξµ,τ and a shift of β equilibrium due to ξe is impossible for
|g| 10−5 . In fact, if the L-non-conserving interactions became efficient only after muonic
and tauonic neutrinos decoupled from the thermal plasma, then the neutrino degeneracy
would still be maintained in muonic or tauonic neutrinos resulting in their larger energy
density at BBN. Majorons, as well, can contribute to additional relativistic degrees of
freedom at BBN. We will discuss these issues below.
Fig. 5. Upper bounds on |g| obtained from the requirement that L-violating reactions, 2ν ↔ 2ν̄ and ν + ν̄ ↔ 2χ ,
are out of equilibrium. Numerical and analytical bounds for the former reaction are shown as solid and
long-dashed lines, respectively. The short-dashed line represents the constraint obtained numerically for the latter
reaction.
5. Neutrino oscillations in primeval plasma
As is established now, neutrino mass eigenstates are related to the flavor eigenstates
through the orthogonal matrix (we neglect a possible CP violation):
νa = Uaj νj , (37)
where a = e, µ, τ and j = 1, 2, 3. We have chosen the parameters so that in the limit of
small mixing νe ≈ ν1 , νµ ≈ ν2 , ντ ≈ ν3 . However, the mixing between neutrinos are known
to be large and they cannot be taken as a dominant single mass eigenstate.
Kinetic equations for the density matrix of oscillating neutrinos can be written as
[23,25]:

i ρ̇ = H(1) , ρ − i H(2), ρ , (38)
where the first commutator term includes the vacuum Hamiltonian and the effective
potential of neutrinos in medium calculated in the first order in Fermi coupling constant,
GF (weak interaction part) and in the second order in the coupling to majoron. The latter
is given by Eq. (26) and does not contain off-diagonal terms if only one flavor coupling
constant dominates. The weak interaction potential contains off-diagonal terms of the same
order of magnitude as diagonal ones because of universal coupling of W and Z bosons
to all neutrino flavors. Their explicit expressions can be found, e.g., in Ref. [7]. In what
follows we will skip upper indices (1) and (2) at H.
The second anti-commutator term in Eq. (38) describes breaking of coherence induced
by neutrino scattering and annihilation as well as neutrino production by collisions in
primeval plasma. It includes imaginary part of the Hamiltonian calculated in the second
order in GF and the fourth order terms in neutrino–majoron coupling g related to neutrino–
neutrino scattering through majoron exchange. If majorons were abundant in the primeval
plasma the processes of neutrino–majoron scattering should also be included.
Since the “atmospheric” neutrino mass difference is much larger than the “solar” one,
we may simplify the problem considering the process in two steps. First, oscillations
between ντ and νµ were switched-on. This started at temperatures about 10–15 MeV and
might lead to equilibration of tauonic and muonic charge asymmetries if these asymmetries
were not erased by reactions (27). Later at T 3 MeV, oscillations would start between νe
and νµ , which is a certain mixture of cντ + sνµ . This process is potentially dangerous for
BBN because it could change number density and spectrum of νe and ν̄e .
If mixing is effective only between two neutrinos, the density matrix is 2 × 2 and kinetic
equations for its components have the form:
iH x∂x ρµµ = Hµτ ρτ µ − Hτ µ ρµτ − iIµµ

(coll)
, (39)
iH x∂x ρτ τ = Hτ µ ρµτ − Hµτ ρτ µ − iIτ(coll)
τ , (40)
iH x∂x ρµτ = (Hµµ − Hτ τ )ρµτ + Hµτ (ρτ τ − ρµµ ) − iΓµτ ρµτ , (41)

where H and x are defined below Eq. (28), Hab = Vab + j =2,3 Uaj Ubj mj /2E, Uaj 2
are matrix elements of the mixing matrix; in the case under consideration Uµ2 = Uτ 3 =
cos θ ≡ c and Uµ3 = −Uτ 2 = sin θ ≡ s. The coherence breaking terms are given by the
usual collision integrals I (coll) in the equations for the diagonal components and by

Γµτ = 1.1 × 10−22 y/x 5 MeV (42)
for the non-diagonal component ρµτ = ρτ∗µ . In the expression for Γµτ we neglected the
contribution from the majoron related processes. The effective potential Vab contains
contributions from the usual weak interactions and from the neutrino–majoron interactions
(26). We assume that the weak part is dominated by the charge asymmetric contribution,
which is true in the case of a large charge asymmetry:

(w)
Vab = 1.5C1 GF T 3
dy (ρab − ρ̄ab ) ≡ B dy (ρab − ρ̄ab ) (43)
with C1 defined in Eq. (6), dy = d 3 y/(2π)3 , ρ̄ is the antineutrino density matrix, and the
coefficient 1.5 comes from (11/4) · 2 ζ(3)/π 2 , i.e., from normalization to photon number
density.
Eqs. (39)–(41) can be solved numerically but we can make reasonable estimates
analytically in the following way (for more details see, e.g., Refs. [1,26]). In the case of
strong coherence breaking, i.e., Γµτ
H , Eq. (41) can be formally solved in the stationary
point approximation, i.e., putting r.h.s. equal zero. In this way we obtain:
Hµτ (ρµµ − ρτ τ )
ρµτ =
Hµµ − Hτ τ − iΓµτ

ρµµ − ρτ τ δm232 iΓµτ − c2 δm232 /2E
≈ Vµτ + s2 1 + , (44)
Vµµ − Vτ τ 4E Vµµ − Vτ τ
where δm232 ≡ m23 − m22 , s2 = sin 2θ and c2 = cos 2θ .
This expression could be substituted into Eqs. (39) and (40) for the diagonal components
but one cannot obtain the closed equations for the latter because the potential Vµτ contains
an integral from ρµτ over momentum, see Eq. (43). We assume for simplicity that only
(w)
weak-interaction potential has noticeable off-diagonal components, i.e., Vµτ Vµτ . It can
be true if, e.g., |gτ τ |
|gµµ |. To express Vµτ through diagonal components we integrate
Eq. (44) and similar equation for ρ̄µτ over momentum and subtract one from the other. To
simplify the expressions let us present the diagonal part of the potential as a sum of charge
symmetric (majoron) and antisymmetric (weak) parts:
(χ)
Vµµ − Vτ τ ≡ V+ + V−(w) , (45)
(w)
where the charge asymmetric part V− can be separated into two terms containing
integrals from neutrino and antineutrino elements of the density matrix (see Eq. (43)),
V−(w) = Vν − V̄ν .
After straightforward calculations we obtain:
(χ) 2
V+ − (Vν − V̄ν )2 s2 δm232 B
Vµτ = (χ) (χ)
dy
V+ (V+ − Vν − V̄ν ) 4E

ρµµ − ρτ τ iΓµτ − δm232 c2 /2E
× (χ) 1+ (χ)
V+ + Vν − V̄ν V+ + Vν − V̄ν

ρ̄µµ − ρ̄τ τ i Γ¯µτ − δm232 c2 /2E
− (χ) 1+ (χ)
. (46)
V+ − Vν + V̄ν V+ − Vν + V̄ν
(χ) (w)
This result is valid if |V+ | |V− |. In the case of the usual weak interactions considered,
(χ) (w)
e.g., in Refs. [7,8] the situation is opposite, |V+ | |V− |, and the character of the
approximate analytical solution is completely different.
(χ) (w) (χ)
For a rough estimate let us assume that |V+ | > |V− | and |V+ | > |δm232 |/2E, while
|Vµτ | < |s2 δm32 |/2E, as follows from Eq. (46). In this case we find
2
2 2
δm32 Γµτ (ρµµ − ρτ τ )
H x∂x ρµµ = − s2 (χ) (χ) 2
− Iµµ
(coll)
. (47)
4E (Vµµ − Vτ τ )
The evolution of the difference of muonic and tauonic charges is determined by

(µτ )
L (µτ )
≡ dy Ly ≡ dy (ρµµ − ρτ τ − ρ̄µµ + ρ̄τ τ ). (48)
The collision integrals disappear from the time derivative of this difference since weak
interactions conserve leptonic charges and majoron induced processes are assumed to be
suppressed, see Section 4 and in particular Eq. (32). The evolution of L is governed by the
equation:
−4 2
dL(µτ ) |g| δm232 dy (µτ )
≈ −7
Ly (49)
dx 1.5 × 10 −3
2.5 × 10 eV 2 y
and for the majoron coupling constant bounded by the conditions presented above, L(µτ )
remains practically constant till nucleosynthesis epoch, i.e., x ∼ 1.
We have performed numerical calculations to solve Eqs. (39)–(41) in a similar way to
Ref. [7]. The coupling constant matrix gab is assumed to take the form: gab = g δaτ δbτ ,
Fig. 6. The evolutions of ξµ and ξτ for several values of |g| for the maximal mixing and δm232 = 2.5 × 10−3 eV2 .
The initial conditions are ξµ = −0.5 and ξτ = 1.0.
for simplicity. The evolution of dimensionless chemical potentials ξµ and ξτ for several
values of |g| are shown in Fig. 6. One can see that the lepton asymmetries of νµ and
ντ equilibrate around T ∼ 12 MeV when |g| = 0. We have checked that the evolution
remains the same for |g| < 10−8 . As |g| increases, the oscillations become less efficient
and completely stop for |g| 10−7 , which well agrees with the analytical estimates. It
should be also noted that the lower bound on |g| obtained from Eq. (49) does not depend
on the magnitude of the initial lepton asymmetries. We have checked that the oscillations
are similarly stuck for |g| 10−7 even if the initial lepton asymmetries are taken to be
smaller. The only difference is the temperature at which the asymmetries equilibrate when
g = 0.
We can repeat the same arguments for oscillations between νe and νµ . The squared
mass difference and mixing angle parameters for the oscillations between νe and νµ are
δm221 ≡ m22 − m21 = 7.3 × 10−5 eV2 and sin2 θ = 0.315 [27]. The coupling constant matrix
gab is similarly approximated to be gab = g δaµ δbµ . Also the effective potential induced
by the energy densities of electrons and positrons is taken into account, since here we
consider oscillations including νe . The lower bound of |g| becomes slightly relaxed due
to the smaller mass difference, as can be seen from Fig. 7. The reason why the lepton
asymmetries of νe and νµ are not equilibrated completely when g = 0 is that the mixing
angle we adopt is not maximal.
Lastly, we comment on the form of the coupling matrix, gab , for which the (νe − νµ )-
transformations are suppressed and, consequently, a large lepton asymmetry of the universe
is allowed. So far we have assumed that gab is diagonal in order to deal with both ντ − νµ
and νe − νµ oscillations similarly. However, in fact, a coupling matrix with more generic
form works as well. We would like to stress that the equilibration of the muonic and tauonic
lepton asymmetries is not harmful to our purpose, as long as these asymmetries are not
Fig. 7. The evolutions of ξe and ξµ for several values of |g| with sin2 θ = 0.315 and δm221 = 7.3 × 10−5 eV2 .
The initial conditions are ξe = 0.1 and ξµ = −0.5.
erased through either oscillations or by the reactions shown in Figs. 3 and 4. Therefore
we do not have to assume the hierarchical structure in gτ τ , gµτ , and gµµ , and they are
constrained only by the astrophysical bounds:
|gab | < 5 × 10−6 (50)
with (a, b) = {(τ, τ ), (µ, τ ), (µ, µ)}. In order to suppress the oscillations between νe and
νµ , at least one of these coupling constants should be larger than ∼ 10−7 , while |gee |,
|geµ |, and |geτ | must be much smaller than ∼ 10−7 . Thus we conclude that the coupling
matrix suitable for our purpose should satisfy:
|gee |, |geµ |, |geτ | 10−7 ,

10−7 Max |gτ τ |, |gµτ |, |gµµ | 5 × 10−6 . (51)
In the simplest class of majoron models, the coupling matrix, gab , is proportional to the
neutrino mass matrix (mν )ab . In particular, for the normal mass hierarchy, m1 < m2 < m3 ,
the reconstructed mass matrix is often parametrized as [28]

0 0 λ λ2 λ λ
g ∝ mν ∝ 0 1 1 or λ 1 1 (52)
λ 1 1 λ 1 1
with λ ∼ 0.2. Therefore the coupling matrix of this type can satisfy the conditions (51). Of
course, in the more involved models, the coupling matrix is not necessarily proportional to
the neutrino mass matrix.
6. Discussions and conclusions
In the preceding sections we have shown that neutrino oscillations in primordial

plasma can be blocked by an introduction of the neutrino–majoron interaction with a
moderate value of coupling constants, |g| 5 × 10−6 . Here let us take up the case of
|g| 10−5 . For a definite discussion, we assume that the elements of coupling constant
matrix whose indices involve electronic neutrino are much smaller than the other elements,
i.e., |gee |, |gea | |gab | with (a, b) = {(τ, τ ), (µ, τ ), (µ, µ)}, so the electron type lepton
number conserves during the relevant BBN epoch. As we have seen in Section 4, the
muonic and/or tauonic lepton number is not conserved for |gab | 10−5 . The fate of the
large degeneracy in νµ and ντ depends on whether this L-violating interaction reached
equilibrium before or after νµ and ντ decoupled from the thermal bath.
First let us consider the former scenario. Since νµ and ντ kept on to be in strong contact
with the primeval plasma when the L-violating interactions were efficient, the muonic
and/or tauonic lepton asymmetry vanished and thereby heated the plasma. At the same time
majorons were abundantly produced and contributed to the effective number of neutrino
species as Nν = 4/7. The CP asymmetric part of the effective potential of neutrino due
to the muonic and tauonic lepton asymmetries vanished, but there would be a similar (but
possibly smaller) term due to the charge asymmetry of νe (see Eqs. (7) and (8)). Also there
exists the contribution the effective potential due to abundant majorons in the plasma (see
Eq. (25)). A further important point is that L-violating processes, shown in Fig. 3, induces
the effective potential between neutrino and antineutrino, since a † b no longer vanishes.
However, neutrino oscillations between νe and νµ can be treated in the same way as before,
since electron type lepton number remains a well conserved quantity. Therefore, we can
deduce from the previous results that neutrino oscillations between νe and νµ are blocked
and ξe remains intact. It should be noted that equilibrium majorons, instead of excessive
νµ and ντ , contribute to the extra radiation, which speeds up the expansion rate, and that
its contribution, Nν = 4/7, well satisfies the constraint shown in Eq. (3).
Next we discuss the other possibility that the L-violating interactions come to thermal
equilibrium after decoupling of νµ and ντ . Majorons are also produced, as in the previous
case. However, since νµ and ντ cannot exchange energy with the primeval plasma, the
total energy of majorons, νµ and ντ must be conserved. Therefore the excessive energy
previously stored in νµ and ντ is redistributed among majorons, νµ , and ντ , and their
distribution would be like Eq. (33) but without any charge asymmetry, i.e., with ξ = ξ̄ . It is
thus clear that Nν remains unchanged before and after L-violating interaction comes into
equilibrium. Our concern here is whether the modified distribution of νµ affects that of νe
through neutrino oscillations. However, the oscillations are blocked on the same ground.
In connection with the second case, there is an interesting possibility that the majoron-
exchange reactions: νµ,τ ν̄µ,τ ↔ νe ν̄e , νµ,τ νµ,τ ↔ νe νe , νµ,τ νµ,τ ↔ ν̄e ν̄e , ν̄µ,τ ν̄µ,τ ↔
νe νe and ν̄µ,τ ν̄µ,τ ↔ ν̄e ν̄e , could produce additional electronic neutrinos and antineutrinos.
These processes can proceed if |gee | and |gea | with a = µ, τ are sizable 10−5 . If this is
the case, the relative abundance of νe and ν̄e with respect to photons and electrons would
be larger than in the standard model. This would lead to a later neutron–proton freezing
and to a smaller number density of survived neutrons, which might compensate the effect
of additional energy density stored in all species of neutrinos and majorons and even
overshoot it. Then we do not need to suppress neutrino oscillations, therefore the coupling
constant matrix gab is allowed to take rather arbitrary values as long as the L-violating
processes come into equilibrium after decoupling of muonic and tauonic neutrinos.
Thus our examination indicates that a conspiracy between the speed-up effect and
the shift of β equilibrium is possible for |gab | 10−5 , as long as |gee |, |gea | |gab |
with a, b = µ, τ . Furthermore, even for |gee |, |gea | 10−5 , such a cancellation might be
possible in a somewhat different way. This considerably extends the allowed parameter
space, although rigorous study might be necessary to obtain further quantitative results.
In this paper we have shown that the hypothetical neutrino–majoron interaction can
suppress neutrino oscillations in the primordial plasma to prevent lepton asymmetries of all
neutrino species from being equilibrated. The exact form of the effective potential induced
by this interaction is calculated. We have found an allowed range of the coupling constant:
10−7 < |g| < 5 × 10−6 (more precisely, see Eq. (51)), which satisfies the astrophysical
bounds and makes the scenario operative. For the coupling constant in this range, νe − νµ
oscillation in the early Universe is blocked, thereby keeping the cosmological lepton
asymmetry of electron type unchanged. The upper bound comes from the requirement that
lepton number is effectively conserved, and the lower bound is obtained from the study of
the evolution of the lepton asymmetries both analytically and numerically, in two flavor
approximation. The constant matrix in the simplest class of majoron models can satisfy
the desired constraints, in the case of the normal mass hierarchy. Furthermore, even for
|g| 10−5 , the large energy density necessary to cancel the effect of ξe can be supplied
by majorons, νµ and ντ , even if muonic and/or tauonic lepton asymmetries were erased.
Thus we conclude that an addition of the majoron field to the standard model can reopen
a possibility that the effect of ξe is compensated by large ξµ,τ (or by the extra energy of
majoron itself), thereby curing a probable discrepancy between the BBN and CMBR.
Acknowledgements
A.D. is grateful to the Research Center for the Early Universe of the University of Tokyo
for the hospitality during the time when this work started. F.T. is grateful to M. Kawasaki
and K. Ichikawa for useful discussions. F.T. would like to thank the Japan Society for the
Promotion of Science for financial support.
Appendix A. Derivation of effective potential due to majoron–neutrino interaction
Here we present the derivation of the effective potentials Eqs. (23) and (25). First
the notations and conventions that we adopt here, are listed. The metric is taken to be
(−, +, +, +). The gamma matrices in the chiral representation are

0 1 0 σi 1 0
γ = −i
0
, γ = −i
i
, γ5 = , (A.1)
1 0 −σi 0 0 −1
where σi are the Pauli matrices. The gamma matrices in the Dirac representation are related
to those in the chiral representation as

µ µ 1 1 1
γDirac = U γchiral U † , U = √ , (A.2)
2 1 −1
and they are given by

1 0 0 σi 0 1
γ = −i
0
, γ =i
i
, γ5 = . (A.3)
0 −1 −σi 0 1 0
The charge conjugation matrix C is defined as

 −i σ2 0 (chiral),
0 −σ
C ≡ iγ 2 γ 0 = 0 σ2
2
(A.4)
 −i (Dirac).
σ 0 2
Hereafter we work with the gamma matrices in the Dirac representation. The momentum
expansion of a free Dirac field ψ(x) with mass m is given by

ψ(x) = dp a(p, s)u(p, s)e−iEp t +ipx + b† (p, s)v(p, s)eiEp t −ipx
s

dp a(p, s)u(p, s) + b† (−p, s)v(−p, s) e−ip t +ipx ,
0
= (A.5)
s

where we defined p0 ≡ p Ep = p p2 + m2 with p = ±1. The sign of p is chosen so
that the positive and negative-energy solutions are reproduced. s = ± describes the sign of
the spin projection eigenvalue, and we have used the notation dp ≡ d 3 p/(2π)3 . The Dirac
spinors u(p, s) and v(p, s) take the form,

1 Ep + m
u(p, s) = χu (s),
2Ep (Ep + m) −p · σ

1 −p · σ
v(p, s) = χv (s). (A.6)
2Ep (Ep + m) Ep + m
The two-component spinors χu (s) and χv (s) are not independent but related to each other
through: χv (s) = −iσ2 χu (s)∗ . If we choose them to be eigenstates of the helicity operator
p · σ /|p|, they are
−iϕ/2 −iϕ/2
e cos θ2 −e sin θ2
χu (+) = , χ u (−) = ,
eiϕ/2 sin θ2 eiϕ/2 cos θ2
−iϕ/2 −iϕ/2
−e sin θ2 −e cos θ2
χv (+) = , χv (−) = , (A.7)
eiϕ/2 cos θ2 −eiϕ/2 sin θ2
where θ and ϕ are defined as p = |p|(sin θ cos ϕ, sin θ sin ϕ, cos θ ). For flipped momentum
−p, {θ, ϕ} should be replaced with {π − θ, ϕ + π}.
In deriving the effective potential, the neutrino field ν(x) is approximated to be a
massless left-handed field:

νi (x) = dp ai (p)up + bi (−p)† v−p e−ip t +ip x ,
0
(A.8)
where

1 + γ5 1 1
up ≡ u(p, −) = √ χu (−),
2 2 1

1 + γ5 i 1
u−p ≡ u(−p, −) = √ χu (+),
2 2 1

1 + γ5 1 1
vp ≡ v(p, +) = √ χu (−),
2 2 1

1 + γ5 i 1
v−p ≡ v(−p, +) = √ χu (+). (A.9)
2 2 1
The annihilation and creation operators a, b, a † and b† satisfy the anticommutation
relations.

ai (p), aj (q)† = (2π)3 δij δ (3) (p − q),

bi (p), bj (q)† = (2π)3 δij δ (3) (p − q),
others = 0. (A.10)
Similarly, the momentum expansion of the free majoron field is

χ(x) = dp aχ (p)uχ (p) + aχ (−p)† vχ (−p) e−ip t +ipx ,
0
(A.11)
where
1
uχ (p) = vχ (p) = . (A.12)
2Ep
The annihilation and creation operators aχ , aχ† satisfy the commutation relations:

aχ (p), aχ (q)† = (2π)3 δ (3) (p − q),

others = 0. (A.13)
Next we derive the effective Hamiltonian Eq. (20). The equation of motion for the
majoron field χ is
i ∗ †

∂µ ∂ µ χ = − gab νaT Cνb + gab νb Cνa∗ . (A.14)
2
To solve this equation, we define the majoron and neutrino fields in momentum space:

d 4p
χ(x) = χ̃(p)eipx ,
(2π)4

d 4p
νa (x) = ν̃a (p)eipx . (A.15)
(2π)4
Then the solution of Eq. (A.14) is

i d 4q ∗

χ̃(p) = 2 gab ν̃a (p − q)T C ν̃b (q) + gab ν̃b (−p + q)† C ν̃a∗ (−q) .
2p (2π)4
(A.16)
Thus the effective Hamiltonian is:

(νν) (νν)
Heff = − d 3 x Leff

1 d 4 p ipx i
i d 4q
=− d 3x e gab ν̃a (p − q)T C ν̃b (q)
2 (2π)4
2 2p2 (2π)4

∗
d 4 r irx d 4 s isx
+ gab ν̃b (−p + q)† C ν̃a∗ (−q) e e
(2π)4 (2π)4
∗

× gcd ν̃c (r)T C ν̃d (s) + gcd ν̃d (−s)† C ν̃c∗ (−r) , (A.17)
where we substituted Eq. (A.16) into the second equation. Note that the numerical factor
1/2 is inserted in front of the last equation. In a first-order perturbative approximation the
neutrino and majoron fields can be set to be free fields:

ν̃i (p) = 2πδ p0 − p Ep νi (p),

χ̃(p) = 2πδ p0 − p Ep χ(p), (A.18)
where
νi (p) = ai (p)up + bi (−p)† v−p ,

χ(p) = aχ (p)uχ (p) + aχ (−p)† vχ (−p). (A.19)
Thus the effective Hamiltonian becomes

1
dp dq dr ds (2π)3 δ (3) (p + q + r + s)e−iEtot t
(νν)
Heff (t) = −
16
F (p, q)F (r, s)
× , (A.20)
p q Ep Eq (1 − p q cos θpq )
where
Etot = p Ep + q Eq + r Er + s Es ,
p·q
cos θpq ≡ ,
|p||q|
∗
F (p, q) ≡ gab νa (p)T Cνb (q) + gab νb (−q)† Cνa∗ (−p). (A.21)
Once the effective Hamiltonian is obtained, we can calculate the effective potential for
neutrinos according to Ref. [23]. All we have to do is to evaluate

Heff (t = 0), aj (p)† ai (p) = −(2π)3 δ (3) (0)[Vp , ρp ]ij , (A.22)
where [Vp ]ab is the effective potential matrix. Using the anticommutation relations (A.10),
we obtain the final result:

(νν)
1 † T
Vp ab
= dq g ρq + ρ Tq g ab . (A.23)
4|p||q|
Next we derive the effective Hamiltonian, which describes neutrino–majoron scattering.
In this case we integrate out the neutrino field instead of majoron field. The equation of
motion for the neutrino is
∗
−iγ 0 γ µ ∂µ νa + iχgab Cνb∗ = 0. (A.24)
The solution is then

ipµ ∗ d 4q
ν̃a (p) = g χ̃(p + q)γ µ γ 0 C ν̃b∗ (q). (A.25)
p2 ab (2π)4
Substituting this result into Eq. (19), we obtain the effective Hamiltonian:

(νχ) i ∗ d 4 p d 4 q d 4 r d 4 s i(p+q+r+s)x
Heff (t) = − gab gbc d x 3
e
2 (2π)4 (2π)4 (2π)4 (2π)4
(q + r)µ
×χ̃(p)χ̃ (q) ν̃¯ c (−r)γ µ ν̃a (s)
(q + r)2

i
= − g † g ab dp dq dr ds (2π)3 δ (3) (p + q + r + s)e−iEtot t

2

(q + r)µ
×χ(p)χ(q)ν̄a (−r)γ µ νb (s) , (A.26)
2q · r q 0 =q Eq ,r 0 =r Er
where we take both neutrino and majoron fields to be on-shell in the second equation.
Applying Eq. (A.22), we obtain the effective potential due to the neutrino–majoron
scattering:

(νχ)
fχ (q) †
Vp ab
= dq g g ab , (A.27)
4|p||q|
(χ)
where the number density of majorons with momentum p, np , is defined as
†
aχ (p) aχ (p ) = (2π)3 δ (3) (p − p ) fχ (p).
Appendix B. Invariant amplitude squared for 2ν ↔ 2ν̄
Here we calculate the invariant amplitude squared for the reaction, 2ν ↔ 2ν̄. Provided
that the diagonal coupling constants are larger than the non-diagonal ones, and that one
of the diagonal coupling constants dominates, we consider neutrinos of the flavor with
the largest coupling. Hereafter we drop the sub-indices for flavor. The diagrams, which
contribute to the reaction, are shown in Fig. 3.
The S-matrix and the invariant amplitude squared are related as
Sβα = 2π(−i)δ (4)(pβ − pα )Mβα ,

|Minv |2 = (2π)−6 (2π)3 2Eα (2π)3 2Eβ |Mβα |2 , (B.1)

α β
where α and β represent symbolically the initial and final states. Each matrix element for
s-, t- and u-channel diagrams is given as
g2 1 T
M(s)
p3 p4 ,p1 p2 = up2 Cup1 vpT4 Cvp3 ,
(2π) (p1 + p2 )
3 2
g2 1 T
M(t )
p3 p4 ,p1 p2 =− vp3 Cup1 vpT4 Cup2 ,
(2π) (p1 − p3 )
3 2
g2 1 T
M(u)
p3 p4 ,p1 p2 = vp4 Cup1 vpT3 Cup2 , (B.2)
(2π) (p1 − p4 )
3 2
where the four momenta satisfy the conservation condition, p1 + p2 = p3 + p4 . If we use

the explicit expressions for up and vp derived in the previous section, we obtain
g2 1 3
M(s)
p3 p4 ,p1 p2 + Mp3 p4 ,p1 p2 + Mp3 p4 ,p1 p2 =
(t ) (u)
√ . (B.3)
(2π)3 E1 E2 E3 E4 4
Substituting this result into Eq. (B.1), the invariant amplitude squared becomes
|Minv |2 = 9|g|4 . (B.4)
Appendix C. Formulas for the collisional integral in the Boltzmann approximation
Here we write down the formulas for the collisional integral in the Boltzmann
approximation for completeness. For the derivation, see Ref. [29].
3 3 3
d y2 d y3 d y4 −y3 −y4 (4)
e δ (y1 + y2 − y3 − y4 ) = 4π 2 e−y1 ,
y2 y3 y4
3 3 3 3
d y1 d y2 d y3 d y4 −y3 −y4 (4)
e δ (y1 + y2 − y3 − y4 ) = 32π 3 . (C.1)
y1 y2 y3 y4
References
[1] A.D. Dolgov, Phys. Rep. 370 (2002) 333.

[2] K. Kohri, M. Kawasaki, K. Sato, Astrophys. J. 490 (1997) 72.
[3] V. Barger, J.P. Kneller, H.S. Lee, D. Marfatia, G. Steigman, Phys. Lett. B 566 (2003) 8.
[4] S.H. Hansen, G. Mangano, A. Melchiorri, G. Miele, O. Pisanti, Phys. Rev. D 65 (2002) 023511.
[5] J.A. Harvey, E.W. Kolb, Phys. Rev. D 24 (1981) 2090;

A.D. Dolgov, D.K. Kirilova, J. Moscow Phys. Soc. 1 (1991) 217;
A.D. Dolgov, Phys. Rep. 222 (6) (1992);
A. Casas, W.Y. Cheng, G. Gelmini, Nucl. Phys. B 538 (1999) 297;
J. March-Russell, H. Murayama, A. Riotto, JHEP 9911 (1999) 015;
J. McDonald, Phys. Rev. Lett. 84 (2000) 4798;
M. Kawasaki, F. Takahashi, M. Yamaguchi, Phys. Rev. D 66 (2002) 043516;
M. Yamaguchi, Phys. Rev. D 68 (2003) 063507;
T. Chiba, F. Takahashi, M. Yamaguchi, Phys. Rev. Lett. 92 (2004) 011301;
F. Takahashi, M. Yamaguchi, hep-ph/0308173, Phys. Rev. D, in press.
[6] SNO Collaboration, Phys. Rev. Lett. 89 (2002) 011301;
SNO Collaboration, Phys. Rev. Lett. 89 (2002) 011302;
KamLAND Collaboration, Phys. Rev. Lett. 90 (2003) 021802;
Super-Kamiokande Collaboration, S. Fukuda, et al., Phys. Lett. B 539 (2002) 179;
B.T. Cleveland, et al., Astrophys. J. 496 (1998) 505;
SAGE Collaboration, D.N. Abdurashitov, et al., Phys. Rev. C 60 (1999) 055801, astro-ph/0204245;
GALLEX Collaboration, W. Hampel, et al., Phys. Lett. B 447 (1999) 127;
GNO Collaboration, C. Cattadori, Nucl. Phys. B (Proc. Suppl.) 110 (2002) 311;
Super-Kamiokande Collaboration, Y. Fukuda, et al., Phys. Rev. Lett. 81 (1998) 1562;
MACRO Collaboration, M. Ambrosio, et al., Phys. Lett. B 434 (1998) 451.
[7] A.D. Dolgov, S.H. Hansen, S. Pastor, S.T. Petcov, G.G. Raffelt, D.V. Semikoz, Nucl. Phys. B 632 (2002)
363.
[8] C. Lunardini, A.Yu. Smirnov, Phys. Rev. D 64 (2001) 073006;
Y.Y.Y. Wong, Phys. Rev. D 66 (2002) 025015;
K.N. Abazajian, J.F. Beacom, N.F. Bell, Phys. Rev. D 66 (2002) 013008.
[9] V. Barger, J.P. Kneller, P. Langacker, D. Marfatia, G. Steigman, Phys. Lett. B 569 (2003) 123.
[10] Y. Chikashige, R.N. Mohapatra, R.D. Peccei, Phys. Rev. Lett. 45 (1980) 1926;
G.B. Gelmini, M. Roncadelli, Phys. Lett. B 99 (1981) 411;
H.M. Georgi, S.L. Glashow, S. Nussinov, Nucl. Phys. B 193 (1981) 297;
A.Yu. Smirnov, Yad. Fiz. 34 (1981) 1547;
J. Schechter, J.W.F. Valle, Phys. Rev. D 25 (1982) 774.
[11] K.S. Babu, I.Z. Rothstein, Phys. Lett. B 275 (1992) 112.
[12] L. Bento, Z. Berezhiani, Phys. Rev. D 64 (2001) 115015.
[13] A.D. Dolgov, Surveys High Energ. Phys. 17 (2002) 91.
[14] D. Nötzold, G. Raffelt, Nucl. Phys. B 307 (1988) 924.
[15] J. Pantaleone, Phys. Lett. B 287 (1992) 128;
S. Samuel, Phys. Rev. D 48 (1993) 1462;
V.A. Kostelecký, J. Pantaleone, S. Samuel, Phys. Lett. B 315 (1993) 46;
V.A. Kostelecký, S. Samuel, Phys. Rev. D 49 (1994) 1740;
V.A. Kostelecký, S. Samuel, Phys. Rev. D 52 (1995) 3184;
S. Samuel, Phys. Rev. D 53 (1996) 5382;
V.A. Kostelecký, S. Samuel, Phys. Lett. B 385 (1996) 159;
S. Pastor, G.G. Raffelt, D.V. Semikoz, Phys. Rev. D 65 (2002) 053011.
[16] R. Tomas, H. Päs, J.W.F. Valle, Phys. Rev. D 64 (2001) 095005.
[17] Y. Farzan, Phys. Rev. D 67 (2003) 073015.
[18] G.G. Raffelt, Stars as Laboratories for Fundamental Physics, University of Chicago Press, Chicago, 1996.
[19] T. Bernatowicz, et al., Phys. Rev. Lett. 69 (1992) 2341.
[20] J.F. Beacom, N.F. Bell, D. Hooper, S. Pakvasa, T.J. Weiler, Phys. Rev. Lett. 90 (2003) 181301;
S. Ando, Phys. Lett. B 570 (2003) 11;
G.L. Fogli, E. Lisi, A. Mirizzi, D. Montanino, hep-ph/0401227.
[21] IceCube Collaboration, J. Ahrens, astro-ph/0305196.
[22] J.F. Beacom, N.F. Bell, Phys. Rev. D 65 (2002) 113009.
[23] G. Sigl, G. Raffelt, Nucl. Phys. B 406 (1993) 423.
[24] A.D. Dolgov, S.H. Hansen, D.V. Semikoz, Nucl. Phys. B 524 (1998) 621.
[25] A.D. Dolgov, Sov. J. Nucl. Phys. 33 (1981) 700, Yad. Fiz. 33 (1981) 1309 (in Russian);
M.A. Rudzsky, Astrophys. Space Sci. 165 (1990) 65;
B.H.J. McKellar, M.J. Thomson, Phys. Rev. D 49 (1994) 2710.
[26] A.D. Dolgov, F.L. Villante, Nucl. Phys. B 679 (2004) 261.
[27] G.L. Fogli, E. Lisi, A. Marrone, D. Montanino, A. Palazzo, A.M. Rotunno, eConf C030626, 2003, THAT05,
hep-ph/0310012.
[28] A.Y. Smirnov, hep-ph/0311259.
[29] A.D. Dolgov, S.H. Hansen, D.V. Semikoz, Nucl. Phys. B 503 (1997) 426.
Nuclear Physics B 688 [FS] (2004) 217–265
Local scale-invariance and ageing in noisy systems

Alan Picone, Malte Henkel
Laboratoire de Physique des Matériaux, 1 Université Henri Poincaré Nancy I, B.P. 239,
F-54506 Vandœuvre lès Nancy Cedex, France
Received 9 February 2004; accepted 23 March 2004
Abstract
The influence of the noise on the long-time ageing dynamics of a quenched ferromagnetic spin
system with a non-conserved order parameter and described through a Langevin equation with a
thermal noise term and a disordered initial state is studied. If the noiseless part of the system is
Galilei-invariant and scale-invariant with dynamical exponent z = 2, the two-time linear response
function is independent of the noise and therefore has exactly the form predicted from the local
scale-invariance of the noiseless part. The two-time correlation function is exactly given in terms of
certain noiseless three- and four-point response functions. An explicit scaling form of the two-time
autocorrelation function follows. For disordered initial states, local scale-invariance is sufficient for
the equality of the autocorrelation and autoresponse exponents in phase-ordering kinetics. The results
for the scaling functions are confirmed through tests in the kinetic spherical model, the spin-wave
approximation of the XY model, the critical voter model and the free random walk.
PACS: 64.60.Ht; 82.20.M; 11.25.Hf
Keywords: Conformal invariance; Schrödinger invariance; Ageing; Phase-ordering kinetics;

Martin–Siggia–Rose theory; Correlation function; Response function
1. Introduction
The study of ageing phenomena as they are known to occur in glassy and non-glassy
systems presents one of the great challenges in current research into strongly coupled
many-body systems far from thermal equilibrium. A common example of this kind of
system is obtained as follows. Consider a magnet at a high-temperature initial state before
E-mail address: henkel@lpm.u-nancy.fr (M. Henkel).

1 Laboratoire associé au CNRS UMR 7556.
doi:10.1016/j.nuclphysb.2004.03.028
218 A. Picone, M. Henkel / Nuclear Physics B 688 [FS] (2004) 217–265
quenching it to a final temperature T at or below its critical temperature Tc > 0. Then

the temporal evolution of the system with T fixed is studied. A key insight has been
the observation that many of the apparently erratic and history-dependent properties of
such systems can be organized in terms of a simple scaling picture [90]. Underlying this
phenomenological picture is the idea that the ageing phenomenon and the related slow
evolution of the macroscopic observables comes from the slow motion of the domain
walls which separate the competing correlated clusters. The domains are of a typical
time-dependent size with length-scale L(t), see [5,9,11,19,23,39] for reviews. In recent
years, much work has been performed on the ageing phenomena of simple ferromagnetic
systems, in the hope that these systems might offer insight useful for the refined study also
of ageing glassy materials. It has turned out that ageing is more fully revealed in two-time
observables, such as the two-time (auto-)correlation function C(t, s) or the two-time linear
(auto-)response function R(t, s) defined as

δφ(t)
C(t, s) := φ(t)φ(s), R(t, s) := , (1.1)
δh(s) h=0
where φ(t) denotes the time-dependent order parameter, h(s) is the time-dependent
conjugate magnetic field, t is referred to as observation time and s as waiting time. One
says that the system undergoes ageing if C or R depend on both t and s and not merely on
the difference τ = t − s. According to the dynamical scaling alluded to above, one expects
for times t, s tmicro and t − s tmicro , where tmicro is some microscopic time scale, the
following scaling forms
C(t, s) = s −b fC (t/s), R(t, s) = s −1−a fR (t/s), (1.2)

such that the scaling functions fC,R (y) satisfy the following asymptotic behaviour:
fC (y) ∼ y −λC /z , fR (y) ∼ y −λR /z (1.3)

as y → ∞ and where λC and λR , respectively, are known as the autocorrelation [27,54] and
autoresponse exponents [77], a and b are further non-equilibrium exponents and z is the
dynamical exponent, where it has been tacitly assumed that the typical cluster size grows
for late times as L(t) ∼ t 1/z . The derivation of such growth laws from dynamical scaling
has been studied in great detail [85]. For a non-conserved order parameter and T < Tc ,
the dynamical exponent z = 2. The exponents λC,R are independent of the equilibrium
exponents and of z [23,39,59]. Since a long time, the equality λC = λR had been taken
for granted but recently, examples to the contrary have been found for either long-ranged
initial correlations in ageing ferromagnets [77] or else in the random-phase sine-Gordon
(Cardy–Ostlund) model [89]. On the other hand, a second-order perturbative analysis of
the time-dependent non-linear Ginzburg–Landau equation reproduces λC = λR [67]. The
precise relationship between λC and λR remains to be understood. If one uses an infinite-
temperature initial state, one has λC = λR d/2 [93].
For ageing ferromagnetic systems with a non-conserved order parameter, the value of
the exponent a depends on the properties of the equilibrium system as follows [47,52].
A system is said to be in class S if its order-parameter correlator Ceq (r) ∼ exp(−|r|/ξ )
with a finite ξ and it is said to be in class L if Ceq (r) ∼ |r|−(d−2+η) , where η is a standard
A. Picone, M. Henkel / Nuclear Physics B 688 [FS] (2004) 217–265 219
equilibrium critical exponent. Then

1/z, for class S,
a= (1.4)
(d − 2 + η)/z, for class L.
For example, in d > 1 dimensions, the kinetic Ising model with Glauber dynamics is in
class S for temperatures T < Tc and in class L at the critical temperature T = Tc . It is
generally accepted that b = 0 for T < Tc and b = a if T = Tc , see, e.g., [39].
The distance from equilibrium is conveniently measured through the fluctuation–
dissipation ratio [20,21]

∂C(t, s) −1
X(t, s) := T R(t, s) . (1.5)
∂s
At equilibrium, the fluctuation–dissipation theorem states that X(t, s) = 1. Ageing systems
may also be characterized through the limit fluctuation–dissipation ratio

X∞ = lim lim X(t, s) . (1.6)
s→∞ t →∞
Below criticality, one expects X∞ = 0, but at T = Tc , the value of X∞ should be universal
according to the Godrèche–Luck conjecture [37,38]. This universality has been confirmed
in a large variety of systems in one and two space dimensions [14,38,51,88]. The order of
the limits is important, since limt →∞ (lims→∞ X(t, s)) = 1 always.
While these statements exhaust the content of dynamical scaling, it may be asked
whether the form of the scaling functions fC,R (y) might be fixed in a generic, model-
independent way through a generalization of that symmetry. Indeed, it has been shown
[46] that an infinitesimal global scale-transformation t → (1 + ε)z t, r → (1 + ε)r with a
constant ε such that |ε| 1 can for any given value of z be extended to an infinitesimal
local scale-transformation where now ε = ε(t, r) may depend on both time and space.2
It can be shown that the local scale-transformations so constructed act as dynamical
symmetries of certain linear field equations which might be viewed as some effective
renormalized equation of motion. Practically more important, assuming that the response
functions of the theory transform covariantly under local scale-transformations, the exact
form of the scaling function fR (y) is found [45,46]
fR (y) = r0 y 1+a−λR /z (y − 1)−1−a , (1.7)

where r0 is a normalization constant.3 Indeed, in deriving this result one actually only
requests that R transforms covariantly under the subalgebra of the infinitesimal local
scale-transformation which excluded time-translations. We say that a theory where the
n-point functions built from certain ‘quasiprimary’ fields transform covariantly under an
algebra of such extended scale-transformations is locally scale-invariant [45,46]. The
prediction (1.7) has been confirmed in a large class of ageing ferromagnets as reviewed
2 This extension of dynamical scaling has an analogue in critical equilibrium systems: there global scale-
invariance can be extended to conformal invariance, see [17,24,44] for introductions.
3 In order to avoid misunderstandings, we recall that (1.7) holds for the total response function as defined in
(1.1) without any subtractions meant to extract an ‘ageing part’.
in [46,50]. The status of the scaling function fC (y) of the spin–spin correlator is less
clear, however. Building on the Ohta–Jasnow–Kawasaki approximation (see [9]) Gaussian
closure procedures [8,83] in the O(n)-model produce approximate forms for fC (y)
but we do not know of any other approach which does not involve some uncontrolled
approximation.
Given the phenomenological success of (1.7), we wish to understand better where such a
supposedly general and exact result derived from a dynamical symmetry and without using
any model-specific properties might come from. In this paper, we shall concentrate on the
important special case z = 2 which describes the phase-ordering kinetics after a quench to a
temperature T < Tc of a ferromagnet with a non-conserved order parameter [85]. We recall
that local scale-transformations are dynamical symmetries of certain differential equations,
such as the free diffusion/free Schrödinger equation for z = 2. Indeed, the maximal
dynamical symmetry of the free diffusion equation ∂t φ = D φ is known since a long time
to be the so-called Schrödinger group [41,63,70], to be defined in Section 2. Also, it is well
established that the same group also describes the dynamic symmetry of non-relativistic
free-field theory [42,57]. It also arises as dynamical symmetry in certain non-linear
Schrödinger equations [31,35,62,81,82], the Burgers equation [56,60] or the equations
of fluid dynamics [74]. If, in addition, D is also considered as a variable, Schrödinger
invariance in d dimension becomes a conformal invariance in d + 2 dimensions [48].
The classification of non-linear equations and of systems of equations admitting as a
dynamical symmetry the Schrödinger group or one of its subgroups (e.g., the Galilei group)
has received a lot of mathematical attention, see [18,30,32–34,72,82]. The extension to
dynamical exponents z = 2 needed for quenches to criticality or for glassy systems will be
left for future work.
However, the setting just outlined is not yet sufficient for the description of ageing
phenomena. Rather, we are interested in the time-dependent behaviour of spin systems
coupled to a heat bath at temperature T . It is usually admitted that after coarse-graining,
this may be modeled in terms of a Langevin equation. If there are no macroscopic
conservation laws, the Langevin equation for the coarse-grained order parameter φ =
φ(t, r) should be model A in the Hohenberg–Halperin classification [53]
∂φ(t, r) δH
= −D + η(t, r) (1.8)
∂t δφ
where H is the classical Hamiltonian, and D stands for the diffusion constant or
equivalently some relaxation rate. Thermal noise is described by a Gaussian random force
η = η(t, r) and is thus characterized by its first two moments

η(t, r) = 0, η(t, r)η(s, r
) = 2DT δ(t − s) δ(r − r
), (1.9)
where T is the bath temperature. It is well known [9,53] that this formalism describes the
relaxation of the system towards its equilibrium state given by the probability distribution
Peq ∼ e−H/T . In addition, initial conditions must be taken into account and are described
in terms of the initial correlation function

a(r − r
) := C(0, 0; r, r
) := φ(0, r)φ(0, r
) (1.10)
and where we already anticipated spatial translation invariance.
Neither the thermal noise nor the initial correlations described by a(r) are included
into the local scale-transformations as studied in [46] which come from systems such as
the free diffusion equation. In this paper, we want to show how both these sources of
fluctuations may be taken into account and we shall explicitly derive the two-time response
and correlation functions. Our analysis will be restricted to the case where z = 2 which,
for instance, is already enough to describe ageing below criticality.
As we shall show, it is useful to slightly generalize the problem and to consider the
kinetics of systems which in the simplest case may be described by a quadratic Hamiltonian
of the form

2
1 ∂φ
H[φ] = dt dr + v(t)φ ,
2
(1.11)
2 ∂r
where v(t) is a time-dependent external potential. Formally, at the level of relativistic free-
field theory, v(t) corresponds to a (time-dependent) mass squared which would measure the
distance from a critical point. Alternatively, v(t) may be viewed as a Lagrange multiplier in
order to ensure the constraint φ(t, r)φ(t, r) = 1 and we shall make this explicit through
the example for the kinetic spherical model in Section 5. In a physically more appealing
way, time-dependent potentials arise when a many-body system is brought into contact
with a heat bath whose temperature T (t) is time-dependent [78,79]. In this paper, we shall
be interested in the dynamics symmetries of Langevin equations derived from a free-field
Hamiltonian (1.11). In particular, we shall compare the situation without (i.e., T = 0) and
with thermal noise (i.e., T > 0). For simplicity, we shall refer to all Eqs. (1.8) obtained
from the Hamiltonian (1.11) as ‘free Schrödinger equations’.
The study of the dynamic symmetries of such free-field theories will yield useful
insights which we expect to extend to physically more realistic interacting field theories
where H would also contain higher than merely quadratic terms. If we identify D −1 = 2im,
it is clear that the order parameter φ is given by a noisy Schrödinger equation in an external
time-dependent potential v(t).
This paper is organized as follows. In Section 2, we first review the basics of
Schrödinger invariance in the absence of thermal noise and without initial correlations and
then show that through a gauge-transformation involving v(t), the entire phenomenology
of ageing and in particular (1.7) can be reproduced. We also consider the selection rules
which follow from Galilei invariance. In Section 3, after having reformulated the problem
in terms of the field-theoretic Martin–Siggia–Rose formalism, we study the effects of
thermal noise and/or initial correlations on free-field theory given by a Hamiltonian (1.11).
In Section 4, these results are extended to any field theory with for T = 0 and a(r) = 0
is Galilei-invariant. We find that the two-time response function R is independent of
both T and a(r) and obtain a new reduction formula (4.9) which relates C to certain
three- and four-point response functions to be evaluated in the noiseless theory and
discuss the scaling of the resulting two-time autocorrelation function. In Sections 5–7,
these results are tested in several exactly solvable systems (with an underlying free-field
theory) undergoing ageing with z = 2, namely, the kinetic spherical model, the XY-model
in spin-wave approximation, the critical voter model and the free random walk. We
conclude in Section 8. Appendix A deals with technical aspects of Gaussian integration and
Appendix B analyses a special four-point response function. In Appendix C we consider a
generalized realization of local scale-invariance and its application to the 1D Glauber–Ising

model
2. Local scale-invariance: a reminder
2.1. Schrödinger invariance
We begin by reviewing the kinematic symmetries of Schrödinger equations with a time-

dependent potential, but without a noise term. A long time ago, Niederer [71] obtained the
maximal kinematic symmetry group of the Schrödinger equation for an arbitrary potential
v = v(t, r) and he also gave a few examples where that group is isomorphic to the maximal
kinematic group Sch(d) of the free Schrödinger equation [63,70]. The group Sch(d) is
called the Schrödinger group [70].
We recall the definition of Sch(d). On the time and space coordinates (t, r) it acts as
(t, r) → (t
, r
) = g(t, r) where
αt + β Rr + vt + a
t −→ t
= , r −→ r
= , αδ − βγ = 1, (2.1)
γt +δ γt +δ
where R is a rotation matrix. The action of Sch(d) on the space of solutions φ of the free
Schrödinger equation is projective, that is, the wave function φ = φ(t, r) transforms into

φ(t, r) −→ (Tg φ)(t, r) = fg g −1 (t, r) φ g −1 (t, r) (2.2)
and the companion function fg is explicitly known [70,71]. The projective unitary
irreducible representations of Sch(d) are classified [76]. We now carry this over to field
theory and consider fields transforming according to (2.2). By analogy with an analogous
terminology in conformal field theory [3], a field φ transforming according to (2.2) and
with g(t, r) given by (2.1) is called quasiprimary [46]. Schrödinger invariance is the z = 2
special case of local scale-invariance, with time-translations added.
Besides the examples given in [71], there exist further noiseless Schrödinger equations
with a maximal kinematic group isomorphic to Sch(d). Consider the noiseless Langevin
equation
∂φ(t, r)
D −1 = φ(t, r) − v(t)φ(t, r) (2.3)
∂t
where D is the diffusion constant. This equation can be reduced to the free Schrödinger
equation D −1 ∂t Ψ (t, r) = Ψ (t, r) through the gauge transformation
t
φ(t, r) = Ψ (t, r) exp −D du v(u) . (2.4)
0
Since the kinematic symmetries of the free Schrödinger equation are well understood and
the realization of the Schrödinger group is explicitly known, the corresponding realization
for the case at hand, similar to (2.2), readily follows.
(0)
It turns out that the only change occurs in the companion function fg . Let fg stand
for the companion function of the free Schrödinger equation, then because of the gauge
transformation Eq. (2.4) we find
t

fg (t, r) = fg(0) (t
, r) exp −D du v(u) , (2.5)
t
where t
= t
(t) has been defined in Eq. (2.1). The generators of the Lie algebra sch1 of
this realization of Sch(1), appropriate for Eq. (2.3) with the potential v(t), read
v(t)
X−1 = −∂t + time drift,
2M
1 x v(t)
X0 = −t∂t − r∂r − + t dilatation,
2 2 2M
v(t) M 2
X1 = −t 2 ∂t − tr∂r − xt + t 2 − r special Schrödinger transformation,
2M 2
Y−1/2 = −∂r space translation,
Y1/2 = −t∂r − Mr Galilei transformation,
M0 = −M phase shift, (2.6)
where we expressed the diffusion constant D −1 = 2M as the ‘mass’ M and x denotes the
scaling dimension of the wave function φ(t, r). Of course, x = d/2 for a solution of the
free Schrödinger equation, but it will be useful to consider arbitrary values of x as well.
The non-vanishing commutators of the Lie algebra sch1 spanned by the generators (2.6)
are

n
[Xn , Xn
] = (n − n
)Xn+n
, [Xn , Ym ] = − m Yn+m ,
2
[Y1/2 , Y−1/2 ] = M0 , (2.7)
where n, n
∈ {±1, 0}, m ∈ {± 12 } and with straightforward extensions to d > 1, see [46].
2.2. Galilei covariance of correlators
When discussing the dynamic symmetries of a time-dependent statistical system, the

requirement of Galilei invariance plays a particular role. Indeed, for a system with local
interactions and which is invariant under space translations, scale transformations with a
dynamical exponent z = 2 and, in addition, Galilei-invariant, it can be shown that there
exists a Ward identity such that the system is also invariant under the ‘special’ Schrödinger
transformation (2.6) [42,43,48].
We shall be particularly interested in the two-point correlator C and the linear response
function R built from the order parameter φ(t, r). Using Martin–Siggia–Rose theory (MSR
theory) which we shall briefly review in Section 3, these may be expressed in terms of φ
and the so-called response field φ̃ as follows

C(t, s; r, r
) := φ(t, r)φ(s, r
) ,

δφ(t, r)
R(t, s; r, r
) :=
= φ(t, r)φ̃(s, r
) ,
δh(s, r ) h=0

C̃(t, s; r, r ) := φ̃(t, r)φ̃(s, r
) ,

(2.8)
where h is the magnetic field conjugate to the order parameter φ. Later, we shall often refer
to t as the observation time and to s as the waiting time.
Generalizing the above definition to v(t) = 0, we say that a field φ is quasiprimary
if its infinitesimal change under sch1 is given by the generators (2.6) with v(t) = 0.
A quasiprimary field φ is characterized by its scaling dimension x and its ‘mass’ M 0. In
turn, if the response field φ̃ associated to φ is also quasiprimary, it has a scaling dimension
denoted by x̃ and the ‘mass’
M̃ = −M 0. (2.9)
This important fact will be used later on. The argument leading to the result (2.9) was
discussed in detail in [48] and will not be repeated here.
If both φ and φ̃ transform as quasiprimary fields of a Schrödinger-invariant theory,
the generators (2.6) can be used to derive restrictions on the form of any multipoint
correlator and in particular determine the two-point functions completely. If Xi is any
of the generators of schd acting on the ith particle in a n-point correlator A{ti , r i } where
i = 1, . . . , n (see (2.8) for n = 2), we have a set of differential equations
(X1 + · · · + Xn )A{ti , r i } = 0. (2.10)

If rotation invariance can be assumed (and we shall implicitly do so throughout this
paper), for the calculation of the two-point functions it is enough to consider the one-
dimensional case and use the generators of Eq. (2.6). Then a straightforward calculation
[80] gives, provided the ‘mass’ of the order parameter is positive M > 0, see, e.g., [43,46]
C̃0 (t, s; r, r
) = 0, C0 (t, s; r, r
) = 0, (2.11)
whereas the response function is basically the gauge-transformed expression of the well-
known zero-potential Gaussian response R, i.e.,
k(t)
R0 (t, s; r, r
) = R(t, s; r, r
),
k(s)

−x M (r − r
)2
R(t, s; r, r ) = δx,x̃ r0 Θ(t − s)(t − s) exp − ,
2 (t − s)
t
1
k(t) := exp − du v(u) , (2.12)
2M
where r0 is a normalization constant and Θ is the Heaviside function which expresses
causality. As they stand, Eqs. (2.11), (2.12) hold for T = 0 and we shall from now on use
the index 0 to remind the reader of this fact.
On the other hand, if the system is not rotation-invariant, we can repeat the same
argument in any fixed direction of space and the non-universal constant M becomes
direction-dependent. Indeed, rotation invariance is broken for phase-ordering systems
defined on a lattice [86,87] for sufficiently small temperatures. Even then, local scale-
invariance still holds in every single space direction, as exemplified in the 2D and 3D Ising
models with Glauber dynamics [49].
A few comments are in order.
1. Eqs. (2.11), (2.12) provide a manifest example of the superselection rule of Galilei
invariance, also known as Bargman superselection rule [2]. Explicitly, if Φi (ti , r i ) are
Galilei-covariant fields, each with a ‘mass’ Mi , Galilei covariance implies [2,43]

Φ1 (t1 , r 1 ) · · · Φn (tn , r n ) = δM1 +···+Mn ,0 F {ti , r i } (2.13)
By physical convention, the ‘masses’ of the fields φ are non-negative, viz. Mi 0.

Furthermore, the response fields φ̃ have negative ‘masses’ M̃i 0 and the result
(2.11) follows.
2. In the introduction, we reviewed the result (1.7) for the autoresponse function, derived
for arbitrary z from a generalization of Schrödinger invariance [45,46]. For z = 2,
Eq. (1.7) coincides with our result (2.12) provided that
1 + a − λR /2
x = x̃ = 1 + a, v(t) = (2M) . (2.14)
t
3. However, there is an important difference in our derivation of the scaling form of

R0 (t, s) with respect to [45,46]: because time-translation invariance is broken in
ageing phenomena, covariance of R0 was required to hold merely under a subalgebra
of schd where the time-translations were excluded. Indeed, in [48], such subalgebras
were studied systematically and a relationship with the parabolic subalgebras of
the (complexified) conformal algebra confd+2 was found. On the other hand, the
realization of schd used in [45,46,48] applies to a vanishing potential v(t) = 0.
Here, we do not follow that point of view. We consider the more general realization
of the entire algebra sch1 with a time-dependent potential v(t) and require that both φ
and φ̃ transform as quasiprimary fields under the whole set of generators (2.6).
That these two different approaches, given the conditions (2.14), yield the same
phenomenology of the scaling of the two-time autoresponse function is our first result
and will be crucial for the developments to follow.
4. At first sight, the result (2.11) that the two-time autocorrelator C(t, s) = φ(t, r) ×
φ(s, r) = 0 may appear strange and, indeed, does not hold true in concrete models. In
the next sections, we shall show that this apparent contradiction comes from the fact
that the noiseless Schrödinger equation does not take the thermal noise into account.
As we shall see, the reformulation of Schrödinger invariance in ageing systems in
terms of a noisy Schrödinger equation with a time-dependent potential allows to arrive
at physically meaningful predictions for correlation functions. Explicit confirmations
in exactly soluble models will be presented.
3. Response and correlation functions for non-interacting Gaussian theories
3.1. The Martin–Siggia–Rose formalism
It is useful to treat noisy Langevin equations in the context of the Martin–Siggia–Rose

(MSR) formalism [23,59,66,91,94]. In equilibrium, the integration over the Gaussian noise
η can be carried out by introducing a response field φ̃.4 It can be shown that the stochastic
Langevin equation (1.8) can be obtained from the following effective action Σ[φ, φ̃]

∂φ δH
Σ[φ, φ̃] = dt dr φ̃ +D
∂t δφ

1
− dt dr dt
dr
η(t, r)η(t
, r
) φ̃(t, r)φ̃(t
, r
). (3.1)
2

This action appears in the generating functional Z = Dφ Dφ̃ e−Σ[φ,φ̃] expressed as a path
integral. In this way, the original dynamical problem in d dimensions has been mapped
onto one of statistical mechanics in d + 1 dimensions.
As long as one is merely interested in equilibrium behaviour and provided the dynamics
is ergodic, there is no need to worry about initial conditions, which might be said to be
specified at a time t = −∞. Here, we are interested in how the equilibrium state is reached
from a given initial state and must include into the action a term describing the initial
preparation of the system. One has
S[φ, φ̃] = Σ[φ, φ̃] + σ [φ, φ̃] (3.2)
where Σ[φ, φ̃] (3.1) describes the ‘bulk’ evolution of the system as derived from the
Langevin equation while σ [φ, φ̃] describes the initial conditions at time t = 0. As already
pointed out by Mazenko [67], it may be written as

1
σ [φ, φ̃] = − dr dr
φ̃(0, r)a(r − r
)φ̃(0, r
), (3.3)
2
Rd ×Rd
where it is implicitly admitted that φ(0, r) = 0 and a(r) is the initial two-point correlator

a(r) := C(0, 0; r + r
, r
) = φ(0, r + r
)φ(0, r
) . (3.4)
From spatial translation-invariance, it follows that a(r) = a(−r) which we shall admit
throughout.
We call the theory described by the action S0 alone the noiseless theory.
For a free field, the noiseless and the thermal parts of the MSR action read (we also
have set D = 1)

∂φ
S0 [φ, φ̃] := dt dr φ̃ − φ + v(t)φ , S[φ, φ̃] := −T dt dr φ̃ 2 (t, r),
∂t
4 In the systematic terminology of [17], this should be rather called a response operator, because φ̃ will
become an operator in a canonical quantization scheme of the action. The notion of a response field should have
been reserved to the canonically conjugate variable of φ̃. Since we shall not use the operator formalism here, we
shall simply, but sloppily, talk about φ̃ as a response field.
Σ[φ, φ̃] = S0 [φ, φ̃] + S[φ, φ̃]. (3.5)

We shall refer to the contribution described by S and σ as the thermal and initial noise,
respectively.
We point out that field-theoretic studies of critical dynamics use a different initial term,
namely [58],

τ0 2
σc [φ, φ̃] = dr φ(0, r) − m(r) . (3.6)
2
Rd
This specifies an initial macroscopic state with spatially varying order parameter
φ(0, r) = m(r) and spatial correlations decaying on a finite scale proportional to τ0 −1 .
Ageing at criticality was studied in the O(n)-model using the ε-expansion with the initial
term σc [12–15,58]. However, the use of σc instead of σ for temperatures below criticality
would lead to contradictions.
Treating the noisy Langevin equation (1.8) as the classical equation of motion of the
field-theory (MSR) action Eqs. (3.2), (3.5), (3.3) has the following advantages.
• Thermal fluctuations and initial conditions are explicitly included. To emphasize this,
notice that the noisy contributions to the MSR action are

S[φ, φ̃] − S0 [φ, φ̃] = − du dr du
dr
φ̃(u, r)κ(u, u
; r − r
)φ̃(u
, r
),
1
κ(u, u
; r) = T δ(u − u
)δ(r) + δ(u)δ(u
)a(r), (3.7)
2
where κ includes both the effects of thermal and of initial-state fluctuations.
• The response field φ̃ describes the thermal noise as can be seen from the equations of
motion derived from the free-field MSR action (3.5)
∂φ(t, r)
= φ(t, r) − v(t)φ(t, r) + 2T φ̃(t, r),
∂t
∂ φ̃(t, r)
− = φ̃(t, r) − v(t)φ̃(t, r). (3.8)
∂t
The first equation (3.8) reduces to the Langevin equation (1.8) provided one makes the
formal identification
η(t, r) = 2T φ̃(t, r). (3.9)

Therefore, at the classical level, the stochastic Langevin equation Eq. (1.8) is described
by two deterministic equations. In our case they are both of Schrödinger type and with
opposite masses for the field φ and the response field φ̃.
• Averages of any n-point function built from the fields φ, φ̃ can be expressed in terms
of the functional integral

F φ(ti , r i ), φ̃(tj , r j )

= Dφ Dφ̃ exp −S[φ, φ̃] F φ(ti , r i ), φ̃(tj , r j ) , (3.10)
with the normalization 1 = 1.

For example, adding a magnetic perturbation δHmag = −hφ to the Hamiltonian (1.11)
and then computing the mean of the order parameter to first order in h, the relation
R(t, s; r, r
) = φ(t; r)φ̃(s, r
) of Eq. (2.8) is easily reproduced.
• The use of Martin–Siggia–Rose formalism makes the machinery of the field-theoretic
renormalization-group available [17,91,94] but we shall not pursue this here.
It will be useful to split the calculation of averages into two steps as follows:

F φ(ti , r i ), φ̃(tj , r j ) = F φ(ti , r i ), φ̃(tj , r j ) exp −S[φ, φ̃] − σ [φ, φ̃] 0 ,

F φ(ti , r i ), φ̃(tj , r j ) 0

= Dφ Dφ̃ exp −S0 [φ, φ̃] F φ(ti , r i ), φ̃(tj , r j ) , (3.11)
where the notation 0 (and more generally the index 0) refers from now on to averages of
the non-fluctuating theory. This allows to make use of the Schrödinger invariance of the
noiseless theory.
3.2. Analytical results for free fields
In this section, we find both response and correlation functions for free-field Martin–
Siggia–Rose theory, as given by Eqs. (3.2), (3.3), (3.5).
3.2.1. Two-point functions without noise

The free Martin–Siggia–Rose action S[φ, φ̃] has a Gaussian structure. We shall write it
as

S[φ, φ̃] = du dr du
dr
Φ(u; r)T Q(u, u
; r, r
)Φ(u
; r
), (3.12)
φ
where Φ = φ̃ is the two-component field built from φ and φ̃, and Φ T stands for its
transpose. The kernel Q reads
Q(u, u
; r, r
)

1 0 δ(u − u
)δ(r − r
)(− − ∂u ) + 12 v(u)
= .
2 δ(u − u
)δ(r − r
)(− + ∂u ) + 12 v(u) −2κ(u, u
; r − r
)
(3.13)
This peculiar form of the Lagrangian density is quite suggestive as regards the Galilei
invariance. When κ = 0, that is in absence of noise, the quadratic form Q is antidiagonal.
This is one way of presenting the Bargman superselection rules and leads, in particular, to
φφ = φ̃ φ̃ = 0 which is a manifestation of Galilei invariance of the noiseless system.
The presence of noise just breaks this symmetry.
In order to study systematically the role of the noise, we shall expand around the non-
fluctuating theory. The correlation functions C0 , C̃0 and the linear response function R0
are
C0 (t, s; r, r
) = 0, C̃0 (t, s; r, r
) = 0,

k(t) −d/2 1 (r − r
)2
R0 (t, s; r, r
) = Θ(t − s) 4π(t − s) exp − . (3.14)
k(s) 4 (t − s)
This follows since the quadratic form Q is antidiagonal the field φ can only be coupled
to φ̃ and the result is just the bare propagator of the theory. It is clear that only the
antidiagonality of Q is important to derive this result and the explicit free-field form (3.12)
is not required for (3.14) to hold.
These results fully agree with the Schrödinger invariance prediction Eqs. (2.12), (2.11)
with the identifications x = x̃ = d/2, M = 1/2 and r0 = (4π)−d/2 .
A further manifestation of the Galilei invariance of the noiseless theory is the fact that
φ · · · φ φ̃ · · · φ̃ 0 = 0 (3.15)

n m
unless n = m as is easily checked. This is a further example of the Bargman superselection
rule (2.13) and will be important in what follows.
3.2.2. Two-point functions in presence of noise

We now find the same two-point functions in the presence of noise.
We begin with the response functions R(t, s; r; r
) which is found from (3.14) by
averaging with the noiseless weight exp(−S0 [φ, φ̃])
R(t, s; r, r
)

= φ(t, r)φ̃(s, r ) exp du dR du dR φ̃(u, R)κ(u, u ; R − R )φ̃(u , r ) ,

0
(3.16)
where κ is given by Eq. (3.7). Formally expanding the exponential and taking the Bargman
superselection rule (3.15) into account, only the term of lowest order remains. We thus
find

k(t) −d/2 1 (r − r
)2
R(t, s; r, r
) = R0 (t, s; r, r
) = Θ(t − s) 4π(t − s) exp − ,
k(s) 4 (t − s)
(3.17)
and we see that R is independent of the noise.
Next, the order-parameter correlation function reads
C(t, s; r, r
)

= φ(t, r)φ(s, r ) exp du dR du dR φ̃(u, R)κ(u, u ; R − R )φ̃(u , r ) .

0
(3.18)
Again, expanding the exponential and using Eq. (3.15), a single term remains and we
readily find

C(t, s; r, r
) = dR du dR
du
κ(u, u
; R − R
)R0(4) (t, s, u, u
; r, r , R, R
),

R0 (t, s, u, u
; r, r , R, R
) := φ(t, r)φ(s, r
)φ̃(u, R)φ̃(u
, R
) 0 ,
(4)
(3.19)
(4)
where R0 is a noiseless four-point function.
Repeating the same arguments as before, it is an easy task to compute the two-point
correlations of response fields. They are
C̃(t, s; r, r
) = 0. (3.20)
We emphasize that Eqs. (3.16), (3.19), (3.20) will hold for any theory satisfying the
Bargman’s superselection rule (3.15). We shall come back to this in Section 4.
The four-point function R0(4) is simple to access for free fields since it factorizes into a
product of two-point functions because of Wick’s theorem. We have
R0 (t, s, u, u
; r, r , R, R
)
(4)
= R0 (t, u; r, R)R0 (s, u

; r , R
) + R0 (t, u
; r, R
)R0 (s, u; r , R). (3.21)
Together with the explicit form of κ, this yields the final result
C(t, s; r, r
) = Cth (t, s; r, r
) + Cpr (t, s; r, r
),

Cth (t, s; r, r ) = 2T du dy R0 (t, u; r, y)R0 (s, u; r
, y),

Cpr (t, s; r, r
) = dy dy
R0 (t, 0; r, y)a(y − y
)R0 (s, 0; r
, y
), (3.22)
where we separated C in a thermal term Cth and an initial term Cpr . We clearly see that
while the only contributions to C come from the noise, R does not depend on it.
We summarize the results obtained so far as follows.
1. It is satisfying that the well-known result C̃ = φ̃ φ̃ = 0 is naturally reproduced, see
(3.20).
2. The independence equation (3.17) of the two-time response function of T and of the
initial correlations goes beyond the usual scaling arguments as reviewed in Section 1.
This explains to some extent the success of the existing confirmations of that prediction
of local scale-invariance.
3. We arrive at an explicit expression for the two-time correlators, which are obtained in
terms of a contraction of two response functions. We also see that the earlier result
C = 0 comes from neglecting both initial and thermal fluctuations.
It is useful to present these results also in momentum space. Using spatial translation
invariance, define the Fourier transform of any two-point function A(t, s; x, y) =
A(t, s; x − y) as

Â(t, s; q) := dr A(t, s; r) exp(−iq · r). (3.23)
Rd
Then Eqs. (3.17), (3.22) become

k(t)
R̂(t, s; q) = exp −q 2 (t − s) Θ(t − s),
k(s)
s
k(t)k(s)
Ĉth (t, s; q) = 2T du 2 exp −q 2 (t + s − 2u) ,
k (u)
0
Ĉpr (t, s; q) = R̂0 (t, 0; q)â(q)R̂0 (s, 0; −q)
k(t)k(s)
= â(q) 2 exp −q 2 (t + s) , (3.24)
k (0)
where in the second line, the convention s < t has been used. In particular, if v(t) = 0 the
two-point correlation function takes an especially simple form Ĉ 0 (t, s; q) where

T T
Ĉ (t, s; q) = â(q) − 2 exp −q 2 (t + s) + 2 exp −q 2 (t − s) .
0
(3.25)
q q
While both response and correlation functions depend on both t and s and therefore
describe an ageing behaviour, there is an equilibrium regime 1 t − s s, t where we
have the following simple expressions:

R̂eq (t, s; q) = exp −q 2 (t − s) ,
T
0
Ĉeq (t, s; q) = 2 exp −q 2 (t − s) . (3.26)
q
More generally, it is not difficult to show that
T −q 2 (t −s) 2 −1
Ĉeq (t, s; q) = e 1 + O (sq ) , (t − s)/s .
q2
In any case, we recover the fluctuation–dissipation theorem T R̂eq (t, s; q) = ∂ Ĉeq (t, s; q)/
∂s in the equilibrium regime as it should be.
Motivated from studies in spin glasses, it is sometimes attempted to separate correlation
and response functions into an equilibrium and an ‘ageing’ part, viz. C = Ceq + Cage ,
R = Req + Rage . In our case, we would have

k(t)
R̂age (t, s; q) = − 1 exp −q 2 (t − s) ,
k(s)

T
0
Ĉage (t, s; q) = â(q) − 2 exp −q 2 (t + s) . (3.27)
q
For v(t) = 0, there is no ‘ageing’ part in the response function.
From these expressions, we can already extract a few general properties of the ageing
process. First, for systems quenched to below their critical temperature Tc > 0 and
described by a MSR Gaussian action (3.1), it is known from the dynamical renormalization
group that the final temperature T < Tc is an irrelevant parameter, and furthermore, T → 0
under renormalization [9,11]. Then the long-time dynamics should be driven by the initial
fluctuations, in agreement with Eq. (3.27) with T = 0. On the other hand, for a critical
quench T = Tc , the situation is different in that both initial and thermal fluctuation may
contribute to the long-time dynamics. From Eq. (3.27) we expect that the small-q behaviour
of the term a(q) − Tc /q 2 will determine the long-time dynamics.
4. Consequences of local scale-invariance of noiseless theories
4.1. MSR formulation
We now generalize the results of the previous section to any theory whose noiseless
MSR action is Schrödinger-invariant. As we shall stay within the context of classical field
theory, the symmetries of the MSR action are the same as for the corresponding Langevin
equation (1.8). The formulation of the MSR action, using a non-conserved order parameter
described by model A dynamics [53] is almost unchanged with respect to Section 3. We
have
S[φ, φ̃] = S0 [φ, φ̃] + S[φ, φ̃] + σ [φ, φ̃], (4.1)

where S[φ, φ̃] and σ [φ, φ̃] are given by Eqs. (3.3), (3.5) and

∂φ δH
S0 [φ, φ̃] = dr dt φ̃ + . (4.2)
∂t δφ
We shall assume throughout that the noiseless action S0 is Schrödinger-invariant and that
it includes an external time-dependent potential v = v(t). For a local effective potential H,
it is known that spatial translation invariance, dilatation invariance (or dynamical scaling)
and Galilei invariance are sufficient for having Schrödinger invariance [48]. Therefore,
from now on the dynamical exponent z = 2.
In order to study the effects of the noise, we first discuss what becomes of the Galilei
invariance of the noiseless theory. If the order parameter φ and the response field φ̃
are quasiprimary, they should transform under a Galilei transformation t → t
= t and
r → r
= r − vt as (see Eq. (2.5))
φ
(t
, r
) = fv (t, r)φ(t, r), φ̃
(t
, r
) = fv−1 (t, r)φ̃(t, r), (4.3)
where the companion function fv reads [70]

M 2
fv (t, r) = exp Mr · v − v t . (4.4)
2
The noisy contributions to the action transform as

S[φ
, φ̃
] − S[φ, φ̃] = −T dt dr φ̃ 2 (t, r) fv−2 (t, r) − 1 ,

1
σ [φ , φ̃ ] − σ [φ, φ̃] = − dr dr
a(r − r
)φ̃(0, r)φ̃(0, r
)
2

× fv−1 (0, r)fv−1 (0, r
) − 1 . (4.5)
Therefore, for fixed temperature and initial conditions, the noise always destroys Galilei
invariance.5 Thus, the only dynamic symmetries of the noisy Langevin equation should be
space translations and dilatations (and possibly space rotations).
Consequently, one would merely expect the following scaling forms

t (r − r
)2
C(t, s, r, r
) = s −b GC , ,
s t −s

t (r − r
)2
R(t, s; r, r
) = s −a−1 GR , , (4.6)
s t −s
where GC,R are undetermined scaling functions. For free-field theories, one would have
a = b = d/2 − 1 by dimensional counting.
4.2. Response and correlation functions for the fluctuating theory
The form of the scaling functions GR,C will now be determined through a generalization
of the expansions carried out in Section 3. The main results will be Eqs. (4.8) and (4.12),
(4.16), respectively.
4.2.1. The two-point response function

Consider a system which initially is at thermal equilibrium with temperature Ti (which
fixes the initial correlator a(r)) and quench it at time t = 0 to the final temperature T = Tf .
The response function is still given by Eq. (3.16) and the expansion of the exponential goes
through as before. Because of the Bargman superselection rule (3.15) which holds because
of the assumed Galilei invariance of S0 only the lowest term survives and we obtain
R(t, s; r, r
) = R0 (t, s; r, r
). (4.7)
The expression of R0 has been derived earlier and using the gauge transform (2.4) we
recover exactly the same result as in Eq. (2.12), namely,

k(t) M (r − r
)2
R(t, s; r, r
) = δx,x̃ r0 Θ(t − s)(t − s)−x exp − . (4.8)
k(s) 2 (t − s)
We stress that, given only Galilei and scale-invariance of the noiseless theory, this result
should hold for any initial and final temperature Ti and Tf . At this stage, nothing has yet
been said on the time-dependence of v(t).
5 The breaking of Galilei invariance through thermal noise can be visualized as follows: consider a system in
contact with a thermal bath at constant and uniform temperature T > 0. If the system moves with respect to the
bath with a constant speed v, the apparent temperature measured in the system will depend on the angle between
the direction of measurement and v.
4.2.2. The two-point correlation function

The two-point correlation function is found from Eq. (3.19). Again, the arguments of
Section 3 go through and we find, using the explicit expression (3.7) for the kernel κ

C(t, s; r, r
) = T du dR R0(3) (t, s, u; r, r
, R)

1
dR dR
a(R − R
)R0 (t, s, 0, 0; r, r , R, R
),
(4)
+ (4.9)
2
where R0(3) is the following three-point response function

R0 (t, s, u; r, r
, R) := φ(t; r)φ(s; r)φ̃(u; R)2 0
(3)
(4.10)
(4)
and R0 was already defined in Eq. (3.19). This central result will be the basis for all
what follows. Consequently, the calculation of C requires the computation of the noiseless
(3) (4)
three- and four-point functions R0 and R0 . This cannot entirely be done, since a
general expression for R0(4) is not yet available. Of course, one might hope that through
an extension of the Schrödinger algebra schd to some infinite-dimensional Lie algebra
techniques analogous to 2D conformal invariance [3] might become applicable but the
formulation of just such an extension is an open problem.
Here, we shall restrict to the case of vanishing initial correlations, that is
a(R) = a0 δ(R), (4.11)
which corresponds to an infinite initial temperature Ti = ∞ and where a0 is a normaliza-
tion constant. In fact, from renormalization group arguments the long-time behaviour of
any system which is prepared in the high-temperature or paramagnetic phase should be
described by this initial condition [9,11]. Then

1
C(t, s) = T du dR R0(3) (t, s, u; R) + dR R0(3) (t, s, 0; R), (4.12)
2
(3) (3)
R0 (t, s, u; r) := R0 (t, s, u; y, y, r + y) = φ(t; y)φ(s; y)φ̃(u; r + y)2 0 . (4.13)
Here, the field φ̃ 2 is a composite field with mass −2M and scaling dimension 2x̃2 . Only
for free fields, one has x̃2 = x̃.
Now, the three-point function of a Schrödinger-invariant theory with v(t) = 0 is well
known since a long time [43,46]. Denoting by R the response function in the case where
v(t) = 0, we have

(3) (3) M t + s − 2u 2 t −s
R0 (t, s, u; r) = R0 (t, s, u) exp − r Ψ 2
r ,
2 (s − u)(t − u) (t − u)(s − u)
−x̃2
R(3)
0 (t, s, u) = Θ(t − u)Θ(s − u)(t − u) (s − u)−x̃2 (t − s)−x+x̃2 , (4.14)
where Ψ is an arbitrary scaling function and M is the ‘mass’ of the field φ. This result is
brought to the case at hand through the gauge transformation (2.4) and we find
(3) k(t)k(s) (3)
R0 (t, s, u; r) = R0 (t, s, u; r). (4.15)
k 2 (u)
Combining (4.12) and (4.15) we obtain, for the first time, a generic prediction for the form
of the two-point correlation function. It is remarkable that under the condition (4.11) only
the noiseless three-point response functions are required in order to predict any two-time
autocorrelator.
It is useful to write down the autocorrelation function in the form C(t, s) = Cth (t, s) +
Cpr (t, s). Here the thermal part Cth and the preparation part Cpr are given by
x̃2 −x−d/2

t
Cth (t, s) = T s d/2+1−x−x̃2 −1
s
1
d/2−x̃2
k(t)k(s) t t/s + 1 − 2θ
× dθ 2 − θ (1 − θ ) Φ ,
k (sθ ) s t/s − 1
0
x̃2 −x−d/2
a0 k(t)k(s) d/2−x̃2−x t d/2−x̃ t t/s + 1
Cpr (t, s) = s −1 Φ ,
2 k 2 (0) s s t/s − 1

Mw 2 2
Φ(w) := dR exp − R Ψ R , (4.16)
2
and we have explicitly used s < t.
As they stand, the above expressions for C(t, s) do not yet necessarily describe a
dynamical scaling behaviour, since the form of k(t) is still completely general. We now
assume, in addition, that we are dealing with a system with dynamical scaling. In order
to reproduce the usual phenomenology of ageing systems, we must have, at least for
sufficiently large times (see Section 2)
k(t) k0 t . (4.17)
Comparing the general form of R(t, s) as given in Eq. (4.8) with the phenomenologically
expected scaling Eqs. (1.2), (1.3), we read off
λR
x = 1 + a, =1+a − , (4.18)
2
and in particular, the scaling function (1.7) is recovered. On the other hand, for C(t, s)
we find the following scaling form, written down separately for the thermal and the initial
term, where y = t/s 1 is the scaling variable
Cth (t, s) = T s −bth fCth (y), Cpr (t, s) = s −bpr fC (y),

pr
(4.19)
pr
where the scaling functions fC and fCth are given by
1
d/2−x̃2 y + 1 − 2θ
fCth (y) = y (y − 1) x̃2 −x−d/2
dθ θ −2
(y − θ )(1 − θ ) Φ ,
y −1
0

pr a0 d/2−x̃2+ x̃2 −x−d/2 y +1
fC (y) = y (y − 1) Φ , (4.20)
2 y −1
with the non-equilibrium exponents

bth = x + x̃2 − 1 − d/2, bpr = x + x̃2 − 2 − d/2. (4.21)
Provided that Φ(1) is finite, the asymptotic behaviour of the scaling functions for y large
can be worked out. We expect fC (y) ∼ y −λC /2 and find
pr
C = λC = 2(x − ).
λth (4.22)
Therefore, comparing this with (4.18), we have shown: For any system with an infinite-
temperature initial state (4.11), quenched to a temperature T < Tc and whose noiseless
part is locally scale-invariant with z = 2, one has
λC = λR . (4.23)
For non-equilibrium critical dynamics (that is T = Tc ) the same conclusion can be drawn
if after renormalization one still has z = 2. While Eq. (4.23) certainly agrees with the
evidence available from models studied either analytically or numerically, we are not aware
of any other general proof of this equality between the autocorrelation and autoresponse
exponents for a fully disordered initial state.
In Sections 5–7, we shall present extensive tests of the prediction (4.19), (4.20), (4.21)
for C and (4.8), (4.18) for R, respectively. The main hypothesis going into it is the
requirement of Galilei invariance of the noiseless theory, while the other conditions appear
to be habitually admitted in the description of physical ageing.
4.2.3. Autocorrelation function in phase-ordering kinetics

In order to understand the result (4.19) for C(t, s) better, we now study the two
contributions separately. First, we consider the ‘preparation’ part Cpr . This term is expected
to describe the late-time behaviour of a system quenched to a temperature T < Tc . Indeed,
renormalization-group arguments show [9] that in this case T is an irrelevant variable and
is renormalized towards zero. Then the thermal contribution Cth vanishes. Therefore, the
non-equilibrium exponent b = bpr is read off from Eq. (4.21). In addition, we know that
b = 0 in the low-temperature phase. This implies
d
x̃2 − x = − λC 0, (4.24)
2
because of a well-known inequality [93]. Only for free fields, one has λC = d/2, otherwise
x̃2 is a new nontrivial exponent. In the scaling limit, we thus have C(t, s) = fC (t/s) where

a0 y+1
fC (y) = y λC /2 (y − 1)−λC Φ . (4.25)
2 y−1
The form of fC (y) still depends on the unknown function Φ(w) which in turn depends
on Ψ (ρ), see (4.16). We attempt to fix its form and reconsider the noiseless response
(3)
function R0 (t, s, 0; r) which describes a response of the autocorrelation C(t, s) =
φ(t)φ(s). It appears to be reasonable requirement that there should be no singularity
in R0(3) when t = s. Using the explicit form Eqs. (4.14), (4.15) this leads to the following
limit behaviour
Ψ (ρ) Ψ0 ρ λC −d/2 , ρ → 0, (4.26)
where Ψ0 is a constant. If this leading term should be still accurate for larger values of ρ,
the following expression for the scaling function Φ(w) is found

(λC ) 2 λC −λC
Φ(w) ≈ Ψ0 Sd w =: Φ0 w−λC , (4.27)
2 M
where Sd is the surface of the unit sphere in d dimensions. Eq. (4.27) should hold in the
w → ∞ limit. Provided this from is still valid for all w, we would obtain the following
simplified form, with y = t/s
2 −λC /2
a0 Φ0 (y + 1)2 −λC /2 2 (y + 1)
C(t, s) ≈ = Meq , T = 0, (4.28)
2 y 4y
2 ,
where we also related the leading constant to the squared equilibrium magnetization Meq
2 f (t/s) with f (1) = 1, see [39].
in order to recover the usual scaling form C(t, s) = Meq C C
4.2.4. Autocorrelation function for critical dynamics

Second, let us turn to the thermal part. It should dominate the autocorrelation for
quenches to the critical point T = Tc , given the initial condition (4.11) and under the
assumption that z = 2 even at criticality. From Eq. (4.21), the exponent b = bth can be
read off and we have
d
x̃2 = b − a + . (4.29)
2
Under the stated conditions, the preparation term drops out at large times and we find
C(t, s) = Tc s −b fC (t/s),
fC (y) = y (y − 1)b−2a−1
1
a−b y + 1 − 2θ
× dθ θ −2 (y − θ )(1 − θ ) Φ . (4.30)
y −1
0
At criticality, one expects the following relationship between the nonequilibrium
exponents a and b
2β d −2+η
a=b= = , (4.31)
νz z
where β, ν, η are well-known equilibrium critical exponents (see Section 1). To understand
Eq. (4.31), recall that C(s, s) ∼ s −b . On the other hand, from the space–time scaling
|r|z ∼ t, one expects the equilibrium correlator to decay as Ceq ∼ |r|−bz and the second
equality in Eq. (4.31) follows. Finally, a = b is a necessary condition for having a non-
vanishing limit fluctuation–dissipation ratio X∞ .
Now, if we let a = b in (4.30), we find
1
−1−a −2 y + 1 − 2θ
fC (y) = y (y − 1) dθ θ Φ , (4.32)
y −1
0
which is the most general form compatible with the standard phenomenological con-
straints. An approximate form of the scaling function Φ(w) may be obtained from the
(3)
requirement that the three-point response function R0 (t, s, u; r) is non-singular for t = s.
We now have x̃2 = d/2 and find Ψ (ρ) Ψ0,c ρ x−d/2 as ρ → 0. This leads to, for w → ∞
Φ(w) ≈ Φ0,c w−1−a , (4.33)
where Φ0,c is a constant. If, in addition, we may use this form also for finite values of w,
we would obtain the simplified form
1
fC (y) ≈ Φ0,c y 1+a−λC /2
dθ θ λC −2−2a (y + 1 − 2θ )−1−a , T = Tc . (4.34)
0
Summarizing, the phenomenological comparison of the autocorrelation function, as

predicted by Schrödinger invariance, and assuming a totally disordered initial state, with
simulational or experimental data will be based on Eqs. (4.25) and (4.32) for quenches to
T < Tc and T = Tc , respectively. In full generality this will allow to obtain information on
the scaling function Φ(w). If, in addition, the heuristic idea of the absence of singularities
(3)
at t = s in the three-point response function R0 should be sufficient to fix the form of this
response function, the simplified forms (4.28), (4.34) apply and the scaling function fC (y)
is completely specified in terms of the exponents a and λC .
5. Tests of local scale-invariance in exactly solvable models
In this and the next two sections we describe phenomenological tests of the predictions
of Schrödinger invariance, that is local scale-invariance with z = 2, which were derived in
Section 4 in concrete physical models of ageing behaviour.
5.1. Kinetic spherical model
The kinetic spherical model is often formulated in a field-theoretic fashion as the

n → ∞ limit of the O(n)-symmetric vector model. For our purposes, it is more useful
to start directly from a lattice system and to take the continuum limit later, following
[22,38,75,77,78,95].
Consider a hypercubic lattice with N sites. At each site r there is a real time-dependent
variable φ(t, r) such that the mean spherical constraint

φ(t, r)2 = N (5.1)
r

holds. The Hamiltonian is H = − r,r
φ(t, r)φ(t, r
) where the sum extends over pairs
of nearest neighbour sites. The (non-conserved) dynamics is given in terms of a Langevin
equation of the type Eq. (1.8)

∂φ(t, r)
= φ(t, n) − 2d + v(t) φ(t, r) + η(t, r) (5.2)
∂t
n(r)
φ(t, r) − v(t)φ(t, r) + η(t, r), (5.3)
where n(r) runs over the nearest neighbours of the site r. In the second line, a formal
continuum limit was taken (for simplicity, all rescalings with powers of the lattice constant
a were suppressed). Finally, η(t, r) is the usual Gaussian noise (see Section 1) and the
Lagrange multiplier v(t) is fixed such that the mean spherical constraint is satisfied.6
Therefore, the kinetic spherical model perfectly fits into the context of a Schrödinger
equation in a time-dependent potential v = v(t) discussed in Section 3. We therefore expect
that the free-field predictions Eqs. (3.17), (3.22) will be fully confirmed.
In order to see this explicitly, we recall the elements of the exact solution of the Langevin
equation (5.2). If we set
t
g(t) = exp 2 du v(u) (5.4)
0
it can be shown [22,38,77] that g(t) is the unique solution of the Volterra integral equation
t
g(t) = A(t) + 2T dt
f (t − t
)g(t
), (5.5)
0
where g(0) = 1 and

d 1
f (t) = Θ(t) e−4t I0 (4t) , A(t) = dq a(r)e−2ω(q)t −iq·r , (5.6)
(2π)d r
B
d
and ω(q) = 2 i=1 (1 − cos qi ) is the lattice dispersion relation, B the Brillouin zone while
I0 is a modified Bessel function. We stress that because of Eq. (5.5), g(t) does not depend
on the order parameter φ(t, r).
We now show that the prediction (3.17), (3.22) for the two-point functions can be fully
reproduced. We begin with the response function. In the spherical model, it is exactly given
by [38,77]
d
δφ(t, r) g(s)

−2(t −s)
R(t, s; r, r ) =
= e Ir i −r
i 2(t − s)
δh(s, r ) h=0 g(t)
i=1
t
−d/2 (r − r
)2
4π(t − s) exp − exp − du v(u)
4(t − s)
s
× Θ(t − s) (5.7)
6 A careful study shows that provided the limit N → ∞ is taken before any long-time limit, the mean spherical
constraint (5.1) and a full, non-averaged, spherical constraint lead to the same results [29].
where in the second line the limit t − s 1 was taken and Ir is a modified Bessel function.
In particular, this reproduces the known autoresponse function [38]
!
R(t, s) = R(t, s; r, r) = f (t − s)/2 g(s)/g(t). (5.8)
This is in exact agreement with Eq. (3.17) and we identify the mass M = 1/2.
We now turn to the correlator but the sake of brevity merely deal with the autocorrelator
explicitly. In the spherical model one has (with t > s) [38,75,77]

C(t, s) = φ(t, r)φ(s, r)

s
1 t +s t +s
=√ A + 2T du f − u g(u) . (5.9)
g(t)g(s) 2 2
0
In order to show how to recover this explicitly from Eq. (3.22), we write again C(t, s) =:
Cpr + Cth and discuss the two terms separately. The first one is, where we use the explicit
form (5.7) of R = R0

g(0) g(0) (r − y)2 (r − y
)2
Cpr = dy dy
(4πt 4πs)−d/2 exp − −
g(t) g(s) 4t 4s
R2d

(2π)−d/2 2 t −q
2 s+i(q+q
)r
=√ dq dq
e−q Jd , (5.10)
g(t)g(s)
R2d
where

Jd := dy dy
a(y − y
)e−iq·y−iq ·y , (5.11)
R2d
and we used g(0) = 1. In order to calculate Jd , we set ζ = y − y

, ζ
= y + y
. Then the
Jacobian |∂(y, y
)/∂(ζ , ζ
)| = 2−d and we have

ζ

ζ
Jd = 2−d dζ dζ
a(ζ )e−i 2 ·(q+q ) e−i 2 ·(q−q )
R2d

= (2π)d δ(q + q
) dζ a(ζ )e−iq·ζ . (5.12)
Rd
We finally obtain

1 1
dq e−q
2 (t +s)
Cpr = √ dζ a(ζ )e−iq·ζ . (5.13)
g(t)g(s) (2π)d
Rd Rd
When the support of a(ζ ) is restricted to the hypercubic lattice, we therefore have, indeed,
for long times, t + s 1
A((t + s)/2)
Cpr √ , (5.14)
g(t)g(s)
in agreement with the first term in (5.9). The second term in (3.22) is analyzed as follows,
for t > s
∞
−d/2 g(u) g(u)
Cth = 2T du dy 4π(t − u)4π(s − u)
g(t) g(s)
0 Rd

(r − y)2 (r − y)2
× exp − − Θ(t − u)Θ(s − u)
4(t − u) 4(s − u)
s

2
2

g(u)
dq dq
dy e−q t −q s+(q +q )u+i(q+q )·r e−i(q+q )·y √
2 2
= 2T du
g(t)g(s)
0 R3d
s
2T −d
dq e−2q
2 ((t +s)/2−u)
=√ du g(u)(2π)
g(t)g(s)
0 Rd
s
2T t +s
√ du g(u)f −u , (5.15)
g(t)g(s) 2
0
where the last line holds for sufficiently large arguments of the function f . Taken together
with Cpr , the expected agreement between the general result Eq. (3.22) and the exact
expression (5.9) for the spherical model is thus recovered. The full space–time correlator
can be checked similarly.
It is interesting to note that for the confirmation of R, see (5.7), we need the condition
t − s 1, while for the confirmation of C, we also need s 1 (which implies t + s 1).
So far, we have not yet used the explicit form of g(t) which follows from Eq. (5.5).
Indeed, it is well known that for long times, one has [38,46,77]
g(t) ∼ t −2 . (5.16)

For the fully disordered initial conditions (4.11), the exponent takes the following values:
(i) for T < Tc , one has = d/4 and (ii) for T = Tc , one has = 1 − d/4 if 2 < d < 4 and
= 0 if d > 4. This is exactly the form expected from a potential of the form v(t) /t,
see Eq. (2.14).
It is instructive to compare also the explicit result for the two-point functions with the
expectations coming from local scale-invariance. For the two-time response function, this
confirmation has already been carried out for both the autoresponse function R(t, s) as well
as the space–time response R(t, s; r) [45,46,77] and need not be repeated here. Therefore,
we concentrate on the two-time autocorrelation function C(t, s). First, we consider the
case T < Tc where from the exact solution it is well known that [38,69]

(y + 1)2 −d/4
C(t, s) = Meq fC (t/s),
2
fC (y) = , (5.17)
4y
and we read off from the y → ∞ limit the exponent λC = d/2. Clearly, this exact result is
in full agreement with the prediction (4.28) of local scale-invariance. Second, we consider
the case T = Tc . Then the exact solution gives [38]

C(t, s) = (4π)−d/2Tc s −d/2+1 fC (t/s),
"
4
(y − 1)−d/2+1y 1−d/4(y + 1)−1 if 2 < d < 4,
fC (y) = d−2 −d/2+1 − (y + 1)−d/2+1
(5.18)
2
d−2 (y − 1) if d > 4.
From this, we read off the exponents a = b = d/2 − 1 for all d > 2, and furthermore,
λC = 3d/2 − 2 for 2 < d < 4 and λC = d for d > 4, respectively. show that Eq. (5.18)
is, indeed, reproduced. Inserting the exponent values into (4.34) it is straightforward to
In particular, we see that in the spherical model the asymptotic forms Eqs. (4.27), (4.33),
respectively, are, indeed, exact as should be expected for a free-field theory.
In conclusion, the exact results of the kinetic spherical model for both two-time
correlation and response functions are in full agreement with the predictions of Schrödinger
invariance.
5.2. XY model in spin-wave approximation
Another system which can be exactly analysed is the kinetic XY model in spin-
wave approximation. As we shall see, it provides an instructive example on the correct
identification of the quasi-primary scaling fields in a given model.
5.2.1. Formulation and observables

The XY model describes the interaction between planar spin variables

cos φ(r)
S(r) = cos φ(r) e1 + sin φ(r) e2 = , (5.19)
sin φ(r)
which are attached to the sites r of a d-dimensional hypercubic lattice and φ(r) is the
phase. The Hamiltonian is

H[φ] = − S(r) · S(r ) = − cos φ(r) − φ(r ) , (5.20)
r,r r,r
where the coupling constant J has been set to unity and the sum runs over nearest
neighbours. The relaxational dynamics is assumed to be described by a Langevin equation.
We prepare the system initially a temperature Ti and quench it at time t = 0 to the final
temperature T = Tf , so that the angular variable obeys [9]
∂φ(t, r) δH[φ]
=− + η(t, r), (5.21)
∂t δφ(t, r)
where η represents an uncorrelated Gaussian noise with zero mean and variance

η(t, r)η(t
, r
) = 2Tf δ(t − t
)δ(r − r
). (5.22)
Here, we shall exclusively study the coarsening dynamics in the low-temperature
regime, that is
Ti , Tf Tc (d), (5.23)
where Tc (d) is the critical temperature of the XY model in d dimensions (if d = 2,

Tc (2) = TKT is the Kosterlitz–Thouless temperature of the transition). Then the so-called
spin-wave approximation [4,84] can be used which amounts to expand H in powers of
φ(r) − φ(r
). Shifting the energy by a constant, the Hamiltonian reads to lowest order

1 2
H[φ] = dr ∇φ(r) . (5.24)
2
In writing this, we have implicitly absorbed the spin-wave stiffness [61,84] into a
redefinition of the temperatures. In 2D, it is known that below TKT , any vortices present
will be tightly bound and for distances larger than the characteristic pair size, the XY model
renormalizes to the Hamiltonian (5.24) [84].7
We are interested in the properties of the two-point functions. It appears natural to define
two-time correlation and linear response functions in terms of the magnetic variables

Γ (t, s; r, r
) := S(t, r) · S(s, r
) = cos φ(t, r) − φ(s, r
) ,
∂S(t, r)
ρ(t, s; r, r
) := lim , (5.25)
h→0 ∂h(s, r
)

where the response is found by adding a term δHmag = r h · S to H. Alternatively, one
may also consider the analogous quantities defined for the angular variables,
∂φ(t, r)
C(t, s; r, r
) := φ(t, r)φ(s, r
) , R(t, s; r, r
) := lim , (5.26)
h→0 ∂h(s, r
)

where a perturbation δHang = r hφ should have been added.
In order that the spin-wave approximation be applicable, we must start from an (almost)
ordered initial state of the system. Therefore, we require the following initial value for the
magnetic correlator, which reads in Fourier space
2πη(Ti ) Ti
â(q) = Ĉ(0, 0; q) = = 2, (5.27)
q2 q
where η(Ti ) is the standard equilibrium critical exponent and the relation 2πη(Ti ) = Ti
valid in the spin-wave approximation was used, see [4,61,84].
5.2.2. Non-equilibrium statistical field theory

As before, we introduce a Martin–Siggia–Rose formalism which characterizes the
system is term of an action S[φ, φ̃, h] depending on the phase field φ and an associated
response field φ̃ and we also include a (possibly space-dependent) magnetic field h =

i hi e i . We decompose the action
S[φ, φ̃] = Σ[φ, φ̃, h] + σ [φ, φ̃] (5.28)
7 For quenches from above T

KT in 2D, vortex configurations also become important and this leads to
logarithmic scaling, see [83] for details.
into a bulk term Σ[φ, φ̃, h] and an initial term σ [φ, φ̃]. These two terms include the
thermal and the initial noise. Explicitly

∂φ
Σ[φ, φ̃, h] = dt dr φ̃ − φ + sin(φ)h1 − cos(φ)h2
∂t

− T dt dr φ̃(t, r)φ̃(t, r) (5.29)
and

1
σ [φ, φ̃] = − dr dr
φ̃(0, r)a(r − r
)φ̃(0, r
), (5.30)
2
where the function a(r) describes the initial conditions according to Eq. (5.27).
We now simplify the general expressions for the two-point functions. There is nothing
to do for the angular correlation function C(t, s; r, r
) = φ(t, r)φ(s, r
) and we start with
the magnetic correlation function Γ which is given by

Γ (t, s; r, r
) = Dφ Dφ̃ cos φ(t, r) − φ(s, r
) exp −S[φ, φ̃] . (5.31)
For a vanishing magnetic field the bulk action Σ[φ, φ̃, 0] is a quadratic form in the fields
φ, φ̃ which are therefore Gaussian. Standard techniques explained in Appendix A lead to

1

2
Γ (t, s; r, r ) = exp − φ(t, r) − φ(s, r ) (5.32)
2

C(t, t; r, r) + C(s, s; r
, r
)
= exp C(t, s; r, r
) − , (5.33)
2
where in the second line the argument of the exponential was expanded. This gives Γ in
terms of angular correlators C.
Next, we consider the response functions. For the angular response R, we quote from
the MSR formalism the standard result

R(t, s; r, r
) = φ(t, r)φ̃(s, r
) . (5.34)
It remains to consider the response of the spin vector S at time t and position r to some
magnetic field h(s, r
) at time s and position r
. From the definition (5.25) we have
ρ(t, s; r, r
)

cos φ(t, r) − cos φ(t, r) sin φ(t, r) − sin φ(t, r)
= lim + , (5.35)
h→0 h1 (s, r
) h2 (s, r
)
where the average · is to be taken with a magnetic field. Expanding the action (5.28) to
first order in both components h1 , h2 of the magnetic field, we find

cos φ(t, r) = cos φ(t, r) + h1 (s, r
) cos φ(t, r) sin φ(s, r
)φ̃(s, r
) ,

sin φ(t, r) = sin φ(t, r) − h2 (s, r
) sin φ(t, r) cos φ(s, r
)φ̃(s, r
) . (5.36)
It follows that the response function can be expressed as

ρ(t, s; r, r
) = φ̃(s, r
) sin φ(t, r) − φ(s, r
) . (5.37)
Since both fields φ, φ̃ are Gaussian, it can further be shown that (see Appendix A)
ρ(t, s; r, r
) = R(t, s; r, r
)Γ (t, s; r, r
), (5.38)
and we see explicitly that the relationship between R and ρ is non-trivial. That no higher
correlators than the magnetic two-point correlation function Γ enter is a consequence of
the Gaussian nature of the theory at hand.
Eqs. (5.33), (5.34), (5.38) are the main results of this subsection. Before we can evaluate
them, we need some information on the validity of the spin-wave approximation.
5.2.3. Remarks on the validity of the spin-wave approximation

We need a criterion informing us up to what point the results on Γ and ρ derived in the
previous subsection within the spin-wave approximation should be reliable.
The correlation function C(t, s; r, r
) has already been obtained above and is given in
Eq. (3.25). For our choice (5.27) of initial conditions, its Fourier transform Ĉ(t, s; q) is
[21,84]
Ĉ(t, s; q) = (Ti − Tf )Ĝ(t + s; q) + Tf Ĝ(t − s; q), (5.39)
where Ĝ is given by
1
Ĝ(u; q) := 2
exp −q 2 (u + Λ2 ) , (5.40)
q
and we have explicitly introduced an UV-cutoff which simulates the lattice spacing (we
shall let Λ → 0 at the end). Therefore, a two-point correlation function φφ is of order
O(Ti , Tf ) whereas a response function φ φ̃ is of order O(1) in the initial and final
temperatures.
In order to discuss further the validity of the spin-wave approximation, we keep the
next-order term as well and consider the Hamiltonian

1 2 4
H[φ] = H0 + dr ∇φ(r) + g4 ∇φ(r) , (5.41)
2
where g4 is some constant. A straightforward calculation shows that to first order in g4 , the
correction to the spin-wave approximation of the two-point correlation function is given
by

3
δC(t, s; r, r ) = du dR φ(t, r)φ(s, r
)φ̃(u, R)∇ ∇φ(u, R)

2 2
Ti Tf
O , , (5.42)
Tc (d) Tc (d)
where the six-point function is factorized into two-point function by Wick’s theorem. As
a result, the spin-wave approximation is a first-order approximation in the initial and
final temperatures. Consistent expressions of two-point functions must be expanded to first
order in Ti , Tf . Higher-order terms in Ti,f calculated within the spin-wave approximation

should not be expected to be reliable.
5.2.4. Correlation and response functions in the spin-wave approximation

Finally, we are ready to list the result for two-time correlation and response functions
in the spin-wave approximation. From the previous subsection we know that C must be of
first order in Ti , Tf . The consistent result for the magnetic correlation function is therefore
obtained by expanding Eq. (5.33) to first order in temperature. Then
1
Γ (t, s; r, r
) 1 + C(t, s; r, r
) − C(t, t; r, r) + C(s, s; r
, r
) . (5.43)
2
A more suggestive form of this is found as follows. We have

S(t, r) · S(s, r
) = cos φ(t, r) cos φ(s, r
) + sin φ(t, r) sin φ(s, r
)
1
1 − φ(t, r)2 φ(s, r
)2 + φ(t, r) φ(s, r
) + · · · , (5.44)
2
where in the second line we performed a low-temperature expansion which must be kept to
second order in φ in order to be of first order in the temperature, since C = φφ = O(T ).
Furthermore, because of the φ → −φ inversion symmetry, φ = 0. Inserting this into
(5.43) we find, of course only in the context of the spin-wave approximation

S(t, r) · S(s, r
) − S(t, r) · S(s, r
) = C(t, s; r, r
), (5.45)
and the relation between Γ and C is finally clarified (the equilibrium version of this is well
known, see [55, Section 4.2.2]).
For notational simplicity, we shall now concentrate on the autocorrelation and
autoresponse functions. First, the angular correlation function is given by Eqs. (5.39) and
(5.40). We have

 2(4π)−d/2
2 1−d/2 + T t − s + Λ2 1−d/2 ,

 d−2 (Ti − Tf ) t + s + Λ
 f
C(t, s) = if d > 2, (5.46)

 (4π) −1 (T − T ) ln t + s + Λ2 − T ln t − s + Λ2 ,
 f i f
if d = 2.
Second, the autoresponse functions are given by

1
ρ(t, s) = R(t, s) 1 + C(t, s) − C(t, t) + C(s, s) ,
2

2 −d/2
R(t, s) = 4π t − s + Λ . (5.47)
These results require a detailed discussion.
1. The 2D XY model was studied in detail in the spin-wave approximation before [4,84]
and we now show that their results, although they might at first sight appear to be
different, agree with Eqs. (5.43), (5.46), (5.47). For notational simplicity, we restrict
to Ti = 0, Tf = T . First, the magnetic autocorrelation function is [4, Eq. (11)]
η(T )/4
Λ4 (t + s + Λ2 )2
Γ (t, s) = S(t, r) · S(s, r) =
(2t + Λ2 )(2s + Λ2 )(t − s + Λ2 )2

η(T ) Λ4 (t + s + Λ2 )2
= exp ln
4 (2t + Λ2 )(2s + Λ2 )(t − s + Λ2 )2

T 1
1+ ln t + s + Λ2 − ln t − s + Λ2 − ln 2t + Λ2
4π 2
1 2 (5.48)
− ln 2s + Λ + ln Λ + O T ,
2 2
2
since the spin-wave approximation is only consistent to lowest order in T . It is now

clear that the above result is reproduced by inserting C(t, s) from Eq. (5.46) with d = 2
into (5.43). Second, the linear spin response in 2D is [4, Eq. (13)]
δS(t, r)
ρ(t, s) = lim
h→0 δh(s, r)
η(T )/4
1 Λ4 (t + s + Λ2 )2
=
4π(t − s + Λ2 ) (2t + Λ2 )(2s + Λ2 )(t − s + Λ2 )2
1
= Γ (t, s), (5.49)
4π(t − s + Λ2 )
in agreement with Eq. (5.47) with d = 2, as it should be.

In 2D, the results for Γ (t, s) and ρ(t, s) were confirmed by a simulational study with
an ordered initial state and Tf < Tc [1].
2. For Ti = Tf , the two-point functions are stationary. This is only to be expected, since
in this case the system was prepared in an equilibrium state and remains there.
3. Ageing occurs when Ti = Tf . Time-translation invariance is broken and we proceed
to analyse the resulting scaling behaviour. Now, there are in principle two equally
appealing sets of variables. First, we may choose to work with the angular correlation
function C and its associated response R. Recalling the scaling forms introduced in
Section 1 (see Eqs. (1.2), (1.3)) we shall characterize them by the exponents a, b, λC ,
λR . Second, we may prefer instead to work with the magnetic correlation function
Γ and its associated response ρ. We shall use the same scaling forms, but for clarity
we shall denote the corresponding exponents by a
, b
, λ
C , λ
R . These exponents are
straightforwardly read off in the ageing regime where t, s, t − s Λ and we collect
the results in Table 1. The exponent λ
C = (ηi + ηf )/2 was already known [84].
We see that the exponents satisfy the equalities a = b and a
= b
expected for
nonequilibrium critical dynamics and point that in 2D, the magnetic autocorrelation
and autoresponse exponents are different: λ
R − λ
C = 2. This effect comes from the
non-disordered initial condition of the spins, as explained first in [77].
Table 1
Ageing exponents of the d-dimensional XY model in the spin-wave approximation. Here ηi,f = η(Ti,f ) =
Ti,f /(2π ) describe the initial and final correlation exponents
Angular correlation and response
a b λC λR
d =2 0 0 0 0
d >2 d/2 − 1 d/2 − 1 d d
Magnetic correlation and response
a
b
λ
C λ
R
d =2 ηf /2 ηf /2 (ηi + ηf )/2 2 + (ηi + ηf )/2
d >2 d/2 − 1 d/2 − 1 d d
4. Having discussed the values of the ageing exponents, we wish to compare the form
of the scaling functions with the predictions of local scale-invariance as derived
in Section 4. This requires, however, the correct identification of the quasiprimary
fields in our system, see Section 2. It is only the quasiprimary fields which are
expected to transform in a simple way under a local scale-transformation and the
transformation laws of more complicated fields built from quasiprimary fields must
derived accordingly.
In the case at hand, it is clear from the complicated structure of the magnetic
correlation and response functions that the magnetic order parameter S(t, r) does not
correspond to a quasiprimary field. Rather, the quasiprimary field should be identified
with the phase φ(t, r). Indeed, the form of the angular response R(t, s) is in perfect
agreement with the prediction (4.8), (4.17) of local scale-invariance. This suggests that
the response field φ̃(s, r) should be quasiprimary as well.
5. Having thus identified φ and φ̃ as quasiprimary fields of the model, it is now clear
that the angular autocorrelation function C(t, s) should be compared to the critical
dynamics correlation function as derived in Section 4. However, a direct comparison
with Eq. (4.34) is not possible, since in its derivation fully disordered initial conditions
were assumed.
We shall therefore proceed in two steps. First, we shall consider the case Ti = 0.
Because of our initial conditions (5.27) the initial correlator then vanishes and
Eq. (4.34) should now hold true. Second, we shall show that in the context of the free-
field theory underlying the spin-wave approximation of the XY model, the restriction
to uncorrelated initial states can be lifted.
We now set Ti = 0 and Tf = T . From the exponents in Table 1, the predicted
autocorrelator scaling function follows from (4.34) as
1
fC (y) = Φ0 dθ (y + 1 − 2θ )−d/2, d 2, (5.50)
0
and we see immediately that this is in agreement with the explicit angular correlator
(5.46), upon identification of Φ0 .
Finally, if we also allow for Ti > 0, there is a contribution to C(t, s) from the initial
condition. We then return to the basic result (4.9) and decompose C(t, s) = Cth + Cpr .
The thermal term Cth was treated before and the preparation term is analysed in
Appendix B, with the result
2(4π)−d/2
Cpr = Ti (t + s)1−d/2 , d > 2, (5.51)
d −2
in complete agreement with (5.46). The case d = 2 is treated similarly.
In conclusion, the two-time autocorrelation and autoresponse functions of the XY model

treated in spin-wave approximation are in perfect agreement with local scale-invariance,
provided the angular variable φ and its associated response field are identified as the
quasiprimary fields of the model.
5.3. Fluctuation–dissipation relations in the XY model
Having checked that both correlation and response functions agree with the local scale-
invariance prediction, we now inquire what can be said on the approach of the model
towards equilibrium. A convenient way to study this is through the so-called fluctuation–
dissipation ratio [21], see Section 1. Since we have seen that in the XY model angular
and magnetic observables behave quite differently, it is convenient to define two distinct
fluctuation–dissipation ratios, namely,

∂Γ (t, s) −1
Ξ (t, s) := Tf ρ(t, s) ,
∂s

∂C(t, s) −1
X(t, s) := Tf R(t, s) . (5.52)
∂s
Of particular interest will be the limit fluctuation–dissipation ratios X∞ defined in Eq. (1.5)
and similarly Ξ∞ . We have seen before that the scaling of autocorrelation and autoresponse
functions is according to the expectations of nonequilibrium critical dynamics. In this case,
according to the Godrèche–Luck conjecture [38], X∞ and Ξ∞ should be universal.
We shall use the available exact results in the XY model to test this conjecture
by studying the dependence of X and Ξ on the ratio α := Ti /Tf of initial and final
temperatures.
5.3.1. Fluctuation–dissipation ratio for magnetic variables

The fluctuation–dissipation ratio Ξ (t, s) obtained from the magnetic correlation and
response functions reads, with y = t/s

1 Ti y − 1 d/2 y − 1 d/2
=1+ 1− − . (5.53)
Ξ (y) Tf y +1 2
For Ti = 0 and d = 2 this was already known [4]. In the quasiequilibrium regime y 1
the fluctuation–dissipation theorem should hold. Indeed, we find limy→1 Ξ (y) = 1 which
confirms that Ξ should be well defined. For large values of y, that is for well separated
times, we have
−d/2
Tf y −1
Ξ (y) , (5.54)
Ti − Tf 2
and therefore Ξ∞ = 0, indeed, a universal constant. It is remarkable that the asymptotic
value of Ξ (y) should be independent of d and that is agrees with the value Ξ∞ = 0
of phase-ordering kinetics in the low-temperature phase with an ordered, non-critical
equilibrium state. This kind of result should be more typical of an ordered ferromagnetic
equilibrium state as it occurs for d > 2 but is not really expected for d = 2 since the
equilibrium 2D XY model is critical even below TKT .
Finally, for large y the asymptotic form of Ξ (y) is independent on whether the system
is cooled or heated. The temperatures merely enter into a scaling amplitude.
5.3.2. Fluctuation–dissipation ratio for angular variables

In the same way, the fluctuation–dissipation ratio for the angular variables is found. It
reads

1 Ti y + 1 −d/2
=1+ 1− . (5.55)
X(y) Tf y −1
As before, in the quasiequilibrium regime t s, X(t, s) = 1. Surprisingly, however, for
large values of y = t/s, the limit fluctuation–dissipation ratio

Ti −1
X∞ = 2 − (5.56)
Tf
depends continuously on α = Ti /Tf . We recall from Table 1 that the non-equilibrium
exponents of the angular variables are all independent of both Ti and Tf and although
the exponents do depend on d, we see from (5.56) that X∞ does not. Taken literally, this
would be an example of a non-universal value of the limit fluctuation–dissipation ratio.
We recall that most ‘physically reasonable’ systems undergoing non-equilibrium critical
dynamics one usually finds 0 X∞ 1/2, see [39] for a review. Motivated from mean-
field theories of spin glasses, it is sometimes suggested that Teff := T /X∞ might be
interpreted as an effective temperature for which the fluctuation–dissipation theorem
would hold. It is hard to see how in this case (X∞ may even become negative) such an
interpretation could be maintained.
6. The critical voter model in d dimensions
We now study a qualitatively different type of application of local scale-invariance

in the so-called voter model, see [64] and references therein. The model describes the
temporal evolution of configurations C of spins σr (t) = ±1 on a d-dimensional hypercubic
lattice Zd . The dynamics is assumed to be given by a master equation

d
P (C; t) = Wr C (r) P C (r) ; t − Wr (C)P (C; t) . (6.1)
dt d
r∈Z
Here the configuration C (r) is obtained from C by inverting the single spin at site r. Finally,
the transition rates for a spin reversal σr → −σr are given by

1
d
1
Wr (C) = 1− (σr σr+ek + σr σr−ek ) , (6.2)
2 2d
k=1
where the ek , k = 1, . . . , d, form an orthonormalized basis of unit vectors on the d-
dimensional hypercubic lattice.
With respect to the kinetic spherical model and the XY model studied previously, the
voter model is different since in general it does not satisfy detailed balance and therefore
will not relax to an equilibrium state. By considering a general kinetic Ising model with
a dynamics respecting the global Z2 -symmetry, it can be shown that the transition rates
(6.2) correspond to the critical point of the so-called linear voter model [73,88]. The
non-equilibrium kinetics of the critical voter model (6.2) has been studied in detail by
Dornic [26]. In d = 1 dimensions the model coincides with the kinetic Glauber–Ising
model at zero temperature (which we shall revisit in Appendix C) but for d > 1 these
two models are different. In particular, it is known that the domain growth of the voter
model is not driven by the minimization of the surface tension between the two phases
[25,26] but which is the mechanism which drives ageing in simple ferromagnets [9,11].
We are interested here in the correlation functions Cr (t) = σr (t)σ0 (t) and Cr (t, s) =
σr (t)σ0 (s) which are easily seen to satisfy the following equations of motion [26,28]
∂ ∂
Cr (t) = 2 r Cr (t), Cr (t, s) = r Cr (t, s), (6.3)
∂t ∂t
subject to the following boundary conditions
C0 (t) = 1, Cr (0) = δr,0 , Cr (t, t) = Cr (t), (6.4)

where r is the discrete Laplacian and where the initial magnetization r σr (0) = 0. For
the autocorrelation function C(t, s) = C0 (t, s) of the critical voter model (6.2) one finds in
the ageing regime, where t, s and t − s are all sufficiently large [26], with y = t/s
 √

 π arctan 2/(y − 1),
2
if d = 1,
−1

C(t, s) = ln(s) ln (y + 1)/(y − 1) , if d = 2,
 s −d/2+1 d d/2 2γd (y − 1)−d/2+1 − (y + 1)−d/2+1 , if 2 < d < 4,

2π d−2
(6.5)
where γd is the probability that a random walk in d dimensions and starting from the origin
never returns. We are not aware of published results on R(t, s) for 2 < d < 4 in this model.
Clearly, time-translation invariance is broken for all d 2.
We wish to compare (6.5) with the prediction Eq. (4.34) of local scale-invariance. The
case d = 1 will be dealt with in appendix C and since for d = 2 logarithmic scaling is
found, the form of local scale-invariance as presented here is not applicable.8 We therefore
concentrate on the dimensions 2 < d < 4. From the asymptotics of C(t, s) in (6.5) we read
off the exponents
d
b= − 1, λC = d (6.6)
2
since the exponent z = 2 is known, see [26,64]. From Eq. (4.34) we expect
1
fC (y) = Φ0,c dθ (y + 1 − 2θ )−d/2
0
Φ0,c
= (y − 1)−d/2+1 − (y + 1)−d/2+1 , (6.7)
d −2
in full agreement with (6.5) and we identify the normalization constant Φ0,c = 2γd (d/
2π)d/2 . Indeed, we see that the form (4.33)—which in principle is only valid asymptoti-
cally—is in fact exact in the critical voter model. This is not surprising in view of the
underlying free-field theory.
In conclusion, for the critical voter model with two competing steady states ageing
occurs. The scaling form of the two-time autocorrelation function is in exact agreement
with the prediction of local scale-invariance. This is the first time such an agreement is
found for a system without detailed balance.
7. The free random walk
Last, but not least, we briefly consider the simplest example of a system undergoing
ageing: the free random walk [21]. The Langevin equation describing the time-evolution
of the order parameter φ reads
∂φ(t)
= h(t) + η(t), (7.1)
∂t
where a deterministic external field h(t) has been added in order to be able to compute
response functions. The Gaussian noise η is characterized as usual by its first two moments,
see Eq. (1.9). Here, we choose the notation such that the relationship to local scale-
invariance becomes evident.
The autocorrelation and linear autoresponse functions were already calculated by
Cugliandolo et al. [21]. They obtained, with the initial condition C(t, 0) = 0

δφ(t) 1
R(t, s) = = φ(t)η(s) = Θ(t − s),
δh(s) h=0 2T
C(t, s) = φ(t)φ(s) = 2T min(t, s). (7.2)
8 An extension of Schrödinger invariance to logarithmic Schrödinger invariance in analogy to logarithmic

conformal invariance, see, e.g., [44], might be needed here.
Clearly, the system undergoes ageing, since C(t, s) does not merely depend on t − s
and furthermore, it stays forever out of equilibrium, since the fluctuation–dissipation ratio
X(t, s) = 1/2 [21].
We wish to check that the results (7.2) are compatible with the predictions of local scale-
invariance. For the autoresponse function, the first equation of (7.2) is clearly compatible
with (4.7), (4.17), (4.18), with the exponents a = −1 and λR = 0. For the autocorrelation
function, we expect from (4.12)

δ 2 φ(t)φ(s)
C(t, s) = T du , (7.3)
δh(u)2 h=0
where in view of the initial condition C(t, 0) = 0 used in Eq. (7.2) we assumed a vanishing
initial correlator.
In order to calculate the above derivative, we solve the Langevin equation for a given
field h and obtain the autocorrelation function
s t

C t, s; [h] = C t, s; [0] + dv dw h(v)h(w), (7.4)
0 0
where C(t, s; [0]) = C(t, s) is of course given by the second Eq. (7.2). The required second
derivative of the autocorrelation function becomes in a field-theoretic formulation some
three-point correlator (see Eq. (4.10)) and it is now easy to see that

(3) δ 2 C(t, s; [h])
R0 (t, s, u) = = 2Θ(t − u)Θ(s − u). (7.5)
δh(u)2 h=0
Inserting this into (7.3), the desired result for C(t, s) in Eq. (7.2) is, indeed, recovered.
We read off the exponents b = −1 and λC = 0. Of course, the exponent equalities a = b
and λC = λR are a necessary requirement for having a non-vanishing limit fluctuation–
dissipation ratio X∞ = 1/2.
In conclusion, the evidence from the two-time autocorrelation and autoresponse
function of the free random walk is fully consistent with local scale-invariance.
8. Conclusions and discussion
We have analysed the ageing behaviour in systems with a non-conserved order

parameter and described by a Langevin equation. Our main assumption was that the
noiseless part of that Langevin equation is Galilei-invariant. Together with dynamical
scaling this hypothesis fixes the dynamical exponent z = 2 and implies for local theories
Schrödinger invariance [48]. There are good reasons for admitting such a hypothesis. For
example, the phase-ordering kinetics of the Glauber–Ising model in d > 1 dimensions
quenched to a temperature T < Tc provides strong evidence that the scaling function
of its space–time response function R(t, s; r, r
) has the form predicted from Galilei
invariance [49]. However, since groups of local scale transformations such as the
Schrödinger group are dynamical symmetries of noiseless differential equations only, the
role of the noise in the Langevin equation or from the initial conditions has to be addressed.
We have carried out such a study, for the important special case where z = 2 and
the initial state is fully disordered. Considering the Langevin equation as the classical
equation of motion of a MSR-type field theory, we have calculated the two-time correlation
and linear response functions by studying how the dynamical symmetry properties of the
noiseless part of that field theory are reflected in these noisy averages. These averages can
be written in form of a perturbative expansion around the noiseless theory and we have
shown that only a finite number of terms in these series contributes. Specifically, we have
found:
1. The two-time linear response function R = φ φ̃ involving only quasiprimary fields
is independent of both the thermal and the initial noise. This explains why the form
(1.7) of scaling function fR (y) of the linear response—previously derived from the
symmetries of the noiseless theory—has been reproduced in many different systems
with T > 0 either exactly or with a considerable numerical precision [12,16,21,38,39,
45–47,49,58,69,77].
2. We obtain the reduction formula Eq. (4.9) which expresses the two-time correlation
function in terms of certain noiseless three- and four-point response functions. For
the uncorrelated initial conditions (4.11) only a single noiseless three-point response
function is needed.
3. The scaling forms of correlation and response functions are governed by the two non-
trivial exponents λC and λR which are, in general, distinct from each other. Given
the initial correlator (4.11), local scale-invariance with z = 2 provides a sufficient
condition for the exponent equality
λC = λR . (8.1)
4. The scaling of the two-time autocorrelator C(t, s) is described by a scaling function

Φ(w) which in turn depends on a scaling function Ψ (ρ) which arises in the three-
point function of quasiprimary fields in Schrödinger-invariant theories. Depending on
whether the thermal or the initial noise is dominant, two distinct scaling forms (4.25)
and (4.32) are found and we have argued that they should describe the cases when the
system is quenched to temperatures T < Tc and T = Tc , respectively (in the latter case
only if z = 2 still holds after renormalization).
5. Schrödinger invariance by itself does not determine the form of Ψ (ρ). We have argued
that the related three-point response function should be non-singular and then find the
asymptotic behaviour for w → ∞

−ϕ d/2 − λC , if T < Tc ,
Φ(w) ∼ w , ϕ= (8.2)
1 + a, if T = Tc .
This suggests the following approximate forms, with y = t/s
"
2 (y + 1)2 /(4y) −λC /2 ,
Meq if T < Tc ,
C(t, s) ≈ 1
Φ0,c y 1+a−λ /2 λ −2−2a (y + 1 − 2θ ) −1−a , if T = Tc .
0 dθ θ
C C
(8.3)
At least, these forms are consistent with the required asymptotic behaviour of fC (y)
as y → ∞, see Section 1. For free-field theories (8.2) holds for all values of w and
then (8.3) becomes exact.
In the past, approximate expressions for the scaling of magnetic correlation functions
were derived from Gaussian closure procedures for kinetic O(n)-models undergoing
phase-ordering kinetics, see [9]. This gives for the magnetic autocorrelation function
at T = 0 [7,8,83]

d/4
n 1 n+1 2 4y
fC,BPT (y) ≈ B ,
2π 2 2 (y + 1)2
d/2
1 1 n+2 4y
× 2 F1 , ; ; , (8.4)
2 2 2 (y + 1)2
where B is Euler’s beta function and 2 F1 a hypergeometric function. It is easy to
see that the scaling function (8.3) with T = 0 is recovered from (8.4) in the n → ∞
limit. In the notation of Section 5.2, Eq. (8.4) implies an exponent λ
C = d/2. This is
a typical value for a free-field theory for T < Tc (which, indeed, describes the O(n)-
model in the n → ∞ limit) but which will, in general, not hold for n < ∞ and thus
(8.4) cannot be expected to represent well the behaviour for y = t/s large.9 One might
wonder whether the long-standing difficulties in arriving at scaling functions which
cover adequately the whole range of values of y should not be related at least partially
to having worked with dynamic variables which might turn out to be not the most basic
ones of the model?
6. We have tested these predictions on the exact solutions of the kinetic spherical model
and the XY model in the spin-wave approximation. In these cases, the exponent a is
given by (1.4). In order to compare the exact model results with the prediction (8.3),
it was necessary to carefully identify the quasiprimary fields of the models. For the
spherical model, the natural magnetic order parameter and its response field could be
used as quasiprimary fields. On the other hand, for the XY model it turned out that the
coarse-grained magnetic moment is not quasiprimary but rather the angular variable is.
These examples underline the importance of the correct identification of the quasipri-
mary fields in a given model.
The explicit results in the XY model also illuminate in a new way the Godrèche–
Luck conjecture on the universality of the limit fluctuation–dissipation ratio. Further
tests on the form of the autocorrelation function in the Glauber–Ising model in d 2
dimensions are presently carried out and will be reported elsewhere.
7. We also showed that the autocorrelator in the critical voter model in 2 < d < 4
dimensions and which does not satisfy detailed balance, again agrees with (8.3). This
is the first example of a new domain of application of local scale-invariance. We also
confirmed (8.3) for the free random walk.
9 Indeed, to leading order in 1/n and for uncorrelated initial conditions [6],
d d −1
λ
C = d2 − 43 (d + 2) 2d d
9 B 1 + 2 ,1 + 2 n .
These four confirmations, although all based on an underlying free-field theory provide
further evidence in favour of a Galilei invariance of the noiseless theory.
8. The scaling of the linear response of the 1D Glauber–Ising model at T = 0 can only
be explained through a generalization of the representations of the Lie algebra of local
scale-invariance. It would be interesting to see whether a similar phenomenon could
be found in different 1D systems undergoing ageing at T = 0.
Our approach has been based in an essential way on the assumption of Galilei invariance
of the noiseless theory. However, Mazenko [67] recently studied phase-ordering kinetics
in the context of the time-dependent non-linear Ginzburg–Landau equation. He carried
out a second-order perturbative calculation around a Gaussian theory which is equivalent
to the Ohta–Jasnow–Kawasaki approximation and reports a deviation of the two-time
autoresponse function R(t, s) from the local scale-invariance prediction (1.7). A similar
difficulty had been observed before by Calabrese and Gambassi who studied non-
equilibrium dynamics (that is T = Tc ) of the O(n)-model, for both model A and model
C dynamics, through MSR field theory [13–15]. Again, already their classical action is
manifestly not Galilei-invariant.10
At face value, Mazenko’s result [67] is in disagreement with the simulational data
obtained from the 2D and 3D Glauber–Ising model with T < Tc and based on the master
equation. These do reproduce (1.7) for the autoresponse function R(t, s) [47,49,52] as well
as the extension for the spatio-temporal response R(t, s; r − r
) [49]. Could this mean that
there are subtle differences between the formulation of stochastic systems either through
a master equation or else through a coarse-grained Langevin equation11 and which affect
the formal Galilei invariance of the theory? Alternatively, if under renormalization the
dynamical exponent z = 2 remains constant, might the theory flow to a fixed point where
asymptotically Galilei invariance would hold?
All in all, based on a postulated extension of dynamical scaling to some local scale-
invariance, we have reformulated the problem of finding the scaling function of the two-
time autocorrelation function of ageing systems as one of a discussion of the properties
of certain three-point response functions of the noiseless theory. The evidence available at
present suggests that this approach might be capable of shedding a new light on the issue.
Finally, it might be of interest to search for extensions for dynamical exponents z = 2
and/or to study ageing systems with a conserved order parameter. Asymptotic information
on the two-point functions in the latter case is now becoming available [40].
Appendix A. On Gaussian integration
We present the details of the calculations of the magnetic two-time correlation function
Γ and its associated linear response function ρ for the d-dimensional XY model in the
10 At criticality, the combined effect of thermal and initial fluctuations leads for interacting theories to a non-
trivial value of z = 2 under renormalization so that our arguments are no longer directly applicable.
11 In 1D and at T = 0, the autocorrelation exponent λ = 1 found in the Glauber–Ising model [37,65] differs
C
from the exactly known exponent λC = 0.6006 . . . [10] determined in the time-dependent Landau–Ginzburg
equation.
spin-wave approximation. They are defined as

Γ (t, s; r, r
) = Dφ Dφ̃ cos φ(t, r) − φ(s, r
) exp −S[φ, φ̃] ,

ρ(t, s; r, r ) = Dφ Dφ̃ φ̃(s, r
) exp −S[φ, φ̃] ,

(A.1)
where S[φ, φ̃] is the free-field Martin–Siggia–Rose action which is Gaussian.
A.1. The correlation function Γ (t, s; r, r

)
Using de Moivre’s identities

1
Γ (t, s; r, r
) = exp i φ(t, r) − φ(s, r
) + exp i φ(s, r
) − φ(t, r)
2
'
1
= exp i du dR J (u, R)φ(u, R)
2
'
(A.2)
+ exp −i du dR J (u, R)φ(u, R) ,
where J (u, R) is given by

J (u, R) = δ(u − t)δ(R − R) − δ(u − s)δ(R − r
). (A.3)
A generally valid result from free-field theory reads, see [94]
'
exp i du dR J (u, R)φ(u, R)

1

= exp − du du dR dR J (u, R) φ(u, R)φ(u , R ) J (u , R ) , (A.4)

2
which is one of the various forms of writing Wick’s theorem. Using the explicit form of
the current in (A.3), Eq. (5.33) follows immediately.
A.2. The response function ρ(t, s; r, r

)
We now focus on the magnetic response function ρ(t, s; r, r

). We decompose the MSR
action
S[φ, φ̃, 0] = S0 [φ, φ̃] + s[φ, φ̃], (A.5)
where

∂φ
S0 [φ, φ̃] = du dr φ̃ − φ ,
∂u

s[φ, φ̃] = − du du
drdr
φ̃(u, r)κ(u, u
; r − r
)φ̃(u
, r
), (A.6)
and
1
κ(u, u
; r − r) = T δ(u − u
)δ(r − r
) + δ(u)δ(u
)a(r − r
). (A.7)
2
The noiseless two-point functions are

−d/2 (r − r
)2
R0 (t, s; r, r ) = φ(t, r)φ̃(s, r ) 0 = Θ(t − s) 4π(t − s) exp − ,
4(t − s)

C0 (t, s; r, r
) = φ(t, r)φ(s, r
) 0 = 0,

C̃0 (t, s; r, r
) = φ̃(t, r)φ̃(s, r
) 0 = 0. (A.8)
Below, we shall need the equal-time response function R(t, s; r, r
)|t =s . To give to this
quantity a value, one may discretize the Langevin equation. This may be done according
to several different schemes, see [59,92]. Here, we shall use the Itô prescription which
amounts to
R0 (t, t; r, r
) = 0. (A.9)
The magnetic response function reads from (A.1)

ρ(t, s; r, r
) = φ̃(s, r
) exp −s[φ, φ̃] 0 (A.10)
where 0 is the average with the exp(−S0 [φ, φ̃]). Expanding the sine and using the
Newton’s binomial identity, we have
∞ 2k+1

ρ(t, s; r, r
) = ρk,p (t, s; r, r
),
k=0 p=0
(−1)k+p+1
ρk,p (t, s; r, r
) =
p!(2k + 1 − p)!

× φ̃(s, r
)φ p (t, r)φ 2k+1−p (s, r
) exp −s[φ, φ̃] 0 , (A.11)
so that ρk,p (t, s; r, r
) is given by a (2k + 2)-point function, containing only one φ̃
contribution and 2k + 1 fields φ. The Bargman superselection rule (A.8) implies that the
only contractions that will lead to non-vanishing averages come from the kth order in the
expansion of the exponential, so that we have
k
(−1)p+1
ρk,p (t, s; r, r
) = dj dj
p!(2k + 1 − p)!k!
j =1

p

× φ̃(s, r )φ (x, t)φ 2k+1−p
(s, r
)φ̃(j )κ(j, j
)φ̃(j
) 0 , (A.12)
where for clarity the notation j (respectively, j
) stands for (uj , r j ) (respectively, (u
j , r j
))
and where integrals run over this set of 2k variables.
Wick’s theorem states that the integrand decomposes into sums of products of two-point
functions. In order for a contraction not to vanish, the field φ̃(s, r
) must contract with one
of the p fields φ(t, r), which leads to
ρk,0 (t, s; r, r
) = 0,
k
(−1)p+1
ρk,p (t, s; r, r
) = dj dj
R0 (t, s; r, r
)
(p − 1)!(2k + 1 − p)!k!
j =1
p−1
× φ (t, r)φ 2k+1−p
(s, r )φ̃(j )κ(j, j
)φ̃(j
) 0 .

(A.13)
Summing over p, the response function ρ(t, s; r, r

) reads
∞

k
1
ρ(t, s; r, r
) = R0 (t, s; r, r
) dj dj
(2k)!k!
k=0 j =1
2k
× φ(t, r) − φ(s, r
) φ̃(j )κ(j, j
)φ̃(j
) 0 . (A.14)
At this stage, it will be interesting to give another equivalent expression of the magnetic
correlation function Γ (t, s; r, r
). Using the same strategy as before, but now expanding
the cosine we find
∞

Γ (t, s; r, r ) = γk (t, s; r, r
),
k=0
(−1)k 2k
γk (t, s; r, r
) = φ(t, r) − φ(s, r
) exp −s[φ, φ̃] 0 . (A.15)
(2k)!
For the same reason as before, because of mass conservation, only the expansion of order
k of the exponential term will contribute and lead to non-vanishing contractions, such that
γk (t, s; r, r
)

k
1 2k
= dj dj
φ(t, r) − φ(s, r
) φ̃(j )κ(j, j
)φ̃(j
) 0 . (A.16)
(2k)!k!
j =1
Comparing (A.14) with (A.16), we thus have
ρ(t, s; r, r
) = R0 (t, s; r, r
)Γ (t, s; r, r
). (A.17)
This is the main result of this appendix and gives Eq. (5.38) in the text.
Appendix B. Scaling form of a special four-point function
We study the scaling of the four-point function

F := R0 (t, s, 0, 0; 0, 0, R, R
) = φ(t, 0)φ(s, 0)φ̃(0, R)φ̃(0, R
) 0
(4)
(B.1)
of a theory in MSR formulation of which the noiseless part is Schrödinger-invariant.
The field φ is assumed to be quasiprimary with scaling dimension x and mass M and
the response field φ̃ should also be quasiprimary with scaling dimension x̃ and mass
M̃ = −M. The covariance conditions on R are the following (we use v(t) = 0)

1 1
t∂t + s∂s + R∂R + R
∂R
+ (x + x̃) F = 0,
2 2

M 2
2

t ∂t + s ∂s + (t + s)x −
2 2
R +R F = 0. (B.2)
2
We shall not discuss here the general solution of these equations. For our purposes it is
enough to observe that if we decompose F in the following symmetrized way
F = G(t, R)G(s, R
) + G(t, R
)G(s, R), (B.3)
which for a free-field theory would follow from Wick’s theorem, then

−x−x̃ M R2
G(t, R) = G0 t exp − (B.4)
2 t
produces a solution to both covariance conditions.
We apply this result to the preparation part Cpr of the autocorrelation function, with the
intention to use the result in the spin-wave approximation of the XY model. From Eq. (4.9)
we have

1
dR dR
a(R − R
)R0 (t, s, 0, 0; 0, 0, R, R
)
(4)
Cpr = (B.5)
2
R2d

Ti (ts)−(x+x̃)/2
= dR dR
dq
(2π)d q2
R3d

M R 2 R 2
× exp iq · (R − R
) − + (B.6)
2 t s
∞
dq
Ti e−q (t +s+z)
2
= dz (B.7)
(2π)d
0 Rd
(4π)−d/2
= Ti (t + s)1−d/2, (B.8)
d/2 − 1
where in going to (B.6) we used the initial condition (5.27) and the result (B.3), (B.4) from
above, next in going to (B.7) we specialized to a free Gaussian field (where x = x̃ = d) of
mass M = 1/2 and the last step we also assumed d > 2.
Eq. (B.8) provides the preparation term, for any initial temperature Ti , as required for
the analysis of the autocorrelation function in the XY model in Section 5.2
Appendix C. On local scale-invariance in the 1D Glauber–Ising model
C.1. Two-point functions in the 1D Glauber–Ising model
In the text we have seen that local scale-invariance implies the following form of the
two-time autoresponse function
1+a−λR /z
t
R(t, s) = r0 (t − s)−1−a . (C.1)
s
In spite of a nice agreement with a large variety of models, this expression is not verified
for the 1D Ising model with Glauber dynamics at T = 0.
The 1D Ising model is described by spins σi = ±1 and the Hamiltonian

H=− σi σi+1 . (C.2)
i=1
The exactly solvable Glauber dynamics [36] may be given through the heat-bath rule,
which gives the probability of finding the spin variables σi (t + 1) in terms of those at
time t
1
P σi (t + 1) = ±1 = 1 ± tanh β σi−1 (t) + σi+1 (t) + hi (t) (C.3)
2
where β = 1/T is the inverse temperature and hi (t) the external magnetic field. Ageing
occurs in this model at T = 0. In the long-time scaling limit, the two-time autocorrelation
and autoresponse functions are [37,65,68]

δσi (t) 1
R(t, s) = = √ , (C.4)
δhi (s) h=0 π 2s(t − s)
2 2
C(t, s) = σi (t)σi (s) = arctan . (C.5)
π t/s − 1
While these results were obtained first for a fully disordered initial state, they remain true
for long-ranged initial conditions σr (0)σ0 (0) ∼ r −ν with ν > 0 [51]. The case ν = 0
corresponds to the case of an initial magnetization m0 . Then the connected part of C(t, s)
as well as R(t, s) are multiplied by 1 − m20 [51,77]. In any case, the forms of the scaling
functions fC,R (y) are unchanged.
Although these two-point functions clearly display dynamical scaling, it is evident that
the scaling form of R(t, s) from (C.4) is incompatible with the form suggested in (C.1).
Local scale-invariance as developed in the text does not hold in the 1D Glauber–Ising
model.
C.2. Generalized realization of the ageing algebra
We now show how Schrödinger invariance can be generalized such that the exact
response function (C.4) can be reproduced. Obviously, time-translation invariance is
broken in ageing systems. Therefore, as already pointed out in [45,46], the dynamical
symmetry cannot be the Schrödinger Lie algebra sch1 which contains the time-translation
generator X−1 = −∂t , but a subalgebra without this generator might be acceptable. We
consider the algebra [48]
age1 := {X0 , X1 , Y−1/2 , Y1/2 , M0 } (C.6)
and keeping the commutation relations (2.7) we now look for a more general realization of
age1 . In this way, we write the generators as {Ξ0,1 , Υ±1/2 , M0 }. These must be of the form
1 x M 2
Ξ0 = −t∂t − r∂r − , Ξ1 = −t 2 ∂t − tr∂r − xt − g(t) − r ,
2 2 2
Υ−1/2 = −∂r , Υ1/2 = −t∂r − Mr,
M0 = −M, (C.7)
where g = g(t) is to be found. The only commutator of age1 constraining g is
[Ξ1 , Ξ0 ] = Ξ1 , (C.8)
which leads to
t∂t g − g = 0, (C.9)
with the solution g(t) = Kt, with K some constant. From (C.7), the dynamical exponent
z = 2. If we were to require, in addition, [Ξ1 , Ξ−1 ] = 2Ξ0 (and thereby go from age1
back to sch1 ), we would recover K = 0. Now, a quasiprimary field φ of age1 will be
characterized by a triplet (x, K, M).
We can now generalize local scale-invariance by requiring that the autoresponse
function R(t, s) formed from a quasiprimary field φ and its associated quasiprimary
response field φ̃ to transform covariantly under the generators Ξ0 and Ξ1 . It is a solution
of the system of linear partial differential equations

x x̃
t∂t + + s∂s + R(t, s) = 0,
2 2
2
t ∂t + (K + x)t + s 2 ∂s + (K̃ + x̃)s R(t, s) = 0, (C.10)
where x̃ and K̃ refer to the response field φ̃. Solving the system (C.10) gives as final result
R(t, s) = s −1−a fR (t/s),

fR (y) = r0 y 1+A−λR /2 (y − 1)−1−A , (C.11)
where the three independent non-equilibrium exponents a, A, λR are
x + x̃
a= − 1, A = a + K + K̃, λR = 2x + 2K. (C.12)
2
In contrast with the previous realization of age1 , K, K̃ = 0 is possible and then a and A
differ from each other.
Comparison with the exact result (C.4) of the 1D Glauber–Ising model now gives
complete agreement and we identify the exponents
1
a = 0, A=− , λR = 1. (C.13)
2
Of course, the values of a and λR have been obtained before [37,65] but A seems to be a
new exponent.
At present, it must remain open whether the unusual properties of the 1D Glauber–Ising
model are related to the fact that Tc = 0 and therefore the critical and low-temperature
properties might have become mixed.
Also, it remains to be seen whether the form of the autocorrelation function can be
understood from the generalized realization of the ageing algebra age1 . We hope to come
back to this elsewhere.
References
[1] S. Abriet, D. Karevski, Eur. Phys. J. B 59 (2004), in press, cond-mat/0309342.

[2] V. Bargmann, Ann. Math. 59 (1954) 1.
[3] A.A. Belavin, A.M. Polyakov, A.B. Zamolodchikov, Nucl. Phys. B 241 (1984) 333.
[4] L. Berthier, P.C.W. Holdsworth, M. Sellitto, J. Phys. A 34 (2001) 1805.
[5] J.P. Bouchaud, in: M.E. Cates, M.R. Evans (Eds.), Soft and Fragile Matter, IOP Press, Bristol, 2000.
[6] A.J. Bray, K. Humayun, T.J. Newman, Phys. Rev. B 43 (1991) 3699.
[7] A.J. Bray, S. Puri, Phys. Rev. Lett. 67 (1991) 2670;
H. Toyoki, Phys. Rev. B 45 (1992) 1965.
[8] A.J. Bray, K. Humayun, Phys. Rev. E 48 (1992) 1609.
[9] A.J. Bray, Adv. Phys. 43 (1994) 357.
[10] A.J. Bray, B. Derrida, Phys. Rev. E 51 (1995) R1633.
[11] A.J. Bray, in: M.E. Cates, M.R. Evans (Eds.), Soft and Fragile Matter, IOP Press, Bristol, 2000.
[12] P. Calabrese, A. Gambassi, Phys. Rev. E 65 (2002) 066120.
[13] P. Calabrese, A. Gambassi, Phys. Rev. B 66 (2002) 212407.
[16] S.A. Cannas, D.A. Stariolo, F.A. Tamarit, Physica A 294 (2001) 362.
[17] J.L. Cardy, Scaling and Renormalization in Statistical Mechanics, Cambridge Univ. Press, Cambridge, 1996.
[18] R. Cherniha, J.R. King, J. Phys. A 33 (2000) 267;
R. Cherniha, J.R. King, J. Phys. A 33 (2000) 7839;
R. Cherniha, J.R. King, J. Phys. A 36 (2003) 405.
[19] A. Crisanti, F. Ritort, J. Phys. A 36 (2003) R181.
[20] L.F. Cugliandolo, J. Kurchan, J. Phys. A 27 (1994) 5749.
[21] L.F. Cugliandolo, J. Kurchan, G. Parisi, J. Physique I 4 (1994) 1641.
[22] L.F. Cugliandolo, D.S. Dean, J. Phys. A 28 (1995) 4213.
[23] L.F. Cugliandolo, in: J.-L. Barrat, J. Dalibard, J. Kurchan, M.V. Feigel’man (Eds.), Slow Relaxation and
Non-Equilibrium Dynamics in Condensed Matter, Les Houches Session, 77 July 2002, Springer, 2003,
cond-mat/0210312.
[24] P. di Francesco, P. Mathieu, D. Sénéchal, Conformal Field Theory, Springer, Heidelberg, 1997.
[25] I. Dornic, H. Chaté, J. Chave, H. Hinrichsen, Phys. Rev. Lett. 87 (2001) 045701.
[26] I. Dornic, Thèse de doctorat, Nice et Saclay, 1998.
[27] D.S. Fisher, D.A. Huse, Phys. Rev. B 38 (1988) 373.
[28] L. Frachebourg, P.L. Kaprivsky, Phys. Rev. E 53 (1996) R3009;
E. Ben-Naim, L. Frachebourg, P.L. Kaprivsky, Phys. Rev. E 53 (1996) 3078.
[29] N. Fusco, M. Zannetti, Phys. Rev. E 66 (2003) 066113.
[30] W.I. Fushchych, R.M. Cherniha, J. Phys. A 18 (1985) 3491.
[31] W.I. Fushchych, M. Serov, J. Phys. A 20 (1987) L929.
[32] W.I. Fushchych, R.M. Cherniha, Ukr. Math. J. 41 (1989) 1161;
W.I. Fushchych, R.M. Cherniha, Ukr. Math. J. 41 (1989) 1456.
[33] W.I. Fushchych, W.M. Shetelen, M.I. Serov, Symmetry Analysis and Exact Solutions of Equations of
Nonlinear Mathematical Physics, Kluwer, Dordrecht, 1993.
[34] W.I. Fushchych, R.M. Cherniha, J. Phys. A 28 (1995) 5569.
[35] P.K. Ghosh, Phys. Rev. A 65 (2002) 012103.
[36] R.J. Glauber, J. Math. Phys. 4 (1963) 294.
[37] C. Godrèche, J.M. Luck, J. Phys. A 33 (2000) 1151.
[38] C. Godrèche, J.M. Luck, J. Phys. A 33 (2000) 9141.
[39] C. Godrèche, J.M. Luck, J. Phys.: Condens. Matter 14 (2002) 1589.
[40] C. Godrèche, F. Krzakala, F. Ricci-Tersenghi, cond-mat/0401334.
[41] J.A. Goff, Am. J. Math. 49 (1927) 117.
[42] C.R. Hagen, Phys. Rev. D 5 (1972) 377.
[43] M. Henkel, J. Stat. Phys. 75 (1994) 1023.
[44] M. Henkel, Phase Transitions and Conformal Invariance, Springer, Heidelberg, 1999.
[45] M. Henkel, M. Pleimling, C. Godrèche, J.-M. Luck, Phys. Rev. Lett. 87 (2001) 265701.
[46] M. Henkel, Nucl. Phys. B 641 (2002) 405.
[47] M. Henkel, M. Paeßens, M. Pleimling, Europhys. Lett. 62 (2003) 644.
[48] M. Henkel, J. Unterberger, Nucl. Phys. B 660 (2003) 407.
[49] M. Henkel, M. Pleimling, Phys. Rev. E 68 (2003) 065101(R).
[50] M. Henkel, A. Picone, M. Pleimling, J. Unterberger, cond-mat/0307649.
[51] M. Henkel, G.M. Schütz, J. Phys. A 37 (2004) 591.
[52] M. Henkel, M. Paeßens, M. Pleimling, Phys. Rev. E 69 (2004), in press, cond-mat/0310761.
[53] P.C. Hohenberg, B.I. Halperin, Rev. Mod. Phys. 49 (1977) 435.
[54] D.A. Huse, Phys. Rev. B 40 (1989) 304.
[55] C. Itzykson, J.M. Drouffe, Théorie Statistique des Champs 1, Editions du CNRS, Paris, 1989.
[56] E.V. Ivashkevich, J. Phys. A 30 (1997) L525.
[57] R. Jackiw, Phys. Today 25 (1) (1972) 23.
[58] H.K. Janssen, B. Schaub, B. Schmittmann, Z. Phys. B 73 (1989) 539.
[59] H.K. Janssen, in: G. Györgyi, et al. (Eds.), From Phase Transitions to Chaos, World Scientific, Singapore,
1992, p. 68.
[60] V.L. Katkov, Zh. Prikl. Mekh. Tekh. Fiz. 6 (1965) 105.
[61] J.M. Kosterlitz, D.J. Thouless, J. Phys. C 6 (1973) 1181.
[62] V.I. Lahno, J. Phys. A 31 (1998) 8511.
[63] S. Lie, Arch. Math. Nat. Vid. (Kristiania) 6 (1882) 328.
[64] T.M. Ligget, Stochastic Interacting Systems: Contact, Voter and Exclusion Process, Springer, Heidelberg,
1999.
[65] E. Lippiello, M. Zanetti, Phys. Rev. E 61 (2000) 3369.
[66] P.C. Martin, E.D. Siggia, H.H. Rose, Phys. Rev. A 8 (1973) 423.
[67] G.F. Mazenko, Phys. Rev. E 69 (2004) 016114.
[68] P. Mayer, P. Sollich, J. Phys. A 37 (2004) 9;
P. Mayer, L. Berthier, J.P. Garrahan, P. Sollich, Phys. Rev. E 68 (2003) 016116.
[69] T.J. Newman, A.J. Bray, J. Phys. A 23 (1990) 4491.
[70] U. Niederer, Helv. Phys. Acta 45 (1972) 802.
[71] U. Niederer, Helv. Phys. Acta 47 (1974) 167.
[72] A.G. Nikitin, R.O. Popovych, Ukr. Math. J. 53 (2001) 1255.
[73] M.J. de Oliveira, J.F.F. Mendes, M.A. Santos, J. Phys. A 26 (1993) 2317.
[74] L. O’Raifeartaigh, V.V. Sreedhar, Ann. Phys. 293 (2001) 215.
[75] M. Paeßens, M. Henkel, J. Phys. A 36 (2003) 8983.
[76] M. Perroud, Helv. Phys. Acta 50 (1977) 233.
[77] A. Picone, M. Henkel, J. Phys. A 35 (2002) 5575.
[78] A. Picone, M. Henkel, J. Richert, J. Phys. A 36 (2003) 1249.
[79] A. Picone, Thèse de doctorat, Nancy, 2004.
[80] A.M. Polyakov, Sov. Phys. JETP Lett. 12 (1970) 381.
[81] R.O. Popovych, N.M. Ivanova, H. Eshraghi, math-ph/0311039.
[82] G. Rideau, P. Winternitz, J. Math. Phys. 34 (1993) 558.
[83] F. Rojas, A.D. Rutenberg, Phys. Rev. E 60 (1999) 212.
[84] A.D. Rutenberg, A.J. Bray, Phys. Rev. E 51 (1995) R1641.
[85] A.D. Rutenberg, A.J. Bray, Phys. Rev. E 51 (1995) 5499.
[86] A.D. Rutenberg, Phys. Rev. E 54 (1996) R2181.
[87] A.D. Rutenberg, B.P. Vollmayr-Lee, Phys. Rev. Lett. 83 (1999) 3772.
[88] F. Sastre, I. Dornic, H. Chaté, Phys. Rev. Lett. 91 (2003) 267205.
[89] G. Schehr, P. Le Doussal, Phys. Rev. E 68 (2003) 046101.
[90] L.C.E. Struik, Physical Ageing in Amorphous Polymers and Other Materials, Elsevier, Amsterdam, 1978.
[91] U.C. Täuber, Critical Dynamics: A Field-Theory Approach to Equilibrium and Non-Equilibrium Scaling
Behaviour, 2004, in press.
[92] N. van Kampen, Stochastic Processes in Physics and Chemistry, second ed., North-Holland, Amsterdam,
1997.
[93] C. Yeung, M. Rao, R.C. Desai, Phys. Rev. E 53 (1996) 3073.
[94] J. Zinn-Justin, Quantum Field Theory and Critical Phenomena, fourth edition, Oxford Univ. Press, Oxford,
1989–2002.
[95] W. Zippold, R. Kühn, H. Horner, Eur. Phys. J. B 13 (2000) 531.
Nuclear Physics B 688 [FS] (2004) 266–290
Completeness of Bethe ansatz for 1D Hubbard

model with AB-flux through combinatorial formulas
and exact enumeration of eigenstates
Akinori Nishino a , Tetsuo Deguchi b
a Institute of Physics, University of Tokyo, 3-8-1 Komaba, Meguro-Ku, Tokyo 153-8902, Japan
b Department of Physics, Ochanomizu University, 2-1-1 Ohtsuka, Bunkyo-Ku, Tokyo 112-8610, Japan
Received 16 January 2004; accepted 1 April 2004
Abstract
For the one-dimensional Hubbard model with Aharonov–Bohm-type magnetic flux, we study
the relation between its symmetry and the number of Bethe states. First we show the existence
of solutions for Lieb–Wu equations with an arbitrary number of up-spins and one down-spin, and
exactly count the number of the Bethe states. The results are consistent with Takahashi’s string
hypothesis if the system has the so(4) symmetry. With the Aharonov–Bohm-type magnetic flux,
however, the number of Bethe states increases and the standard string hypothesis does not hold. In
fact, the so(4) symmetry reduces to the direct sum of charge-u(1) and spin-sl(2) symmetry through
the change of AB-flux strength. Next, extending Kirillov’s approach [J. Sov. Math. 30 (1985) 2298,
J. Sov. Math. 36 (1987) 115], we derive two combinatorial formulas from the relation among the
characters of so(4)- or (u(1) ⊕ sl(2))-modules. One formula reproduces Essler–Korepin–Schoutens’
combinatorial formula for counting the number of Bethe states in the so(4)-case. From the exact
analysis of the Lieb–Wu equations, we find that another formula corresponds to the spin-sl(2) case.
PACS: 71.10.Fd; 02.30.Ik; 03.65.-w
1. Introduction
Low-dimensional physics has attracted a great interest of theoretical and experimental

physicists for almost a half century [16,22,23]. Among them, the Bethe ansatz method,
which was originally developed as a non-perturbative method for diagonalizing one-
E-mail addresses: nishino@gokutan.c.u-tokyo.ac.jp (A. Nishino), deguchi@phys.ocha.ac.jp (T. Deguchi).
doi:10.1016/j.nuclphysb.2004.04.004
A. Nishino, T. Deguchi / Nuclear Physics B 688 [FS] (2004) 266–290 267
dimensional spin- 21 isotropic Heisenberg spin chain [1], opened a new realm of mathe-
matical physics. Roughly speaking, Bethe’s work consists of three parts: (i) Bethe states
are introduced and Bethe equations, which are sufficient conditions for the Bethe states
being eigenstates, are derived; (ii) the existence of solutions for the Bethe equations are
discussed in simple cases, and the general forms of solutions are conjectured, which are
called string hypothesis; (iii) under the string hypothesis, the formula for counting the num-
ber of the Bethe states is provided, which leads to the combinatorial completeness of Bethe
ansatz. At present the Bethe ansatz method is applied to various kinds of one-dimensional
spin chains and strongly correlated electron systems [16,22,30].
In Bethe’s work, the combinatorial formula for counting the number of Bethe states
possesses a wealth of mathematical implications. In the case of the isotropic Heisenberg
spin chain, the Bethe states constructed by finite-valued solutions of Bethe equations
do not produce all the eigenstates. In fact the system has sl(2) symmetry and the
Bethe states are sl(2)-highest [6]. The eigenstates other than highest weight vectors, i.e.,
sl(2)-descendant states, are constructed by applying the lowering operator to the Bethe
states. Mathematically, the number of Bethe states is interpreted as the multiplicity of
irreducible components in the tensor products of two-dimensional highest weight sl(2)-
modules. Bethe’s formula is also extended to the generalized Heisenberg spin chains
with higher spins or sl(n) symmetry, for which the powerful tools such as Q-systems are
introduced [12,13].
The application of Bethe ansatz method to the one-dimensional Hubbard model was
given by Lieb and Wu [21]. The Bethe equations for the Hubbard model are often called
Lieb–Wu equations. Takahashi’s string hypothesis asserts that, in the thermodynamic limit,
the solutions of Lieb–Wu equations are approximated by string solutions, and the number
of Bethe states is estimated under the hypothesis [28,29]. The Hubbard model with even
sites has so(4) symmetry [7,32,33]. Essler, Korepin and Schoutens proved that, when the
system has the so(4) symmetry, the Bethe states are so(4)-highest [5]. They also showed in
a combinatorial way that all the eigenstates are obtained by taking the so(4)-descendants
of the Bethe states into account [3]. On the other hand the completeness of Bethe ansatz
for the system with odd sites, which has just sl(2) symmetry related to the spin degrees of
freedom, has not been discussed.
In this article, we study the Bethe states in the one-dimensional Hubbard model. In
particular, we deal with the system on a ring with Aharonov–Bohm-type magnetic flux.
The system has so(4) symmetry only at special values of the AB-flux strength and the so(4)
symmetry reduces to spin-sl(2) symmetry for other values. More precisely, for a generic
value of the AB-flux strength, the so(4) symmetry breaks into the direct sum of the charge-
u(1) and the spin-sl(2) symmetry. Varying the AB-flux strength, we investigate solutions
of Lieb–Wu equations. Here we recall that all the enumeration of Bethe states we have
mentioned above are based on the string hypothesis. However, the violation of the string
hypothesis is numerically observed: for the spin- 21 isotropic Heisenberg spin chain, some
of the string solutions reduce to real solutions when the number of sites is large, which
is called redistribution phenomenon [4,8,9]. Thus, without making any approximation, we
discuss the existence of Bethe ansatz solutions with the AB-flux. In particular, we show
the existence of solutions of the Lieb–Wu equations with an arbitrary number of up-spins
and one down-spin. Here we employ a graphical approach [2,24]. We exactly count the
268 A. Nishino, T. Deguchi / Nuclear Physics B 688 [FS] (2004) 266–290
number of solutions in the case and verify that the enumeration with the string hypothesis
is correct only in the so(4)-case. We find that the Lieb–Wu equations for the system with
only the spin-sl(2) symmetry have more solutions than those in the so(4)-case. Next we
study the combinatorial completeness of Bethe ansatz. We obtain the relation among the
characters of so(4)-modules through the power series identities similar to Kirillov’s [12,
13], which gives a new proof of Essler–Korepin–Schoutens’ combinatorial completeness of
Bethe ansatz [3]. We also introduce a new combinatorial formula derived from the relation
among the characters of (u(1) ⊕ sl(2))-modules. The formula suggests the combinatorial
completeness of Bethe ansatz for the system only with the spin-sl(2) symmetry, which has
not been discussed in the literature.
The Bethe ansatz solutions with the AB-flux should be quite important in the low-
dimensional physics of the one-dimensional Hubbard model. As pointed out by Kohn,
the low frequency conductivity is directly related to the shift of the energy levels due to
twisted boundary conditions [15]. Sharpening Kohn’s argument on electron systems in
any dimensions, Shastry and Sutherland discussed effective mass of the one-dimensional
Hubbard model through the twisted boundary conditions [25]. Furthermore, Kawakami
and Yang obtained an explicit expression for the effective mass of the electric conductivity
for the one-dimensional Hubbard model [10,11]. For the Bethe ansatz solutions with the
twisted boundary conditions, there are other aspects such as persistent current associated
with the AB-flux.
The article is organized as follows: in Section 2, we review the symmetry of the
one-dimensional Hubbard model and the Bethe ansatz method. We also describe how to
construct eigenstates other than the Bethe states. In Section 3 we prove the existence of
solutions for the Lieb–Wu equations with one down-spin and exactly count the number of
Bethe states in varying the strength of AB-flux. In Section 4 we study the combinatorial
formulas for counting the Bethe states in terms of the characters of so(4)- and (u(1) ⊕
sl(2))-modules. The final section is devoted to summary and concluding remarks.
2. Bethe ansatz method
2.1. Hubbard model and Lieb–Wu equations
We introduce the Hubbard model on an L-site ring with Aharonov–Bohm-type

†
magnetic flux Φ. Let cis and cis , (i ∈ Z/LZ, s ∈ {↑, ↓}) be the creation and annihilation
†
operators of electrons satisfying {cis , cj t } = {cis , cj†t } = 0 and {cis , cj†t } = δij δst , and
†
define the number operators by nis := cis cis . We consider the Fock space V of electrons
with the vacuum state |0 (dim V = 4 ). The one-dimensional Hubbard model is described
L
by the following Hamiltonian acting on V :

√ √
Hφ = − e −1φ cis
†
ci+1,s + e− −1φ ci+1,s
†
cis
1iL s=↑,↓
1

1

+U ni↑ − ni↓ − , (2.1)
2 2
1iL
where we assume φ := Φ/L ∈ R/2πR and U > 0. It is clear that Hφ = Hφ+2π .

Furthermore Hφ has the same energy spectra as those of Hφ+ 2π . Hence we often restrict
L
the region of φ to 0 φ < 2π L in what follows.
Let N be the number of electrons and M that of down-spins. We assume 0 2M
N L. Let {ki | i = 1, 2, . . . , N}, (Re(ki ) ∈ R/2πR) denote a set of wavenumbers of N
electrons and {λα | α = 1, 2, . . . , M} that of rapidities of M down-spins. Given a set of spin
configuration {si | i = 1, 2, . . . , N} with N − M up-spins and M down-spins, the Bethe
state with {ki , λα } has the following form:
φ

|k, λ; sN,M = ψk,λ (x; s)cx†1 ,s1 cx†2 ,s2 · · · cx†N ,sN |0. (2.2)
{1xi L}
The coefficients ψk,λ (x; s) in (2.2) are explicitly given in [31]. The Bethe states (2.2) are
eigenstates of the Hamiltonian (2.1) if {ki , λα } satisfy the following equations:
√
√ λβ − sin(ki + φ) − −1U/4
−1ki L
e = √ ,
1βM β
λ − sin(ki + φ) + −1U/4
√ √
λα − sin(ki + φ) − −1U/4 λα − λβ − −1U/2
√ = √ , (2.3)
1iN α
λ − sin(ki + φ) + −1U/4 β(=α) λα − λβ + −1U/2
which are coupled non-linear equations called Lieb–Wu equations [21]. The Lieb–Wu
equations have not been solved analytically. However it predicts some important results
on thermodynamic properties of the system through Takahashi’s string hypothesis [16,28–
30]. In terms of the solutions {ki , λα } of the Lieb–Wu equations (2.3), energy eigenvalues
are written as
φ φ
Hφ |k, λ; sN,M = E|k, λ; sN,M ,
1
E = −2 cos(ki + φ) + U (L − 2N). (2.4)
4
1iN
Recall that the Bethe states (2.2) give only the eigenstates with 0 2M N L. In order
to construct other eigenstates, we need to consider the symmetries of the system.
φ
Hereafter we shall sometimes abbreviate the superscript φ in |k, λ; sN,M .
2.2. Symmetries
The U -independent symmetries of the Hubbard model are classified in [7]. First we
review the symmetries connected with the spin and charge degrees of freedom [26,27,32,
33]. Define the following operators related to the spin degrees of freedom:
1 †
Sz := (ni↑ − ni↓ ), S+ := ci↑ ci↓ , S− := (S+ )† . (2.5)
2
1iL 1iL
They give a representation of the algebra sl(2) on the Fock space V . Since all the operators
{Sz , S± } commute with Hφ (2.1) for an arbitrary value of φ, it is said that the system has
spin-sl(2) symmetry. One finds another representation of sl(2) related to the charge degrees
of freedom,
1
ηz := (1 − ni↑ − ni↓ ),
2
1iL
√
η+ := e −1(2φ+π)i ci↓ ci↑ , η− := (η+ )† . (2.6)
1iL
Note that all the operators in (2.6) are commutative with {Sz , S± } (2.5). It is easy to see
that the operator ηz also commutes with Hφ for an arbitrary value of φ. However other
operators η± commute with Hφ only for the special values of φ satisfying L2 + Lφ π ∈ Z.
Thus the system has charge-sl(2) symmetry if L2 + Lφ π ∈ Z, and it reduces to charge-u(1)
symmetry given by ηz for other values of φ. Combining the above two kinds of sl(2)
symmetries, we see that the system has so(4)( sl(2) ⊕ sl(2)) symmetry if L2 + Lφ π ∈ Z,
and u(1) ⊕ sl(2) symmetry otherwise. For simplicity, we call the former so(4)-case and the
latter sl(2)-case. For the later discussion, we define the Casimir operators for each sl(2),
1 1
η2 := (η+ η− + η− η+ ) + ηz2 , S 2 := (S+ S− + S− S+ ) + Sz2 ,
2 2
which are employed to see the dimension of each representation.
Next we introduce the following three operators:
†
Ts := Pi↑; i↓ , Tph := cis + cis ,
1iL 1iL s=↑,↓

Tr := Pis; L−i,s ,
s=↑,↓
2
1i L−1
†
where Pis; j t := 1 − (cis − cj†t )(cis − cj t ) and x
denotes the greatest integer in x. One
notices that Ts induces a spin-reversal transformation, Tph a particle–hole transformation
and Tr a reflection of the lattice. Direct calculation shows
−1
Ts−1 Hφ Ts = Hφ , Tph Hφ Tph = Hπ−φ , Tr−1 Hφ Tr = H−φ .
For the system with even L, i.e., a bipartite lattice, we also introduce

Tb := (−1)n2i,s , Tb−1 Hφ Tb = Hφ+π .
1i L2 s=↑,↓
Combining these operations, we obtain the following transformation properties of the

Hamiltonian Hφ (2.1):
−1 −1
Ts−1 Hφ Ts = Hφ , Tph Tr Hφ Tr Tph = Hφ+π , for even and odd L,
−1 −1 −1
Tph Tr Tb Hφ Tb Tr Tph = Hφ , for even L. (2.7)
One notices that for both even and odd L, the system has spin-reversal symmetry. While
the system with even L has particle–hole symmetry brought by the transformation Tb Tr Tph ,
the system with odd L does not have particle–hole symmetry for generic values of φ except
the special values satisfying L2 + Lφ
π ∈Z
2.3. Construction of non-Bethe states
We construct eigenstates that are not included in the Bethe states (2.2). First we consider
the system with so(4) symmetry, i.e., both spin-sl(2) and charge-sl(2) symmetries. Here
we recall that the AB-flux parameter φ takes a special value: L2 + Lφ
π ∈ Z for even or odd L.
Then, as shown in [5] for even L with φ = 0, the Bethe states (2.2) characterized by finite-
valued solutions of the Lieb–Wu equations (2.3) are so(4)-highest, i.e., η+ |k, λ; sN,M =
S+ |k, λ; sN,M = 0. Since
η2 |k, λ; sN,M = η(η + 1)|k, λ; sN,M ,

S 2 |k, λ; sN,M = S(S + 1)|k, λ; sN,M , (2.8)
with η = 12 (L − N) and S = 12 (N − 2M), the Bethe state |k, λ; sN,M is the highest weight
vector of an (L − N + 1)(N − 2M + 1)-dimensional highest weight so(4)-module. Here
we note that η + S is an integer for even L [33], while it is a half-integer for odd L. The
so(4)-descendant states of the Bethe state |k, λ; sN,M ,
(η− )n (S− )m |k, λ; sN,M (0 < n L − N, 0 < m N − 2M), (2.9)

are also eigenstates of Hφ (2.1) (see Fig. 1). They are energy degenerate with |k, λ; sN,M .
It is easy to see that such application of the lowering operators η− and S− to the Bethe
states produces other eigenstates than those with 0 2M N L. Essler, Korepin and
Schoutens counted the number of Bethe states (2.2) and their so(4)-descendant states (2.9)
under the string hypothesis to show the combinatorial completeness of Bethe ansatz [3].
Next we consider the system only with the charge-u(1) and the spin-sl(2) symmetries.
The Bethe states (2.2) are sl(2)-highest and satisfy only the second relation in (2.8), which
φ
means that the Bethe state |k, λ; sN,M is the highest weight vector of an (N − 2M + 1)-
dimensional highest weight sl(2)-module. The sl(2)-descendant states of the Bethe states
φ
(S− )m |k, λ; sN,M (0 < m N − 2M), (2.10)
Fig. 1. Bethe states and non-Bethe states for L = 5. The Bethe state with (N, M) = (3, 1) denoted by a closed
circle produces five so(4)-descendants denoted by open circles if the system has so(4) symmetry. As the so(4)
symmetry reduces to sl(2) symmetry, the six eigenstates form three doublets of sl(2). Then we need one more
Bethe state with (5, 2). Note that the eigenstate with (7, 3) is not a Bethe states and is constructed through the
transformation Tr Tph .
are eigenstates of Hφ (2.1) (see Fig. 1). One notices here that the Bethe states (2.2) and
their sl(2)-descendants (2.10) do not produce the eigenstates with L < N 2L since
the application of the lowering operator S− does not change the number of electrons.
Such eigenstates with L < N 2L are constructed as follows: (i) by applying the
φ
transformation Tb Tr Tph to the Bethe state |k, λ; s2L−N,L−M (L < N 2M 2L) if L
φ+π
is even, or by applying Tr Tph to the Bethe state |k, λ; s2L−N,L−M (L < N 2M 2L)
obtained by the Lieb–Wu equations (2.3) with φ + π instead of φ if L is odd, we get the
lowest weight vector of a highest weight sl(2)-module; (ii) the application of the raising
operators (S+ )n , (0 < n 2M − N) to the lowest weight vector produces other degenerate
eigenstates.
Even if we are interested only in the system with φ = 0, the Bethe ansatz method needs
the system with φ = π for odd L. One of the main purposes in this article is therefore to
discuss whether or not the above procedure produces all the eigenstates.
3. Lieb–Wu equations with one down-spin
By employing a graphical approach, we exactly count the number of finite-valued

solutions for the Lieb–Wu equations (2.3) in the case when the system contains an arbitrary
number of up-spins and one down-spin, i.e., N 2 and M = 1. We remark that such exact
analysis is presented in [2,24] for the case φ = 0. In this section, we assume 0 φ < 2π L
since Hφ has the same energy spectra as those of Hφ+ 2π . In the case M = 1, the string
L
hypothesis [28] predicts that two types of solutions exist; one is the solution with only real
wavenumbers {ki } and another includes two complex wavenumbers. We investigate such
types of solutions below.
3.1. Real k solutions
First we consider the real solutions. For M = 1, the Lieb–Wu equations (2.3) reduce to
√
√
−1ki L λ − sin(ki + φ) − −1U/4
e = √ (i = 1, 2, . . . , N),
λ − sin(ki + φ) + −1U/4
√
λ − sin(ki + φ) − −1U/4
√ = 1. (3.1)
1iN
λ − sin(ki + φ) + −1U/4
These are equivalent to the following equations:

U ki L √
sin(ki + φ) − λ = cot , exp −1 ki L = 1. (3.2)
4 2
1iN
We investigate the real solutions for the first equation

U qL
sin(q + φ) − λ = cot . (3.3)
4 2
Fig. 2. The generic behaviour of sin(q + φ) − λ and U4 cot( qL2 ) in (3.3) under the condition U L > 8 for L = 6.
Horizontal dashed lines divide six branches as specified by (3.4). Each branch has an intersection which leads to
a solution of (3.3).
In the interval 0 q < 2π , its right-hand side has L branches

2π 1 2π 1 2j − 1

− <q <
+ ,
∈ j = 1, 2, . . . , L . (3.4)
L 2 L 2 2
Note that, with a given

satisfying (3.4), the first equations in (3.1) are rewritten as

λ − sin(ki + φ)
ki L = 2π
− 2 arctan ,
U/4
which are convenient to relate the
to the (half-)integers appearing in the string
hypothesis [28]. By regarding λ as a real parameter, we seek a solution q of (3.3) in one of
the branches (3.4). From the graphical discussion (Fig. 2), the solution in the branch
is
uniquely determined for arbitrary λ under the following condition:

d d U qL UL
−1 = min sin(q + φ) − λ > max cot =− .
0q<2π dq 0q<2π dq 4 2 8
Hence, for U > L8 , the solution of (3.3) can be written as an increasing function of λ,
i.e., q = q
(λ). Given a non-repeating set {
i | 1 i N} ⊂ { 2j2−1 | 1 j L} of the
L
branches, the second equation in (3.2) is satisfied when 2π i q
i (λ) ∈ Z. The behaviour
of the solution q = q
(λ) tells us that
L 1

lim q
i (λ) =
i ± . (3.5)
λ→±∞ 2π 2
1iN 1iN

Thus there exist N − 1 values of λ giving the following integer values for L
2π i q
i (λ):

1

m∈
i − + j
j = 1, 2, . . . , N − 1 .
2
1iN
Note that such {λ} and integers {m} are in one-to-one correspondence due to dqdλ
(λ)
> 0.
As a consequence, the solutions {ki = q
i (λ), λ} are specified by the set of indices {
i , m}.
L
The number of possible {
i , m} is given by ( N )(N − 1). Here we note that the number is
consistent with the formula Z(L; N, M) in [3,28].
L
Proposition 3.1. For U > 8
L, the Lieb–Wu equations (2.3) with M = 1 have ( N )(N − 1)
real solutions [2,24].
3.2. (k–Λ)-string solutions
Next we consider the solutions including a couple of complex wavenumbers. We assume

the form of solutions as
ki ∈ R/2πR (i = 1, 2, . . . , N − 2),
√ √
kN−1 = ζ − −1ξ, kN = ζ + −1ξ,
where 0 ζ < 2π and ξ > 0. Note that kN−1 and kN form a complex conjugate pair
which is referred to as (k–Λ–2)-string. Then the first set of equations in (3.1) are rewritten
in terms of real variables as follows

U ki L
sin(ki + φ) − λ = cot (i = 1, 2, . . . , N − 2), (3.6a)
4 2
U sin(ζ L)
sin(ζ + φ) cosh ξ − λ = , (3.6b)
4 cosh(ξ L) − cos(ζ L)
U sinh(ξ L)
cos(ζ + φ) sinh ξ = − . (3.6c)
4 cosh(ξ L) − cos(ζ L)
On the other hand, the second equation in (3.1) is equivalent to the following condition:
2π
ki + 2ζ = m, with m = 0, 1, . . . , NL − 1. (3.7)
L
1iN−2
In the same way as the previous case of Section 3.2, if we consider a solution of each
Eq. (3.6a) in one of the branches (3.4), and the solution can be written as a function of λ.
Given a set {
i | 1 i N − 2} of non-repeating indices specifying the branches (3.4),
we express the solutions of (3.6a) as ki = q
i (λ) (1 i N − 2). Then, from the relation
(3.7), the ζ is also written as a function of λ,
π 1
ζ = ζ (λ) := m − q
i (λ), (3.8)
L 2
1iN−2
for fixed {
i } and m.
For an illustration, we consider (3.6b) and (3.6c) in the case N = 2. Since ζ does not
depend on λ in the case, Eqs. (3.6b) and (3.6c) decouple into the following:

π U
λ = sin m + φ cosh ξ, sinh ξ = − f (2) (ξ ), (3.9)
L 4 cos( πL m + φ)
where

sinh(ξ L) coth(ξ L/2), for m ∈ 2Z,
f (2) (ξ ) := =
cosh(ξ L) − (−1)m tanh(ξ L/2), for m ∈ 2Z + 1.
We seek a solution of the second equation in (3.9) through graphical discussion. Since
f (2)(ξ ) > 0 when ξ > 0, we need the condition π2 < πL m + φ < 3π 2 so that the second
equation of (3.9) has a solution. If such m is even, it is straightforward that the second
equation in (3.9) determines a unique solution ξ(> 0) since limξ →∞ f (2) (ξ ) = 1. For odd
m, the equation has a unique solution if the condition

d(sinh ξ ) U df (2) UL
1= (0) < min − (0) = ,
dξ m∈2Z+1 4 cos( πL m + φ) dξ 8
π
2 < πL m+φ< 3π
2
is satisfied. The number of allowed values of m here depends on L and φ,

{ L2 − L
π φ + j | j = 1, 2, . . . , L − 1}, 2 − π φ ∈ Z,
L L
m∈
{ L2 − L
π φ
+ j | j = 1, 2, . . . , L}, 2 − πφ ∈
L L
/ Z.
(2) (2)
Let ξm denote the solution specified by the m. By substituting the solution ξm into the
first equation in (3.9), one immediately obtains λ. Recall that, in the cases of L2 − L
π φ ∈ Z,
i.e., even L, (respectively, odd L) and φ = 0 or L , (respectively, φ = 2L or 2L ), the
π π 3π
system has the so(4) symmetry. Thus, as the so(4) symmetry reduces to the spin-sl(2)
symmetry through the change of AB-flux strength φ, the number of solutions for the Lieb–
Wu equations increases.
Let us consider the case N > 2. By inserting (3.8) into (3.6c), we have
U
sinh ξ = − fa (ξ ), (3.10a)
4 cos(ζ (λ) + φ)

sinh(ξ L) L
fa (ξ ) := , a = (−)m cos q
i (λ) . (3.10b)
cosh(ξ L) − a 2
1iN−2
Note that |a| 1. One sees that, for a fixed a, the function fa (ξ ) has the following
properties: (i) fa (0) = 0, df dξ (0) 2 , limξ →∞ fa (ξ ) = 1; (ii) if a < 0, fa (ξ ) is
a L
monotonically increasing and concave; (iii) if a > 0, fa (ξ ) has a single positive maximum
at ξ (> 0) and a single turning point at ξ (> ξ ). These properties are sufficient to discuss
the solution ξ of (3.10) for arbitrary λ. From the graphical discussion similar to the case
N = 2 (see Fig. 3), this determines a unique ξ as a function of λ under the condition

d(sinh ξ ) U dfa UL UL
1= (0) < min − (0) = min = ,
dξ |a|1 4 cos(ζ + φ) dξ |a|1 4(1 − a) 8
π
2 <ζ +φ< 3π
2
U
Fig. 3. The behaviour of sinh ξ and − 4 cos(ζ +φ) fa (ξ ) in (3.10a) under the conditions U L > 8 and
π < ζ + φ < 3π . The ξ and ξ respectively denote the maximum and the turning point of − U
2 2 4 cos(ζ +φ) fa (ξ ) in
the cases 0 < a < 1.

if and only if π2 < ζ (λ) + φ = πL m + φ − 12 i q
i (λ) < 3π2 . By using (3.5), it is sufficient
to have a unique solution for (3.10) that the integer m satisfies

1 L L 1 3L L

i + + − φ<m<
i − + − φ,
2 2 π 2 2 π
1iN−2 1iN−2
that is,

i (
i + 12 ) + 2 − π φ + j | j = 1, 2, . . . , L − N + 1 ,
L L L
−L
π φ ∈ Z,
m∈ 2
i (
i + 2 ) + 2 − π φ
+ j | j = 1, 2, . . . , L − N + 2 ,
1 L L L
2 −L
πφ ∈
/ Z.
(3.11)
(2)
Note that limλ→±∞ ξ(λ) = ξ , which is well-defined for the above m. We see
m− i (
i ± 12 )
that, for the values of m given in (3.11), Eq. (3.6b) with ξ(λ) and ζ(λ)

π 1
λ = sin m+φ− q
i (λ) cosh ξ(λ)
L 2
i

U sin( L2 i q
i (λ))
−
4 cos( L2 i q
i (λ)) − (−)m cosh(ξ(λ)L)

=: g q
i (λ) , ξ(λ) , (3.12)
determines λ. In fact, since q

i (λ) and ξ(λ) are continuous functions of λ and the function
g satisfies the following:

2π 1 (2)
lim g q
i (λ) , ξ(λ) = g
i ± ,ξ .
λ→±∞ L 2 m− i (
i ± 12 )
Here g is a continuous and finite function with respect to λ. Hence there exists a solution
λ for Eq. (3.12).
L
Proposition 3.2. For U > L8 , the Lieb–Wu equations (2.3) with M = 1 have ( N−2 )(L −
N + 1) (k–Λ–2)-string solutions if the system has the so(4) symmetry, and they have
L
( N−2 )(L − N + 2) (k–Λ–2)-solutions, otherwise.
One notices that, only for the system with the so(4) symmetry, the number of (k–Λ–
2)-solutions is consistent with the string hypothesis [28]. Note that, for 0 < U < L8 , some
of the (k–Λ–2)-strings may disappear. In Appendix A, we numerically investigate the case
N = 2 and show that, for 0 < U < L8 , the (k–Λ–2)-strings with odd m disappear, while
additional real solutions appear [5]. For the system with only the spin-sl(2) symmetry,
the Lieb–Wu equations have more (k–Λ–2)-solutions than those expected by the string
hypothesis.
3.3. Completeness of Bethe ansatz for L = 3
Applying the above results, we now show that all the eigenstates can be constructed
through the Bethe ansatz method in a simple case. We consider the case L = 3 and
φ = 2L
π 3π
, 2L when the system does not have so(4) symmetry but the spin-sl(2) symmetry.
We note that the completeness of eigenstates in this situation has not been discussed in the
literature.
φ
The number of Bethe states |k, λ; sN,M (0 2M N 3) is exactly calculated as
follows: the case M = 0 is trivial since the eigenstates are those of lattice free fermion
system; for the cases (N, M) = (2, 1) and (3, 1), we have obtained the following formulas
from Propositions 3.1 and 3.2:
3 N−1
, for real solutions,
(Bethe states) = N 3 15−N
N−2 1 , for (k–Λ–2)-string solutions.
φ
Since each Bethe state |k, λ; sN,M corresponds to the highest weight vector of a highest
weight sl(2)-module, we should count their sl(2)-descendant states
φ
(S− )n |k, λ; sN,M (0 < n N − 2M).
The eigenstates with 4 N 6, which are not Bethe states nor their sl(2)-descendant
states, are constructed through the transformation Tr Tph as we have described in
φ+π
Section 2.3. Indeed, by applying Tr Tph to the Bethe state |k, λ; s6−N,3−M (4 N
2M 6) of the system described by the Hamiltonian Hφ+π , we get the eigenstate
φ+π
Tr Tph |k, λ; s6−N,3−M of Hφ with 4 N 2M 6 which is the lowest weight vector
Table 1
π , 3π
Enumeration of eigenstates for L = 3 and φ = 2L 2L
N M 6−N 3−M Type of solutions (Bethe) sl(2) sym. (state)
0 0 real 1 1 1
1 0 real 3 2 6
2 0 real 3 3 9
2 1 real 3 1 3
2 1 (k–Λ–2)-string 3 1 3
3 0 real 1 4 4
3 1 real 2 2 4
3 1 (k–Λ–2)-string 6 2 12
4 2 2 1 real 3 1 3
4 2 2 1 (k–Λ–2)-string 3 1 3
4 3 2 0 real 3 3 9
5 3 1 0 real 3 2 6
6 3 0 0 real 1 1 1
64
of a highest weight sl(2)-module. We also count the eigenstates

φ+π
(S+ )n Tr Tph |k, λ; s6−N,3−M (0 < n 2M − N).
Table 1 indeed shows that we obtain 64 = 43 = dim V eigenstates, which give a complete
system of the Fock space V .
4. Combinatorial completeness of Bethe ansatz
The Bethe ansatz method was first introduced in the case of one-dimensional spin- 21
isotropic Heisenberg spin chain [1]. Bethe assumed the string hypothesis and estimated the
number Z(N; M) of solutions for the Bethe equations with M down-spins on an N -site
chain as
P + M
m m
Z(N; M) = , (4.1)
Mm
{M
n } m1
M= nMn
Mn denotes the number of n-strings composing a solution and Pn = N − 2M +

where
2 m(>n) (m − n)Mm . He obtained the following summation formula:

N N
Z(N; M) = − (4.2)
M M −1
which implies that the number Z(N; M) of Bethe states is interpreted as the multiplicity
of (N − 2M + 1)-dimensional irreducible sl(2)-modules in the tensor product of N two-
dimensional irreducible sl(2)-modules. By taking into account that the Bethe states are
sl(2)-highest and generate (N − 2M + 1) sl(2)-descendant states, the completeness of
Bethe ansatz is shown in a combinatorial way as

(N − 2M + 1)Z(N; M) = 2N ,
0M N/2
where x
denotes the greatest integer in x.
It is known that, in general, solutions of Bethe equations do not have the nature assumed
in the string hypothesis [4,8,9]. Indeed, for N > 21, some of the 2-string solutions are
redistributed to real solutions that are not counted in Z(N; M) [4]. Hence, when we
actually employ the formula Z(N; M) to show the completeness of Bethe ansatz, we must
regard such redistributed real solutions as 2-string solutions.
We apply the techniques developed in [12,13] to the Hubbard model with so(4)
symmetry. Indeed a new proof for Essler–Korepin–Schoutens’ combinatorial completeness
of Bethe ansatz [3] is obtained as a corollary of the relation among the characters of so(4)-
modules. Moreover, based on the results in the previous section, we propose the conjectural
formula related to the combinatorial completeness of Bethe ansatz for the system with
only the charge-u(1) and spin-sl(2) symmetry. In both cases, we obtain the formulas
corresponding to (4.2), which has not been established even for the so(4)-case.
4.1. Kirillov’s power series and Q-system
First we give three lemmas introduced in the case of one-dimensional isotropic

Heisenberg spin chain [12,13]. Detailed proofs are given in [12,13]. Let a1 , a2 , . . . , al be
a set of integers for l L. Define a set of formal power series {ϕn (zn , zn+1 , . . . , zl ) | 1
n l} by
ψn (z) := (1 − z)−an +2M−1 ,

ϕn (zn , zn+1 , . . . , zl ) := ψn (zn )ϕn+1 (1 − zn )−2 zn+1 , . . . , (1 − zn )−2(l−n) zl .
Lemma 4.1. The power series ϕn (zn , . . . , zl ) has the following expression:
Pm (am ) + Mm
ϕn (zn , . . . , zl ) = znMn · · · zlMl ,
Mm
Mn ,...,Ml 0 nml
for 1 n l,
where

Pn (an ) = an − 2M + 2 (m − n)Mm .
m>n
Proof. The case n = l is given by the formula

α +m
−α−1
(1 − z) = zm .
m
m0
Then the case 1 n < l is proved by induction on n.

(k)
Introduce the variables {zn | 1 n l, 0 k n} through
(k−1) −2(n−k) (k−1)
zn(0) = zn , zn(k) = 1 − zk zn , for 1 k n.
Lemma 4.2. The power series ϕ1 (z1 , . . . , zl ) is rewritten in terms of the variables {zn(k) },
−a +2M−1
ϕ1 (z1 , . . . , zl ) = 1 − zn(n−1) n .
1nl
Proof. Since
(k−1) (k−1)
ϕk zk , . . . , zl
(k−1) (k−1) −2 (k−1) (k−1) −2(l−k) (k−1)
= ψk zk ϕk+1 1 − zk zk+1 , . . . , 1 − zk zl
−a +2M−1
= 1 − zk(k−1) k (k)
ϕk+1 zk+1 , . . . , zl(k) ,
for 1 k l − 1, the lemma is proved.
Define the polynomials {Qn = Qn (t) | n ∈ Z0 } through the recursion relation,
Qn+2 = Qn+1 − tQn , Q0 = Q1 = 1. (4.3)
Lemma 4.3.
(i) Q2n = Qn+1 Qn−1 + t n , for n 1, (4.4a)

−an +2M−1
(ii) ϕ1 t, t 2 , . . . , t l = Qn+1 Q−2
n Qn−1 , (4.4b)
1nl
(1 − v)n−1 (1 − v n+1 ) v
(iii) Qn t (v) = , where t (v) := . (4.4c)
(1 − v 2 )n (1 + v)2
Proof.
(i) Use induction on n.

(0) (k)
(ii) Set zn = zn = t n (1 n l) in {zn }. One obtains the following relations:
1 − zk(k−1) = Qk+1 Q2k Qk−1 ,

zn(k) = Q−2(n−k)
k+1 Qk2(n−k−1) t n , for 1 k n.
Combining these with Lemma 4.2, we can prove (4.4b).
(iii) One can directly verify that the Qn (t (v)) satisfies the recursion relation (4.3) with
t = t (v). 2
The relations (4.4a) are called the Q-system of type sl(2), which is a key object in [12].
Indeed the expression (4.4b) produces an identity among the characters of sl(2)-modules.
The Q-system also plays a significant role in the combinatorial identities associated with
the XXZ-Heisenberg spin chain and its generalizations [14,17–19].
4.2. Combinatorial formulas
Using the above lemmas, we discuss the Hubbard-case. In the similar way, we
define ϕn (zn , zn+1 , . . . , zl ) and Pn (an ) by replacing an and 2M with an and N in
ϕn (zn , zn+1 , . . . , zl ) and Pn (an ), respectively. Define

ϕ(s, t) := (1 + s)L ϕ1 s 2 t, s 4 t 2 , . . . , s 2l t l ϕ1 t, t 2 , . . . , t l .
The following is straightforward from Lemma 4.1:
L P (a ) + M Pn (an ) + Mn
n n n
ϕ(s, t) =
Nr Mn Mn
0Nr L 1nl
{Mn ,Mn }

× s Nr +2 m1 mMm t m1 m(Mm +Mm ) .
Then the coefficient of s N t M (0 2M N L) in ϕ(s, t) is expressed by

Z L, {an , an }; N, M

L Pn (an ) + Mn Pn (an ) + Mn
:= ,

Nr Mn Mn
{Nr ,Mn
,Mn } 1nl

r +2 nMn
N=N
M= n(Mn +Mn )
where
the sum runs over
all configurations {Nr , Mn , Mn 0} such that N = Nr +

2 n0 nMn and M = n1 n(Mn + Mn ). We calculate explicit forms for the power
series ϕ(s, t) after taking the special values of {an } and {an }.
Introduce

L Pn + Mn Pn + Mn
Z(L; N, M) = , (4.5a)

Nr Mn Mn
{Nr ,Mn,Mn } n1
N=N
r +2 nMn
M= n(Mn +Mn )

L Pn + Mn + n Pn + Mn
Z̃(L; N, M) = , (4.5b)
Nr Mn Mn
{Nr ,Mn,Mn } n1
N=N
r +2 nMn
M= n(Mn +Mn )
where

Pn = L − N + 2
(m − n)Mm , Pn = N − 2M + 2 (m − n)Mm .
m>n m>n
Note that Pn , Pn 0 due to 0 2M N L.

The Z(L; N, M) (4.5a) is the very number of Bethe states for the Hubbard model
estimated under the string hypothesis [3,28,30]. In terms of the string hypothesis, Nr
denotes the number of real k’s, Mn the number of (Λ–n)-strings, and Mn the number
of (k–Λ–2n)-strings. We have verified in the previous section that, if the system has so(4)
symmetry, the number

L N −1 L L−N +1
Z(L; N, 1) = + ,
N 1 N −2 1
gives the correct number of solutions for the Lieb–Wu equations (2.3) with M = 1. In fact
the first and second terms in the above Z(L; N, 1) correspond to the following two cases:
N real k’s (Nr = N ) with a (Λ–1)-string (M1 = 1), and N − 2 real k’s (Nr = N − 2) with
a (k–Λ–2)-string (M1 = 1), respectively.
We now propose the Z̃(L; N, M) (4.5b) as a formula counting the number of Bethe
states for the system with charge-u(1) and spin-sl(2) symmetries. Indeed

L N −1 L L−N +2
Z̃(L; N, 1) = + ,
N 1 N −2 1
is consistent with the number and the string-type of solutions for the Lieb–Wu equa-
tions (2.3) with M = 1 in the sl(2)-case. The first term corresponds to the case of N real
k’s (Nr = N ) with a (Λ–1)-string (M1 = 1), and the second term to the case of N − 2
real k’s (Nr = N − 2) with a (k–Λ–2)-string (M1 = 1). We note that, for L < N 2L, we
φ+π
interpret Z̃(L; N, M) as the number of the lowest weight vectors Tr Tph |k, λ; s2L−N,L−M .
To derive Z̃(L; N, M) (4.5b) for M 2 from Takahashi’s string center equations [28,30],
we need to appropriately extend the region of the allowed (half-)integers characterizing the
Bethe states.
In both the so(4)- and sl(2)-cases, one must also take a redistribution phenomenon into
consideration [3,4,8,9]; what it means here will be more clear in Appendix A.
Proposition 4.4. We have the following identities:

(i) (1 + u)L (1 + uv)L 1 − u2 v (1 − v) = Z(L; N, M)uN v M ,
0N2L+2
0ML+1

(ii) (1 + u)L (1 + uv)L (1 − v) = Z̃(L; N, M)uN v M . (4.6)
0N2L
0ML+1
Proof. By using Lemma 4.3, we have

−a +N−1
ϕ(s, t) = (1 + s)L Qn+1 Qn−2 Qn−1 n
1nl
−an +2M−1
× Qn+1 Q−2
n Qn−1 ,
where Qn (t) := Qn (s 2 t). Through the change of variables
u(1 + v) v (1 − u2 v)(1 − v)
s= , t= , ds dt = du dv,
1 + u2 v (1 + v)2 (1 + u2 v)2 (1 + v)2
the coefficient of s N t M in ϕ(s, t) is calculated as

Z L, {an , an }; N, M
ds dt
= Res ϕ(s, t)
s=0,t =0 s N+1 t M+1

(1 + u)L (1 + uv)L −a +N−1
= Res Qn+1 Qn−2 Qn−1 n
u=0,v=0 (1 + u v)
2 L
1nl
−an +2M−1
× Qn+1 Q−2
n Qn−1

(1 + u2 v)N+1 2M+2 (1 − u v)(1 − v)
2 du dv
× (1 + v)
(1 + v)N+1 (1 + u2 v)2 (1 + v)2 uN+1 v M+1

(1 + u)L (1 + uv)L (1 − u2 v)(1 − v)
= Res
u=0,v=0 (1 + u2 v)L (1 + v)N

du dv
2n n bn n bn
× 1−u v 1−v ,
uN+1 v M+1
1nl
where we have introduced
bn = −an + 2an−1 − an−2 ,

bn = −an + 2an−1

− an−2 (1 n l),
with a−1 = a0 = a−1 = a = 0. Note that, in the third equality, we have employed the
0

assumption l L. Setting an1 = L and an1 = N , we have −b1 = b2 = L, −b1 = b2 =

N and bn3 = bn3 = 0, which proves the identity (i) in (4.6). And, setting an1 = L + n

and an1 = N , we have b1 = −L − 1, b2 = L, −b1 = b2 = N and bn3
= bn3 = 0,
which proves the identity (ii). 2
Through the change of variables u = xy −1 and v = y 2 , we have

L
x + x −1 + y + y −1 x − x −1 y − y −1

= Z(L; N, M)x −L+N−1 y −N+2M−1 , (4.7a)
0N2L+2
0ML+1
L
x + x −1 + y + y −1 y − y −1

= Z̃(L; N, M)x −L+N y −N+2M−1 . (4.7b)
0N2L
0ML+1
First we consider the relation (4.7a). If we express the left-hand side of (4.7a) as F (x, y),
we find the property F (x −1 , y) = F (x, y −1 ) = −F (x, y). Then the first relation (4.7a) is
rewritten as
L
x + x −1 + y + y −1
x L−N+1 − x −L+N−1 y N−2M+1 − y −N+2M−1
= Z(L; N, M) , (4.8)
x − x −1 y − y −1
0NL
0M N/2
where x
denotes the greatest integer in x. Let Vn be an (n + 1)-dimensional irreducible
sl(2)-module associated with the charge-sl(2) symmetry and Vn that associated with the
spin-sl(2) symmetry. We introduce an (n + 1)(m + 1)-dimensional irreducible so(4)-
module by the tensor product Vn,m = Vn ⊗ Vm . Through the representation (2.5) and
(2.6) of so(4), the Fock space V of the L-site system is isomorphic to the tensor product
(V1,0 ⊕ V0,1 )⊗L as an so(4)-module. Note that

n n m m
η2 |v = + 1 |v, S 2 |v = + 1 |v, for |v ∈ Vn,m ⊂ V .
2 2 2 2
We now decompose the Fock space V into the direct sum of Vn,m . From the relations (2.8),
each Bethe state |k, λ; sN,M corresponds to the highest weight vector belonging to
VL−N,N−2M = V2η,2S . The characters of the so(4)-module Vn,m are calculated as
x n+1 − x −n−1 y m+1 − y −m−1
ch Vn,m = .
x − x −1 y − y −1

To be precise, x = eΛ1 and y = eΛ1 where both Λ1 and Λ1 are the fundamental weight of
sl(2) and they are orthogonal to each other. One notices that, in terms of the characters, the
identity (4.8) can be rewritten as

(ch V1,0 + ch V0,1 )L = Z(L; N, M) ch VL−N,N−2M . (4.9)
0NL
0M N/2
Theorem 4.5 (Multiplicity formula). The multiplicity of the irreducible component

VL−N,N−2M in the tensor product (V1,0 ⊕ V0,1 )⊗L is given by Z(L; N, M),

(V1,0 ⊕ V0,1 )⊗L = Z(L; N, M)VL−N,N−2M . (4.10)
0NL
0M N/2
Next we turn to the relation (4.7b). If we express the left-hand side of (4.7b) as F̃ (x, y),
we find F̃ (x, y −1 ) = −F̃ (x, y). This gives
L y N−2M+1 − y −N+2M−1
1 + x y + y −1 + x 2 = Z̃(L; N, M) x N .
y − y −1
0N2L
0M N/2
(4.11)
(N)
Let ⊂ V be a subspace of the Fock space V with N electrons and let
V (N) ⊂ V (N) Vm
be an (m + 1)-dimensional irreducible sl(2)-module related to the spin-sl(2) symmetry in
(N)
V (N) . The Vm can be regarded as a 1 × (m + 1)-dimensional irreducible (u(1) ⊕ sl(2))-
module by employing the representation (2.5) of sl(2) and taking u(1) = C(L − 2ηz ) with
(2.6). Here

m m
(L − 2ηz )|v = N|v, S |v =
2
+ 1 |v, for |v ∈ Vm(N) ⊂ V .
2 2
The Fock space V of the L-site Hubbard model is isomorphic to the tensor product
(V0(0) ⊕ V1(1) ⊕ V0(2) )⊗L as a (u(1) ⊕ sl(2))-module. We decompose the Fock space V
(N)
into the direct sum of Vm . From the first relation in (2.8), each Bethe state |k, λ; sN,M
(N) (N) (N)
is the highest weight vector belonging to VN−2M = V2S . The characters of Vm are
calculated as
y m+1 − y −m−1
ch Vm(N) = x N .
y − y −1
In terms of the characters, the identity (4.11) can be rewritten as
(0) (1) (2) L
(N)
ch V0 + ch V1 + ch V0 = Z(L; N, M) ch VN−2M . (4.12)
0N2L
0M N/2
(N)
Theorem 4.6 (Multiplicity formula). The multiplicity of the irreducible component VN−2M
in the tensor product (V0 ⊕ V1 ⊕ V0 )⊗L is given by Z̃(L; N, M),
(0) (1) (2)
(0) (1) (2) ⊗L

(N)
V0 ⊕ V1 ⊕ V0 = Z̃(L; N, M)VN−2M . (4.13)
0N2L
0M N/2
Corollary 4.7 (Combinatorial completeness). We have

(i) dim V = (L − N + 1)(N − 2M + 1)Z(L; N, M),
0NL
0M N/2

(ii) dim V = (N − 2M + 1)Z̃(L; N, M). (4.14)
0N2L
0M N/2
Proof. Consider the limit x, y → 1 in the identities (4.8) and (4.11). 2
The identity (i) in Corollary 4.7 reproduces the combinatorial completeness of Bethe
states for the Hubbard model with so(4) symmetry obtained by Essler, Korepin and
Schoutens [3]. The factor (L − N + 1)(N − 2M + 1) in (i) corresponds to the dimension of
the highest weight so(4)-module VL−N,N−2M with the highest weight vector |k, λ; sN,M .
In Essler–Korepin–Schoutens’ proof of (i) in Corollary 4.7, they take the sum on N
after taking that on M. In our proof, the sums on N and M are taken “simultaneously” in
the level of characters.
The factor (N − 2M + 1) of the identity (ii) in Corollary 4.7 for 0 N L is the

(N)
dimension of the highest weight (u(1) ⊕ sl(2))-module VN−2M with the highest weight
vector |k, λ; sN,M . For L < N 2L, the factor (N − 2M + 1) should be interpreted as
(N)
the dimension of the highest weight (u(1) ⊕ sl(2))-module VN−2M with the lowest weight
φ+π
vector Tr Tph |k, λ; s2L−N,L−M . If even L, the identity (ii) can be rewritten as

dim V = 21−δN,L (N − 2M + 1)Z̃(L; N, M),
0NL
0M N/2
by considering the particle–hole symmetry of the system (2.7). Thus we speculate that the
identity (ii) in Corollary 4.7 accounts for the combinatorial completeness of Bethe states
for the system with the charge-u(1) and the spin-sl(2) symmetries.
The identities (4.7) also enabled us to get the explicit form of Z(L; N, M) through the
binomial theorem.
Corollary 4.8. We obtain the summation formulas for Z(L; N, M) and Z̃(L; N, M),

L+2 L L L+2
(i) Z(L; N, M) = − ,
M N −M M −1 N −M +1

L L L L
(ii) Z̃(L; N, M) = − . (4.15)
M N −M M −1 N −M +1
5. Summary and concluding remarks
In the framework of Bethe ansatz, we have studied the Hubbard model with the
AB-flux that controls the symmetry of the system. In Section 3 we have shown the
existence of solutions for Lieb–Wu equations with an arbitrary number of up-spins and
one down-spin. We have found that the number of (k–Λ–2)-solutions increases as the
so(4) symmetry reduces to the spin-sl(2) symmetry (Proposition 3.2). The number of Bethe
states is consistent with the string hypothesis only in the so(4)-case. In Section 4 we have
investigated the combinatorial formulas giving the combinatorial completeness of Bethe
states. We have shown that the number of Bethe states can be interpreted as the multiplicity
of irreducible components in the tensor products of so(4)-modules (Theorem 4.5). Essler–
Korepin–Schoutens’ combinatorial formula is reproduced by the relation (4.9) among
the characters of so(4)-modules (Corollary 4.7). An advantage of our approach is that
we can obtain the summation formula (4.15) for Z(L; N, M). We have also proposed
a new combinatorial formula derived from the relation (4.12) among the characters of
(u(1) ⊕ sl(2))-modules. The formula is related to the combinatorial completeness of
Bethe ansatz in the sl(2)-case (Corollary 4.7). It should be remarked that, in Section 3,
we have not proved the uniqueness of solutions. Although the Lieb–Wu equations may
have other solutions that we have not expected, the combinatorial formulas introduced in
Section 4 give an evidence that, for M = 1, we have found solutions enough to verify the
combinatorial completeness of Bethe ansatz. The problem is open for M 2.
The combinatorial completeness of Bethe ansatz has not been discussed for the one-
dimensional isotropic Heisenberg spin chain with twisted boundary conditions. Kirillov’s
identity [12] also produces the following formula:
Pn + Mn + n N
Z̃(N; M) = = ,
Mn M
{M
n } n1
M= nMn

where Pn = N − 2M + 2 m(>n) (m − n)Mm . It is expected that, if the redistribution
phenomenon of solutions for Bethe equations [4,8,9] is taken into consideration, the
formula corresponds to the system with twisted boundary conditions. We remark that the
formula also appears in the different context [20].
Acknowledgements
The authors would like to thank Prof. A. Kuniba and Dr. T. Takagi for valuable
discussion. They also would like to thank Prof. M. Wadati and Dr. M. Shiroishi for helpful
comments. One of the authors (A.N.) appreciates the Research Fellowships of the Japan
Society for the Promotion of Science for Young Scientists. The present study is partially
supported by the Grant-in-Aid for Encouragement of Young Scientists (A) No. 14702012.
Appendix A. Redistribution phenomenon
In Section 3, we exactly show the existence of solutions of Lieb–Wu equations with

M = 1 for U > L8 . There, real solutions have been specified by non-repeating indices
Fig. 4. The real parts of solutions for Lieb–Wu equations with L = 20 and N = 2. The dashed lines correspond
to the centers ζ of (k–Λ–2)-string solutions with even m, and the dots express the centers ζ of (k–Λ–2)-string
solutions with odd m and their redistribute real solutions.
Fig. 5. The imaginary parts ξ of solutions for Lieb–Wu equations with L = 20 and N = 2.
{
i | i = 1, 2, . . . , N} and m. But, for 0 < U < L8 , real solutions with repeating indices
may appear at the same time as (k–Λ–2)-string solutions disappear. Such redistribution of
type of solutions is observed in the isotropic Heisenberg model for a large number of sites
[4,8,9]. Here we numerically investigate such phenomenon for the Hubbard model with
L = 20 and N = 2 [3].
As we have already mentioned, (k–Λ–2)-string solutions with odd m may disappear for
0 < U < L8 (see Fig. 3). For each m, the critical value of U is exactly given by [3],

8 π 8
U (m)
= − cos m < .
L L L
Plotted on Fig. 4 are the center ζ = πL m (m = 11, 12, . . ., 29) of (k–Λ–2)-string solutions
and redistributed real solutions {k1 , k2 } for L = 20 and N = 2 in varying the value of U .
Plotted on Fig. 5 are their imaginary parts. The (k–Λ–2)-string solutions with odd m
disappear for U < U (m) and, at the same time, real solutions with repeating indices
{
1 ,
2 ; m} = { m2 , m2 ; m} appear. Note that, as U → 0, all the (k–Λ–2)-string solutions on
Fig. 4 approach to the wavenumbers of lattice free fermion system. Thus, when we apply
the combinatorial formulas in Corollary 4.7 to the case, we must count the number of Bethe
states by regarding the real solutions with repeating indices as (k–Λ)-string solutions.
References
[1] H. Bethe, Zur theorie der metalle, Z. Phys. 71 (1931) 205–226.

[2] T. Deguchi, F.H.L. Essler, F. Göhmann, A. Klümper, V.E. Korepin, K. Kusakabe, Thermodynamics and
excitations of the one-dimensional Hubbard model, Phys. Rep. 331 (2000) 197–281.
[3] F.H.L. Essler, V.E. Korepin, K. Schoutens, Completeness of the so(4) extended Bethe ansatz for the one-
dimensional Hubbard model, Nucl. Phys. B 384 (1992) 431–458.
[4] F.H.L. Essler, V.E. Korepin, K. Schoutens, Fine structure of the Bethe ansatz for the spin- 12 Heisenberg
XXX model, J. Phys. A: Math. Gen. 25 (1992) 4115–4126.
[5] F.H.L. Essler, V.E. Korepin, K. Schoutens, New eigenstates of the 1-dimensional Hubbard model, Nucl.
Phys. B 372 (1992) 559–596.
[6] L.D. Faddeev, L.A. Takhtadzhyan, Spectrum and scattering of excitations in the one-dimensional isotropic
Heisenberg model, J. Sov. Math. 24 (1984) 241–267.
[7] O.J. Heilmann, E.H. Lieb, Violation of the noncrossing rule: the Hubbard Hamiltonian for benzene, Ann.
Natl. Acad. Sci. 172 (1971) 583–617.
[8] A. Ilakovac, M. Kolanović, S. Pallua, P. Prester, Violation of the string hypothesis and the Heisenberg XXZ
spin chain, Phys. Rev. B 60 (1999) 7271–7277.
[9] G. Jüttner, B.D. Dörfel, New solution of the Bethe ansatz equations for the isotropic and anisotropic spin- 12
Heisenberg chain, J. Phys. A: Math. Gen. 26 (1993) 3105–3120.
[10] N. Kawakami, S.K. Yang, Comment on “Twisted boundary conditions and effective mass in Heisenberg–
Ising and Hubbard rings”, Phys. Rev. Lett. 65 (1990) 3063.
[11] N. Kawakami, S.K. Yang, Conductivity in one-dimensional highly correlated electron systems, Phys. Rev.
B 44 (1991) 7844–7851.
[12] A.N. Kirillov, Combinatorial identities, and completeness of eigenstates of the Heisenberg magnet, J. Sov.
Math. 30 (1985) 2298–3310.
[13] A.N. Kirillov, Completeness of states of the generalized Heisenberg magnet, J. Sov. Math. 36 (1987) 115–
128.
[14] A.N. Kirillov, N.A. Liskova, Completeness of Bethe’s states for the generalized XXZ model, J. Phys. A:
Math. Gen. 30 (1997) 1209–1226.
[15] W. Kohn, Theory of the insulating state, Phys. Rev. A 133 (1964) 171–181.
[16] V.E. Korepin, F.H.L. Essler, Exactly Solvable Models of Strongly Correlated Electrons, in: Advanced Series
in Mathematical Physics, vol. 18, World Scientific, 1994.
[17] A. Kuniba, T. Nakanishi, The Bethe equation at q = 0, the Möbius inversion formula, and weight
multiplicities, I: the sl(2) case, in: M. Kashiwara, T. Miwa (Eds.), Physical Combinatorics, Birkhauser.
[18] A. Kuniba, T. Nakanishi, The Bethe equation at q = 0, the Möbius inversion formula, and weight
multiplicities, II: the Xn case, J. Algebra 251 (2002) 577–618.
[19] A. Kuniba, T. Nakanishi, Z. Tsuboi, The Bethe equation at q = 0, the Möbius inversion formula, and weight
(r)
multiplicities, III: the XN case, Lett. Math. Phys. 59 (2002) 19–31.
[20] A. Kuniba, M. Okado, T. Takagi, Y. Yamada, Vertex operators and partition functions in box–ball systems,
RIMS Kokyuroku 1302 (2003) 91–107.
[21] E.H. Lieb, F.Y. Wu, Absence of Mott transition in an exact solution of the short-range, one-band model in
one dimension, Phys. Rev. Lett. 20 (1968) 1445–1448.
[22] D.C. Mattis, The Many-Body Problem, An Encyclopedia of Exactly Solved Models in One Dimension,
World Scientific, 1993.
[23] A. Montorsi, The Hubbard Model, World Scientific, 1992.
[24] A. Nishino, T. Deguchi, Bethe-ansatz studies of energy-level crossings in the one-dimensional Hubbard
model, Phys. Rev. B 68 (2003) 075114.
[25] B.S. Shastry, B. Sutherland, Twisted boundary conditions and effective mass in Heisenberg–Ising and
Hubbard rings, Phys. Rev. Lett. 65 (1990) 243–246.
[26] M. Shiroishi, H. Ujino, M. Wadati, SO(4) symmetry of the transfer matrix for the one-dimensional Hubbard
model, J. Phys. A: Math. Gen. 31 (1998) 2341–2358.
[27] M. Shiroishi, M. Wadati, Integrable boundary conditions for the one-dimensional Hubbard model, J. Phys.
Soc. Jpn. 66 (1997) 2288–2301.
[28] M. Takahashi, One-dimensional Hubbard model at finite temperature, Prog. Theor. Phys. 47 (1972) 69–82.
[29] M. Takahashi, Low-temperature specific-heat of one-dimensional Hubbard model, Prog. Theor. Phys. 52
(1974) 103–114.
[30] M. Takahashi, Thermodynamics of One-Dimensional Solvable Models, Cambridge Univ. Press, Cambridge,
1999.
[31] F. Woynarovich, Excitations with complex wavenumber in a Hubbard chain: states with one pair of complex
wavenumbers, J. Phys. C: Solid State Phys. 15 (1982) 85–96.
[32] C.N. Yang, η pairing and off diagonal long range order in a Hubbard model, Phys. Rev. Lett. 63 (1989)
2144–2147.
[33] C.N. Yang, S.C. Zhang, SO(4) symmetry in a Hubbard model, Mod. Phys. Lett. B 4 (1990) 759.
CUMULATIVE AUTHOR INDEX B681–B688
Abel, S.A. B682 (2004) 183 Blumenfeld, B. B686 (2004) 3

Abel, S.A. B685 (2004) 150 Bobisut, F. B686 (2004) 3
Ahn, C. B683 (2004) 177 Bœhm, C. B683 (2004) 219
Alishahiha, M. B686 (2004) 53 Böhm, G. B682 (2004) 585
Alonso Izquierdo, A. B681 (2004) 163 Bolognesi, S. B686 (2004) 119
Anderson, L.B. B686 (2004) 285 Bonciani, R. B681 (2004) 261
Anoka, O.C. B686 (2004) 135 Bouchez, J. B686 (2004) 3
Anoka, O.C. B687 (2004) 3 Boyd, S. B686 (2004) 3
Anselmi, D. B687 (2004) 124 Branco, G.C. B686 (2004) 188
Anselmi, D. B687 (2004) 143 Braun, V.M. B685 (2004) 171
Aoki, H. B684 (2004) 162 Bueno, A. B686 (2004) 3
Aoki, K. B688 (2004) 3 Bunyatov, S. B686 (2004) 3
Arnaudon, D. B687 (2004) 257
Arnowitt, R. B682 (2004) 347 Caffo, M. B681 (2004) 230
Astier, P. B686 (2004) 3 Camilleri, L. B686 (2004) 3
Cardini, A. B686 (2004) 3
Attal, R. B684 (2004) 369
Carmelo, J.M.P. B683 (2004) 387
Autiero, D. B686 (2004) 3
Castro-Alvaredo, O. B682 (2004) 551
Auzzi, R. B686 (2004) 119
Castro-Alvaredo, O. B687 (2004) 303
Avan, J. B687 (2004) 257
Cattaneo, P.W. B686 (2004) 3
Cavasinni, V. B686 (2004) 3
Babu, K.S. B686 (2004) 135
Cervera-Villanueva, A. B686 (2004) 3
Babu, K.S. B687 (2004) 3
Challis, R.C. B686 (2004) 3
Bajnok, Z. B682 (2004) 585
Chapovsky, A.P. B686 (2004) 205
Baldisseri, A. B686 (2004) 3 Chen, H. B683 (2004) 196
Baldo-Ceolin, M. B686 (2004) 3 Choi, K. B687 (2004) 101
Banner, M. B686 (2004) 3 Chukanov, A. B686 (2004) 3
Bassompierre, G. B686 (2004) 3 Collazuol, G. B686 (2004) 3
Becker, M. B683 (2004) 67 Conforto, G. B686 (2004) 3
Beisert, N. B682 (2004) 487 Constantin, D. B683 (2004) 67
Bellucci, S. B684 (2004) 321 Conta, C. B686 (2004) 3
Beneke, M. B685 (2004) 249 Contalbrigo, M. B686 (2004) 3
Beneke, M. B686 (2004) 205 Cousins, R. B686 (2004) 3
Benslama, K. B686 (2004) 3 Crampé, N. B687 (2004) 257
Besson, N. B686 (2004) 3 Cucchieri, A. B687 (2004) 76
Bianchi, M. B685 (2004) 65 Cudell, J.R. B682 (2004) 391
Bianchi, M. B686 (2004) 261 Czyż, H. B681 (2004) 230
Bird, I. B686 (2004) 3
Bjerrum-Bohr, N.E.J. B684 (2004) 209 Dai, J. B684 (2004) 75
Bloch, J.C.R. B687 (2004) 76 Dall’Agata, G. B682 (2004) 243
0550-3213/2004 Published by Elsevier B.V.

doi:10.1016/S0550-3213(04)00296-2
292 Nuclear Physics B 688 (2004) 291–295
Daniels, D. B686 (2004) 3 Gehrmann-De Ridder, A. B682 (2004) 265

Datta, A. B681 (2004) 31 Geiser, A. B686 (2004) 3
D’Auria, R. B682 (2004) 243 Geng, C.Q. B684 (2004) 281
de Forcrand, P. B686 (2004) 85 Geppert, D. B686 (2004) 3
Degaudenzi, H. B686 (2004) 3 Geyer, B. B684 (2004) 351
Deguchi, T. B688 (2004) 266 Ghinculov, A. B685 (2004) 351
Delfino, G. B682 (2004) 521 Ghosh, P.K. B681 (2004) 359
Del Prete, T. B686 (2004) 3 Giannakis, I. B681 (2004) 120
Derkachov, S. B681 (2004) 295 Gibin, D. B686 (2004) 3
De Santo, A. B686 (2004) 3 Gili, V. B685 (2004) 3
D’Hoker, E. B688 (2004) 3 Giudice, G.F. B685 (2004) 89
Dignan, T. B686 (2004) 3 Gninenko, S. B686 (2004) 3
Di Lella, L. B686 (2004) 3 Göckeler, M. B688 (2004) 135
Djouadi, A. B681 (2004) 31 Godley, A. B686 (2004) 3
do Couto e Silva, E. B686 (2004) 3 Gogoladze, I. B686 (2004) 135
Doikou, A. B687 (2004) 257 Gogoladze, I. B687 (2004) 3
Dolgov, A.D. B688 (2004) 189 Gomez-Cadenas, J.-J. B686 (2004) 3
Dumarchez, J. B686 (2004) 3 González León, M.A. B681 (2004) 163
Dutta, B. B682 (2004) 347 Gosset, J. B686 (2004) 3
Gößling, C. B686 (2004) 3
Eden, B. B681 (2004) 195 Gottwald, S. B685 (2004) 171
Ellis, M. B686 (2004) 3 Gouanère, M. B686 (2004) 3
Engelhardt, M. B685 (2004) 227 Grant, A. B686 (2004) 3
Evslin, J. B686 (2004) 119 Graziani, G. B686 (2004) 3
Grinza, P. B682 (2004) 521
Fayet, P. B683 (2004) 219 Grzadkowski, B. B686 (2004) 165
Feldman, G.J. B686 (2004) 3 Grzelińska, A. B681 (2004) 230
Feldmann, Th. B685 (2004) 249 Guchait, M. B681 (2004) 31
Feng, H. B683 (2004) 168 Guglielmi, A. B686 (2004) 3
Ferrari, R. B686 (2004) 3
Ferrère, D. B686 (2004) 3 Hagner, C. B686 (2004) 3
Ferroglia, A. B681 (2004) 261 Hardmeier, A. B682 (2004) 150
Flaminio, V. B686 (2004) 3 Harmark, T. B684 (2004) 183
Forde, D.A. B684 (2004) 125 Hatsuda, M. B681 (2004) 152
Frampton, P.H. B687 (2004) 31 Havare, A. B682 (2004) 457
Frappat, L. B687 (2004) 257 Heinrich, G. B682 (2004) 265
Fraternali, M. B686 (2004) 3 Henkel, M. B688 (2004) 217
Fré, P. B685 (2004) 3 Hernando, J. B686 (2004) 3
Frenkel, J. B685 (2004) 393 Ho, I.-L. B684 (2004) 281
Freyhult, L. B681 (2004) 65 Horsley, R. B688 (2004) 135
Friedrich, R. B687 (2004) 279 Hou, H.-S. B683 (2004) 196
Fring, A. B682 (2004) 551 Hu, B. B682 (2004) 347
Fring, A. B687 (2004) 303 Hubbard, D. B686 (2004) 3
Fukuma, M. B682 (2004) 377 Hurst, P. B686 (2004) 3
Hurth, T. B685 (2004) 351
Gaillard, J.-M. B686 (2004) 3 Hyett, N. B686 (2004) 3
Gaiotto, D. B688 (2004) 70
Gangler, E. B686 (2004) 3 Iacopini, E. B686 (2004) 3
García Fuertes, W. B681 (2004) 163 Imai, T. B686 (2004) 248
Gardi, E. B685 (2004) 171 Intriligator, K. B682 (2004) 45
Gargiulo, F. B685 (2004) 3 Isidori, G. B685 (2004) 351
Gates Jr., S.J. B683 (2004) 67 Iso, S. B684 (2004) 162
Gegelia, J. B682 (2004) 367 Itoyama, H. B686 (2004) 155
Gehrmann, T. B682 (2004) 265 Itzhaki, N. B684 (2004) 264
Nuclear Physics B 688 (2004) 291–295 293
Itzhaki, N. B688 (2004) 70 Letessier-Selvon, A. B686 (2004) 3

Ivanov, E. B684 (2004) 321 Levy, J.-M. B686 (2004) 3
Li, G.-L. B687 (2004) 220
Jaikumar, P. B683 (2004) 264 Linch III, W.D. B683 (2004) 67
Janke, W. B682 (2004) 618 Linssen, L. B686 (2004) 3
Johnston, D.A. B682 (2004) 618 Lipatov, L.N. B685 (2004) 405
Joseph, C. B686 (2004) 3 Liu, J.T. B681 (2004) 120
Juget, F. B686 (2004) 3 Ljubičić, A. B686 (2004) 3
Long, J. B686 (2004) 3
Kalkkinen, J. B687 (2004) 279 Lukyanov, S.L. B683 (2004) 423
Kanno, H. B686 (2004) 155 Lunghi, E. B682 (2004) 150
Karakhanyan, D. B681 (2004) 295 Lupi, A. B686 (2004) 3
Kawai, H. B683 (2004) 27 Lyubushkin, V. B686 (2004) 3
Kenna, R. B682 (2004) 618
Kent, N. B686 (2004) 3 Ma, W.-G. B683 (2004) 196
Khan, A. B686 (2004) 75 Marchetti, A. B686 (2004) 261
Khoze, V.V. B682 (2004) 217 Marchionni, A. B686 (2004) 3
Khveshchenko, D.V. B687 (2004) 323 Martelli, F. B686 (2004) 3
Kim, I.-W. B687 (2004) 101 Martynov, E. B682 (2004) 391
Kirsanov, M. B686 (2004) 3 Mastrolia, P. B681 (2004) 261
Kirschner, R. B681 (2004) 295 Mastrolia, P. B688 (2004) 165
Klebanov, I.R. B684 (2004) 264 Mateos Guilarte, J. B681 (2004) 163
Klimov, O. B686 (2004) 3 Mazumdar, A. B683 (2004) 264
Kokkonen, J. B686 (2004) 3 McArthur, I.N. B683 (2004) 3
Konishi, K. B686 (2004) 119 Méchain, X. B686 (2004) 3
Konishi, Y. B682 (2004) 465 Meessen, P. B684 (2004) 235
Kono, Y. B682 (2004) 377 Mendes, T. B687 (2004) 76
Körs, B. B681 (2004) 77 Mendiburu, J.-P. B686 (2004) 3
Korunur, M. B682 (2004) 457 Merrell, W. B683 (2004) 67
Kostov, I.K. B683 (2004) 309 Meyer, J.-P. B686 (2004) 3
Kotikov, A.V. B685 (2004) 405 Mezzetto, M. B686 (2004) 3
Kovacs, S. B684 (2004) 3 Miller, D.J. B681 (2004) 3
Kovzelev, A. B686 (2004) 3 Mishra, S.R. B686 (2004) 3
Krasnoperov, A. B686 (2004) 3 Misiak, M. B683 (2004) 277
Kraus, P. B682 (2004) 45 Miwa, A. B682 (2004) 377
Krivonos, S. B684 (2004) 321 Moch, S. B688 (2004) 101
Kuroki, T. B683 (2004) 27 Moorhead, G.F. B686 (2004) 3
Kürzinger, W. B688 (2004) 135 Moortgat, F. B681 (2004) 31
Kuzenko, S.M. B683 (2004) 3 Morita, T. B683 (2004) 27
Kyae, B. B683 (2004) 105 Mosaffa, A.E. B686 (2004) 53
Movshev, M. B681 (2004) 324
Lacaprara, S. B686 (2004) 3 Mück, A. B687 (2004) 55
Lachaud, C. B686 (2004) 3 Mülsch, D. B684 (2004) 351
Lakić, B. B686 (2004) 3 Mussardo, G. B687 (2004) 189
Laliena, V. B683 (2004) 455
Langfeld, K. B687 (2004) 76 Nagao, K. B684 (2004) 162
Lanza, A. B686 (2004) 3 Nath, P. B681 (2004) 77
Laporta, S. B688 (2004) 165 Naumov, D. B686 (2004) 3
La Rotonda, L. B686 (2004) 3 Necco, S. B683 (2004) 137
Larsen, A.L. B686 (2004) 75 Nédélec, P. B686 (2004) 3
Laveder, M. B686 (2004) 3 Nefedov, Yu. B686 (2004) 3
Leal, H. B687 (2004) 323 Nehme, A. B682 (2004) 289
Lechtenfeld, O. B684 (2004) 321 Nevzorov, R. B681 (2004) 3
Lee, C.-A. B683 (2004) 105 Nguyen-Mau, C. B686 (2004) 3
294 Nuclear Physics B 688 (2004) 291–295
Nishino, A. B688 (2004) 266 Rodejohann, W. B687 (2004) 31

NOMAD Collaboration B686 (2004) 3 Román, J.M. B683 (2004) 387
Notari, A. B685 (2004) 89 Rossi, G. B685 (2004) 65
Rubbia, A. B686 (2004) 3
Obers, N.A. B684 (2004) 183 Rückl, R. B687 (2004) 55
Orestano, D. B686 (2004) 3 Rulik, K. B685 (2004) 3
Ortín, T. B684 (2004) 235 Ryzhov, A.V. B682 (2004) 45
Ouyang, P. B684 (2004) 264
Owen, A.W. B682 (2004) 183 Sadri, D. B687 (2004) 161
Sagnotti, A. B682 (2004) 83
Pando Zayas, L.A. B682 (2004) 3 Sakaguchi, M. B681 (2004) 137
Pastore, F. B686 (2004) 3 Sakaguchi, M. B684 (2004) 100
Peak, L.S. B686 (2004) 3 Sakai, K. B682 (2004) 465
Penc, K. B683 (2004) 387 Salvatore, F. B686 (2004) 3
Pennacchio, E. B686 (2004) 3 Sarma, D. B681 (2004) 351
Pessard, H. B686 (2004) 3 Sato, T. B682 (2004) 117
Petcov, S.T. B687 (2004) 31 Schahmaneche, K. B686 (2004) 3
Petti, R. B686 (2004) 3 Scherer, S. B682 (2004) 367
Phillips, J. B683 (2004) 67 Schierholz, G. B688 (2004) 135
Phong, D.H. B688 (2004) 3 Schindler, M.R. B682 (2004) 367
Picone, A. B688 (2004) 217 Schmidt, B. B686 (2004) 3
Pilaftsis, A. B687 (2004) 55 Schmidt, T. B686 (2004) 3
Pirjol, D. B682 (2004) 150 Schofield, B.W. B685 (2004) 150
Placci, A. B686 (2004) 3 Schwarz, A. B681 (2004) 324
Pleiter, D. B688 (2004) 135 Sconza, A. B686 (2004) 3
Polesello, G. B686 (2004) 3 Serban, D. B683 (2004) 309
Pollmann, D. B686 (2004) 3 Sevior, M. B686 (2004) 3
Polyarush, A. B686 (2004) 3 Shafi, Q. B683 (2004) 105
Ponsot, B. B683 (2004) 309 Sheikh-Jabbari, M.M. B687 (2004) 161
Popov, B. B686 (2004) 3 Shi, K.-J. B687 (2004) 220
Poulsen, C. B686 (2004) 3 Shigemori, M. B682 (2004) 45
Profumo, S. B681 (2004) 247 Shimada, H. B685 (2004) 297
Siegel, W. B681 (2004) 152
QCDSF Collaboration B688 (2004) 135 Siegel, W. B683 (2004) 168
Quandt, M. B685 (2004) 227 Signer, A. B684 (2004) 125
Signer, A. B686 (2004) 205
Ragoucy, E. B687 (2004) 257 Sillou, D. B686 (2004) 3
Raidal, M. B685 (2004) 89 Silva-Marcos, J.I. B686 (2004) 188
Rakow, P.E.L. B688 (2004) 135 Smith, J. B682 (2004) 421
Rastelli, L. B684 (2004) 264 Sogut, K. B682 (2004) 457
Rastelli, L. B688 (2004) 70 Soler, F.J.P. B686 (2004) 3
Ravindran, V. B682 (2004) 421 Sommovigo, L. B682 (2004) 243
Rebelo, M.N. B686 (2004) 188 Song, W.Y. B687 (2004) 101
Rebuffi, L. B686 (2004) 3 Sonnenschein, J. B682 (2004) 3
Reinhardt, H. B685 (2004) 227 Sorin, A. B685 (2004) 3
Remiddi, E. B681 (2004) 230 Sotkov, G. B687 (2004) 189
Remiddi, E. B681 (2004) 261 Soyez, G. B682 (2004) 391
Remiddi, E. B688 (2004) 165 Sozzi, G. B686 (2004) 3
Ren, H.-C. B681 (2004) 120 Splittorff, K. B683 (2004) 467
Rico, J. B686 (2004) 3 Stanev, Y.S. B685 (2004) 65
Riemann, P. B686 (2004) 3 Stanishkov, M. B683 (2004) 177
Riotto, A. B685 (2004) 89 Steele, D. B686 (2004) 3
Riva, V. B687 (2004) 189 Steinhauser, M. B683 (2004) 277
Roda, C. B686 (2004) 3 Stiegler, U. B686 (2004) 3
Nuclear Physics B 688 (2004) 291–295 295
Stipčević, M. B686 (2004) 3 Vinogradova, T. B686 (2004) 3

Stolarczyk, Th. B686 (2004) 3 Vitchev, E.S. B683 (2004) 423
Strumia, A. B685 (2004) 89 Vogt, A. B688 (2004) 101
Sun, Y.-B. B683 (2004) 196
Wang, W. B683 (2004) 48
Takács, G. B682 (2004) 585 Wang, X.-J. B683 (2004) 363
Takahashi, F. B688 (2004) 189 Weber, F.V. B686 (2004) 3
Takayama, Y. B686 (2004) 248 Weigert, H. B685 (2004) 321
Tareb-Reyes, M. B686 (2004) 3 Weisse, T. B686 (2004) 3
Taylor, G.N. B686 (2004) 3 Wheeler, J.T. B686 (2004) 285
Taylor, J.C. B685 (2004) 393 Wilson, F.F. B686 (2004) 3
Tereshchenko, V. B686 (2004) 3 Winton, L.J. B686 (2004) 3
Toharia, M. B686 (2004) 165 Wu, T.H. B684 (2004) 281
Toropin, A. B686 (2004) 3 Wu, Y.-S. B683 (2004) 363
Touchard, A.-M. B686 (2004) 3 Wu, Y.S. B684 (2004) 75
Tovey, S.N. B686 (2004) 3 Wyler, D. B682 (2004) 150
Tran, M.-T. B686 (2004) 3
Trigiante, M. B685 (2004) 3
Yabsley, B.D. B686 (2004) 3
Tsesmelis, E. B686 (2004) 3
Yaguna, C.E. B681 (2004) 247
Tsulaia, M. B682 (2004) 83
Yamamoto, M. B683 (2004) 177
Yao, Y.-P. B685 (2004) 351
Ulrichs, J. B686 (2004) 3
Yavartanoo, H. B686 (2004) 53
Uzawa, K. B683 (2004) 122
Yee, H.-U. B686 (2004) 31
Vacavant, L. B686 (2004) 3 Yetkin, T. B682 (2004) 457
Vafa, C. B682 (2004) 45 Yi, P. B686 (2004) 31
Valdata-Nappi, M. B686 (2004) 3 Yoshida, K. B681 (2004) 137
Valuev, V. B686 (2004) 3 Yoshida, K. B683 (2004) 122
Vaman, D. B682 (2004) 3 Yoshida, K. B684 (2004) 100
van der Bij, J.J. B681 (2004) 261 Yue, C. B683 (2004) 48
van Neerven, W.L. B682 (2004) 421 Yue, R.-H. B687 (2004) 220
Vannucci, F. B686 (2004) 3
Varvell, K.E. B686 (2004) 3 Zaccone, H. B686 (2004) 3
Vaulà, S. B682 (2004) 243 Zamolodchikov, A.B. B683 (2004) 423
Veltri, M. B686 (2004) 3 Zanderighi, G. B686 (2004) 205
Verbaarschot, J.J.M. B683 (2004) 467 Zei, R. B686 (2004) 3
Vercesi, V. B686 (2004) 3 Zerwas, P.M. B681 (2004) 3
Vermaseren, J.A.M. B688 (2004) 101 Zhang, R.-Y. B683 (2004) 196
Vettorazzo, M. B686 (2004) 85 Zhou, P.-J. B683 (2004) 196
Vidal-Sitjes, G. B686 (2004) 3 Zuber, K. B686 (2004) 3
Vieira, J.-M. B686 (2004) 3 Zuccon, P. B686 (2004) 3

Nucl - Phys.B v.688

Uploaded by

Copyright:

Available Formats

Nucl - Phys.B v.688

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Nucl - Phys.B v.688

Uploaded by

Copyright:

Available Formats

Nuclear Physics B 688 (2004) 3–69

Received 29 March 2004; accepted 1 April 2004

PACS: 11.25.Db; 11.25.Hf

1.1. Outline of the construction

1.2. Organization and summary of main formulas

AC [δ, ε](pε ) = AM [δ] exp iπ(τ̂ε + ∆τε )pε2

ϑi (0, τε )4 ϑ[δi+ ](0, Ω)2ϑ[δi− ](0, Ω)2

∆τε = − d 2 z d 2 w χz̄+ Sδ (z, w)χw̄+ ωε (z)ωε (w) − ωI (z)ωJ (w)∂I J Rε (Ω̂) .

2 Throughout, we shall use the notation ∂ ≡ ∂/∂Ω , ∂

Hε = I, T2 , S2 , S2 T2 , S22 , S22 M1 T2 × Hε0 . (1.12)

dµC [ε](pL ) = d Ω Ξ6 [δ]ϑ[δ]2 ϑ[δ + ε]2 . (1.16)

= d Ω α|δΞ6 [δ]ϑ[δ]2ϑ[δ + ε]2 .

4 For brevity, the notation Ω = τ , Ω = τ and Ω = τ will be used throughout.

2. Z2 orbifold conformal field theory

2.1. Spin structures and ϑ-functions

Let Σ be an orientable compact genus h Riemann surface without boundary on which

κ|ρ ≡ exp 4πi(κ · ρ − ρ · κ ) = ±1. (2.3)

ϑ[κ](ζ, Ω) ≡ exp πi(n + κ )Ω(n + κ ) + 2πi(n + κ )(ζ + κ ) . (2.4)

2.2. Z2 twisting; unramified double covers

The configurations of a scalar x(z) (respectively spinor ψ(z)) field on Σ with a Z2 -

The Prym period is a complex number satisfying Im τε > 0.

2.3. The Dijkgraaf–Verlinde–Verlinde partition function

Z[ε](Ω, R) = Z qu [ε](Ω) exp πi pL τε − pR

where the momentum sum is taken over the lattice

2.4. Genus one Z2 -twisted propagators and partition functions

2.5. Genus two Z2 -twisted propagators

Bi (z, w) = Sδ + (z, w)Sδ − (z, w), i = 2, 3, 4 (2.25)

Bε (z, w) = Sδ + (z, w)Sδ − (z, w) + bi ωε (z)ωε (w). (2.26)

bi = −4πi∂τε ln ϑi (0, τε ). (2.27)

δzz τε = 2πiωε (z)2 . (2.31)

Combining these results readily yields (2.27).

2.6. The Prym differential and the Schottky relations

The constants of proportionality in this relation will be determined below.

To do so, the hyperelliptic parametrization of genus 2 Riemann surfaces is used.8 The

νa odd ↔ branch point pa ,

sA (z)2 = (z − pa1 )(z − pa2 )(z − pa3 ),

The Schottky relations

we have a refined form of the genus 2 Thomae identities,

The Prym differential in terms of Szegö kernels

Therefore, the normalized Prym differential is simply

= −σ (µi , µj )π ϑk (0, τε ) ωε (z)ωε (w).

3. Chiral splitting for Z2 orbifold theories

Chiral splitting is the process of decomposition of a chirally symmetric amplitude into

3.1. Chirally symmetric amplitudes

The compactified space–time C will be viewed as an N = 1 super conformal field

sM2 = EM A , ΩM /{Gauge Symmetries}. (3.2)

This integral can be written explicitly by choosing a (3|2)-dimensional slice S of

3.2. Chiral splitting of general amplitudes in Z2 -twisted sectors

3.3. The chiral measure in the Z2 -twisted sectors

= 1− d z d w χz̄ χw̄ S(z)S(w) x ZC [δ; ε] exp iπ τ̃ε pε2 . (3.15)

AC [δ; ε](pε ) = ∗ exp iπ τ̃ε pε2

by these methods, and we obtain,

Substituting this result into (3.17) yields

AC [δ; ε](pε ) = (Ω̂) (Ω̂) exp iπ τ̃ε pε2

ϑ(z1 + z2 − w0 − ∆)E(z1 , z2 )σ (z1 )σ (z2 )

AC [δ, ε] pεµ = AM [δ] exp iπ τ̃ε pε2

3.4. The super-Prym period

as well as the defining relations,

i ν0 |µi ϑ[δ]2

ϑ[δj+ ](0, Ω)2ϑ[δj− ](0, Ω)2