Nucl - Phys.B v.788

Nuclear Physics B 788 (2008) 120
The -resonance in a finite volume

Vronique Bernard a , Ulf-G. Meiner b,c , Akaki Rusetsky b,,1
a Universit Louis Pasteur, Laboratoire de Physique Thorique, 3-5, rue de lUniversit, F-67084 Strasbourg, France
b Universitt Bonn, Helmholtz, Institut fr Strahlen und Kernphysik (Theorie), Nuallee 14-16,
D-53115 Bonn, Germany

c Forschungszentrum Jlich, Institut fr Kernphysik (Theorie), D-52425 Jlich, Germany
Received 14 February 2007; received in revised form 14 June 2007; accepted 13 July 2007
Available online 8 August 2007
Abstract
We study the extraction of -resonance parameters from lattice data for small quark masses, corresponding to the case of an unstable . To this end, we calculate the spectrum of the correlator of two -fields in a
finite Euclidean box up-to-and-including O( 3 ) in the small scale expansion using infrared regularization.
On the basis of our numerical study, we argue that the extraction of the parameters of the -resonance (in
particular, of the mass and the pionnucleondelta coupling constant) from the measured volume dependence of the lowest energy levels should be feasible.
2007 Elsevier B.V. All rights reserved.
PACS: 12.38.Gc; 12.39.Fe; 11.10.St
Keywords: Lattice QCD; Baryon resonances; Chiral Lagrangians; Field theory in a finite volume
This research is part of the EU Integrated Infrastructure Initiative Hadron Physics Project under contract number
RII3-CT-2004-506078. Work supported in part by DFG (SFB/TR 16, Subnuclear Structure of Matter) and by the EU
Contract No. MRTN-CT-2006-035482, FLAVIAnet.
* Corresponding author.
E-mail addresses: bernard@lpt6.u-strasbg.fr (V. Bernard), meissner@itkp.uni-bonn.de (U.-G. Meiner),
rusetsky@itkp.uni-bonn.de (A. Rusetsky).
1 On leave of absence from: High Energy Physics Institute, Tbilisi State University, University 9, 380086 Tbilisi,
Georgia.
0550-3213/$ see front matter 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.nuclphysb.2007.07.030
V. Bernard et al. / Nuclear Physics B 788 (2008) 120
1. Introduction
Monte Carlo simulations in lattice QCD enable one to calculate the spectrum of low-lying
hadrons from first principles and thus to obtain important information about the long-range dynamics of quarks and gluons. In the past, a major effort has been concentrated on the study of
the ground-state particle spectrum. However, recently there has been a lot of work on the spectroscopy of the excited nucleon states as well [19]. For the status of lattice calculations of the
baryon spectrum, see, e.g., the recent reviews [10,11]. We note that several long-standing puzzles
in the field are still awaiting a resolution. In particular, it is worth to mention the problem of the
level ordering of the negative-parity nucleon excitation N (1535) and the positive-parity Roper
resonance N (1440) as well as the structure of the (1405).
The -resonance is the most important baryon resonance. Its mass is close to the mass of
the nucleon, and it couples strongly to nucleons, pions and photons. It is clear that a systematic
study of the properties of the -resonance in lattice QCD could lay a solid theoretical basis for
understanding the low-energy QCD dynamics in the one-baryon sector in lattice QCD. Evaluating the mass of the -resonance has already been addressed in the last few years. For illustration,
we consider the recent calculation of the -mass in quenched QCD using a tadpole-improved
anisotropic action and FLIC fermions [8]. In this paper, the local interpolating field with quantum
numbers of the + has been chosen in the following manner

1
= abc 2 uaT C d b uc + uaT C ub d c ,
3
(1)
where a, b, c denote color indices and C is the charge conjugation matrix. With this interpolating
field one further calculates the two-point correlation function at zero spatial momentum and at
large Euclidean time, extracting the mass spectrum. More precisely, in Ref. [8] the calculation
of the ratio m /mN (as well as the similar ratios for some other excited baryon states) has been
performed down to the quark masses corresponding to (M /M )2 = 0.4. At such quark masses,
the dependence of the data on (M /M )2 is very smooth (almost linear). However, a large gap
in the quark (pion) mass has still to be bridged if the data are extrapolated down to the physical
value of (M /M )2 0.03.
The above example highlights all features of the lattice calculations of the parameters of
baryon resonances (and the -resonance, in particular), which we wish to put under scrutiny in
the present paper:
a) As already mentioned, the distance from the lowest data point to the physical value of the
quark mass is still very large and this will lead to large extrapolation errors. In case of a
stable particle, one would argue that performing calculations at smaller quark masses will
reduce this uncertainty until, eventually, the simulations are done close to or at the physical
value of the quark mass. However, in the case of a resonance, the threshold value of the
quark mass exists after which, e.g., the starts to decay into a pion and a nucleon. It is
evident that, below the threshold, the method which was described above cannot be applied
straightforwardly. Does this mean that, in the case of a resonance, there is always a (not so
small) gap in the lattice data, where one should rely only on chiral extrapolation?
b) Lattice data are always real. Does this mean that one gets the real part of the resonance pole
mass as a result of a chiral extrapolation below the decay threshold?
c) Can one determine the decay width of a resonance by combining the method described above
with the chiral extrapolation?
Fig. 1. A schematic representation of the avoided level crossing in the presence of a narrow resonance. The center-of-mass
momentum of a two-particle pair is plotted against the size of a box L (arbitrary units). It is seen that in the vicinity of
the resonance momentum pR a peculiar behavior of the energy levels is observed. The exact position of the resonance in
this case could be easily pinned down by measuring the energy levels on the lattice.
Note also that the Monte Carlo simulations, e.g., in Ref. [8] have been carried out in the
quenched approximation. It is however clear that unstable systems at small quark masses, which
will be considered below, can be meaningfully discussed only in the context of lattice data based
on simulations with dynamical fermions.
To summarize, we want to ask whether there exists a modification of the above method that
enables one (at least, in principle) to carry out calculations of the properties of an excited state
at such quark masses when this state becomes unstable. Moreover, one would like to eventually
walk all way down to the physical quark mass, excluding the extrapolation error altogether (like
this appears possible for ground-state hadrons). It would also be very instructive to see in detail
what happens in the vicinity of the threshold, when one crosses it while performing a chiral
extrapolation of the spectrum. Finally, one should investigate which quantities are most sensitive
to the resonance parameters at small quark masses, and how accurately one could extract the
mass and the width of the resonance by calculating these quantities on the lattice.
The question of identification of hadron resonances on the lattice has already been addressed
in the past. For example, we would like to mention the papers [1219]. In general, the method
proposed originally by Lscher considers the extraction of the two-body scattering phase shifts
from the energy levels, calculated in a finite Euclidean box. In particular, the signatures of unstable particles have been studied. It is demonstrated that in the presence of a narrow resonance
the dependence of the energy spectrum of the system on the box size L exhibits a very peculiar
behavior near the resonance energy, where the so-called avoided level crossing takes place, see
Fig. 1. Note that usually this name is used to describe an abrupt rearrangement of the structure of
the energy levels of a system, which takes place near the resonance energy when the scattering
phase passes through /2. From Fig. 1 it is clear that the position of a narrow resonance can
be readily identified by measuring the energy levels of a system at a finite volume and locating
horizontal plateaus. Moreover, the minimal distance that separates the energy levels near the
avoided crossing, is determined by the decay width of a resonance and, consequently, the latter
quantity can be also extracted from the same lattice data. A very nice qualitative discussion of
the avoided level crossing is given, e.g., in Refs. [15,16,20].
An approach, related to the Lschers method, is used in Ref. [20] to study the structure of the
energy levels in the two-pion system carrying the quantum numbers of the -meson. To this end,
these energy levels are calculated in an effective field theory (EFT) with a phenomenological
lowest-order coupling. Furthermore, in Ref. [21] it has been shown that the presence of
a narrow excited state above the threshold modifies the simple exponential decay law of the
time-sliced two-point function. The decay width within this approach is extracted not from the
two-point function, but directly from the decay amplitudes (see also [22,23]). In addition, we
point out the recent investigation [24], where it is proposed to reconstruct the spectral density in
the two-point function by using the maximum entropy method. This approach, in principle, also
has the capability to address the problem of unstable states in lattice calculations.
Note also that Lschers approach has been recently applied to study nucleonnucleon phase
shifts at low-energy, as well as two-body shallow bound states [2529]. To this end, in Ref. [25],
Lschers master formula which relates the scattering phase shifts to the energy level displacements caused by the NN interaction, is re-derived within a non-relativistic EFT. Note that at large
box sizes, which are used in the study of the scattering processes, the characteristic center-ofmass momenta are small and hence working within the non-relativistic EFT is justified.
An approach, which we use in the present paper, is closely related to the one of Refs. [12
16,20]. To be precise, we investigate the energy levels of the system with the quantum numbers
of the -resonance in a finite box with the size L. This is achieved by calculating the selfenergy of the in the small scale expansion (SSE)2 at a finite volume and finding the poles of
the propagator, which yields the parameterization of the spectrum of the Hamiltonian in a finite
Euclidean box as a function of the variables M and L, as well as the physical masses m (M ),
mN (M ) and the N coupling constant, denoted gN . We further investigate the behavior
of energy levels with respect to these parameters. The main question which is addressed here is
whether the dependence of the energy levels on the above-mentioned parameters is pronounced
enough in order to enable one to perform an accurate fit to determine the mass and width of
the -resonance. Here we would like to mention that, as it turns out later, the width of is so
large that the avoided level crossing is almost completely washed out from the energy spectrum.
Therefore, the nice procedure shown in Fig. 1 does not immediately apply and, in order to be
able to accurately extract the parameters of the -resonance even in this case, one has to find an
optimum fitting strategy. The discussion of this issue constitutes the main content of this work.
The paper is organized as follows. In Section 2 we consider the formalism that is used to
calculate the self-energy of the -resonance in the SSE, including a discussion of the infrared
regularization procedure in a finite volume. Section 3 is devoted to the numerical calculations
and the description of the fitting algorithm and the presentation of our results. In Section 4 we
introduce analytic parameterizations of the energy spectrum at finite volume, which might be
useful for performing fits to the lattice data. Finally, Section 5 contains our conclusions.
2 The framework of including the delta is laid out in detail in Refs. [3032]. We make use here of the formulation
presented in the last of these references. Note that the SSE is a phenomenological extension of chiral perturbation theory
in which the nucleon mass splitting is counted as an additional small parameter. This quantity, however, does not
vanish in the chiral limit. Thus, to investigate the chiral limit of QCD, one must enforce resonance decoupling, as
discussed, e.g., in detail in Ref. [33]. For the following discussion, this subtlety plays no role.
2. The formalism
2.1. Extraction of the energy spectrum
Let us consider the two-point function of two -fields, given by Eq. (1), in a Euclidean box
and choose the rest frame. After projecting out the spin- 32 state, this two-point function in momentum space develops poles at pn = (iEn , 0). In coordinate space, the same two-point function
at large Euclidean times t will be given by a sum of exponentials

0| (x) (0)|0L
(zn ) exp(En t),
(2)
n
where the subscript L indicates that the calculations are done at a finite volume L L L Lt .
We further assume that the size of the box in the time direction Lt is much larger than in the
spatial direction L (eventually, Lt ). Consequently, studying the behavior of this two-point
function at large Euclidean times, one is able to extract the energy levels En = En (L) of the positive parity, spin- 32 states, which are functions of the box size L. In case of a stable -state, finitesize corrections to the lowest energy level vanish exponentially, E1 (L) m = exp(const L)
and in the large-L limit this level yields the value of the stable -mass. This is, however, not the
case for the decaying , when the dependence of the energy levels on L is governed by a power
rather than by an exponential law. The question is, whether one can extract the parameters of the
-resonance from the measured dependence of En (L) on L.
2.2. The Lagrangian and Feynman rules
In order to parameterize the volume-dependent energy levels of the system in terms of the
-resonance parameters, we calculate the same two-point function in the SSE at finite volume. These calculations are similar to the recent study of the volume-dependent nucleon mass
[3438] except for the fact that in the present paper we deal with an unstable particle. The calculations are performed by using the effective chiral Lagrangian, which explicitly contains pion,
nucleon and degrees of freedom [32,39] and coincides with the Lagrangian used at infinite
volume. The relevant terms (in Minkowski space) are displayed below:

F2
U U + U + U + + LN + L + LN ;
4
(1)

(1)
(2)
(3)
(2)
(3)
LN = LN + LN + LN + = N N + N + N + N ,
g
0
(1)
N = iD
/ mN + A u/ 5 ,
2

c3
c2
(2)
u u D D + h.c. +
u u + ,
N = c1 +
0
2
4(mN )2
L=
N
N 3
(3)
N = B1 0 + + B0 0 + ;
i i
L(1)
N = gN O w N + h.c.;
LN = L(1)
N + ,
(1)
(2)
(3)
L = L + L + L +

(1) (2) (3)
j
= i O ij + ij + ij + O ,
(3)
(4)
(5)
(1)

1
0
0
3/2
3/2
g iD
ij = O
O
/ ij m ij , iD
/ ij m ij
2
g1 ij
/ 5 g ,
+ u
2

a3
(2)
a2

u u ij + ,
u u Dik
Dkj + h.c. +
ij = g a1 + ij
0
2
4(m )2
(3)

ij = g ij B1 0 + + B0 30 + ,
(6)
where denotes the trace in flavor space. Throughout, we work in the isospin limit mu =
md = m.
The building blocks for the above Lagrangian are defined by
U = u2 ,

1
u , u ,
2
+ = u u + u u,
s = m1
+ ,

Dij = ij D iij k k D ,
uij = ij u
u = iu U u ,
= 2B(s + ip),

1
wi = i u ,
2
D = +
(7)
and
O = g
2
.
d
(8)
In these formulae, U = exp{i /F }, N and i represent the pion, nucleon and in0
terpolating fields, respectively. Further, mN , m are the masses of the nucleon and the in
0
the chiral limit, and 0 =m mN is the -nucleon mass difference in this limit. The pion
mass is given by M2 = 2B m
+ O(m
2 ). Furthermore, F , gA , gN denote the pion decay
constant, the nucleon axial-vector constant and the N coupling constant in the chiral limit
and ci , BiN , ai , Bi denote various low-energy constants (LECs). Finally, the projectors onto the
isospin- 32 and isospin- 12 subspaces are defined by
1
1
3/2
1/2
3/2
1/2
ij = ij i j ,
(9)
ij = i j ,
ij + ij = ij .
3
3
The free -propagator in d-dimensional space can be read off from the quadratic part of the
Lagrangian

d d p ipx 0
0,i
0,j
i0|T (x) (0)|0 =

(10)
e
S (p) 3/2,ij
(2)d
with
(d 2)p p
p p
1

+
0
0
d
1
m /
p
(d 1)(m )2
(d 1) m
1/2
1/2
1 3/2
1
=
P12 + P21
+
P
0
0
m /
p
d 1 m

d 2
0 1/2
/ + m P22 ,
0 2 p
(d 1) m
0
S
(p) =
0
(11)
where the projectors onto the spin- 32 and spin- 12 states are defined as

P 3/2
1/2
P12

1/2
P21
1/2
P22
1/2
P11
= g
1
1
d 4 p p
(/
p p + p p
/)
,

d 1
d 1 p2
(d 1)p2
1
=
(p p p
/ p ),
d 1p2
1
=
(/
p p p p ),
d 1p2
p p
=
,
p2

1/2
= g P 3/2 P22 .
These projectors obey the following relations

1/2
1/2 1/2
= j k Pil ,
Pij Pkl
3/2 1/2 1/2 3/2
P
P
= Pij P
= 0,
ij
3/2 3/2 3/2
P
P
= P
.
(12)
i, j, k, l = 1, 2,
(13)
2.3. The poles of the propagator

The propagator after inclusion of the self-energy diagrams (see Fig. 2) becomes

d d p ipx L
i0|T i (x) j (0)|0 =
e
S (p) 3/2,ij ,
(2)d
where
3/2
1
L
S
(p) =
+ ,
P
0
m /
p + L (p)
0

L (p) = p
/ 1L p 2 + m 2L p 2 ,
(14)
(15)
and the ellipses stand for the terms with spin- 12 projectors. The poles of the propagator are given
by the zeros of the denominator
0

0
m /
(16)
p +p
/ 1L p 2 + m 2L p 2 = 0.
We further introduce the quantities

iL p 2 = iL p 2 Re i p 2 ,
i = 1, 2,
(17)
where (p) = p
/ 1 (p 2 ) + m 2 (p 2 ) denotes the self-energy part of the -resonance, calculated in the infinite volume (note that this quantity is complex below the decay threshold).
Fig. 2. The self-energy of the at O( 3 ) in the SSE.
Introducing now the physical -mass (the real part of the pole mass) through the conventional
definition (in the infinite volume)

0
m =m 1 + Re 1 m2 + Re 2 m2 + O p 4 ,
(18)
it can be seen that, at the order we are working, Eq. (16) can be rewritten by using the physical
mass instead of the mass in the chiral limit

m p
(19)
/ +p
/ 1L p 2 + m 2L p 2 = 0.
and choosing
Performing the analytic continuation to Euclidean space p = (p0 , p) = (i p 4 , p)
the center-of-mass frame p = (, 0), from Eq. (19) we finally get

2

2
m2 1 + 2L 2 + 2 1 1L 2 = 0.
(20)
2.4. Calculation of the self-energy at O( 3 )
Only pionnucleon and pion loops, which are shown in Fig. 2, contribute to the quantities
iL (2 ). The counterterm contribution is independent on L and thus drops out from the final
expression. One obtains

2
2
L
L
iL 2 = i,N
(21)
+ i,
, i = 1, 2,
where
2
2 gN
N 2

L
=
W 2 W 3N 2 ,
1,N
2
F
2

g
mN N 2
L
2 = N
2,N
W
F 2 m 2
and

2

5g12
L
1,
=
T0 + m2 + 2 W 0 2 + m2 2 W 1 2
2
12F

2

2
2 2
2 2
2
2
3m
W
m
W
,
+

2
3
3m2
3m2

2

5g12
L
2,
=
T0 + m2 + 2 W 0 2 22 W 1 2
2
12F

4 2 42 2
W2 +
W3 .
3
3m2
(22)
(23)
In the above expressions, the quantity T0 corresponds to the pion tadpole and W iN (2 ),
W i (2 ) to the N and scalar loop functions, respectively (note that the tadpoles emerge
from the one-loop diagrams shown in Fig. 2 after simplifying the numerators). An exact definition of these quantities is given below. Further, in these expressions one may take d = 4 because
all ultraviolet divergences are contained in the infinite-volume integrals which, at the end, are
included in the definition of the infinite-volume mass m . Further, at the order we are working,
the quantity F can be replaced by the pion decay constant F . At the end, the self-energy part in
a finite volume at O( 3 ) is expressed in terms of physical quantities only.
In the calculations at finite volume the loop integrals are replaced by infinite sums over discrete lattice momenta kn = 2n/L, n Z 3 , in spatial dimensions (see, e.g., [40]). To ease the
notation, below we give the loop functions with the infinite-volume part included. In order to
arrive at the quantities that enter Eqs. (22), (23), one has first to isolate the infinite-volume part
and then subtract it, e.g.,

T0 = T0 T0 L .
(24)
Consequently, the expressions that are given below, are ultraviolet divergent and imply the application of some kind of regularization. Here we do not refer to a particular regularization
procedure explicitly, since the divergent infinite-volume part will be always subtracted before
the actual calculations are performed.
The one-loop integrals that contribute to the self-energy of the -resonance, are listed below
(X = N, ). These are tadpole diagrams

T0 =
T0X =
dk4 1
1
,
2 L3 n M2 + k 2
dk4 1
1
,
2 L3 n m2X + k 2

p p
p p
dk4 1 k k
T2 =
,
N2 +

2
2
2 L3 n M2 + k 2
p
p
p p
p p X
T =
N2X +
2
p
p 2 2
dk4 1 k k
,
2 L3 n m2X + k 2
(25)
as well as the mesonbaryon loop functions

W0X 2 =
p W1X

dk4 1
1
,
2 L3 n (M2 + k 2 )(m2X + (p k)2 )

dk4 1
k
,
2 L3 n (M2 + k 2 )(m2X + (p k)2 )
2 p p X 2
p p
X
+
S
W
2
p 2
p 2 2

k k
dk4 1
=
,
3
2
2
2 L n (M + k )(m2X + (p k)2 )

p p p X 2
( p + p + p )W3X 2 +
S3
p 2

k k k
dk4 1
,
=
3
2
2
2 L n (M + k )(m2X + (p k)2 )
(26)
10
with the Euclidean 4-momentum k = (k4 , kn ). After subtracting the infinite-volume piece, the
following recurrence relations between various functions can be obtained:

1
N 2 = M2 T0 T2 ,
3

1
N 2X = m2X T0X T2X ,
3
2
2
2

T
T X
+ mX M X
W 0 2 0 2 + 0 2 ,
W 1X 2 =
2
2
2
2

1
2 , m2X , M2 W 0X 2 2 + m2X M2 T0
W 2X 2 =
2
12

2 m2X + M2 T0X ,

2 + m2X M2 X 2

1
W 2 2 M2 T0 T2 m2X T0X + T2X ,
W 3X 2 =
22
6
(27)
where (x, y, z) = x 2 + y 2 + z2 2xy 2yz 2zx denotes the usual triangle function.
2.5. Calculation of the scalar integrals
The calculation of the tadpole graphs is straightforward and is carried out by using standard
techniques (see, e.g., [36,37]). We demonstrate the method for the case of the pion tadpole. Using
dimensional regularization to tame the ultraviolet divergence in Eq. (25), one gets
T0
1
= 3
L

=
dk4
2

d d1 k

1
2n
d1
L
M2 + k42 + k2 n

1
dd k
eiLkj ,
(2)d M2 + k42 + k2
j
(28)
where the Poisson formula

+

(x n) =
n=
+

e2inx
(29)
n=
has been used to arrive at the second equality. Further, in this sum the term with j = 0 corresponds
to the infinite-volume integral. Separating this term, we finally obtain

T0
1
1
dd k
+
(2)d M2 + k42 + k2
4 2 Lj
j
=0

dk4 e

Lj M2 +k42

M 2 K1 (M Lj )
= T0 L + 2
,
M Lj
4
j
=0
where j = |j| =
j12 + j22 + j32 and K (z) denotes the modified Bessel function.
(30)
11
The final result for the tadpoles is given by

M 2 K1 (M Lj )
,
T0 = 2
M Lj
4
T0X =
j
=0
M4
T2 = 2
4
j
=0
K2 (M Lj )
j
=0
(M Lj )2
m2X K1 (mX Lj )
,
mX Lj
4 2
T2X =
m4X K2 (mX Lj )
.
4 2
(mX Lj )2
(31)
j
=0
In the calculation of the mesonbaryon loop functions one has to distinguish between two cases.
In the loop, the variable 2 is below threshold. Using the Feynman parameterization, one
may combine two denominators and then use the same technique as for the calculation of the
tadpole contribution. As a result, one gets

1
W 0 2 =
8 2
1
dx
0

K0 Lj g x, 2 ,
j
=0

g x, 2 = (1 x)M2 + xm2 + x(1 x)2 .
(32)
Note that for 2 close to m2 the function g (x, 2 ) never vanishes in the integration region.
In contrast to the above example, the N loop cannot be calculated by using the same method,
because the variable 2 can now be above the decay threshold N . To calculate this
quantity, the following trick has been used. First, in order to avoid the ultraviolet divergence in
the infinite sum over momenta, we have subtracted the integral at some scale 2 = 2 below
threshold (one subtraction is enough for the convergence, but double subtraction enables one
to achieve faster convergence). Now in the subtraction terms one is allowed to use the same
technique as in the calculation of the loop, because 2 is below threshold. This strategy is
illustrated below in detail. The quantity W 0N (2 ) which we are looking for is split into several
terms

W 0N 2 = W0N 2 Re W0N 2 L = H1 2 + H2 2 + H3 2 ,

d
2
N

H1 2 = W0N 2 W0N 2 + 2 2
W
2 2 ,
d2 0
=

2

d
H2 = W 0N 2 + 2 2
,
W N 2
d2 0
2 =2

H3 2 = Re W0N 2 W0N 2

d
2
N

+ 2 2
(33)
W
.
0

2
d
2 =2 L
The first term is the twice-subtracted infinite momentum sum, where the integration over k4 is
explicitly performed

2 1 EN + E
1
1
H1 2 = 2 2
,
3
2
2
2
L n 2EN E + (EN + E ) ( + (EN + E )2 )2

EN = m2N + k2n , E = M2 + k2n .
(34)
12
The second expression corresponds to the subtraction term

H2
1
=
8 2
1
dx
0

K0 Lj gN x, 2
j
=0
x(1 x)Lj

2
K1 Lj gN x,
,

2 gN (x, 2 )

gN x, 2 = (1 x)M2 + xm2N + x(1 x)2 .
(35)
Furthermore, the remainder is included into the third term, which contains only quantities evaluated in the infinite volume

2 m2N + M2 + B
2 + m2N M2 + B
B
+
ln
ln
H3 2 =
32 2 2
2 + m2N M2 B
2 m2N + M2 B

2 m2N + M2
2 + m2N M2
B
+
arctan
+
arctan
B
B
16 2 2

2
2
2
2
2
2
2
( )(mN M ) mN

ln 2
1+
16 2 2
22 2
M
2 (m2N + M2 ) + (m2N M2 )2
2 B

2 + m2N M2
2 m2N + M2
arctan
+ arctan
,
B
B

B = 1/2 2 , m2N , M2 , B = 1/2 2 , m2N , M2 .
(36)
2.6. Remarks
(a) In order to study baryon resonances, the volume should be taken much larger than in the
case of stable particles since for excited states, the volume-dependent effects decrease only as
powers of L and not by an exponential law as for stable particles. More precisely, the parameter L
should be large enough, so that one could neglect all contributions of the type exp(const M L)
(see, e.g., [40,41]) as compared to the corrections that decrease according to the power law.
(b) The Lagrangians, given by Eqs. (3), (4), (5) and Eq. (6) do not contain the so-called
off-shell pieces that do not contribute to observable quantities like the S-matrix elements or
transition currents. They could in principle contribute to the self-energy of the , which is not an
on-shell quantity. However, eliminating these terms at the diagrammatic level corresponds to
canceling one of the propagators in the loop. The pertinent diagram turns into a tadpole, which
vanishes exponentially. Therefore, for large values of L, where the exponential factors can be
neglected, the off-shell couplings do not contribute to the self-energy.
(c) The same line of reasoning can be used to show that the contribution from the spin- 12
components of the -propagator is irrelevant at finite volume.
(d) As seen from Eq. (32), the whole L-dependent part of the contribution from the
loop is exponentially suppressed at large L. The same is true for all tadpoles. Even if we have
retained them in the final expressions for completeness, the L-dependent part thereof can be
safely neglected at any stage of the calculation. Indeed, we have checked numerically that the
contribution from the loop is very small and does not affect the results.
13
(e) Covariant SSE calculations in the baryon sector at infinite volume are performed, in particular, by using infrared regularization (IR), which leads to a consistent power counting in the
presence of the (large) baryon mass. This method has been applied in the calculations at a finite
volume as well (see, e.g., [36]) by merely extending the integration interval over the pertinent
Feynman parameter from [0, 1] to [0, [. In this work we apply infrared regularization only to
the infinite-volume part of each Feynman diagram. The finite-volume corrections obey the counting rules automatically. It can be easily seen that, when only stable particles are considered, the
difference between the volume-dependent particle masses obtained by using the above two prescriptions, vanishes at large L as exp(const mX L) with X = N, and is therefore physically
irrelevant (this is, e.g., the case of the volume-dependent nucleon mass, considered in Ref. [36]).
3. Numerical results
Solving Eq. (20) for different values of L numerically we get the dependence of the energy
levels En (L) of the system, placed in a Euclidean box of size L. At this order, the energy levels
depend on two parameters, namely the mass m and the coupling constant gN . The term
containing the coupling constant g1 is exponentially suppressed and we assume that mN , M , F
are determined independently in the same lattice calculations. We wish to investigate whether one
can determine m and gN at a reasonable accuracy from a fit to the energy levels.
The results of calculations of the energy levels for the physical values of all parameters
are shown in Fig. 3 (solid lines). We use the following values for the particle masses: M =
140 MeV, mN = 940 MeV and m = 1210 MeV (the real part of the pole mass) and for the pion
decay constant F = 92.4 MeV. Further, we fix the physical value of the N coupling con-
Fig. 3. The dependence of the energy levels in a finite box on the box size L for different values of the coupling constant g N . The avoided level crossing, which is clearly seen at small values of g N (dashed lines), is washed out for
the physical value of this coupling constant (solid lines). For comparison, we also display the free energy levels (dotted
lines). It is seen that the energy levels in the presence of the interaction interpolate between different free energy levels.
As expected, an abrupt change emerges in the vicinity of the resonance energy (the resonance position corresponds to
the solid horizontal line).
14
Fig. 4. The lowest energy level for different input values of m and the physical values of g N and M .
2 /4 = 13.5 GeV2
stant as gN = 1.2 (at M = 140 MeV). This corresponds to the value g
(gN = F g ) obtained by using the pole approximation in dispersion relations, see Ref. [42].
Finally, we take g1 = 2.0 from Ref. [39] (nothing changes if we take g1 = 0). As we see, the coupling of the to the N -system is so strong that the nice structure with the avoided level crossing
has been almost completely washed out. It will resurface again, if the input value of gN is
drastically reduced by hand (dashed lines in the same figure). This property, however, cannot be
meaningfully used in the fitting procedure. It is clear that, in order to perform the fit, another
strategy, not linked to the identification of the resonance energy from the position of the avoided
level crossing, should be looked for.
In order to find such a strategy, we continue to study the structure of the energy levels, as well
as the dependence on all available parameters. In particular, we start with varying m keeping all
other parameters fixed. Consider for instance the lowest energy level shown in Fig. 4, where the
dependence of the energy eigenvalue on L is plotted for different input value of m . It is seen that
the curves are almost linear in the variable L = M L from the interval shown, with the tangent
that monotonically decreases with decreasing m . This property can be used for extracting m .
More precisely, one has to measure E1 on the lattice for different L and fit the curve, treating m
as a free parameter. The lattice data decide which of the curves in Fig. 6 has to be chosen. Since
each curve corresponds to a particular value of the m , the fit to the lowest level determines this
parameter unambiguously (provided gN is known).
Next, we consider the possibility of the determination of the N coupling constant. To this
end, one has to find a quantity that could be maximally sensitive to gN and try to extract this
coupling by a fitting procedure. Bearing in mind the level structure in the presence of a very
narrow resonance (avoided level crossing), we expect that the difference of two energy levels
E2 (L) E1 (L) can strongly depend on the width of the resonance. In Fig. 5 we plot the quantity
E2 (L) E1 (L) against the variable L at a fixed value of m and varying the parameter gN .
The avoided level crossing in this quantityat small values of the parameter gN is seen as a
sharp minimum near the value of L where the crossing takes place and the value of the function
15
Fig. 5. The difference of the first two energy levels for different values of g N and the physical value of M . For small
values of the coupling constant there is a dip, corresponding to the avoided level crossing. A plateau is clearly visible
even for the physical value of g N .
at the minimum determines the width. As evident from Fig. 5, even at the physical value of gN
one may observe a remnant of the avoided level crossinga plateau, which disappears if gN
increases further. Given a rather pronounced dependence of the quantities plotted in Fig. 5 on
the input value of gN , one may expect that fitting would allow one to determine this coupling
constant with a reasonable accuracy.
We are now in a position to describe our proposal for determining the parameters of the resonance m and gN from the lattice data, which is expected to work, even if the width of
the resonance is not very small. In brief, we propose to fit the first few energy levels, measured
on the lattice, to the calculated energy levels, which are parameterized by the free parameters
m and gN . Further, it could be advantageous to carry out this procedure iteratively. Fix
first the decay constant to some input value and determine the -mass from the L-dependence
of the lowest energy level, see Fig. 4. With the newly determined -mass plot the difference
E2 (L) E1 (L) and fit the parameter gN to the data, see Fig. 5. Repeat the procedure until
convergence is achieved.
Finally, we wish to comment on the dependence of the energy levels on the other parameter
which is at our disposal, namely the quark (pion) mass. Since this structure depends only on
the physical nucleon and -masses, in the present (exploratory) study of the problem we have
restricted ourselves to the O( 2 ) expressions of the baryon masses

phys 2
phys 2
phys
phys
mN = mN 4c1 M2 M
(37)
,
m = m 4a1 M2 M
and take c1 = 0.9 GeV1 , a1 = 0.3 GeV1 [39] (note that c1 and a1 enter only through
the nucleon and -masses in the infinite volume, so the choice of particular numerical values for
these constants does not affect our conclusions). Moreover, we neglect the pion mass dependence
of the coupling constant F , since the pertinent correction arises at higher order. The above
simple case perfectly models the real situation: the mass difference between m and mN + M
16
Fig. 6. The lowest energy level for different values of M and the physical value of g N . The curves correspond to
.
the following values of 0 = m mN M : (a) solid line: 0 = 130 MeV, (b) dashed line: 0 = 78 MeV, (c) dotted
line: 0 = 20 MeV.
monotonically decreases with the increase of M and vanishes at around M = 210 MeV. Below
threshold, the structure of the energy levels and the dependence on the parameters m , gN is
similar to the case with M = 140 MeV. After crossing the threshold from below, the is stable
and one expects that the finite-volume corrections to the lowest energy level get exponentially
suppressed.
Fig. 6 clearly illustrates this pattern. In this figure, the lowest-order energy level is plotted at
three different values of the pion mass and the physical value of the decay constant. As we see, the
curve is rather smooth in all cases and monotonically flattens as the decay threshold is approached
from below. This property can be discussed in a more quantitative fashion. Namely, fitting the
level energy in Fig. 6 E 1 = E1 /mN to the variable L within the interval 3 L 8.3 with a
for the different values of the mass gap 0 = m mN M we
linear function E 1 = A + B L,
get: (A, B) = (1.34, 0.019) for 0 = 130 MeV, (A, B) = (1.30, 0.009) for 0 = 78 MeV,
(A, B) = (1.25, 0.006) for 0 = 20 MeV. It is seen that, while the parameter A remains almost
stable, the linear coefficient decreases monotonically. It is important to observe that there is no
sign of irregularity in the quantity E1 (L), when one crosses the threshold. Finally, we would like
to mention that here the distance at which the chiral extrapolation takes place is not bound from
below and could be taken smaller than the distance to the decay threshold. Consequently, the
extrapolation error could be reduced.
4. Analytic parameterization
The dependence of the energy levels En (L) on the parameters m , gN and L is given by the
numerical solution of Eq. (20). If one implements the above fitting procedure in practice, it would
be useful to have a simplified algebraic expression for the energy levels, where the dependence on
the parameters is explicit. Below we demonstrate how such an expression can be derived. Starting
from Eqs. (21)(23), we neglect all contributions that are exponentially suppressed in L. Namely,
17
the loop is neglected altogether as well as various tadpoles. The functions i (2 ), which
appear in the self-energy part, are then proportional to the loop function W 0N (2 ), defined in
Eq. (26). Then, Eq. (20) simplifies to

m 2
2
2
2
2

g2
1 2
2 ( , mN , M ) N
W0 2 .
= N
+
m
M
(38)
2
2
F 2 2
12
Further, since the is a P -wave state, the lowest singularity in the self-energy corresponds to
the contribution of the term with n2 = 1 (it can be easily checked that the singularity at n2 = 0
cancels with the factor (2 , m2N , M2 ) in the numerator). Isolating the singularity at n2 = 1 in
the function W 0N (2 ), one may write

E (1)
1
6
+ R 0N 2 ,
W 0N 2 = 3
(1)
(1)
2
(1)
2
L 2E E + (E )
(39)
where
(1)
EN =
m2N + (2/L)2 ,
E(1) =
M2 + (2/L)2 ,
(1)
E (1) = EN + E(1) ,
(40)
and the function R 0N (2 ) is regular in the vicinity of 2 = (E (1) )2 . Further, Eq. (20) can be
rewritten as

g(L)
m 2 =
(41)
+ r 2 ,
(1)
2
E
where
2
(1)
2

2

gN
1
1
(42)
E + mN M2 E (1) , m2N , M2
(1) (1)
3
F 2 16(E (1) )3
L EN E
and the function r( 2 ) is regular at 2 = E (1) .

If now we use the approximation r( 2 ) = 0, we arrive at a quadratic equation, whose
solution

2

1
(1)
E1,2 (L) = m + E m E (1) + 4g(L)
(43)
2
gives the position of the first two energy levels. In Fig. 7 these levels are plotted at the physical
value of gN , as well as for the case of a smaller value. It is amusing that such a simple
solution reproduces the gross features of the exact solution quite well. For example, the avoided
level crossing is clearly visible at smaller values of gN . Finally, we note that in Ref. [16] a
very similar equation is derived in a simple two-channel quantum-mechanical model.
The approximate solution, given by Eq. (43), contains an explicit dependence on the parameters L, m and gN . For this reason, it is easier to use it for performing the fit to the lowest
energy levels. Of course, from the quantitative point of view, the quality of the approximation
is still not satisfactory. It is however obvious that the accuracy can be systematically improved
by including the next nearby singularities as well as regular contributions. We do not display
the pertinent formulae here. It is clear that, at the end, the accuracy of the analytic parameterization, which is used to analyze the real lattice data, should be matched to the precision of this
data.
g(L) =
18
Fig. 7. Approximate solution for the first two energy levels, Eq. (43), for physical (left panel) and reduced (right panel)
values of the coupling constant g N . The approximate solution qualitatively reproduces the features of the exact solution. Namely, the avoided level crossing is clearly visible in the right panel, as well as an almost stable -state at
M = 200 MeV (left and right panels).
5. Conclusions
(a) In this paper we present the results of calculations of the pole structure of the correlator
of two -fields in a finite Euclidean box. It has been argued that, calculating the first few energy
levels En (L) in terms of the resonance parameters m and gN within the SSE and fitting the
lattice data at finite volume, an extraction of these parameters at a reasonable accuracy may be
possible. This statement constitutes the main result of the present paper.
(b) The main question that remains is: how the result will be affected by higher-order corrections in the chiral expansion? It can be for instance shown that, parameterizing pionnucleon
scattering matrix by the pure s-channel -pole term, Eq. (20) can be rewritten in a form similar
to the Lschers master formula, which expresses the displacement of an energy level through
the scattering phase shift. The higher-order corrections in the SSE contribute to the non-resonant
background and are expected to be moderate. Of course, such heuristic arguments cannot be a real
substitute of explicit calculations. Note that such calculations at O( 4 ) are already in progress
and the results will be reported elsewhere, including a detailed error analysis in SSE at this order [43]. One expects that these calculations shed light on the question of the convergence of the
chiral expansion.
(c) Our approach is closely related to the method proposed originally by Lscher and developed in number of subsequent publications [1216,25]. Note, however, that the Lscher formula,
relating the energy levels of a system in a finite volume to the scattering phase shifts, is valid
in general beyond the chiral expansion, thus avoiding the above-mentioned problem of convergence. On the other hand, our result contains an explicit parameterization of the energy
levels in terms of m , gN and can be used, in addition, to study the quark mass dependence of the energy levels. This is important because the first lattice data which will appear
below N threshold, will probably still correspond to the pion mass higher than the physical
value.
We now note that, due to the condition M L 1, the characteristic 3-momenta of the system
p M and the non-relativistic approach must be applicablethe processes with a mass gap
M or higher are suppressed exponentially. One therefore expects that, for a sufficiently large L
these two approaches overlap and the results are complementary to each other. At present, we are
investigating the problem in detail within non-relativistic EFT, aiming to explicitly demonstrate
19
this relationship that, in turn, will enable one to choose an optimum strategy for determining the
resonance parameters from the lattice data in the future [44].
Acknowledgements
The authors would like to thank G. Colangelo, J. Gasser, C. Haefeli, M. Savage, G. Schierholz,
R. Sommer and U.-J. Wiese for interesting discussions.
References
[1] C.M. Maynard, D.G. Richards, UKQCD Collaboration, Nucl. Phys. B (Proc. Suppl.) 119 (2003) 287, heplat/0209165.
[2] D.G. Richards, M. Gckeler, R. Horsley, D. Pleiter, P.E.L. Rakow, G. Schierholz, C.M. Maynard, LHPC Collaboration, Nucl. Phys. A (Proc. Suppl.) 109 (2002) 89, hep-lat/0112031.
[3] C. Gattringer, et al., BGR Collaboration, Nucl. Phys. B 677 (2004) 3, hep-lat/0307013.
[4] S. Sasaki, Prog. Theor. Phys. Suppl. 151 (2003) 143, nucl-th/0305014.
[5] S. Sasaki, T. Blum, S. Ohta, Phys. Rev. D 65 (2002) 074503, hep-lat/0102010.
[6] J.M. Zanotti, et al., CSSM Lattice Collaboration, Phys. Rev. D 65 (2002) 074507, hep-lat/0110216.
[7] W. Melnitchouk, et al., Phys. Rev. D 67 (2003) 114506, hep-lat/0202022.
[8] L. Zhou, F.X. Lee, Phys. Rev. D 74 (2006) 034507, hep-lat/0604023.
[9] K. Sasaki, S. Sasaki, Phys. Rev. D 72 (2005) 034502, hep-lat/0503026.
[10] C. McNeile, hep-lat/0307027.
[11] D.B. Leinweber, W. Melnitchouk, D.G. Richards, A.G. Williams, J.M. Zanotti, Lect. Notes Phys. 663 (2005) 71,
nucl-th/0406032.
[12] M. Lscher, Commun. Math. Phys. 105 (1986) 153.
[13] M. Lscher, Nucl. Phys. B 354 (1991) 531.
[14] M. Lscher, Nucl. Phys. B 364 (1991) 237.
[15] M. Lscher, DESY-88-156, Lectures given at Summer School Fields, Strings and Critical Phenomena, Les
Houches, France, 28 June5 August 1988.
[16] U.-J. Wiese, Nucl. Phys. B (Proc. Suppl.) 9 (1989) 609.
[17] K. Rummukainen, S.A. Gottlieb, Nucl. Phys. B 450 (1995) 397, hep-lat/9503028.
[18] N.H. Christ, C. Kim, T. Yamazaki, Phys. Rev. D 72 (2005) 114506, hep-lat/0507009.
[19] C.H. Kim, C.T. Sachrajda, S.R. Sharpe, Nucl. Phys. B 727 (2005) 218, hep-lat/0507006.
[20] T.A. DeGrand, Phys. Rev. D 43 (1991) 2296.
[21] C. Michael, Nucl. Phys. B 327 (1989) 515.
[22] R.D. Loft, T.A. DeGrand, Phys. Rev. D 39 (1989) 2692.
[23] L. Lellouch, M. Lscher, Commun. Math. Phys. 219 (2001) 31, hep-lat/0003023.
[24] T. Yamazaki, N. Ishizuka, Phys. Rev. D 67 (2003) 077503, hep-lat/0210022.
[25] S.R. Beane, P.F. Bedaque, A. Parreno, M.J. Savage, Nucl. Phys. A 747 (2005) 55, nucl-th/0311027.
[26] S.R. Beane, P.F. Bedaque, A. Parreno, M.J. Savage, Phys. Lett. B 585 (2004) 106, hep-lat/0312004.
[27] S.R. Beane, P.F. Bedaque, T.C. Luu, K. Orginos, E. Pallante, A. Parreno, M.J. Savage, hep-lat/0612026.
[28] S.R. Beane, P.F. Bedaque, K. Orginos, M.J. Savage, Phys. Rev. Lett. 97 (2006) 012001, hep-lat/0602010.
[29] S. Sasaki, T. Yamazaki, hep-lat/0610081.
[30] E. Jenkins, A.V. Manohar, Phys. Lett. B 259 (1991) 353.
[31] E. Jenkins, M.E. Luke, A.V. Manohar, M.J. Savage, Phys. Lett. B 302 (1993) 482, hep-ph/9212226;
E. Jenkins, M.E. Luke, A.V. Manohar, M.J. Savage, Phys. Lett. B 388 (1996) 866, Erratum.
[32] T.R. Hemmert, B.R. Holstein, J. Kambor, J. Phys. G 24 (1998) 1831, hep-ph/9712496.
[33] V. Bernard, H.W. Fearing, T.R. Hemmert, U.-G. Meiner, Nucl. Phys. A 635 (1998) 121, hep-ph/9801297;
V. Bernard, H.W. Fearing, T.R. Hemmert, U.-G. Meiner, Nucl. Phys. A 642 (1998) 563, Erratum.
[34] W. Detmold, M.J. Savage, Phys. Lett. B 599 (2004) 32, hep-lat/0407008.
[35] P.F. Bedaque, H.W. Griehammer, G. Rupak, Phys. Rev. D 71 (2005) 054015, hep-lat/0407009.
[36] A. Ali Khan, et al., QCDSF-UKQCD Collaboration, Nucl. Phys. B 689 (2004) 175, hep-lat/0312030.
[37] S.R. Beane, Phys. Rev. D 70 (2004) 034507, hep-lat/0403015.
[38] G. Colangelo, A. Fuhrer, C. Haefeli, Nucl. Phys. B (Proc. Suppl.) 153 (2006) 41, hep-lat/0512002.
20
[39]
[40]
[41]
[42]
[43]
[44]
V. Bernard, T.R. Hemmert, U.-G. Meiner, Phys. Lett. B 622 (2005) 141, hep-lat/0503022.
J. Gasser, H. Leutwyler, Nucl. Phys. B 307 (1988) 763.
M. Lscher, Commun. Math. Phys. 104 (1986) 177.
G. Hhler, in: H. Schopper (Ed.), LandoltBrnstein, vol. 9, Springer, Berlin, 1983.
V. Bernard, D. Hoja, U.-G. Meiner, A. Rusetsky, in preparation.
V. Bernard, M. Lage, U.-G. Meiner, A. Rusetsky, in preparation.
The soft-energy region in the radiative decay

of bound states
Pedro D. Ruiz-Femena
Max-Planck-Institut fr Physik (Werner-Heisenberg-Institut), Fhringer Ring 6, 80805 Mnchen, Germany
Received 17 May 2007; accepted 20 July 2007
Available online 27 July 2007
Abstract
The orthopositronium decay to three photons is studied in the phase-space region where one of the phoThe NRQED
tons has an energy comparable to the relative three-momentum of the e+ e system ( m).
computation in this regime shows that the dominant contribution arises from distances 1/ m, which
allows to treat the Coulomb interaction perturbatively. The small-photon energy expansion of the 1-loop
decay spectrum from full QED yields the same result as the effective theory. By doing the threshold expansion of the 1-loop QED amplitude we confirm that the leading term arises from a loop-momentum region
where q 0 q2 /m . This corresponds to a new non-relativistic loop-momentum region, which has to be
taken into account for the description of a non-relativistic particleantiparticle system that decays through
soft photon emission.
PACS: 36.10.Dr; 11.10.St; 12.38.Bx; 13.30.Ce; 13.20.Gd
1. Introduction
Effective field theories (EFTs) and asymptotic expansions have become standard tools in the
description of the radiative decay of bound states. Their success lies on their ability to select the
relevant dynamics in the different kinematic regions defined by the relative size of the radiated
photon energy as compared to the bound state mass. Methods based on the use of operator product expansions and effective Lagrangians do not only provide more efficient ways to perform the
computations but also allow to extract physical interpretations which may go unnoticed in a full
theory approach.
E-mail address: ruizfeme@mppmu.mpg.de.
doi:10.1016/j.nuclphysb.2007.07.013
22
P.D. Ruiz-Femena / Nuclear Physics B 788 (2008) 2146
Apart from being of physical relevance in itself, the positronium system provides a testing
ground for the EFT concepts and techniques devised in the study of radiative decay amplitudes, that could eventually be applied to the description of the more intricate radiative decays of
quarkonia. In this paper we illustrate which are the characteristic features of the radiative decay
spectrum when the energy of the photon lies in the soft-energy region ( m) through the
study of the 3 annihilation amplitude of the positronium spin triplet ground state (orthopositronium: o-Ps). The soft-energy region is accessible in the o-Ps 3 decay because one of the
final state photons can have an arbitrarily small energy, the other two being hard photons with
energy m. The process can be then viewed as the radiative version of the o-Ps 2 decay.
In a previous paper [1] the o-Ps differential decay spectrum was calculated in the region where
the photon energy is comparable to the positronium binding energy, m 2 . Binding effects
were included for the first time in the Ps structure function using non-relativistic QED (NRQED)
[2], and it was found that they are essential to reach agreement with the Lows theorem prediction
for the 0 behaviour of this decay [1,3]. Only the leading term in the multipole expansion of
the radiated photon field (dipole approximation) was required for the calculation, in accordance
with the NRQED velocity counting rules for photons with m 2 [4]. Based on general considerations about the power-counting of higher-order multipoles in the non-relativistic description
of the o-Ps decay process, it was then argued by Voloshin [5] that the dipole approximation used
for photon energies m 2 should also hold for a description of the o-Ps photon spectrum in
the whole range m, thus enlarging the validity region of the formula given in Ref. [1]. In
particular, it was shown by Voloshin that the
expansionof the above-mentioned formula in the
region m 2 m is actually a series in m/ rather than in integer powers of .
The origin of such unnatural expansion has to be traced back to the fact that the main contri
bution in the EFT calculation arises from distances in the e+ e system of order r 1/ m,
that are much smaller than the size of the Ps atom r 1/m. Translated into momentum
space, this implies that there is a kinematic region where the relative 3-momentum of the nonrelativistic e+ e pair obeys the non-relativistic relation p 0 p2 /m . This scaling does not
correspond to any of the known modes that have been identified in the common applications of
non-relativistic EFTs, and should be compared to the scaling of the familiar potential modes,
p 0 p2 /m m 2 , characteristic of the heavy particles that form the bound state, and that of the
soft modes, p 0 p2 /m.
In view of these novel features, it is worth questioning if the non-relativistic EFT framework
devised to describe the o-Ps decay for m 2 in Ref. [1] also provides the proper expansion
when extended to the whole m region, as advocated by Voloshin. It is the main purpose of
this work to confirm that this is indeed the case by comparing the analytic results for the EFT
and full QED spectrum in the m region, where, as we argue, the usual perturbative QED
expansion can already be applied. Such comparison has become feasible after a recent evaluation
of the 1-loop QED o-Ps 3 phase-space distribution [6], that we use to determine the QED
spectrum for soft photon energies. The agreement between both computations shows that the softenergy region provides a regime where the EFT and the perturbative QED calculations can have
a smooth matching. Moreover, we shall confirm by explicitly computing the o-Ps 3 decay
spectrum in NRQED beyond the dipole approximation that the (/m)k suppression advocated
by Voloshin for the higher-order multipoles is of application in the soft-energy region.
It is also the aim of this work to show that the momentum scale p0 p2 /m that rules
the behaviour of the corrections to the NRQED o-Ps decay amplitude, corresponds to a new
loop-momentum region that has to be taken into account for a successful application of the
threshold expansion method [7] to QED and QCD loop diagrams involving a particleantiparticle
23
system decaying through the emission of a soft photon. In the conventional loop expansion, soft
photon radiation from a heavy particleantiparticle system introduces a soft energy component
m in the zero component of the heavy particle momenta. Non-relativistic poles in massive
propagators of the form (p 0 p2 /2m )1 can be thus found in these loops, giving rise to
a contribution from the loop-momentum region p0 p2 /m . It is a well-known fact that
momentum regions that have not been considered previously may become relevant for some
kinematic configurations or at higher loop order in the asymptotic expansion of integrals based
on the method of regions [21], and looking for missing regions is a required task in order to
check the robustness of the method. Likewise, it is of conceptual importance to understand the
role of this new momentum scale in the EFT frameworks that describe the radiative decays of
heavy particleantiparticle systems in the soft-energy region, which we discuss in this paper.
The outline of the paper is as follows. In Section 2 we detail the EFT calculation of the
o-Ps 3 decay amplitude without multipole expanding the electromagnetic interaction. We
also show how relativistic corrections that contribute to the decay amplitude in the whole photon energy range m can be accommodated in the EFT framework. In particular, the recoil
of the intermediate virtual e+ e pair, which yields also an /m correction, is included in our
calculation. In Section 3 we perform the 0 limit of the 1-loop QED spectrum directly from
the 1-loop phase-space distribution computed recently by Adkins [6]. The 1-loop QED amplitude in this regime is also obtained with the asymptotic expansion of the graphs near threshold
in Section 4. A discussion on the EFT description of the new loop-momentum region found is
postponed to the end of the latter section. Finally in Section 5 we suggest the use of the threshold
expansion within the NRQCD factorization framework to compute the short-distance coefficients
in quarkonium radiative decays at soft energies. Appendix A collects the exact formulas for the
multipole calculation of the NRQED spectrum, while Appendix B shows another example of the
asymptotic expansion method applied to a radiative amplitude.
2. NRQED computation of the decay spectrum for m
The NRQED framework used in Ref. [1] provides a systematic way to compute bound state
effects in the o-Ps decay spectrum when the photon energy is much smaller than the electron
mass. Contrary to the usual relativistic approach where the 3-photon annihilation is considered to
take place at very short distances as compared to the range of the electromagnetic binding force
between e+ e , the non-relativistic description takes into account that there is a long-distance
part in the o-Ps decay process when one the photons in the final state is not hard ( m). In
the latter case, the decay proceeds in two steps: the low-energy photon is first radiated from the
bound state, making a transition from the C-odd ground state o-Ps (3 S1 ) to a C-even positronium
state, which subsequently decays into two photons (see the EFT graph of Fig. 1(a)). The decay
amplitude of the intermediate C-even e+ e state has no long-distance contribution, since both
photons must be hard. The emission of the low energy photon from o-Ps is described by the
Coulomb Hamiltonian of the e+ e system in interaction with a quantized electromagnetic field:
H = H0 + Hint ,
P2
p2
+ HC , HC =
,
4m
m
r
e
e
Hint = p1 A(x1 ) + p2 A(x2 ) B(x1 ) B(x2 ),
m
m
H0 =
(1)
24
(a)
(b)
Fig. 1. NRQED graphs for the o-Ps 3 decay. The allowed intermediate states n correspond to those of the spectrum
of H0 . The zig-zag line represents the low-energy photon. The black dot in (a) denotes either a p A or a B interaction.
Graph (b) represents the radiation of the low-energy photon directly from the 2 annihilation vertex.
where we have used the center of mass variables (r |x|)

x 1 + x2
,
2
with the subindices 1, 2 referring to the electron and the positron, respectively, and , being
the Pauli matrices acting on the electron and positron spinors ( = e/2m). The terms shown in
the Hamiltonian Hint are the leading ones in the non-relativistic expansion. Relativistic effects
can be included through higher dimensional operators suppressed by powers of 1/m (see, e.g.,
[8]), but will not be needed for the purposes of this work.
Since the computation of the photon spectrum in Ref. [1] was intended for photon energies
comparable to the positronium binding energy, m 2 , the interaction Hamiltonian was used
in the dipole approximation limit. This approximation amounts to evaluating the vector potential A in the center of mass of the positronium system (i.e., at x = 0), which is fully legitimate
for radiated photons with wavelengths much larger than the characteristic size of the positronium
atom (a = 2a0 = 2/m). Higher order multipoles arise as a Taylor series in the relative coordinate x and are suppressed by powers of /a under the assumption that the relevant amplitudes
one has to compute involve integrations to spatial extents of order a. For photons with larger
energies, m 2 , this premise will invalidate the use of the multipole expansion.
However, in the case of the o-Ps system that undergoes a radiative transition before decaying,
and as it has been properly pointed out by Voloshin [5], the scale (m)1 does not constraint
the maximum photon energy for which we can apply the multipole expansion. The reason is that
after the soft photon is radiated we have to consider all possible e+ e intermediate states with
the right quantum numbers. The propagation of these e+ e states is described by the Greens
function obtained from the Hamiltonian HC ,

2
G(x, y, ) = (x y),
HC +
(2)
m
p = (p1 p2 )/2,
x = x 1 x2 ,
P = p 1 + p2 ,
X=
at energy
2 /m = E0 2 /4m
(3)
(see Eq. (8) below). Since the annihilation of the intermediate e+ e pair into two photons takes
place at small distances, the amplitude for the full process is thus given by a convolution of the
o-Ps ground state and the Greens function. In coordinate space, the Greens function has a characteristic size ruled by the exponential factor exp(r). We note that for soft photons of energy
m 2
m, this exponential factor constrains the relevant integration region to distances of
order 1 m, much smaller than the spatial extent of the initial positronium atom. Therefore, the characteristic distance that enters in the multipole expansion of the soft-photon emission
25
in the o-Ps 3 decay amplitude is determined by the falloff of the intermediate Greens function rather than by
the size of the Ps atom, so the series of multipoles is actually an expansion in
powers of r /m which can be used as long as m. Note also that for the intermediate
pair, an iterative computation of the Coulomb Greens function, G = Gf + Gf VC Gf + , with
Gf the free Greens function, shows that adding a Coulomb exchange generates a term

m
m
d 3 x VC Gf
1, for m 2 m,
as Gf m exp(r)/r. Therefore the Coulomb interaction can be treated perturbatively in the

virtual e+ e system after soft photon radiation.
It is illustrative to show how the latter is realized in the actual computation of the o-Ps decay
spectrum without multipole expanding the electromagnetic potential of the radiated photon. For
this purpose let us now consider the NRQED calculation of Ref. [1] and keep the full xi dependence of the various terms in the Hamiltonian (1). We start with the p A electric amplitude.1
Using time-ordered perturbation theory (TOPT), the amplitude corresponding to the graph of
Fig. 1(a) reads

0|A(2 ) |n, P in, P| ie
m {p1 A(x1 ) p2 A(x2 )}|o-Ps
.
Me =
(4)
Eo En,P
n,P
The sum above extends over all discrete and continuum states |n, P of the spectrum of the
unperturbed Hamiltonian H0 , that can be written as the direct product of a plane wave with the
c.m. momentum P and a wavefunction describing the relative motion. In configuration space,
x1 x2 |n, P = eiPX n (x), where n (x) is an eigenstate of HC with eigenvalue En , and we have
omitted the spin wavefunction. The energy of this intermediate (virtual) e+ e state is En,P =
En + P2 /4m, and E0 = m 2 /4 is the o-Ps ground state energy.
The first matrix element in the r.h.s of Eq. (4) gives the short-distance part of the annihilation. The quantity A(2 ) is the e+ e 2 amplitude in QED calculated as an expansion in the
momenta of the leptons:

(W0 + W1 p + W2 P)p + O p2 ,
A(2 ) p
(5)
where p , p
are the Pauli spinor fields that annihilate the electron and positron with relative
momentum with respect the c.m. p and p, respectively. Only the first order in the momentum
expansion will be needed in this work. Expressions for W0 , W1 can be found in Ref. [1], while
W2 = (W1 + ( 1 2 ) )/2, with 1 , 2 the polarizations of the outgoing hard photons. The
W2 P term in Eq. (5) is an operator which depends on the total momentum of the e+ e pair. In
principle, we need to consider the latter because in the o-Ps decay the intermediate e+ e pair can
recoil after the soft photon is radiated. Inserting the operators in Eq. (5) into the short-distance
amplitude one gets:

0|A(2 ) |n, P = Tr W0 n (0) iW1 n (y) y=0 + W2 Pn (0) .
(6)
The trace above is taken over spinor indices,and is the spin wavefunction of the state |n, P ,
which can be ina spin-0 state ( 1/ 2 ), or in a spin-1 state with polarization vector
( / 2 ).
1 We refer to the amplitudes as of electric type if there is a change in the parity between the atomic states (of
magnetic type otherwise).
26
The matrix element for the radiation of the soft photon with three-momentum k and polarization vector (we pick a gauge where photon polarizations are purely transverse,
i ki = 0,
i0 = 0) can be calculated easily:

n, P|p1 A(x1 ) p2 A(x2 )|o-Ps

= i(2)3 (3) (P + k) 0
d 3 x n (x) eikx/2 + eikx/2 0 (x),
(7)
with the initial o-Ps state x1 x2 |o-Ps 0 (x), once we set X = 0. 0 (x) is the ground state
wave function in position space
0 (x) =
1
a 3
ex/a ,
and 0 is the spin-1 polarization vector. The photon wavefunctions eikx/2 arise from the emission of the photon from the electron or from the positron line, respectively. The p A interaction
can change the orbital angular momentum but not the spin of the e+ e pair, which thus remains
in a spin-1 state. The delta function in Eq. (7) sets the recoil momenta of the intermediate e+ e
system to k, so that En,P = En + 2 /4m is its energy.2 Taking into account the sum over
polarizations,

i j = ij ,
we can write

e
Me = Wi1
0
m
+

3
d x

ie
W2 k
0
m
yi n (y)n (x)

eikx/2 + eikx/2 0 (x)
En,k + E0 y=0

n (0) (x)
d 3x
En,k + E0

eikx/2 + eikx/2 0 (x).
(8)
The term W0 in Eq. (6) has disappeared from the electric amplitude because it only contributes
to the 2 annihilation of a spin-singlet (parapositronium: p-Ps) state. Likewise, the second line
in Eq. (8) vanishes because the vector integration can only be proportional to the photon momentum k and k = 0. Therefore we are left only with the term proportional to W1 , where
the term between brackets can be written in terms of the Coulomb Greens function at energy
2 /m = E0 2 /4m,
i

y n (y)n (x)

(9)
= yi G(y, x; ) y=0 = 3xi G1 (0, r; ).
En,k + E0 y=0
n
The derivative acting on the Greens function picks out the = 1 component of its partial wave
decomposition in y = 0,

(2 + 1)(xy) P (x y/xy)G (x, y, k).
G(x, y, k) =
(10)
=0
(See Appendix C of Ref. [1] for explicit expressions of the partial waves G .) Therefore only
P -wave states contribute in the sum over virtual fluctuations in Eq. (4). Using 0 (x) =
2 Let us recall that in TOPT three-momentum is conserved at the vertices and virtual states are always on-shell.
0 (x)xi /ar, we can rewrite the electric amplitude as

M = e0 (0) Wi1
j D ij (k),
(11)
with
3
D (k) =
ma
ij

d 3x
27

xi xj ikx/2
+ eikx/2 G1 (0, r; )er/a .
e
r
(12)
The general structure of D ij (k) reads

i j
kk
.
D ij (k) = D() ij + D()
(13)
2
The term proportional to D()

vanishes when contracted with the photon polarization . The
coefficient D() can be projected out by contracting with the tensor ij ki kj /2 , and after
performing the angular integration, we obtain:

r
r
12
24
8
3
r/a
sin
cos
dr r G1 (0, r; )e
.
D() =
(14)
ma
2
2
(r)3
(r)2
Note that all recoil effects up to this order are properly accounted for by the 2 /4m term in the
parameter that enters the Greens function, and are naturally suppressed with respect leading
terms in the non-relativistic regime.
In addition, gauge invariance requires that the momentum operators in the short-distance amplitude A(2 ) are replaced by the covariant derivatives pi eA(xi ) [1]. This generates a purely
local term that has to be added to the electric amplitude calculated with the interaction Hamiltonian of Eq. (1). The corresponding graph is depicted in Fig. 1(b). This contribution reads

0|A(2 ) |o-Ps = e0 (0) W1 ,
(15)
0
and the total electric amplitude is then equal to

Me = e0 (0) W1 1 + D() .
(16)
The expression for D() in Eq. (13) allows us to identify the typical distances that control the
falloff of the integrand. The = 1 Greens function,
G1 (0, r; ) =
m 3 r
e (2 )U (2 , 4, 2r),
3
m
1
=
,
2
a
(17)
contains the exponential factor er . On the other hand, we have the factor er/a from the wavefunction of the initial state. Therefore, distances in the integrand for the electric amplitude are
constrained by the exponential factor

exp ( + 1/a)r ,
(18)
and the qualitative behaviour of D() depends on the scaling of the soft photon energy:
m 2 ( m): both exponential factors are of the same size, constraining the integration variable to r a. The multipole expansion is applicable here as the product r , and it
is realized in the full result (14) by expanding the trigonometric functions between brackets:

r
r
12
(r)2
24
sin
cos
=
1
+ O 4 r 4 ,
(19)
3
2
2
2
40
(r)
(r)
28
and thus
D() = de () + de(1) () +

= de () 1 + O 2 /a 2 ,
(20)
where the leading term is given by

8
dr r 3 G1 (0, r; )er/a .
de () =
ma
(21)
The quantity
1 + de () ae ()
gives the electric amplitude in the dipole limit, which was calculated in Ref. [1] (Eq. (60) therein)
using an interaction Hamiltonian written in a gauge invariant form in terms of the electric field.3
A formula for ae () in terms of a hypergeometric function was given in Ref. [5], and can also
be found in Appendix A. In this region the Coulomb interaction among the e+ e pair after
soft photon radiation has a strength given by O(1), and is resummed to all orders through
the = 1 Coulomb Greens function. Note that the usual QED perturbative expansion does not
accommodate binding effects and cannot be applied to obtain the photon spectrum in this region.
For smaller photon energies, m 2 the parameter 1 and ae () 2/m 2 , in agreement
with Lows theorem
prediction for this decay [1].
m ( m ): since now 1/a, the falloff of D() is ruled by the exponential
factor from the Greens function, so the characteristic distance
is r 1 1/m 1/2 . The argument of the trigonometric functions is constrained to r /m 1/2 , which allows to
multipole expand the full result also for radiated photons of this energy. The photon wavefunctions can be expanded out from the beginning and the leading term in such expansion is again
given by de () above, although higher terms yield now corrections proportional to (/m)n :
D() = de () + de(1) () +

= de () 1 + O(/m) .
A new feature arises in the whole region m 2 m because the parameter
small and allows to further expand the Coulomb Greens function inside de ():
0
G1 (0, r; )
m 1 + r r
f
e
G1 (0, r; ),
12 r 3
(22)
is
(23)
i.e., the leading term of the expansion in of the P -wave Coulomb Greens function is equal
to the P -wave projection of the free Greens function. The latter shows that we can treat the
Coulomb interaction as a perturbation in this region, and higher order terms in the -expansion
of G1 correspond to insertions of the Coulomb potential in the diagrammatic picture. Using the
result (23) and expanding out the o-Ps wave function er/a , one obtains the leading term in of
de (),
3 One has to approximate mE to make contact with the result in Refs. [1,5], since recoil corrections were
0
not considered there.

8
4
f
de () =
dr r 3 G1 (0, r; ) + = + O 2
ma
3

m 2
2 m
1
+ + O 2 ,
=
3
8m
8
29
(24)
which agrees with the -expansion of the exact result of de (), which can be found in Eq. (A.9)
of Appendix A. An /m correction to the lowest order result is generated automatically in the
O() term of Eq. (24) from the recoil correction which is included in . Higher order terms in the
expansion correspond to the neglected terms in the expansion of G1 and the o-Ps wavefunction.
(1)
The next term in the expansion of the photon wavefunctions, de () in Eq. (20), also produces
f
a leading term proportional to (/m) /2 /m, which is obtained by replacing G1 G1
r/a
and e
1:

(r)2
8
dr r 3 G1 (0, r; )er/a
de(1) () =
ma
40

2

2
2

2
5 f
2
dr r G1 (0, r; ) + =
=
(25)
+O 2 .
5ma
15
The relation
2
1 2
=
+
=
m 1 + 4m
m
(26)
can be used in Eq. (25) to write de(1) () as an expansion in /m. The exact result for de(1) () can
be found in Appendix A. For completeness we also give there the exact result of the full electric
amplitudes D().
Corrections of O(/m) can also arise considering higher powers in the momentum expansion
of the annihilation amplitude A(2 ) . The higher order terms in p will pick intermediate e+ e
pairs with higher angular momentum L when they are transformed into derivatives acting on the
Greens function. The scaling of this part of the electric amplitude is of order
p
0|p
m
L1
G|x
L1
0|pG|x ,
L1
i.e., shows a suppression (/m) 2 with respect the L = 1 case, since p m. On
the other hand, the radiative transition to the intermediate e+ e L-wave state, 3 S1 + 3 LJ ,
will select an L-tensor in the multipole expansion of the photon wavefunctions, that will yield
L1
a (r)L1 (/m) 2 extra factor with respect the lowest order transition amplitude (L = 1).
Combining both factors we see that higher order terms in the expansion of A(2 ) for the electric
amplitude are thus down by an overall factor (/m)L1 . In order to compute these corrections it
is convenient to write the expanded amplitude of A(2 ) in terms of the spherical harmonics built
from the vector p, so they can project the required L-wave component of the Greens function
directly. A representation of the spherical harmonics in terms of the Cartesian coordinates of the
vector is specially suited for this task, and can be found in Ref. [9].
Let us now turn to the magnetic contribution. We will not detail all the steps as we did for the
electric amplitude, but just quote the result:
30
Mm =

0|A(2 ) |n, P in, P|i{ B(x1 ) + B(x2 )}|o-Ps
Eo En,P
n,P
= i

e
0 (0)W0 A(),
0
m
with

A() = 8
dr rG0 (0, r; )er/a sin

r
,
2
(27)
(28)
and = (k )/. Photon emission through a B term can induce a transition from a triplet
to a singlet spin configuration. The constant term in the annihilation amplitude A(2 ) allows only
for intermediate e+ e states with quantum numbers 1 S0 , which are contained in the = 0 component of the Greens function. As in the case of the electric amplitude, the falloff of the Greens
function fixes the characteristic distance for the multipole expansion to r 1 , so that we can
always expand in r for m. Under the dipole approximation, the photon wavefunctions
are dropped and the radiative matrix element reduces to the scalar product between the ground
state and the rest of Coulomb wavefunctions, which are mutually orthogonal. In that limit, only
the 1 S0 (p-Ps) ground state contributes [1], whose energy differs from that of the ground state
o-Ps only if the hyperfine splitting is considered. The latter is only relevant for the study of the
o-Ps decay spectrum at very low photon energies, m 4 [1], and can be neglected here. The
multipole expansion of the magnetic amplitude A() then reads:

(r)2
A() = 4 dr r 2 G0 (0, r; )er/a 1
+
24
(1)
() +
= am () + am

= am () 1 + O(/m) ,
(29)
where the magnetic amplitude in the dipole approximation,

am () =
1
,
1 + /4m
(30)
is trivially equal to one if recoil corrections are neglected. If we take into account higher terms
in the expansion of the photon wavefunction, transitions between the ground state and S-wave
radial excitations are possible, and the Coulomb interaction becomes relevant. For photons with
energies in the range m 2 m the Coulomb interaction can be treated perturbatively and
(1)
the first O() correction arises in the subleading term in the multipole expansion, am (), and
(1)
is proportional to (/m) /2 /m (see Eq. (A.15)). The exact result for am () and for
A() are given in Appendix A.
According to Eqs. (20), (29), the NRQED calculation of the o-Ps decay amplitude at leading
order in the multipole expansion reads:

W0
am () .
M = Me + Mm = e0 (0) W1 ae () + i
(31)
0
0
m
From a knowledge of the amplitude, we can derive an expression for the o-Ps differential photon
spectrum at low energies ( m) written in terms of the electric and magnetic amplitudes [1]:

7
m 6
d
2
2
=
x |am | + |ae | , x /m.
(32)
dx
9
3
31
Let us now focus on the region m. It is clear that since in this region Coulomb effects
can be accounted for perturbatively, one can apply the conventional QED loop expansion. The
leading (LO) and next-to-leading (NLO) terms in of the o-Ps 3 NRQED amplitude when
x are obtained by taking
ae = 1
2
+ O(),
3 x
am = 1 + O(),
and the NRQED spectrum in this region, up to NLO, thus reads:

d
2m 6
14
=
5x x + O 2 .
dx
27
3
(33)
(34)
The LO term of the NRQED amplitude (31) matches the 0 limit of the tree-level QED
calculation first done by Ore and Powell [10], as it was shown inRef. [1]. The same is of course
also true for the LO photon spectrum. The comparison for the x correction requires the oneloop result of the QED series. In Section 3 we will derive the soft photon limit of the 1-loop
QED spectrum obtained from an analytic expression of the phase-space
distribution given in
Ref. [6], and check that the leading term when 0 agrees with the x correction in the
NRQED spectrum shown above.4 In Section 4 we will compute the leading term of the 1-loop
QED amplitude for soft photon energy by means
of the threshold expansion, and show that it
arises from a loop-momentum region p m, and that it indeed agrees with the NRQED
result. Concerning higher orders in the /m expansion for the O( 0 ) and O() terms discussed
above, any possible mismatch between the QED and the NRQED amplitudes would require that
we introduce a short-distance e+ e 3 local contribution in the NRQED Hamiltonian.5
3. One-loop QED spectrum for m 2 m
An analytic evaluation of the 1-loop o-Ps 3 decay amplitude has not become available
until recently [6]. In the later work, Adkins has provided a compact form for the one-loop phasespace distribution, which allows to obtain the one-loop correction to the energy spectrum upon
integration:
d1
m 7
=
dx1
36 2
1
dx2
1x1
F (x1 , x2 ) + permutations ,
x1 x2 x3
(35)
where xi = i /m (i = 1, 2, 3), with i the energy of the photons, which satisfy that x1 + x2 +
x3 = 2. Analytic expressions for the function F (x1 , x3 ) can be found in the appendix of Ref. [6].
They involve a number of intermediate functions of the variables xi , including dilogarithm functions. Therefore an analytic integration of the formula (35) may be difficult to obtain. However,
for an evaluation of the low-energy spectrum one does not need to perform the full integration
before taking the limit x1 0. A Taylor series in x1 can be obtained by first taking the small
photon energy limit of the integrand in Eq. (35) up to the desired accuracy, and afterwards doing
4 A comparison between the effective theory O() spectrum and a numerical calculation of the one-loop QED result
[11] has been done by Voloshin [5], showing a good agreement.

5 Let us remind that the NRQED Hamiltonian is constructed to reproduce the QED amplitude neglecting binding
effects.
32
the (trivial) integration. Since also the variables 1 x2 and 1 x3 are small when x1 0, some
care is needed in order to perform this limit.
The two essential variables that parameterize the 1 3 phase-space have been chosen in
Eq. (35) to be x1 and x2 , where x1 is the variable of the differential spectrum. Therefore x1
x2 1. Note that in the x2 integration the limits imply that 1 x2 < x1 , so we can write 1 x2
x1 t, with t [0, 1], and change the integration variable to t :
d1
m 7
=
dx1
36 2
1
dt
0
F (x1 , t)
,
(1 x1 t)(1 x1 (1 t))
(36)
where
F (x1 , t) = F (x1 , 1 x1 t) + permutations.
(37)
The integrand in Eq. (36) is symmetric under the change t (1 t), which is indeed equivalent
to changing x2 x3 . With the redefinition above we can Taylor expand the integrand in the
variable x1 while keeping the variable t fixed, and the resulting terms turn out to be well behaved
in the x1 0 limit. Up to the second non-trivial order it yields a relatively simple phase-space
distribution:
F (x1 , t)
(1 x1 t)(1 x1 (1 t))

16
3 4t (1 t) x1
=
3

3/2
+ 2t (1 t) 3 2 48 log 2 + 40 + 2 + 16 log 2 60 x1 + O x1 ,
(38)
that can also be written back in terms of x2 if needed. Accordingly, the 1-loop QED correction
to the low-energy spectrum reads (writing x1 x)

2

3/2
3
35
14
d1 2m 6
.
x
+
=
x
+
O
x
(39)
dx
3
4
2
27 2
We see that the x term above agrees with the corresponding one in the NRQED formula,
Eq. (34). The x correction also matches the corresponding term in the EFT computation, which
was included in the formula (34) by Voloshin [5] considering the O() correction to the electron
gyromagnetic ratio and the O() corrections to the annihilation amplitude e+ e 2 . The
latter correction can be accounted for in a straightforward way if the NRQED spectrum is written
in terms of local four-fermion operators which effectively integrate out all possible states of
hard photons [1]. In this description, the coefficients multiplying the squares of the electric and
magnetic amplitudes in the photon spectrum (32), are identified with the imaginary part of the
matching coefficients of S- and P -wave four-fermion annihilation operators (see Eqs. (21), (74)
of Ref. [1]). These coefficients can be found up to O( 3 ) in Ref. [12], and yield a part of the
x correction shown above. The remaining part, due to the radiative correction to the electron
gyromagnetic ratio, is accounted for in the NRQED Lagrangian naturally as the 1-loop matching
coefficient of the B interaction.
To conclude, we wish to point out that higher order corrections in the x-expansion of the
1-loop low-energy photon spectrum can be computed straightforwardly from the analytic phasespace distribution given by Adkins through the expansion procedure outlined here.
33
4. Threshold expansion of the QED 1-loop diagrams

The threshold expansion technique is a prescription developed by Beneke and Smirnov [7] for
the asymptotic expansion of loop integrals involving heavy massive particles close to threshold
(i.e., that are moving at small velocities). It provides the velocity expansion of an amplitude up to
a certain order as a set of simpler integrals than those present in the original Feynman diagram.
The method relies on the so-called strategy of regions, which replaces the integration over the
whole space of loop momenta to the integration only over some specific regions. The first and
crucial step is to identify the relevant momentum regions in the loop integrals. These in principle
follow from the pole structure of the Feynman propagators, that is characterized by the relevant
scales that appear in the problem. For amplitudes involving a non-relativistic pair, three scales
enter the dynamics: the mass of the particle, m, the relative 3-momentum mv and the nonrelativistic energy mv 2 . Accordingly, four different loop momentum regions were identified to
give non-zero contributions in Ref. [7]:
hard: q0 q m,
soft:
q0 q mv,
potential: q0 q2 /m mv 2 ,
ultrasoft: q0 q mv 2 .
(40)
In a second step the original integration region must be decomposed into a sum of integrals,
one for every region above, and a Taylor expansion in the parameters that are small in each regime
performed. Every integral, containing just one scale, thus contributes only to a single power in the
velocity expansion, which can be determined easily before integration. The procedure requires
the use of dimensional regularization, even if the original integration is finite.6 Following these
heuristic rules, the authors of Ref. [7] reproduced the exact v expansion of some one-loop and
two-loop examples. Although a formal proof of the validity of the asymptotic expansion close to
threshold has not been given, the perfect agreement in the examples supports their use in general
one-loop diagrams.
We now wish to apply the threshold expansion to the 1-loop o-Ps 3 amplitude when one
of the photons has an energy m, where a conventional perturbative QED calculation is
still valid. The momentum of the emitted soft photon, k m, introduces a further scale in the
amplitude and, as we will show below, gives rise to a new region that contributes to the exact
result. It corresponds to a loop momentum
q0 q2 /m ,
(41)
which arises from the heavy particle propagatorsand indeed resembles the potential region but
with a different scaling for the 3-momentum q m. It thus corresponds to a small momentum
region (q m) which lies in between the hard relativistic region and the small momentum
regions shown in Eq. (40), and we shall refer to it in the following as soft-radiation region.
The physical origin of the new region has been discussed in the NRQED description of the o-Ps
decay in Section 2.
6 Other analytic regularization procedures can be employed, since the important point is that scaleless integrals are put
to zero.
34
(a)
(b)
Fig. 2. Ladder and double-vertex graphs contributing to the o-Ps decay amplitude to order . Momenta flow in the same
direction as the arrows.
The 1-loop graphs contributing to the o-Ps 3 amplitude can be classified into one of
the following categories [6,13]: self-energy, vertex corrections, double-vertex, ladder and annihilation contributions. The leading term in the /m expansion is given by the ladder and
double-vertex graphs, shown in Figs. 2(a) and 2(b), respectively. The momenta flowing through
the fermion propagators have been written in the figure. The positronium atom four-momentum is
P = k1 + k2 + k3 , with components P = (M, 0) in the o-Ps c.m. frame, and we can take M 2m
safely for the discussion that follows. Note that the assignment of momenta in the figure implies
that the electron and positron inside the bound state have incoming momenta P /2 = (m, 0). The
evaluation of the 1-loop diagrams for the o-Ps decay with e+ e strictly at threshold is a valid
procedure to obtain the matching coefficients of the NRQED annihilation operators that give the
o-Ps total decay rate, as it has been shown in Ref. [13]. Setting the velocity to zero automatically
kills off the contributions of the small momentum regions shown in Eq. (40), so that only hard
and soft-radiation momentum regions will survive in the QED 1-loop o-Ps amplitude.7 In order
to show that the threshold expansion technique for diagrams with a soft photon emission does not
only apply for this particular kinematic configuration, we will consider in Appendix B a more
general case with P 2 = 4m2 , where several small momentum regions exist at the same time.
Let us consider the ladder graph shown in Fig. 2(a), where the soft photon is the one attached
to the incoming electron. Therefore k1 m, while k2 , k3 are hard photons scaling as m. Two
integrals turn out to be relevant in the 1 0 limit of the ladder amplitude. The first one is the
5-point scalar integral:

[d D q]
I0 =
.
2
2
2
2
2
2
2
2
2
q [(q + P /2) m ][(q P /2) m ][(q + P /2 k1 ) m ][(q P /2 + k3 ) m ]
(42)
The integration measure is defined as [d D q] d D q/i(2)D , in D = 4 2
dimensions, and the
standard +i
prescriptions are implicitly understood in the propagators. When the loop momentum is hard, we can expand the fermion propagator in k1 . In this way, only the scale m enters
into the propagators in the hard region. Taking into account that [d D q] m4 , the leading term
(h)
in the contribution from the hard region is estimated to be I0 m6 . However, for 1 0 the
dominant contribution to the integral above comes from a loop-momentum region where q m.
Since the momentum that flows through the fermion propagator connecting the two hard photons
7 Loosely speaking, in the approach of Ref. [13] the non-relativistic dynamics is properly taken into account by the
NRQED part of the calculation.
35
is always hard, when q m we can expand

1
1
q P 2q k3
=
+
[(q P /2 + k3 )2 m2 ] P k3
(P k3 )2
(43)
so that the propagator effectively shrinks to a local annihilation vertex. Likewise, in the remaining
massive propagators we can drop the q02 terms with respect q P = 2mq0 . The leading contribution of I0 in the small momentum region is thus

1
[d D q]
(small)
I0
. (44)
=
2m3
(q02 q2 )(2mq0 q2 )(2mq0 q2 )(2mq0 2m1 q2 )
We have also dropped the term q k1 2mq0 . Closing the integration contour in the upper
complex q0 -plane, we pick the contribution from the pole of the massive propagator at q0 =
(small)
reads
q2 /2m + i
. The contribution of the residue of this pole to I0

n
3
1
i
1
(s-r)
=
d q 2 2 2
(m1 ) 2 + O(
)
I0 =
(45)
16m2 3
(q ) (q + m1 ) 64m2 3
with n = D 1. In deriving Eq. (45) we have also expanded the photon propagator in q02 , since
q02 q2 once we pick the residue from
the fermion pole. The integral in Eq. (45) is finite and it is
dominated by loop momentum q m, which corresponds to the scaling of the soft-radiation
loop momentum region introduced above.8 As in the hard region, the contribution from the softradiation region to the final result can be readily estimated before making any integration by first
expanding the integrand according to the hierarchy q0 q2 /m m1 . For the leading term the
propagators yield a factor m2 (m1 )4 , while [d D q] m1 (q2 )5/2 m1 (m1 )5/2 , so that
(s-r)
we obtain an overall scaling I0 m3 (m1 )3/2 , as found in Eq. (45). The exact result of the
integral I0 can be easily derived from the 4- and 5-point functions computed in Ref. [6]. We have
(s-r)
checked that the leading term in the 1 0 limit of the exact expression for I0 agrees with I0
(small)
above. There is also a contribution to I0
from the pole of the massless propagator at q0 =
|q| + i
. After picking up the corresponding residue the remaining integration is dominated by
loop-momenta |q| 1 . The characteristic loop-momentum in this region is thus the same as for
the soft region in Eq. (40), if we consider v . The contribution from this pole has an overall
scaling [d D q]/(mq0 )5 m5 11 , i.e., larger than the hard region one, although suppressed by
m1 with respect the soft-radiation region. We do not give the result because we will not keep
terms of the similar order that arise from other diagrams.
The second integral that we shall need for the leading term in the 1 0 limit of the ladder
amplitude is the 4-point tensor integral

D
qi qj
ij
,
d q 2
I0 =
(46)
q [(q + P /2)2 m2 ][(q P /2)2 m2 ][(q + P /2 k1 )2 m2 ]
with the indices i, j = 1, 2, 3. We have already expanded out the hard fermion propagator in
ij
order to write down I0 , since the leading contribution will come up from the q m regime. As
we shall see when we put back the numerators of the amplitudes, part of the q i q j term arises
when a q j is brought to the numerator of the integrand from the subleading term in the q m
expansion of the hard propagator shown in Eq. (43). Strictly, one should do the tensor reduction
8 The integral in Eq. (45) with replaced by
1
QCD was also found in a non-perturbative enviroment [24].
36
of the integral and afterwards apply the threshold expansion to the scalar integrals that appear.
However, since the tensor structure in the integral in Eq. (46) only involves the spatial indices, we
shall take the path of expanding the integrand according to the different loop-momenta directly
without prior reduction to scalar integrals. In the small loop-momentum region, the propagators
ij (small)
in I0
are identical to those in I0(small) , and thus they share the same pole structure. The
expansion of the propagators in the soft-radiation region leads to:

n
1
i
qi qj
ij
ij (s-r)
=
=
+ O(
).
d q 2 2 2
I0
(47)
8m
(q ) (q + m1 ) 96m m1
After dropping the k1 -dependence in the fermion propagator, the reduction of the tensor integral
ij
is trivial. The contribution from the gluon pole in I0 can be shown to scale as 1 /m3 , and it
is thus suppressed with respect the soft-radiation contribution.
For the double vertex diagram of Fig. 2(b) the leading term in the 1 0 limit comes from
the integral

[d D q]
,
I1 =
(48)
q 2 [(q P /2)2 m2 ][(q + P /2 k1 )2 m2 ]
after expanding the hard region propagator as done in Eq. (43). The expansion of I1 in the small
loop-momentum gives an analogous expression to (44) but with one fermion propagator less.
The soft-radiation region arises from the fermion pole at q0 = q2 /2m + i
, and yields

n
1
1
i
1
(s-r)
I1 =
(49)
=
d q 2 2
(m1 ) 2 + O(
).
4m
16m
q (q + m1 )
On the other hand, the contribution from the gluon pole in the small loop-momentum scales as
m2 , so it is again suppressed with respect the soft-radiation region.
Let us now include the numerators of the propagators and vertex factors in order to calculate
the full amplitude. We can use the identity

s (p)
vs (p)vs (p)
p
/ +m
s us (p)u
= 0
+ s0
(50)
p 2 m2 + i
p Ep + i
p + Ep i

with Ep = m2 + p2 m + p2 /(2m), to pick the numerators associated to the non-relativistic
fermion poles that are relevant when the loop-momentum is small, and afterwards perform the
non-relativistic expansion of the spinors in p/m. This will automatically lead us to the usual
non-relativistic Feynman rules for the vertices and propagators. The identity (50) effectively separates particle and antiparticle contributions, as it is conventional in the time-ordered perturbation
theory. The spinors in Eq. (50) have non-relativistic normalization, more precisely:

p

Ep + m
Ep + m m+E
s
s
p
(p)
=
,
v
,
us (p) =
(51)
s
p
2Ep
2Ep
s
m+Ep s
where s is the spin label. Let us illustrate how the non-relativistic expansion is performed for the
ladder diagram. The structure of the fermion propagators in the small loop-momentum region
has been shown in Eq. (44). Take, for example, the fermion propagator at the left of the soft
photon vertex, with momentum q + P /2. The expansion of this propagator for q m yields
2mq 0 q2 . The latter agrees (up to a normalization factor) with the first denominator that is
obtained from the identity (50),which reads q 0 + P 0 /2 Eq q 0 q2 /2m. Therefore we will
associate the electron spinors s us (q)u s (q) to this denominator in order to derive the leading
(a)
37
(b)
Fig. 3. The soft photon vertex (M1 ) and the 2 annihilation (M23 ) parts in the ladder graph.
term in the small momentum region. In the same way, the expanded fermion propagators after
soft photon emission and after the k3 vertex correspond to electron and positron contributions,
respectively, in the splitting of Eq. (50). For the massive line connecting the k2 to the k3 vertices,
we need to keep the full covariant propagator because the momentum flowing through it is always
hard. Taking into account the particle and antiparticle spinor wavefunctions from the adjacent
propagators, the non-relativistic reduction of this part of the ladder diagram is indeed equivalent
to the e+ e 2 annihilation amplitude of an incoming electron with 3-momentum q k1 and
an incoming positron with 3-momenta q (see Fig. 3(b)),
M23 = ie2 vs (q)/
3
/ u (q k1 ) + {2 3},
q/ P
/ /2 + k/3 m 2 s
(52)
where we have also included the contribution from the graph with the hard photons permuted.
Recall that we found already the non-relativistic expansion of the e+ e 2 amplitude in
Section 2. Indeed, the expansion of Eq. (52) for the hierarchy m q k1 , characteristic of the
soft-radiation region, yields the same leading and subleading terms already shown in Eq. (5), so
that
M23 = s (W0 + W1 q)s + O().
(53)
The size of the terms in the expansion of M23 above should be estimated considering the soft
radiation region scaling, q m1 m 1/2 . The reason why we only need to keep the two
first terms in M23 shall become clear when we put together all the pieces in the amplitudes.
Similarly, the non-relativistic expansion of the soft-photon vertex is also done easily once we
pick the particle contribution from the propagators (see Fig. 3(a)):
M1 = ieu s (q k1 )/
1 ur (q)

e
e
= i q 1 s r +
(54)
(k1 1 )s r + O 3/2 .
m
2m
We can check that the non-relativistic reduction of the amplitude M1 agrees with the Feynman
rule that is derived for the e e + vertex from the non-relativistic interaction Hamiltonian,
Eq. (1). Finally, the photon exchange between the initial e+ e reduces to a Coulomb-like potential:
Mc =
0
ie2
ie2
u r (q) u(0)v(0)
vs (q) = 2 r s + O ,
2
q
q
(55)
in the region q0 q2 /m m, and with , the Pauli spinors of the initial e+ e . We can now
put together the different pieces, Mc , M1 and M23 , including the denominators and taking into

account that s s s = s s s = 1. The leading term of the threshold expansion of the ladder
amplitude in Fig. 2(a) (plus the graph with the photons k2 and k3 interchanged) is thus equal to
38
W1 q (q ) iW0 1
2
2
2
0 q /2m)(q0 + q /2m)(q0 1 q /2m)

m

m
W0
= e W1
+ ie
.
1
3 1
m
2 1
Mladder =
e3
m
d nq
(q2 )(q
(56)
The integrals appearing in Eq. (56) are proportional to the soft-radiation contributions of the
ij
integrals I0 and I0 , computed in Eqs. (45) and (47), respectively. We can proceed in a similar
way to obtain the leading term in the threshold expansion of the double-vertex diagram, Fig. 2(b):
Mdou.ver. =

ie3
W0 1
2m

W0
= ie
1
m
2
[d n q]
(q2 )(q

m
,
1
+ q2 /2m)(q
1 q2 /2m)
(57)
where we have required the contribution from the soft-radiation region of I1 , Eq. (49). For the
double-vertex amplitude there is no contribution from the q 1 term in the soft-photon vertex
because the incoming electron is static. The W0terms from the ladder and double-vertex graphs
cancel each other and we are only left with a m/1 proportional to W1 . It is straightforward
to verify that the contribution from the ladder graphs where the soft photon vertex is attached
in the lower part of the diagram (i.e., to the incoming positron line) gives the same W1 term.
Therefore, the leading term in 1 0 limit of the 1-loop e+ e 3 QED amplitude reads

2
M1-loop = e W1
3
m
,
1
(58)
which agrees with the leading O() term in the NRQED computation of this amplitude, Eqs. (31)
and (33), once we include the o-Ps wavefunction factor.
A comment on the rest of 1-loop diagrams that had been not considered is appropriate. First
one should note that all diagrams where the soft photon k1 is placed among the two hard ones
scale as O(1) because the momenta flowing through the fermion propagators that connect two
photon vertices are always hard. With respect the annihilation contribution, the scaling can be
inferred from that of the ladder graph if we replace the t -channel photon exchanged between the
initial e+ e by an s-channel photon exchange. The replacement carries a q2 /m2 suppression
in the non-relativistic expansion, which is of order 1 /m for the soft-radiation contribution.
Regarding the self-energy and vertex correction topologies, they correspond to typical relativistic
corrections whose effects are not expected to be enhanced when we explore the non-relativistic
regime. Indeed it can be easily checked from a quick analysis of the propagators in the small loopmomentum region that for the integrals in the vertex and self-energy corrections, we can always
avoid the fermion poles so that only the pole in the photon propagator at q 0 q contributes. In
this way, the small loop-momentum expansion of the 2- and 3-point scalar integrals give results
which are down by (1 /m)2 and 1 /m, respectively, with respect the hard region contributions
to the same integrations.
It is interesting to note that the expressions (56), (57) correspond to the amplitudes that would
be written down from the diagrams displayed in Fig. 4(a) using the usual NRQED Feynman rules
39
(a)
(b)
Fig. 4. (a) NRQED diagrams that contribute to the O( m/1 ) term of the e+ e 3 decay amplitude. The zig-zag
lines represent soft photons; (b) local vertex that describes real soft photon radiation from the heavy fermionantifermion
pair in vNRQED. The 2 label denotes that the vertex arises from the integration of two hard photons.
with free fermions as asymptotic states,9 and after we drop the k1 term in the fermion propagator:
1
1

.
q 0 1 (q k1 )2 /2m q 0 1 q2 /2m
(59)
Keeping the full NRQED propagator and expanding afterwards leads of course to the same leading order term, but unnecessarily complicates the calculation. In the usual NRQED counting for
bound state systems, the heavy fermion and antifermion inside the bound state have energies
E m 2 and relative 3-momentum q m, so they are classified as potential degrees of freedom. As we explained in Section 2, the radiation of a soft photon from the e+ e allows for
higher virtualities in the relative 3-momentum, of order q m1 . Therefore the heavy particles propagating in the loops of Fig. 4(a) do not obey the potential scaling (otherwise the massive
propagator (59) for a potential fermion will reduce to just 1/1 and we will miss the contribution
from the soft-radiation region). The fact that we have to further expand the NRQED propagators
in the soft-radiation regime is an unwanted feature because it jeopardizes power-counting.
However, we can take advantage from the fact that soft radiation takes place at distances 1/ m,
smaller that the typical size of the Coulomb bound state, to integrate out the contributions stemming from the soft-radiation momentum region. From the point of view of the potential fermions,
soft radiation is a short-distance process, just like the 2 -annihilation, and can be described by
a local vertex. In the vNRQED/vNRQCD [14] formalism (with the external sources for the hard
photons integrated out) the decay spectrum in the soft-energy can be calculated from the imaginary part of the matrix element of a 6-field operator as the one shown in Fig. 4(b). The matching
coefficient of such operator can be obtained by comparing with the QED result in the 1 m
region: at tree-level it is just a constant 3 (the limit x 0 of the OrePowell decay spectrum
[10], divided by the x factor from the phase-space), while at O() it is proportional to 4 m/.
Note that this picture is similar to the NRQCD factorization approach [12] used to describe the
production and annihilation of heavy quarkonium, that we briefly discuss in the next section.
5. Relevance for heavy quarkonium
The threshold expansion technique applied above for the o-Ps 3 decay, can also be useful
to calculate the direct photon spectrum of quarkonium radiative decays in the soft-energy region.
9 Recall that we used bound state perturbation theory in coordinate space to compute the NRQED amplitude in Section 2. In the m region the leading O() correction in the latter computation could be obtained by replacing the
intermediate Coulomb Greens function by the free one and neglecting the o-Ps wavefunction dependence (see Eq. (24)).
This is effectively equivalent to consider the Coulomb interaction as a perturbation, so that in this regime the NRQED
computation can be done using as unperturbed states either free propagating fermions (as in Fig. 4(a)) or e+ e Coulomb
eigenstates (as in Section 2).
40
Heavy quarkonium decay rates can be understood within the NRQCD factorization approach
[12], which allows to separate the short distance physics related with the heavy quarkantiquark
annihilation process from the long distance bound state dynamics. An operator product expansion
can be written down for the direct photon spectrum in, for example, decays:

d
(60)
Cn (x) |On | ,
=
dx
n
where the sum above extends over all QQ[n]

configurations that can be found inside the quarkonium, and the Cn (x) are short distance Wilson coefficients that can be determined from the
annihilation cross section of the on-shell QQ[n]

pair as a perturbative series in s (mb ). The
On in the long-distance matrix element are NRQCD operators which are organized in powers
of the relative velocity of the heavy quarks. At leading order only the color-singlet operator
O1 (3 S1 ), that creates and annihilates a quarkantiquark pair in a color-singlet 3 S1 configuration,
contributes. The nonperturbative NRQCD matrix element is related to the wave function at
the origin

2

i

|O1 3 S1 | = |p i p p
(61)
p | = 2Nc (0) ,
where now p , p refer to heavy quark and antiquark fields respectively. The matching coefficient at leading order in s comes from the tree-level QQ gg annihilation [15],

32 s2 Q2b 2 x (1 x)x 2(1 x)2 log(1 x) 2(1 x) log(1 x)
C1 (x) =
+
+
,
27 m2b
x
x2
(2 x)2
(2 x)3
(62)
which has identical x-dependence to that of the OrePowell spectrum for o-Ps [10]. Indeed the
factorization approach applies trivially to QED bound states, where everything is calculable perturbatively and the bound state wavefunction reduces to the Coulomb one. The leading order
coefficient for Ps decays, C 1 (x), is obtained by the trivial replacements s2 Q2b /m2b 2 /m2
and 32/27 16/9 in Eq. (62). In the soft-energy region, it reads

40 3
x + O x2 .
C 1 (x) =
(63)
2
27 m
The O() correction to the o-Ps 3 amplitude for soft photon energies computed in the previous sections can also be included, so that

0
8 3
14
,
x
+
O
x
5x
C 1NLO (x) =
(64)
27 m2
3
that we can plug in Eq. (60) to obtain the o-Ps photon spectrum to NLO in the soft-energy region
(we also need to replace c and NC 1 in Eq. (61)). A similar result to that of Eq. (64)
could be obtained for the O(s3 ) short-distance coefficients Cn (x) in heavy quarkonium decays
by doing the threshold expansion of the corresponding 1-loop QCD diagrams. The latter could
provide a determination of the O(s ) corrections to the quarkonium photon spectrum in the softenergy range of photon energies. To our knowledge, s corrections to the coefficient C1 (x) in
Eq. (62) are only known numerically [16]. We should be aware though that the fragmentation
contributions to the photon spectrum in quarkonium decays (i.e., those in which the photon is
emitted from the decay products) become important in the low-x region [17,18], namely for
x 0.3 in decays, and have to be properly taken into account for a computation of the full
41
spectrum. At lower energies, x s2 , the emission of the photon can produce transitions to virtual
bound states and does not longer belong to the short-distance part of the decay [1]. The standard
NRQCD factorization also breaks down at large values of the photon energy [19], where one
needs to consider also collinear degrees of freedom [20].
6. Conclusions
The region of photon energies m in the radiative decay of a heavy fermionantifermion
pair has been analyzed in this work through the study of the three-photon decay of ground state
orthopositronium. The soft-energy region effectively separates the regime where binding effects
become essential, from the hard-energy region, where the details of the bound state dynamics
are irrelevant. The NRQED framework is able to yield the correct o-Ps spectrum from energies
< m 2 , where it properly accounts for bound state effects, up to energies m, where the
binding can be neglected and the calculation is simplified by replacing the Coulomb Greens
function with the free one. We have found agreement in the comparison of the o-Ps spectrum in
the soft-energy region as computed with NRQED and with conventional perturbation theory in
QED. For the NRQED computation, the leading order approximation in the multipole expansion
of the photon field yields already the dominant contribution in the whole energy range m,
as it had been argued by Voloshin [5]. We have explicitly calculated the contribution from higher
order multipoles to the o-Ps decay amplitude in NRQED, and discussed the size of relativistic
corrections coming from other sources in the soft-energy region.
The 1-loop QED result of the o-Ps spectrum in the soft-energy region has been obtained by
two different methods. First by expanding the analytic expression for the phase-space distribution obtained by Adkins [6], and second by means of the threshold expansion technique. With
the latter method it has been shown that the dominant contribution arises from a new loopmomentum region q 0 q2 /m m. This momentum region arises naturally when the heavy
fermionantifermion system decays emitting a soft photon and has to be added to the rest of momentum regions which are known to be relevant for the analysis of heavy particleantiparticle
loop diagrams. The identification of momentum regions that have not been previously considered
is mandatory for the success of the asymptotic expansion method in a wider range of kinematic
situations. It is also of conceptual importance for the construction of EFTs for these systems
with well-defined power-counting rules and able to reproduce the correct low-energy behaviour.
Acknowledgements
I have benefited from numerous discussions on the subject of this work with A. Hoang and
A. Manohar. I also thank them for their comments on the manuscript. I am particularly grateful
to G. Adkins for sharing the details of his calculation and for very useful correspondence. Part of
this work was done at the Instituto de Fsica Corpuscular de Valencia (IFIC). This work has been
supported in part by the EU Contract No. MRTN-CT-2006-035482 (FLAVIAnet).
Appendix A. Closed formulas for the NRQED o-Ps 3 amplitude
We give first some useful formulae for the computations that follow. The partial waves of the
Coulomb Greens function can be written in terms of the associated Laguerre polynomials [22].
42
Setting one of the arguments to zero they read,
G (0, x, ) =

L2+1
(2x)
m
n
(2)2 ex
,
2
(n + + 1 )(2 + 1)!
(A.1)
n=0
with [23]
(x) =
L2+1
n

n

(1)r n + 2 + 1
r=0
r!
nr
xr .
(A.2)
A related integral is

m
Jn,
() =
dz ez L2+1
(z)zm = (1)m
n
dm 0
J (),
d m n,
(A.3)
that we shall need only for the = 0 and = 1 cases:

1 n+1
0
Jn,0 () =
+ 1,
( 1)n+3
n2 5
+
(
3)
+
+ n + 3.
(A.4)
2
2
n+1
Let us start with the electric amplitude D(), Eq. (14). Using the representation (A.1) and making
the change of variables z = 2r, it can be written as:
3

1
1
0
D() = 16
(A.5)
iJn,1 () +
J () + h.c.,
n+2
4 n,1
0
() =
Jn,1
n=0
with
1 + i
2
4
(recall from Section 2 that = 1/a). With the help of the results (A.3), (A.4) and the identity,
=

xn
2 F1 (a, 1; a + 1; x)
=
,
n+a
a
(A.6)
n=0
we arrive to a closed expression for D() in terms of a hypergeometric function,

3
D() = 162

(2 + 1)
i( 1) 4
2 F1 (2 , 1; 3 ; )
+
+ h.c.,
2
4
(A.7)
where we have defined = ( 1)/. The expansion in (/)2 of the result (A.7) is equivalent to the one which is obtained by directly expanding the trigonometric functions inside the
integral representation of D(), see Eq. (19). The leading term (dipole approximation) yields a
representation of de (), Eq. (21), in terms of an hypergeometric function,

4 (1 + 2) 8 2 ( 1)
1
de () =
(A.8)
+
.
2 F1 2 , 1; 3 ;
3 (1 + )2
3 ( 2)(1 + )3
+1
43
The representation above agrees with the corresponding one for ae () = 1 + de () found in [5].
The expansion of de () for small is legitimate in the m 2 m region, and follows
immediately from (A.8),

4
8
de () = + (1 log 2) 2 + O 3 ,
(A.9)
3
3
where higher order terms in account for further Coulomb interactions. The subleading term in
the multipole expansion of the electric amplitude corresponds to the (/)2 term in the expansion
of D() and so on. Since (/)2 /m (see Eq. (26)), multipoles of rth order get suppressed
by the overall factor (/m)r . A formula for the rth multipole is obtained in a more direct way
through the -expansion of the trigonometric functions inside D() prior to integration. It reads
2r
r +1
J 2r+3 (),
de(r) () = (1)r+1
(A.10)
(2r + 3)! 4
n + 2 n,1
n=0
(r)
with = (1 + )/2. A closed expression for de () can be obtained with the aid of Eqs. (A.3),
(A.4) and a generalization of the identity (A.6),
r

nr n
(a)1 Lir (x) + r, + (a)r1 2 F1 (a, 1; a + 1; x).
x =
n+a
n=0
(A.11)
=1
For example, the next-to-leading multipole (r = 1) reads

2
1 2
3 2
de(1) () =
15 (1 + )4
1

2
4(1 + )
1
+ 2
2 F1 2 , 1; 3 ;
+1
( 3 + 2)(1 + )
2

2
=
+ O 2 2 / 2 ,
15
(A.12)
whose leading order term in the small expansion agrees with our previous calculation using the
free Greens function in Eq. (25).
Let us also give the exact formula for the full magnetic amplitude A(). From Eq. (28), we
obtain
1
im
J 1 () + h.c.
2
n + 1 n,0
n=0

im
F
(1
,
1;
2
;
)
+ h.c.
1+
=
2 1
2
(1 + )
A() =
(A.13)
As in the case of the electric amplitude, the multipole expansion of A() is realized through a
series in (/)2 . The exact result for the magnetic dipole term, am (), has been already shown
in Eq. (30). The next-to-leading term in the multipole expansion reads

1
1 m 2
(1)
am () =
2
2
12
(1 )3 (1 + )

1
2
3
2
.
3 + + 5 8 2 F1 1 , 1; 2 ;
(A.14)
+1
44

(1)
The expansion of am () for small gives

2
2
4
(1)
,
1+
am () =
1 +O
4m
4m
3
(A.15)
where we have made use of the relation between 2 and given in Eq. (26). Finally, we give the
general form of the rth order magnetic multipole:

1
(1)r m 2r
(r)
am
(A.16)
() =
J 2r+2 (),
(2r + 1)! 4 2 4
n + 1 n,0
n=0
which scales as (/m)r .

Appendix B. Threshold expansion of a triangle diagram with a soft radiated photon
In this appendix we show a further example of the application of the asymptotic expansion
method near threshold to a 1-loop diagram with massive lines and a soft momentum component
radiated off. The diagram is shown in Fig. 5. This example is similar to the 1-loop ladder diagram
considered in Section 4, but we shall not take here the total incoming momentum P 2 equal to
4m2 , but instead define y = m2 P 2 /4 (m)2 . The parameter y, which appears in the massive
propagators, will give rise to a non-zero contribution from the potential region, which was not
present in the ladder diagram evaluated at threshold. For simplicity, we have substituted the
Coulomb-ladder interaction by an effective production vertex. We shall take the soft photon
momentum k m, and keep terms in the asymptotic expansion up O().
If we consider scalar propagators, the integral of Fig. 5 is given by

[d D q]
I2 =
(B.1)
,
(q 2 + q P y)(q 2 q P y)(q 2 + q P 2q k P k y)
where we have chosen a routing of the external momentum P through the massive lines of the
graph which allows to use the scaling arguments defined in Eqs. (40) and (41). When the loop
momentum is hard (q m), we can expand the propagators in the small variables y and k.
Keeping terms up to O() the expansion in the hard region reads:

2q k + P k
[d D q]
(h)
I2 =
+
1+ 2
(q 2 + q P )2 (q 2 q P )
q +q P

1
=
(B.2)
+ O 2 ,
1+
3m
32m2 2
and higher order terms can be calculated straightforwardly. The first region that we shall consider
when the loop momentum is small is the soft-radiation region, that we already found in the 1-loop
diagrams contributing to the o-Ps 3 amplitude. According to the hierarchy q 0 q2 /m
m we expand the propagators retaining terms up to O():

[d D q]
I2(sr) =
2
2
(q + q0 P0 y)(q q0 P0 y)(q2 + q0 P0 P0 )

y q02
y + 2q0 q02
y q02
+
+
1+
q2 + q0 P0 y q2 q0 P0 y q2 + q0 P0 P0

4(q k)2
+
+
(q2 + q0 P0 P0 )2
1
32m2

m
y
1+
+
+ O 3/2 .
2m 6m
45
(B.3)
The integrals that result from Eq. (B.3) are calculated

easily with standard methods. The leading
term in the soft-radiation region is enhanced by 1/ with respect the hard region. The potential region, where q 0 q2 /m y/m, also contributes when the loop momentum is small. The
expansion of the integrand gives:

[d D q]
1
q2 + q0 P0 y
(p)
+
I2 =
1
+
P k
P k
(q2 + q0 P0 y)(q2 q0 P0 y)

y
1
=
(B.4)
+ O 2 .
32m2
The term P k dominates the massive propagator depending on k, which shrinks to a point in the
potential region. The subleading term between parentheses in Eq. (B.4) arises from the expansion
of the latter propagator, but gives a vanishing contribution because the numerator is proportional
to one of the propagators in front. Subleading terms in the expansion of the propagators not
depending on k generate corrections which scale as (q0 )2 /(q2 + q0 P0 y) 2 , that we do
not retain.
The soft loop-momentum region leads to a scaleless q-integration after the q dependence has
been expanded out from the denominators according to the scaling q0 q m:

[d D q]
(s)
I2 =
(B.5)
+ = 0,
(q0 P0 )(q0 P0 )(q0 P0 P0 )
and the same holds for the ultrasoft region. Contributions from the soft and ultrasoft regions can
arise in diagrams with massless propagators, like the 1-loop ladder graph of Section 4. Apart
from the latter, the analysis of the threshold expansion of the 1-loop ladder diagram for P 2 = 0
only leads to trivial modifications to the integrations for the various regions found here, and can
be obtained easily.
Finally, the exact result for I2 when k 2 = 0 can be read off the list of integrals in Ref. [6]:

2
1
1
(P k)2
P
L
L
,
I2 = 2
(B.6)
2
P
k
8
m
m2
with

2

s
.
L(s) = 2 arctan
4s
(B.7)
One can verify that the expansion of the exact result for m up to terms of O() is reproduced by the sum of the contributions of the regions in (B.2), (B.3) and (B.4).
Fig. 5. Scalar triangle diagram.
46
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
A.V. Manohar, P. Ruiz-Femenia, Phys. Rev. D 69 (2004) 053003, hep-ph/0311002.

W.E. Caswell, G.P. Lepage, Phys. Lett. B 167 (1986) 437.
P.D. Ruiz-Femenia, Nucl. Phys. B (Proc. Suppl.) 152 (2006) 200, hep-ph/0410211.
M. Luke, A.V. Manohar, Phys. Rev. D 55 (1997) 4129, hep-ph/9610534;
P. Labelle, Phys. Rev. D 58 (1998) 093013, hep-ph/9608491;
A. Pineda, J. Soto, Nucl. Phys. B (Proc. Suppl.) 64 (1998) 428, hep-ph/9707481;
A. Pineda, J. Soto, Phys. Rev. D 59 (1999) 016005, hep-ph/9805424;
M. Luke, A. Manohar, I. Rothstein, Phys. Rev. D 61 (2000) 074025, hep-ph/9910209.
M.B. Voloshin, Mod. Phys. Lett. A 19 (2004) 181, hep-ph/0311204.
G.S. Adkins, Phys. Rev. A 72 (2005) 032501, hep-ph/0506213.
M. Beneke, V.A. Smirnov, Nucl. Phys. B 522 (1998) 321, hep-ph/9711391.
A.V. Manohar, Phys. Rev. D 56 (1997) 230, hep-ph/9701294.
A.H. Hoang, P. Ruiz-Femenia, Phys. Rev. D 74 (2006) 114016, hep-ph/0609151.
A. Ore, J.L. Powell, Phys. Rev. 75 (1949) 1696.
G.S. Adkins, Ann. Phys. 146 (1983) 78.
G.T. Bodwin, E. Braaten, G.P. Lepage, Phys. Rev. D 51 (1995) 1125, hep-ph/9407339;
G.T. Bodwin, E. Braaten, G.P. Lepage, Phys. Rev. D 55 (1997) 5853, Erratum.
G.S. Adkins, R.N. Fell, J. Sapirstein Ann. Phys. 295 (1982) 136.
A.V. Manohar, I.W. Stewart, Phys. Rev. D 62 (2000) 014033, hep-ph/9912226;
A.V. Manohar, I.W. Stewart, Phys. Rev. D 62 (2000) 074015, hep-ph/0003032;
A.H. Hoang, I.W. Stewart, Phys. Rev. D 67 (2003) 114020, hep-ph/0209340.
K. Koller, T. Walsh, Nucl. Phys. B 140 (1978) 449;
S.J. Brodsky, D.G. Coyne, T.A. DeGrand, R.R. Horgan, Phys. Lett. B 73 (1978) 203.
M. Kramer, Phys. Rev. D 60 (1999) 111503, hep-ph/9904416.
S. Catani, F. Hautmann, Nucl. Phys. B (Proc. Suppl.) 39BC (1995) 359, hep-ph/9410394.
F. Maltoni, A. Petrelli, Phys. Rev. D 59 (1999) 074006, hep-ph/9806455.
I.Z. Rothstein, M.B. Wise, Phys. Lett. B 402 (1997) 346, hep-ph/9701404.
S. Fleming, A.K. Leibovich, Phys. Rev. D 67 (2003) 074035, hep-ph/0212094;
X. Garcia i Tormo, J. Soto, Phys. Rev. D 69 (2004) 114006, hep-ph/0401233.
V.A. Smirnov, Phys. Lett. B 465 (1999) 226, hep-ph/9907471.
L.C. Hostler, J. Math. Phys. 11 (1970) 2966.
M. Abramowitz, I.A. Stegun, Handbook of Mathematical Functions, Dover, 1970.
N. Brambilla, A. Pineda, J. Soto, A. Vairo, Phys. Lett. B 580 (2004) 60, hep-ph/0307159.
Two-loop formfactors in theories with mass gap

and Z-boson production
A. Kotikov a,b , J.H. Khn c , O. Veretin b,d,
a Bogoliubov Laboratory of Theoretical Physics, JINR, 141980 Dubna, Russia
b II Institute fr Theoretische Physik, Universitt Hamburg, 22761 Hamburg, Germany
c Institut fr Theoretische Teilchenphysik, Universitt Karlsruhe, 76128 Karlsruhe, Germany
d University of Petrozavodsk, 185910 Petrozavodsk, Karelia, Russia
Received 20 May 2007; received in revised form 18 July 2007; accepted 20 July 2007
Abstract
The non-factorizable two-loop corrections to the formfactor both for a U (1) U (1) and a SU(2) U (1)
gauge theory with massive and massless gauge bosons respectively are evaluated at arbitrary momentum transfer q 2 . The asymptotic behaviour for q 2 is compared to a recent calculation of Sudakov
logarithms. The result is an important ingredient for the calculation of radiative corrections to Z-boson
production at hadron and lepton colliders.
1. Introduction
Precise measurements of cross sections for the production of massive and massless gauge
bosons were one of the central topics of LEP experiments. At the LHC similar reactions, namely
the production of W and Z-bosons, singly or in pairs, with or without additional quark or gluon
jets, will be crucial for precise studies of the electroweak and strong interactions. Single W - and
Z-boson production will be used for the determination of parton distributions and eventually even
for luminosity calibrations. A future linear collider, operating in the GIGA-Z mode, will measure
the properties of the Z-resonance with unprecedented precision. All these measurements will
rely on the theoretical knowledge of radiative corrections to better than one percent accurracy,
* Corresponding author at: II Institute fr Theoretische Physik, Universitt Hamburg, 22761 Hamburg, Germany.
E-mail address: veretin@mail.desy.de (O. Veretin).

doi:10.1016/j.nuclphysb.2007.07.018
48
A. Kotikov et al. / Nuclear Physics B 788 (2008) 4762
perhaps even down to the level of several permille. QCD and electroweak radiative corrections,
as well as their interplay, thus will be crucial for the interpretation of these results.
QCD corrections to single W - and Z-production are identical to those for the DrellYan
process and have been evaluated in two-loop approximation in [1,2], those for Higgs boson
production in [3]. Electroweak corrections for the on-shell process were computed some time
ago (see, e.g., [4] and references therein). The next step evidently requires to combine QCD and
electroweak effects, resulting in non-factorizable terms of order weak s . For the inclusive Z
decay rate these terms have been calculated for final states with up-, down-, and bottom-quarks
[57] and turned out to be relevant for the precise determination of the strong coupling constant.
However, these results cannot be directly applied to the production process and to more differential distributions. For the Z-boson such corrections for high pT distribution have been obtained
in [8].
In the present paper we describe conceptial developments and concrete results which are important ingredients for the complete evaluation of these non-factorizable terms of order weak s .
In particular we consider those amplitudes which correspond to vertex diagrams with a virtual
gluon attached to one-loop electroweak corrections. These are relevant for the mixed corrections of order weak s to Z-boson production, and for hadronic Z decay. Essentially the same
diagrams are also important ingredients for the combination of photonic and weak corrections to
Z production in electronpositron collisions and to leptonic Z decays.
Our study identifies the infrared singular as well as the finite parts, investigates the structure
of these singularities and shows how they can be combined with real radiation to arrive at a finite
result. The infrared finite remainder will be presented in analytical form in terms of generalized
polylogarithms.
The form factor will also be investigated in the Sudakov limit M 2 /q 2 1. In the special
case of an Abelian theory the result coincides with the one of [9] (see also [10]) and allows
to contrast the logarithmic approximation with the complete result. The calculational method
relies on an approach that has already been successfully employed in a number of cases [11,12].
General considerations restrict the structure of the final result to a sum of basis functions (in
our casegeneralized harmonic polylogarithms up to fourth degree) with specific arguments
and prefactors. Calculating on one hand directly a large number of terms in the low q 2 expansion
with the technique of large mass expansion, expanding the basis functions on the other hand, and
equating the results, the coefficients in front of the basis functions can be determined. In a final
step most of the basis functions are transformed into Nielsen polylogarithms, leading to a fairly
compact result whose asymptotic behaviour can be analyzed in a straightforward manner.
To facilitate the discussion, we present, in a first step, in Section 2, the results for a
U (1) U (1) theory with one of the gauge bosons taken to be massive, the other one massless. The explicit analytical result confirms the factorization of the infrared singularities and
allows to identify the infrared-finite remainder. In Section 3 the formalism will be extended to
a massive non-Abelian theory and applied to the complete set of virtual corrections of order
weak s , contributing to Z-boson production and decay. The triple-boson coupling leads to additional diagrams with additional generalized polylogarithms, which cannot easily be transformed
into Nielsons polylogarithms. However, they can be evaluated numerically with high precision
[13] and their asymptotic behaviour is under control. The paper concludes with a brief summary.
Much of the formulae and calculational details will be collected in Appendices A, B, C.
49
2. Abelian theory
For definiteness and simplicity we will, in a first step, consider the form factor in a ficticious
U (1) U (1) theory with one massive and one massless gauge boson and with coupling constants
and respectively.
For the Abelian theory the form factor F will be defined as matrix element of an external
current

F (q, M) = dx eixq |J (x)|.
(1)
Here and denote on-shell massless fermions of momenta p and p = p + q, respectively,
M the mass of the gauge boson.
In a perturbative expansion

m n (m,n)
f
(q, M, ),
F (, , q, M, ) =
(2)
4
4
m,n=0
one needs to evaluate the expansion coefficients f (m,n) . In Born and one-loop approximation
they are given by
f (0,0) = 1,

7 2 2 + 3z
2
2(1 + z)2
(1
+
z)
log(z) +
,
Li
f (1,0) = +
2
2 z
z
6
z2
2

3
3
14
2
16
+
+
8
+
,
f (0,1) =
2
2
3
2
3
q 2
2
(3)
(4)
(5)
where z = q 2 /M 2 + i0, n = (n) is the Riemann -function. The infrared singularities are
controlled by dimensional regularization in d = 4 2 dimension and is related to the scale
parameter of dimensional regularisation by 2 = 2 4 exp( ). In the euclidean region q 2 < 0,
so that no imaginary parts appear in the above formulae.
The two-loop result for the massless case, f (0,2) , can be found, e.g., in [14,15]. The twoloop result for the fully massive case, f (2,0) , is only known in the large q 2 limit [9,16] and [17]
in Abelian and non-Abelian theories, respectively. The evaluation of the mixed corrections is
drastically simplified by the fact that the infrared singularities factorize within infrared evolution
equation approach [16,18], which gives in our case

F , , q 2 /M 2 , = Fmassless ( , q, )F , , q 2 /M 2 ,
(6)

with Fmassless = ( /4)n f (0,n) (q, ) denoting the formfactor for the massless theory and F
being free from infrared singularities. The function F can again be expressed as double series,
and the coefficients depend on the ratio z = q 2 /M 2 only. The terms F (m,0) = f (m,0) coincide
by definition with those valid for the massive U (1)-theory. The evaluation of the nonfactorizable
part of the two-loop contribution

F (1,1) q 2 /M 2 ,
(7)
will be the central result of this section. The Feynman diagrams necessary for this computation
have two thresholds: at q 2 = 0 and at q 2 = M 2 . The analytical structure of vertex diagrams of
this type has been explored in [11]. The coefficients of an expansion in q 2 /M 2 (and M 2 /q 2 )
50
can always be expressed as combinations of so-called harmonic sums [19] or more generally
nested harmonic sums [20]. These sums correspond to (generalized) polylogarithms [21,22] of
arguments q 2 /M 2 and their generalizationsharmonic polylogarithms [23] (see also [24]).
This structure suggests the following method for the evaluation of Feynman integrals. First, using the method of large mass expansion [25], one calculates a large number of coefficients of
the series in q 2 /M 2 . From the basis functions (polylogarithms) one then constructs an ansatz
(see Appendix A) with unknown coefficients ri . Equating ansatz and series one obtains a unique
answer for parameters ri . This method has been applied earlier [11,26] to various scalar vertex
masterintegrals. (In a different context the method has also been applied in [27].) Here it is applied to amplitudes deduced from a set of realistic Feynman diagrams representing a physical
process and leading to amplitudes with irreducible numerators and shrunken lines.
A few comments on this procedure are in order. First, the main problem is to write down the
correct prefactors in the ansatz. Empirically one finds that the presence of a numerator or the
absence of a line may lead to the additional factors M 2 /q 2 or (M 2 /q 2 )2 in front of polylogarithms.1 Therefore such factors should also be included in the ansatz. Second, only five functions
could not be represented as Nielsen polylogarithms with the argument q 2 /M 2 . These remaining functions belong to the class of harmonic polylogarithms [23] discussed in more detail in
the Appendix B.
Instead of expanding the amplitude in q 2 /M 2 one could find the differential equation (see
[28]) for a diagram and again apply an ansatz based on polylogarithms. This approach has recently been used for similar two-loop vertex diagrams in [29].
Altogether 16 one-particle-irreducible two-loop vertex diagrams contribute to the formfactor.
These diagrams can be obtained from the one-loop one shown in Fig. 1(a) by adding one gluon
line. The two-loop, one-particle reducible diagrams which are obviously products of one-loop
diagrams contribute to the term f (0,1) f (1,0) and are not repeated here. We also do not display
the contributions to the fermionic wave function renormalization, which receives contributions
from additional 6 diagrams. For the generation of the input the program DIANA [30] has been
used, for the evaluation and expansion a program written in FORM [31]. The evaluation of the
Dirac traces has lead to about 700 different integrals. For most of them the asymptotic expansion
was performed up to order 45 which required in total several hours of CPU time on a Pentium IV
processor. For the remaining, most complicated cases (nonplanar diagram) up to 60 expansion
coefficients had to be computed. For this purpose the parallel version of FORM [32], running on
an SGI machine with multiprocessor SMP architecture, was used.
The function (z) can be cast into the following form (here and below z = q 2 /M 2 + i0)

(1 + z)2 2
(z) =
6L + 242 243 log(1 + z) + 4L2 6L 202 log2 (1 + z)
2
z
8
+ log3 (1 + z)L + 8 log2 (1 + z)Li2 (z) 12 log(1 + z)Li2 (z)
3
16 log(1 + z)Li3 (z) + 16 log(1 + z)S1,2 (z) 16Li2 (z)2 4Li2 (z)L2
8Li22 (z) + 16Li3 (z)L 24Li4 (z) 12S1,2 (z) + 16S1,2 (z)L

+ 16S1,3 (z) 16S2,2 (z) + 24h(z) 48H3 (z) + 8H2 (z) + 32H4 (z)
1 In a series representation such multiplications lead to shifts of the summation index in c . Indeed, if z = q 2 /M 2 then,
n

e.g., 1z n=1 cn zn = c1 + n=1 cn+1 zn and so on.
51
1 + 3z + z2
32Li2 (z)2 8Li2 (z)L2 8Li2 (z)Li2 (z) + 32Li3 (z)L
z2
1 z2
48Li4 (z) + 32S2,2 (z) +
72 log(1 z)2 + 18 log(1 z)L2
z2
2 + 3z
+ 36 log(1 z)Li2 (z) + 36Li2 (z)L + 72S1,2 (z) +
(322 + 123 )
z
16 + 23z 2
34 + 51z
L+
L
z
z
2(1 z)(13 + 27z)
4(3 + 4z)
log(1 z) +
Li2 (z)L
z2
z2
4(11 + 9z)
2(1 + z)(3 + 5z)
log(1 + z)L + Li2 (z) +
Li2 (z)
2
z
z
+
4(3 + 2z 3z2 )
4(9 + 4z 6z2 )
2(8 z)
Li
(z)
Li3 (z) +
,
3
z
z2
z2
(8)
where L = log(q 2 /M 2 ), a = (a) is the Riemann -function, Sa,b (z) are Nielsen polylogarithms [21]. The functions h(z) and Hj (z) are defined and discussed in Appendix B.
Note that the result (8) depends only on one VermaserenRemiddi function (we choose
H2 (z) H1,0,0,1 (z)) and a combination of Nielsen polylogarithms (see Appendix A). This
observation should be correct also for any results which can be representaed as a combination of
massless Feynman integrals and/or diagrams with one-massive particle thresholds.
3. Z-production
For definitenes and simplicity, consider, in the next step, Z-boson production in quark
antiquark annihilation. To fix the notation, we recapitulate the one-loop results. The weak corrections to the Born term can be split into those involving the exchange of W - and Z-bosons,
(Fig. 1(a)) and those involving the triple-boson coupling (Fig. 1(b)). The combination of photonic and QCD corrections follows essentially from the two-loop QED or QCD results and will
not be addressed here.
For a light quark the form factor can be decomposed as follows

1 + 5 2
1 5 2
F q 2 =
F R q +
FL q .
2
2
(9)
At the Born level the expressions for the form factors FR and FL are given by
FR = i
e
gR ,
sW
FL = i
e
gL ,
sW
(10)
2 /c and g = (I Qs 2 )/c being the right- and left-handed couplings of

with gR = QsW
W
L
3
W
W
a quark to the Z-boson. Here I3 is the third component of the isospin of a quark, Q its electric
charge and sW = sin W and cW = cos W denote sine and cosine of the weak mixing angle,
respectively.
Including radiative corrections and adopting a form similar to Eq. (6) the formfactors can be
cast into the following form
52

s (0,1)
e
FR = i
gR3 A q 2 /m2Z , m2Z
f
1 + CF
gR +
2
sW
4
4s
W

s
+ CF
g 3 q 2 /m2Z ,
2 R A
4 4sW

s (0,1)
e
FL = i
f
1 + CF
gL
sW
4

2 2 2 gL 2 2

2 2

I3
3
2
2
+
A q /mW , mW + cW NA q /mW , mW
gL A q /mZ , mZ +
2
2
2
4sW

2 2
s
I3
gL
3
2
2
2
2
q
+
q
+
c
q
, (11)
/m
/m
/m
g
+ CF
A
W
NA
Z
W
W
L A
2
4 4sW
2
2
where first factors in the brackets in the above equations represent the QCD corrections. Terms
given by A and NA account for the one-loop electroweak corrections. The Abelian part A is
defined by the diagram of the Abelian type (Fig. 1(a)) and obviously closely related to f (1,0)
defined in Eq. (4). The unrenormalized result2 is given by

1

2 2 + 3z
log(z)
A z, M 2 = ln M 2 / 2 4 +
z
z

2(1 + z)2
2
+
(1
+
z)
.
Li
2
6
z2
(12)
The non-Abelian part NA (z, M2 ) receives corrections from both diagrams of Figs. 1(a) and 1(b).
It is given by

NA z, M 2 = 2A z, M 2 2 z, M 2 ,
(13)

3
2
4
2
1 4 2
1 l 1+
l ,
z, M 2 = + 3 ln M 2 / 2 2 + 1 +
(14)
z
z
z
2z z
with

1 4/z + 1 + i0
l = ln
.
1 4/z 1 + i0
(15)
(a)
(b)
Fig. 1. Diagrams, contributing to the vertex Zq q (a) and (b). The two-loop diagrams are obtained by attaching one virtual
gluon in all possible ways. The case (b) represents the non-Abelian part and contributes to NA (z). Diagram (a) with W
exchange contributes both to A (z) and NA (z). The cross indicates the coupling to the external Z boson.
2 We shall not discuss issues related to renormalization, since the non-factorizable part, which is the quantity of interest
in this paper, is independent of the renormalization scheme.
53
The function (z) can be taken from [34] (see also [3537] and references therein for one-loop
calculations in the Standard Model). We do not include the terms from the renormalization of the
coupling and the Z-boson wave function3 which follow from textbook prescriptions and will not
be considered in this work.
Evaluated for arbitrary q 2 = MZ2 , the above results are gauge dependent and are presented in
Feynman gauge. For the offshell case they can be considered as building blocks for a complete
calculation.
The functions A (z) and NA (z), representing the non-factorizable terms of O(s ), are
written in a form completely analogous to the electroweak one-loop terms. The function A (z)
has been given in the previous section. The non-Abelian part NA (z) involves new functions
generalized polylogarithms. Our result in Feynman gauge reads
A (z) = (z),
(16)
NA (z) = 2A (z)
8 5z
1 + 2z
+4
+ 48 2 H0,r,r,1 (z) 12Hr,r,1 (z) + 8H0,r,r (z)
z
z
2
6 + 2z 3z
(4 z)(4 + 3z)
+4
Hr,r (z) 6
gHr,1 (z)
z2
z2

12 11z2
16 + 23z
(4 z)(6 + 7z)
gHr (z) +
Li3 (z)
2
82 + 2L2 4
2
z
z
z2
(1 z)(13 + 34z)
5 + 9z
66 + 49z
log(1 z) 16
Li2 (z) 4
L
+2
z
z
z2
(1 z2 )
+
96 log(1 z)2 + 24 log(1 z)L2
z2

+ 48 log(1 z)Li2 (z) + 48Li2 (z)L + 96S1,2 (z)
(1 + 4z + z2 )
32Li2 (z)2 8Li2 (z)L2 8Li22 (z)
z2

+ 32Li3 (z)L 48Li4 (z) + 32S2,2 (z) .
(17)
The function NA (z) receives contributions not only from diagrams of Fig. 1(b) but also from
those of Fig. 1(a) with the exchange of W -boson. The functions H... are considered in more detail
in Appendix C. For the special case q 2 = M 2 one finds

16
1
A (1 + i0) = 14 + 722 l2 642 l22 l24 + 222 283 + 164 128Li4
3
2

32
+ i 85 + 32l2 + 24l22 l23 + 142 1203
(18)
3
= 2.1073 19.0331i,
(19)
32
70
184
NA (1 + i0) = 16 1442 l2 + 1282 l22 + l24 + 2 +
3 2364
3
3

3

1
16
1
+ 26 + 256Li4
84 Ls2
Ls2
2
3
3
3
3
3
3 Hence the function (z) considered in [34] differs by subtracting the term 3/ 3 ln(M 2 /
2 ) 1/2. Furthermore,
2
a typo in [34] has been corrected, flipping the signs of the terms proportional to l and l .
54
2

64
+ 96 Ls2
+ i 54 64l2 48l22 + l23 282 + 483
3
3
= 7.5880 + 16.7194i,
with l2 = log 2. Substituting the actual masses of the W - and Z-bosons
we find:
(z = m2Z /m2W
(20)
(21)
= 1.2856)
A (1.2856 + i0) = 1.3598 30.4095i,
(22)
NA (1.2856 + i0) = 10.1248 + 35.0336i.
(23)
In the limit q 2 the function , as given by Eq. (8) coincides with the result of [9] where
the power unsuppressed logarithmic and constant part have been evaluated. For the leading and
the first power suppressed term we find
(z) = (3 242 + 483 ) log(z) 2 + 402 843 + 144
1
+ (26 + 82 ) log2 (z) + (120 162 + 1283 ) log(z)
z

1
188 82 83 + 1164 + O 2 .
z
(24)
In Fig. 2 the exact result is compared with the Sudakov approximation and with the approximation including the first power-suppressed term. For electroweak interactions the mass of the
gauge boson can be taken to be of order 100 GeV, the characteristic energy of order one to two
TeV. For one TeV the relative error of the Sudakov approximation (the logarithmic plus constant
term) amounts to 15%, at 2 TeV it is reduced to 2.5%.
Fig. 2. Non-factorizable two-loop correction to the Abelian formfactor in the Euclidean regime (z = q 2 /m2 ). The solid
line represents the exact result, the dashed line the Sudakov approximation and the dash-dotted line includes the power
suppressed terms.
55
Assuming that this behaviour is characteristic for the general case, power suppressed terms
start to become important for M 2 /s below 1/100. Furthermore inclusion of the first subleading
term only does not lead to a significant improvement of the approximation.
4. Conclusions
Using the technique of asymptotic expansions and the knowledge of the general structure of
integrals we evaluated analytically the two-loop formfactor in a U (1) U (1) theory with one
massive and one massless gauge boson and constant terms have been evaluated. The Sudakov
approximation gives a good description for M 2 /s below 1/100. For lower energies mass suppressed terms must be included.
In the Sudakov limit full agreement is obtained with [9], where the logarithmic and constant
terms had been evaluated obtained. We furthermore perform the same calculation for a SU(2)
U (1) theory and derive the non-factorizable part of the two-loop formfactor in the Standard
Model. As an application we evaluate the mixed virtual O(s ) radiative correction for Drell
Yan production of the Z-boson.
Acknowledgements
We thank M. Kalmykov for useful comments, discussions and communication and M. Tentyukov for his help with DIANA. We acknowledge T. Gehrmann for information about the
numerical program hplog and B. Jantzen for the carefull reading the paper. This work was
supported by BMBF under grants Nos. 05HT6VKA, 05HT4GUA4 and HGF grant No. NG-VH008. A.K. is supported in part by an Alexander von Humboldt Foundation (a renewed academic
stay in Germany) and RFBR Foundation through grant No. 05-02-17645-a.
Appendix A
In this appendix we give the details of the ansatz which is used to obtain the analytical result
for the amplitudes.
Generally the result of the finite part of expansion for a given diagram J looks like
J = Mg
n
max

zn an + bn ln z + cn ln2 z ,
(A.1)
n=1
where z = q 2 /M 2 and g is the mass dimension of the diagram. In (A.1) we substitute a linear
combination
r1 1 (z) + r2 2 (z) + + rN N (z)
(A.2)
where rj are unknown rational numbers and j (z) are basis functions.
In our case the basis functions are chosen as follows. In the Abelian case it is enough to
include all harmonic polylogarithms of the weight one to four. This includes
ln(1 z),
Li2 (z),
Li3 (z),
Li4 (z),
S1,3 (z),
S2,2 (z),
with Sa,b being Nielsen polylogarithm. In addition one should add the following harmonic polylogarithms
H1,0,1 (z),
H1,0,1,1 (z),
H1,0,0,1 (z),
H0,1,0,1 (z),
56
which are considered in Appendix B. In the non-Abelian part (see Fig. 1(b)) new functions appear
which are listed in formulae (C.1)(C.6).
We should also note that the above functions could be multiplied by additional factors 1/z or
1/z2 . All these combinations should be included in the ansatz. We also observed that functions
proportional to 1/z3 and higher inverse powers of z do not appear.
Appendix B
Here we consider the asymptotic behaviour of the most complicated basis functions in the
limit z = q 2 /m2 . Most of basis functions can be expressed in terms of Nielsen polylogarithms and then the standard transformations formulae can be applied to go from argument
z to 1/z (see [21,22]). Therefore we will consider here only the five special cases, mentioned
previously, where complications arise.
In our calculation the following functions appear in addition to usual Nielsen polylogarithms:
h(z) = H1,0,1 (z),
H1 (z) = H1,0,1,1 (z),
H2 (z) = H1,0,0,1 (z),
H3 (z) = H1,1,0,1 (z),
H4 (z) = H0,1,0,1 (z),
where Ha,...,d (z) are harmonic polylogarithms defined in [23].
These functions correspond to the alternating Taylor series in z:
h(z) =

S2 (n 1)
n=1
H1 (z) =
n=1
H2 (z) =
(z)n ,
S2,1 (n 1)
(z)n ,
n

S3 (n 1)
n=1
H3 (z) =
(z)n ,

S3 (n 1) + S2,1 (n 1) S1 (n 1)S2 (n 1)
n=1
H4 (z) =

S2 (n 1)
n=1
n2
(z)n ,
(z)n ,

with finite harmonic sums Sa (n) = nj=1 1/j a and Sa (n) = nj=1 (1)j /j a and S2,1 (n) =
n
j
2
j =1 (1) S1 (j )/j . It is interesting to note that the function H1 cancels in the final result (8)
for the formfactor but is present in the particular integrals.
Following [11] it is not difficult to write down simple integral representations for the above
series, e.g.,
z
h(z) =
0
dx
Li2 (x),
1+x
(B.1)
z
H1 (z) =
0
z
H2 (z) =
0
57
dx
S1,2 (x),
1+x
(B.2)
dx
Li3 (x),
1+x
(B.3)
z
H3 (z) = log(1 + z)h(z)

0
z
H4 (z) = log(z)h(z)
0
dx
Li2 (x) log(1 + x),
1+x
dx
Li2 (x) log(x).
1+x
(B.4)
(B.5)
Now the integrals can be expressed in terms of Nielsen polylogarithms of nonlinear arguments
and only one harmonic polylogarithm function H2 (this choice being not unique, however). We
have4

1
h(z) = S1,2 z2 S1,2 (z) S1,2 (z) + ln(1 + z)Li2 (z),
2

1
1
H1 (z) = log(1 + z)S1,2 (z) + S1,3 z2 S1,3 (z) + (z),
4
2

2
1
H3 (z) = log(1 + z) S1,2 z S1,2 (z) S1,2 (z)
2

1
1
1
+ log2 (1 + z)Li2 (z) + S1,3 z2 S1,3 (z) (z),
2
4
2
2
1
H4 (z) = S2,2 z S2,2 (z) S2,2 (z) + log(1 + z)Li3 (z) H2 (z),
4
where
(B.6)
(B.7)
(B.8)
(B.9)

15
1
1
4 + log3 s log z + log2 s Li2 (s) Li2 (s)
8
6
2

log s Li3 (s) Li3 (s) + Li4 (s) Li4 (s),
(z) =
with s = (1 z)/(1 + z).

In order to find the asymptotic behaviour for z one needs to use the standard formulae
for polylogarithms and for the function H2 (z) H1,0,0,1 (z) the inversion formula (A.6) from
[33]. It is important to take care of imaginary parts, therefore we approach the cut in q 2 -plane
from above, which means that z is replaced by z + i0. Thus we obtain

1
3
1
1
3
2
h(z + i0) = log z + 22 log z 3 +
log z log z + 22
6
2
z
2

1
1
1
1
1
2
log z 2 + log z +
+O 2 ,
+ i
(B.10)
2
2
z
z
z
4 Eq. (B.6) is taken from [11]. Equations similar to (B.7)(B.9) have been considered also in [33,38]. In particular,
Eq. (B.7) coincides with (A.29) of [33].
58
1
3
57
log4 z 2 log2 z + 3 log z + 4
24
2
16

1 1
1
3
2
+
log z + log z 32 log z + 3 32 1
z 6
2

7
1
1
1
3
2
+ i log z + 2 log z 3 +
log z log z + 2
6
8
z
2

1
+O 2 ,
z
1
5
H2 (z + i0) = log4 z + 2 log2 z 4
24
8

1
1
1
+
log3 z log2 z + (22 1) log z + 22 2
z
6
2

3
1 1
1
1
+ i
log3 z 3 +
log2 z + log z + 1 + O 2 ,
6
4
z 2
z
1
3
3
H3 (z + i0) = log4 z + 2 log2 z 3 log z 4
24
2
16

1
3
1
+
log3 z + (22 + 1) log z 3 + 1
z
6
2

1
1
7
1 1
1
3
2
+ i
log z 2 log z + 3 +
log z 2 1
6
2
8
z 2
2

1
+O 2 ,
z
1
3
7
H4 (z + i0) = log4 z + 2 log2 z 3 log z + 4
24
2
8

1 1
+
log2 z + 2 log z 22 + 2
z 2

1
3
1
1
1
+ i
log3 z 2 log z + 3 + ( log z 2) + O 2 .
6
2
2
z
z
H1 (z + i0) =
(B.11)
(B.12)
(B.13)
(B.14)
Finally we used the program hplog [13] to check numerically the asymptotic behaviour of the
H -functions.
Appendix C
In this appendix we consider the H -functions contributing to the non-Abelian part of the
formfactor. For the definitions and recursive constructions of these functions we refer to [29].
However, for completeness we give here explicitly the definitions of the functions which appear
in our calculation. The following six new functions arise in the evaluation of the two-loop nonAbelian formfactor:
z
Hr (z) =
dt1
,
t1 (t1 + 4)
(C.1)
z
Hr,r (z) =
0
dt2
t2 (t2 + 4)
z
dt2
t2 (t2 + 4)
Hr,1 (z) =
0
z
Hr,r,1 (z) =
0
z
H0,r,r (z) =
0
t2
t2
0
dt3
t3 (t3 + 4)
dt3
1 + t3
z
H0,r,r,1 (z) =
0
dt4
t4
t3
0
t4
0
dt1
,
t1 (t1 + 4)
(C.2)
dt1
,
1 + t1
t3
0
(C.3)
dt2
t2 (t2 + 4)
dt2
t2 (t2 + 4)
dt3
t3 (t3 + 4)
59
t2
0
t3
0
t2
0
dt1
,
1 + t1
(C.4)
dt1
,
t1 (t1 + 4)
dt2
t2 (t2 + 4)
t2
0
(C.5)
dt1
.
1 + t1
(C.6)
In the formula (17) for the non-Abelian part the functions with odd number of indices r
appear always with the factor
1
g(z) =
.
1 4/z
(C.7)
It is easy to check that Hr and Hr,1 cannot be expanded in the Taylor series of small arguments (they have a branche point at zero), but the combinations gHr and gHr,1 can.
The integral representations given above are not very suitable for the analysis and numerics.
The ultimate task would be to relate them to the usual (harmonic) polilogarithms. In order to
do this one should choose a right variable. From the previous experience [11] it is known that
for the diagrams, possessing a branch point at q 2 = 4m2 , the appropriate variable is given by
(z = q 2 /m2 )
1 z/(z 4)
y=
(C.8)
.
1 + z/(z 4)
In terms of y the g-factor (C.7) is expressed as
g(z) =
1y
,
1+y
(C.9)
and the required H -functions take form

Hr (z) = log y,
1
Hr,r (z) = log2 y,
2

1
1
1
Hr,1 (z) = log2 y + Li2 y 3 Li2 (y) 2 ,
2
3
3
1
3
2
H0,r,r (z) = log y + log(1 y) log y 2Li3 (y) + 2 log yLi2 (y) + 23 ,
6

1
1
2
1
Hr,r,1 (z) = log3 y + 2 log y + 3 + Li3 (y) Li3 y 3 ,
6
3
3
9
(C.10)
(C.11)
(C.12)
(C.13)
(C.14)
60
H0,r,r,1 (z)

1
1
2
89
1
log4 y 2 log2 y 3 log y +
4 Li4 (y) + Li4 y 3
=
24
6
3
108
27

2
2
1 3
+ 2S1,3 (1 y) 2 Li2 (1 y) + 2 ln(1 y) 3 + Li3 (y) Li3 y
3
3
9

1 3
+ 2Li2 (y) Li2 (y) Li2 y
(C.15)
+ 2N1 (1) 2N1 (y),
3
where
y
N1 (y) =

dt
Li2 (t) ln 1 t + t 2 ,
t
N1 (1) =
11
4 .
54
(C.16)
As it is seen from the above fomulae, the H -functions with index r can be rewritten in
terms of harmonic polylogarithms but of nonlinear argument y.
In the limit when z + + i0 we obtain

2
1
1
+O 2 ,
g(z)Hr (z) = log z + (2 log z 2) + i 1
(C.17)
z
z
z

1
10
20
g(z)Hr,1 (z) = log2 z 2 + log2 z 2 log z 2 1
2
3
3

1
1
+ i log z (2 log z 2) + O 2 ,
(C.18)
z
z

2
1
1
2 log z
+ i log z +
+O 2 ,
Hr,r (z) = log2 z 32
(C.19)
2
z
z
z
1
1
H0,r,r (z) = log3 z 32 log z + 23 + (2 log z + 2)
6
z

1
1
2
2
+O 2 ,
+ i log z + 2
(C.20)
2
z
z

1
10
2
1
20
log2 z + 2 + 1
Hr,r,1 (z) = log3 z 2 log z + 3 +
6
3
3
z
3

1
4
2
1
+ i log2 z + 2 + log z + O 2 ,
(C.21)
2
3
z
z
1
5
2
H0,r,r,1 (z) =
log4 z 2 log2 z + 3 log z + 74
24
3
3

20
1
2
2
log2 z + 2 + 2 3 log z + 1 2
+
z
3
3
3

1
4
2
+ i log3 z + 2 log z 3
6
3
3

2
1
2
1
2 log z 2 2 + 3 2 log z + O 2 .
+
(C.22)
z
3
3
z
And finally we give the values of H -functions at the particular point z = 1:
g(1)Hr (1) = ,
3 3
(C.23)
g(1)Hr,1 (1) =
2 Ls2 ( 3 )
,
3
3
1
Hr,r (1) = 2 ,
3

4
2
,
H0,r,r (1) = 3 Ls2
3
3
3
1
Hr,r,1 (1) = 3 ,
9
2
7
2
H0,r,r,1 (1) = 4 +
,
Ls2
12
3
3
61
(C.24)
(C.25)
(C.26)
(C.27)
(C.28)
where n is the Riemann -function and Lsn (x) is the log-sine integral defined as
x
Lsn (x) =

t
logn1 2 sin
dt.
2
(C.29)
In particular the constant Ls2 ( 3 ), sometimes denoted as Clausens integral Cl2 ( 3 ) (see, e.g.,
[22]), is given by

= 1.014941606409653625 . . . .
Ls2
(C.30)
3
References
[1] 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.
[2] P.J. Rijken, W.L. van Neerven, Phys. Rev. D 51 (1995) 44.
[3] R.V. Harlander, W.B. Kilgore, Phys. Rev. D 68 (2003) 013001.
[4] U. Baur, O. Brein, W. Hollik, C. Schappacher, D. Wackeroth, Phys. Rev. D 65 (2002) 033007.
[5] A. Czarnecki, J.H. Kuhn, Phys. Rev. Lett. 77 (1996) 3955.
[6] R. Harlander, T. Seidensticker, M. Steinhauser, Phys. Lett. B 426 (1998) 125.
[7] J. Fleischer, F. Jegerlehner, M. Tentyukov, O. Veretin, Phys. Lett. B 459 (1999) 625.
[8] J.H. Kuhn, A. Kulesza, S. Pozzorini, M. Schulze, Nucl. Phys. B 727 (2005) 368.
[9] B. Feucht, J.H. Kuhn, A.A. Penin, V.A. Smirnov, Phys. Rev. Lett. 93 (2004) 101802.
[10] B. Jantzen, V.A. Smirnov, Eur. Phys. J. C 47 (2006) 671.
[11] J. Fleischer, A.V. Kotikov, O.L. Veretin, Nucl. Phys. B 547 (1999) 343.
[12] B.A. Kniehl, A.V. Kotikov, A. Onishchenko, O. Veretin, Nucl. Phys. B 738 (2006) 306.
[13] T. Gehrmann, E. Remiddi, Comput. Phys. Commun. 144 (2002) 200.
[14] R.J. Gonsalves, Phys. Rev. D 28 (1983) 1542.
[15] W.L. van Neerven, Nucl. Phys. B 268 (1986) 453.
[16] J.H. Kuhn, A.A. Penin, V.A. Smirnov, Eur. Phys. J. C 17 (2000) 97.
[17] B. Jantzen, J.H. Kuhn, A.A. Penin, V.A. Smirnov, Phys. Rev. D 72 (2005) 051301;
B. Jantzen, J.H. Kuhn, A.A. Penin, V.A. Smirnov, Phys. Rev. D 74 (2006) 019901, Erratum;
B. Jantzen, J.H. Kuhn, A.A. Penin, V.A. Smirnov, Nucl. Phys. B 731 (2005) 188;
B. Jantzen, J.H. Kuhn, A.A. Penin, V.A. Smirnov, Nucl. Phys. B 752 (2006) 327, Erratum.
[18] V.S. Fadin, L.N. Lipatov, A.D. Martin, M. Melles, Phys. Rev. D 61 (2000) 094002.
[19] A. Gonzalez-Arroyo, C. Lopez, F.J. Yndurain, Nucl. Phys. B 153 (1979) 161;
A. Gonzalez-Arroyo, C. Lopez, Nucl. Phys. B 166 (1980) 429;
D.I. Kazakov, A.V. Kotikov, Nucl. Phys. B 307 (1988) 721;
D.I. Kazakov, A.V. Kotikov, Theor. Math. Phys. 73 (1988) 1264, Teor. Mat. Fiz. 73 (1987) 348.
[20] J.A.M. Vermaseren, Int. J. Mod. Phys. A 14 (1999) 2037;
J. Blumlein, S. Kurth, Phys. Rev. D 60 (1999) 014018.
62
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
A. Devoto, D.W. Duke, Riv. Nuovo Cimento 7N6 (1984) 1.

L. Lewin, Polylogarithms and Associated Functions, North-Holland, Amsterdam, 1981.
E. Remiddi, J.A.M. Vermaseren, Int. J. Mod. Phys. A 15 (2000) 725.
A.B. Goncharov, Math. Res. Lett. 5 (1998) 497.
F.V. Tkachov, Preprint INR P-0332, Moscow, 1983;
F.V. Tkachov, Preprint INR P-0358, Moscow, 1984;
K.G. Chetyrkin, Teor. Math. Phys. 75 (1988) 26;
K.G. Chetyrkin, Teor. Math. Phys. 76 (1988) 207;
K.G. Chetyrkin, Preprint, MPI-PAE/PTh-13/91, Munich, 1991;
V.A. Smirnov, Commun. Math. Phys. 134 (1990) 109;
V.A. Smirnov, Renormalization and Asymptotic Expansions, Birkhuser, Basel, 1991;
V.A. Smirnov, Applied Asymptotic Expansions in Momenta and Masses, Springer Tracts in Modern Physics,
vol. 177, Springer, Berlin, Germany, 2002.
J. Fleischer, A.V. Kotikov, O.L. Veretin, Phys. Lett. B 417 (1998) 163.
R. Harlander, P. Kant, JHEP 0512 (2005) 015.
A.V. Kotikov, Phys. Lett. B 254 (1991) 158;
E. Remiddi, Nuovo Cimento A 110 (1997) 1435.
U. Aglietti, R. Bonciani, Nucl. Phys. B 668 (2003) 3;
U. Aglietti, R. Bonciani, G. Degrassi, A. Vicini, Phys. Lett. B 595 (2004) 432;
U. Aglietti, R. Bonciani, G. Degrassi, A. Vicini, Phys. Lett. B 600 (2004) 57;
U. Aglietti, R. Bonciani, G. Degrassi, A. Vicini, JHEP 0701 (2007) 021.
M. Tentyukov, J. Fleischer, Comput. Phys. Commun. 132 (2000) 124.
J.A.M. Vermaseren, Symbolic Manipulation with FORM, Computer Algebra, Amsterdam, Netherland, 1991.
M. Tentyukov, D. Fliegner, M. Frank, A. Onischenko, A. Retey, H.M. Staudenmaier, J.A.M. Vermaseren,
cs.sc/0407066.
A.I. Davydychev, M.Y. Kalmykov, Nucl. Phys. B 699 (2004) 3.
B. Grzadkowski, J.H. Kuhn, P. Krawczyk, R.G. Stuart, Nucl. Phys. B 281 (1987) 18.
M. Bohm, H. Spiesberger, W. Hollik, Fortschr. Phys. 34 (1986) 687.
D.Y. Bardin, G. Passarino, The Standard Model in the Making: Precision Study of the Electroweak Interactions,
International Series of Monographs on Physics, vol. 104, Clarendon, Oxford, UK, 1999, p. 685.
F.A. Berends, W.L. van Neerven, G.J.H. Burgers, Nucl. Phys. B 297 (1988) 429;
F.A. Berends, W.L. van Neerven, G.J.H. Burgers, Nucl. Phys. B 304 (1988) 921, Erratum.
M. Czakon, J. Gluza, T. Riemann, Phys. Rev. D 71 (2005) 073009;
M. Czakon, J. Gluza, T. Riemann, Nucl. Instrum. Methods A 559 (2006) 265.
N = 8 non-BPS attractors, fixed scalars

and magic supergravities
Sergio Ferrara a,b,c , Alessio Marrani b,c,d,
a Physics Department, Theory Unit, CERN, CH 1211, Geneva 23, Switzerland
b INFNLaboratori Nazionali di Frascati, Via Enrico Fermi 40, 00044 Frascati, Italy
c Department of Physics and Astronomy, University of California, Los Angeles, CA, USA
d Museo Storico della Fisica e Centro Studi e Ricerche Enrico Fermi, Via Panisperna 89A, 00184 Roma, Italy
Received 11 June 2007; accepted 30 July 2007

Available online 7 August 2007
Abstract
We analyze the Hessian matrix of the black hole potential of N = 8, d = 4 supergravity, and determine
its rank at non-BPS critical points, relating the resulting spectrum to non-BPS solutions (with non-vanishing
central charge) of N = 2, d = 4 magic supergravities and their mirror duals. We find agreement with the
known degeneracy splitting of N = 2 non-BPS spectrum of generic special Khler geometries with cubic
holomorphic prepotential. We also relate non-BPS critical points with vanishing central charge in N = 2
magic supergravities to a particular reduction of the N = 8, 18 -BPS critical points.
1. Introduction
After their discovery some time ago [15], extremal black hole (BH) attractors have been
object of intensive study in the last years [631]. Such a flourishing development mainly can be
essentially traced back to new classes of solutions to the attractor equations corresponding to
non-BPS horizon geometries.
It has been recently realized that the effective black hole potential VBH of N 2-extended,
d = 4 supergravities exhibits various species of critical points, whose supersymmetry-preserving
and stability features depend on the set of electric and magnetic BH charges.
* Corresponding author at: INFNLaboratori Nazionali di Frascati, Via Enrico Fermi 40, 00044 Frascati, Italy.
E-mail addresses: sergio.ferrara@cern.ch, ferrara@physics.ucla.edu (S. Ferrara), marrani@lnf.infn.it (A. Marrani).

doi:10.1016/j.nuclphysb.2007.07.028
64
S. Ferrara, A. Marrani / Nuclear Physics B 788 (2008) 6388
For what concerns the case N = 2, critical points fall into three distinct classes: ( 12 -)BPS and
two non-BPS classes, depending whether the N = 2 central charge Z vanishes or not at the BH
event horizon. The BPS critical points are known to be always stable (and thus to give rise to
actual attractor solutions), as far as they are points at which the metric of the scalar manifold is
positive-definite [5].
The stability not guaranteed in the non-BPS cases, in which the Hessian is generally degenerate, i.e. it exhibits some flat directions. For example, for N = 2 supergravities whose vector
multiplets scalar manifold is endowed with special Khler (SK) d-geometries1 of complex dimension nV , it was shown in [10] that the rank of the 2nV 2nV Hessian matrix of VBH (whose
real form is the scalar mass matrix) at the non-BPS Z = 0 critical points has (at most) rank
nV + 1 (corresponding to strictly positive eigenvalues), with (at least) nV 1 flat directions
(i.e. vanishing eigenvalues).
Such a splitting nV + 1/nV 1 of the non-BPS Z = 0 spectrum has been confirmed in [21],
where the N = 2 attractor equations were studied in the framework of the homogeneous symmetric SK geometries, which (apart from the case of the irreducible sequence based on quadratic
prepotential) are actually particular d-geometries.
In N > 2-extended, d = 4 supergravities the BPS spectrum is degenerate, too. As pointed
out in [36], the BPS splitting into non-degenerate (with strictly positive eigenvalues) and flat
(with vanishing eigenvalues) directions can be explained respectively in terms of the would-be
vector multiplets scalar and hypermultiplets scalars of the N = 2 reduction of the considered
E7(7)
N > 2 theory. For example, in N = 8, d = 4 supergravity (based on the coset manifold SU(8)
) the
70 70 Hessian of VBH at the (non-degenerate) 18 -BPS critical points has rank 30; its 30 strictly
positive and 40 vanishing eigenvalues respectively correspond to the 15 vector multiplets and to
the 10 hypermultiplets of the N = 2, d = 4 spectrum obtained by reducing N = 8 supergravity
according to the following branching of the 70 (four-fold antisymmetric) of SU(8):
SU(8) SU(6) SU(2);
70 (15, 1) (15, 1) (20, 2),
(1.1)
where SU(6) SU(2) is nothing but the symmetry of the 8 8 N = 8 central charge matrix ZAB
(skew-diagonalizable in the so-called normal frame [44]) at the considered non-degenerate
1
8 -BPS critical points. 15, 15 and 20 respectively are the two-fold antisymmetric, its complex
conjugate and the three-fold antisymmetric of SU(6). In general, the rank of the non-singular N1 BPS Hessian of VBH in 2 N 8-extended, d = 4 supergravities is [36] (N 2)(N 3) + 2nV ,
where nV stands for the number of matter vector multiplets (for N = 6, nV = 1 even though
there are no vector matter multiplets, because the extra singlet graviphoton counts as a matter
field).
The present paper is devoted to the study of the degeneracy of the non-BPS Hessian of VBH
in N = 8, d = 4 supergravity, and of the corresponding N = 2 theories obtained by consistent
truncations. Since such N = 2 theories contain vector multiplets and hypermultiplets which are
some subsets of the kinematical reduction N = 8 N = 2 given by Eq. (1.1), the massive and
massless modes of the N = 2 non-BPS (Z = 0) Hessian must rearrange following the pattern of
degeneracy of the parent N = 8 supergravity, when reduced down to N = 2 theories.
1 Following the notation of [32], by d-geometry we mean a SK geometry based on an holomorphic prepotential funcA B C
tion of the cubic form F (X) = dABC X X 0 X (A, B, C = 0, 1, . . . , nV ).
65
The plan of the paper is as follows. In Section 2 we review the N = 2, d = 4 magic models
which can be obtained by consistent reduction of N = 8, d = 4 supergravity. Thence, Section 3
deals with the N = 8 (non-singular) 18 -BPS and non-BPS critical points of VBH ; in particular,
Section 3.1 reports known results on the N = 8, d = 4 attractor equations and the( symmetries
of the)ir solutions, whereas Section 3.2 concerns the Hessian matrix of VBH both at 18 -BPS and
non-BPS critical points. Thus, in Section 4 we consider the N = 2 descendants of the N = 8,
1
1
8 -BPS critical points; they divide in N = 2, 2 -BPS and non-BPS Z = 0 classes, whose spectra
are both studied and compared. In Section 5 we perform the same analysis for the descendants
of the N = 8 non-BPS critical points of VBH , i.e. for the N = 2 non-BPS Z = 0 class of critical
points of VBH,N =2 . We show that the interpretation of the mass degeneracy splitting of N = 8
spectra in terms of N = 2 multiplets requires a different embedding of the N = 2 R-symmetry
SU(2)H in the R-symmetry SU(8) of the parent N = 8 theory, depending on the structure and
on the eventual supersymmetry-breaking features of the considered class of solutions to attractor
equations. Our analysis also yields the interpretation, in terms of the U -duality symmetry E7(7)
of N = 8, d = 4 supergravity, of the splitting nV + 1/nV 1 of the 2nV eigenvalues of the
N = 2 non-BPS Z = 0 Hessian matrix for generic SK d-geometries of complex dimension nV ,
found in [10]. Finally, Section 6 contains some general remarks, as well as an outlook of possible
future developments.
2. N = 8 and N = 2 magic supergravities
G
N = 8, d = 4 supergravity is based on the 70-dimensional coset H
, where the (continuous)
U -duality group G is E7(7) and its maximal compact subgroup (m.c.s.) H is SU(8), which is
also the (local) R-symmetry of the N = 8, d = 4 supergravity. The vector and hyper multiplets
content of an N = 2, d = 4 reduction of N = 8, d = 4 supergravity is given by a pair

(nV , nH ) dimC

GV
GH
, dimH
,
HV
HH
nV 15, 2nH 20,
(2.1)
GH
V
where G
HV and HH respectively stand for the SK vector multiplets scalar manifold and for
the quaternionic Khler hypermultiplets scalar manifold. Clearly, in order for the N = 8
N = 2 truncation to be consistent, the isometry groups GV and GH of the two non-linear
-models should commute and should be both (proper) subgroups of G = E7(7) . We denote
HV = m.c.s.(GV ) and HH = m.c.s.(GH ). Moreover, HV always contains a factorized commuting U (1) subgroup, which is promoted to global symmetry (as the Gs) when nV = 0; on the other
hand, HH always contains a factorized commuting SU(2) subgroup, which is promoted to global
symmetry (as the Gs) when nH = 0. As previously mentioned, nV = 15 and nH = 10 correspond
to the reduction (1.1) of N = 8 supergravity, determining two N = 2 supergravities, one based
E6(2)
SO (12)
GH
V
on G
HV = SU(6)U (1) with (nV , nH ) = (15, 0), and the other one based on HH = SU(6)SU(2) with
(nV , nH ) = (0, 10).
In the following treatment we will consider only N = 2 maximal supergravities, i.e. N = 2
theories (obtained by consistent truncations of N = 8 supergravity) which cannot be obtained by
a further reduction from some other N = 2 theory, which are also magic. They are called magic,
since their symmetry groups are the groups of the famous Magic Square of Freudenthal, Rozenfeld and Tits associated with some remarkable geometries [57,58]. From the analysis performed
66
Table 1
G = E7(7) -based) N = 8, d = 4
Data of the magic N = 2, d = 4 supergravities obtained as consistent truncation of ( H
SU(8)
supergravity
GV
HV
J3H
J3C
GV
GH
HV
HH
SO (12)
SU(2)
SU(6) U (1)
SU(2, 1)
SU(3) SU(3)
U (1)
SU(3, 3)
J3R
Sp(6, R)
stu
SU(1, 1)
SO(2, 2)
R
J3,M
SU(1, 1)
C
J3,M
U (1)
G2(2)
SU(3) U (1)
SO(4, 4)
U (1)
SO(2) SO(2)
F4(4)
U (1)
E6(2)
SU(2) U (1)
GH
HH
SO (12)
SU(6)U (1)
(nV , nH )
(15, 0)
SU(3,3)
SU(3)SU(3)U (1)
(9, 1)
SU(2,1)
SU(2)U (1)
SU(2) SU(2)
SO(4) SO(4)
Sp(6,R)
SU(3)U (1)
(6, 2)
G2(2)
SO(4)
SU(1,1)
SO(2,2)
U (1) SO(2)SO(2)
(3, 4)
SO(4,4)
SO(4)SO(4)
USp(6) SU(2)
SU(6) SU(2)
SU(1,1)
U (1)
(1, 7)
F4(4)
USp(6)SU(2)
E6(2)
SU(6)SU(2)
(0, 10)
in [35,37,62], only six N = 2, d = 4 maximal magic supergravities2 exist which can be obtained
by consistently truncating N = 8, d = 4 supergravity; they are given3 by Table 1. The models
have been denoted by referring to their SK geometry. J3H , J3C and J3R stand for three of the
four N = 2, d = 4 magic supergravities which, as their 5-dimensional versions, are respectively
2 By E
7(p) we denote a non-compact form of E7 , where p (#non-compact #compact) generators of the group
[59,60]. In such a notation, the compact form of E7 is E7(133) (dimR E7 = 133).
3 With a slight abuse of language we include among magic supergravities the stu model, related to the Jordan algebra
R 2 = R R R, which is the n = 0 element of the sequence R 2+n of reducible Euclidean Jordan algebras
of degree 3. R denotes the one-dimensional Jordan algebra and n+2 denotes the Jordan algebra of degree 2 associated
with a quadratic form of Lorentzian signature (see e.g. Table 4 of [21], and references therein).
GV
SO(2,2)
2
( SU(1,1)
Due to the group isomorphism SO(2)SO(2)
U (1) ) , the scalar manifold H of the stu model, corresponding
V
SO(2,2+n)
to the element n = 0 of the reducible SK cubic sequence SU(1,1)
U (1) SO(2)SO(2+n) (n N {0, 1}, dimC = n + 3)
SU(1,1) 3
is nothing but ( U (1) ) .
SO(4,4)
3
The image of ( SU(1,1)
U (1) ) through c-map is given by the 4-dimensional (in H) quaternionic manifold SO(4)SO(4) ,
GH
which is the H of the stu model. Consistently, it is nothing but the element n = 0 of the quaternionic sequence
H
SO(4+n,4)
SU(1,1)
SO(2,2+n)
SO(4+n)SO(4) (n N {0}, dimH = n + 1), image of U (1) SO(2)SO(2+n) through c-map (see e.g. Table 4
of [38], and [39]).

GH
SU(2,1)
H
Finally, the 1-dimensional (in H) quaternionic manifold SU(2)U
(1) , corresponding to the HH of the model J3 ,
is the so-called universal hypermultiplet, given by the c-map of the case nV = 0, i.e. of pure N = 2, d = 4 supergravity, which (among the homogeneous SK geometries) is defined as the n = 0 limit of the rank-1 sequence of quadratic
SU(1,n)
irreducible SK manifolds U (1)SU(n)
(n N, dimC = n) [40].
67
defined by the three simple Jordan algebras J3H , J3C and J3R of degree 3 with irreducible norm
forms, namely by the Jordan algebras of Hermitian 3 3 matrices over the division algebras of
quaternions H, complex numbers C and real numbers R [4956].
Since E7(25) is a non-compact form of E7 (as E7(7) is, as well), the magic N = 2, d = 4
supergravity defined by the simple Jordan algebra J3O over the octonionic division algebra O,
E
7(25)
having vector multiplets scalar manifold E6(78)
SO(2) (dimC = 27), cannot be obtained from
N = 8, d = 4 supergravity. Beside the analysis performed in [21], Jordan algebras have been
recently connected to extremal black holes also in [61].
M subscript denotes the model obtained by performing a d = 4 mirror map (i.e. the composition of two c-maps in d = 4) from the original manifold; such an operation maps a model
with content (nV , nH ) to a model with content (nH 1, nV + 1), and thus the mirror of J3H , with
7(5)
does not exist, at least in d = 4.
(nV , nH ) = (1, 16) and quaternionic manifold SO(12)SU(2)
The stu model [23,47,48] is self-mirror: stu = stuM .
3. N = 8, d = 4 critical points and Hessian

In Section 3.1 we will review the solutions to the attractor equations of N = 8, d = 4
supergravity, mainly following [19] (see [34] for a recent review of attractor mechanism in
N 2-extended, d = 4 supergravities). Thence, in Section 3.2 we will consider the related critical spectrum given by the Hessian of VBH ; while the non-singular 18 -case was investigated in [36]
(see also [41]), the non-BPS case was hitherto unknown.
3.1. Solutions to attractor equations
The black hole potential of N = 8, d = 4 supergravity (based on the real coset
reads as follows [5,43] (A, B = 1, . . . , 8 throughout):
1
VBH = ZAB Z AB ,
2
E7(7)
SU(8) )
[42]
(3.1)
where ZAB (and its complex conjugate Z AB ) is the central charge matrix (and its conjugate),
E7(7)
)) real
sitting in the two-fold antisymmetric complex 28 of E7(7) . It depends on 70(= dimR ( SU(8)
i
scalars (i = 1, . . . , 70 throughout, unless otherwise noted), where the local SU(8) symmetry
was used to remove 63 scalars from the representation 133 of scalars in E7(7) .
The SU(8)-covariant derivatives [43] of the central charge matrix are defined by the Maurer
E7(7)
Cartan equations for SU(8)
:
1
Di ZAB = Z CD PABCD,i
2
1
Di Z AB = ZCD P,iABCD ,
2
(3.2)
7(7)
, sitting in the 70 (four-fold
where PABCD = Pi,[ABCD] d i is the 70 70 vielbein 1-form of SU(8)
antisymmetric) of the stabilizer SU(8), and satisfying to the self-dual reality condition
1
1
P ABCD = ABCDEF GH PEF GH PABCD = ABCDEF GH P EF GH ,
(3.3)
4!
4!
ABCDEF GH being the rank-8 completely antisymmetric RicciLevi-Civita tensor of SU(8).
By using Eqs. (3.2) and (3.3), and by exploiting the invertibility (non-singularity) of PABCD,i ,
68
the criticality conditions for VBH can be rewritten as [5,19,43]

1
Z [AB Z CD] + ABCDEF GH Z[EF ZGH ] = 0,
(3.4)
4!
which are usually referred to as the N = 8, d = 4 attractor equations. They are purely algebraic
in the (ZAB , Z AB ), and they hold for all non-singular (i.e. with VBH = 0) critical points of VBH
E7(7)
in SU(8)
at which PABCD,i is invertible.
The local SU(8) symmetry allows one to go to the so-called normal frame [44]. In such
a frame, ZAB and the unique CartanCremmerJulia quartic invariant
J4[42,45] of the fundamental representation 56 of E7(7) respectively read as follows ( 01 1
is the 2-dimensional
0
symplectic metric):
1
z1 0
0
0
0
0
0
0 0 2
0
0 i/4
0 z2 0
;
ZAB,normal =
e
0
0
0 z3 0
0
0 3
0
0
0 z4
0
0
0 4
zi i ei/4 C,
i R + ,
1 2 3 4 0,
i = 1, 2, 3, 4,
[0, 8).
(3.5)

J4,normal = (1 + 2 )2 (3 + 4 )2 (1 2 )2 (3 4 )2
+ 81 2 3 4 (cos 1).
(3.6)
Note that ZAB,normal has an (SU(2))4 symmetry. The N = 8 attractor equations (3.4) acquire the
following simple form [19]:

z1 z2 + z3 z4 = 0;
(3.7)
z1 z3 + z2 z4 = 0;
z2 z3 + z1 z4 = 0.
As expected from the analysis of [33,46], N = 8, d = 4 extremal black hole attractor equations (3.7) have only 2 distinct classes of non-singular solutions ( 18 -BPS for J4 > 0, non-BPS
for J4 < 0):
3.1.1.
1
8 -BPS
1 = 1 -BPS R+
0,
8
1 -BPS [0, 8),

8
2, 1 -BPS = 3, 1 -BPS = 4, 1 -BPS = 0. (3.8)

8
, with
The corresponding orbit of supporting BH charges in the 56 of E7(7) is O 1 -BPS = E7(7)
6(2)
8

4
2
> 0 and classical entropy SBH, 1 -BPS = J4,normal, 1 -BPS = 1
.
J4,normal, 1 -BPS = 1
8
8 -BPS
8 -BPS
As implied by Eq. (3.8), ZAB,normal, 1 -BPS has symmetry enhancement (SU(2))4 SU(6)
8
SU(2) = m.c.s.(E6(2) ). Notice that 1 -BPS is actually undetermined.
8
3.1.2. Non-BPS
1,non-BPS = 2,non-BPS = 3,non-BPS = 4,non-BPS = non-BPS R+
0,
non-BPS = .
(3.9)
69
E
The corresponding orbit of supporting BH charges in the 56 of E7(7) is Onon-BPS = E7(7)

, with
6(6)

4
J4,normal,non-BPS = 16non-BPS < 0 and classical entropy SBH,non-BPS = J4,normal,non-BPS =
2
4non-BPS
. The deep meaning of the extra factor 4 in SBH,non-BPS as compared to SBH, 1 -BPS can
8
be clearly explained when considering the so-called stu interpretation of N = 8 regular critical
points [19]. As implied by Eq. (3.9), ZAB,normal,non-BPS has symmetry enhancement (SU(2))4
USp(8) = m.c.s.(E6(6) ); indeed
ZAB,normal,non-BPS = ei 4 non-BPS AB ,
(3.10)
where AB is the USp(8) metric:

AB
(3.11)
.

Thus, as pointed out at the end of the Introduction of [21], the symmetry of ZAB,normal gets
E7(7)
given by the non-singular solutions of N = 8, d = 4
enhanced at the particular points of SU(8)
attractor equations (3.7). In general, the invariance properties of the non-singular solutions to
attractor equations are given by the m.c.s. of the stabilizer of the corresponding supporting BH
charge orbit.
3.2. Critical spectra
Let us now consider the Hessian of VBH . By further covariantly differentiating VBH , one
gets [36]
1
Hij Di Dj VBH = ZCD Z AB P,jCDEF PABEF,i = Hj i .
2
(3.12)
3.2.1. 18 -BPS
By recalling Eq. (3.8), it can be computed that (a, b = 3, . . . , 8) [36]

1
ZCD Z AB P,jCDEF PABEF,i 1 -BPS
8
2

= 2 21 -BPS P,j12ab P12ab,i 1 -BPS
Hij, 1 -BPS =
8
1
= 21 -BPS 12abEF GH [PEF GH,j P12ab,i ] 1 -BPS .
(3.13)
8
12 8
As observed in [36], the pattern of degeneracy of the modes of Hij, 1 -BPS can be understood
8
by noticing that the very structure of the non-singular 18 -BPS solution (3.8), in which only one
eigenvalue of the skew-diagonal matrix ZAB,normal is not vanishing, yields that the N = 8 theory
effectively reduces to an N = 2 theory. Consequently, the degeneracy splitting of the eigenvalues
of Hij, 1 -BPS will respect the multiplicity of the N = 2 scalar degrees of freedom: the flat
8
directions will correspond to the N = 2 hypermultiplet content, whereas the non-flat directions
(with strictly positive eigenvalues) will correspond to the N = 2 vector multiplet content.
The crucial point is the choice of the kinematical reduction N = 8 N = 2. As previously
mentioned, in the non-singular 18 -BPS case it is performed through the branching of 70 of SU(8)
along the 18 -BPS enhanced symmetry SU(6) SU(2) given by Eq. (1.1), yielding:
70
(i) 2nV = 30 strictly positive directions (massive Hessian modes), corresponding to 15 complex N = 2 vector multiplets scalars, sitting into the (15, 1) (15, 1) of SU(6) SU(2), and
parameterized by the 30 real components Pabcd ; and
(ii) 4nH = 40 flat directions (massless Hessian modes), corresponding to 10 quaternionic
N = 2 hypermultiplets scalars, sitting into the (20, 2) of SU(6) SU(2), and parameterized by
the 40 real components4 {P1abc , P2abc }.
Thus, at N = 2, 12 -BPS critical points of VBH,N =2 Eq. (1.1) can be written as follows:
m=0
m=0

70 (15, 1) (15, 1)
vectors scalars
m=0

(20, 2)
(3.14)
hypers scalars
Under the branching (1.1) PABCD decomposes as PABCD {P1abc , P2abc , Pabcd }. As it
holds true in general (also at non-BPS non-singular critical points), the N = 2 vector and hyper scalar degrees of freedom are respectively singlets and doublets of the N = 2 R-symmetry
SU(2)R,N =2 SU(2)H , which in general lies inside the whole N = 8 R-symmetry SU(8).
Thus, in the non-singular N = 8, 18 -BPS case all N = 2 vector multiplets scalar degrees of
freedom of Hij are massive, while all its N = 2 hypermultiplets scalar degrees of freedom are
massless; this can be understood by observing that the preservation of 4 supersymmetric degrees
of freedom forces such two different kind of N = 2 degrees of freedom to follow separated mass
degeneracy patterns.
3.2.2. Non-BPS
The same can be intuitively guessed not to hold in the (non-singular) non-BPS case, where no
supersymmetric degrees of freedom are preserved by the critical solution. In fact, what actually
happens is that, for what concerns the mass degeneracy splitting, the N = 2 vector and hyper
scalar degrees of freedom of Hij mix together, in a way which follows the various possibilities
yielded by all the maximal magic N = 2, d = 4 supergravities which are consistent truncations
of N = 8, d = 4 supergravity (given by Table 1).
Indeed, by recalling Eqs. (3.9) and (3.10), it can be computed that

1
Hij,non-BPS = ZCD Z AB P,jCDEF PABEF,i non-BPS
2

1 2
4 ABCDEF GH
= non-BPS
P[ABCD|,i P|EF GH ],j

2
27

1
+ 32
. (3.15)
PABCD,i PEF GH,j [AB CD] [EF GH ]
18
non-BPS
In this case, the relevant branching of the 70 of the stabilizer SU(8) is along the non-BPS
enhanced symmetry USp(8):
SU(8) USp(8);
70 42 27 1,
(3.16)
where 42, 27 and 1 respectively are the four-fold antisymmetric (traceless), two-fold antisymmetric (traceless) and the singlet of USp(8). Under the branching (3.16) PABCD decomposes as
4 Notice that, due to the self-dual reality condition (3.3), P
12ab can be re-expressed in terms of the other independent
component of PABCD .
71
follows:

PABCD PABCD , PAB , P 0 ;
1 of USp(8):
P 0 214 PABCD [AB CD] ;
27 of USp(8): PAB 32 PABCD CD 3P 0 AB , PAB = P[AB] ,

PAB AB = 0;
42 of USp(8): PABCD PABCD P[AB CD] P 0 [AB CD] ,
PABCD = P[ABCD] , PABCD CD = 0.
(3.17)
By using such an USp(8)-covariant decomposition of PABCD , the result (3.15) can be rewritten
as follows:
4

1 2
2
AB
13 0 0
Hij,non-BPS = non-BPS
(3.18)
,
P ,j PAB,i + 2 P,i P,j
2
3
non-BPS
where the barred quantities have definitions and properties analogue to the ones in Eq. (3.17), to
which they are related by the self-dual reality condition (3.3), too.
Thus, one sees that the non-BPS kinematical reduction N = 8 N = 2 performed through
the branching of 70 of SU(8) along the non-BPS enhanced symmetry USp(8) given by Eq. (3.16)
yields a different mass degeneracy splitting with respect to the 18 -BPS case treated above. Indeed,
as evident from Eq. (3.18), Hij,non-BPS is split in:
(i) 28 strictly positive directions (massive Hessian modes), sitting into the 27 1 of USp(8),
and parameterized by the 27 + 1 real components PAB and P 0 ; and
(ii) 42 flat directions (massless Hessian modes), sitting into the 42 of USp(8), and parameterized by the 42 real components PABCD .
Thus, at N = 8 non-BPS critical points of VBH Eq. (3.16) can be written as follows:
m=0
m=0
m=0

70 42 27 1 .
(3.19)
As we will see below, the identification of the massive and massless Hessian modes with the
N = 2 vector multiplets and hypermultiplets scalars is model-dependent.
However, from the splitting nV + 1/nV 1 found in [10] (holding for generic SK d-geometries), we can state the following result for non-BPS Z = 0 critical points of all N = 2, d = 4
supergravities listed in Table 1: given a pair (nV , nH ) describing the multiplets content of the
model, 4nH + nV 1 massless real modes sit in the 42 of USp(8), while nV real massive modes
sit in the 27 of USp(8) (the remaining 1 real massive mode sitting in the singlet 1 of USp(8)).
4. N = 8, 18 -BPS critical points and their N = 2 descendants
As pointed out above, N = 8, 18 -BPS critical points of VBH have symmetry SU(6) SU(2)R ,
where SU(2)R is the SU(2) factor of the N = 8 R-symmetry SU(8) which commutes with
SU(6). The 70 70 18 -BPS Hessian matrix Hij, 1 -BPS of VBH has rank 30, corresponding to
8
the N = 8 N = 2 kinematical decomposition (1.1). It is worth noticing that, under the same
branching, the 56 fundamental representation of the N = 8 U -duality group G = E7(7) decomposes into representation of the 18 -BPS symmetry SU(6) SU(2)R as follows:
56 (15, 1) (15, 1) (1, 1) (1, 1) (6, 2) (6, 2),
(4.1)
72
which consistently gives 16 electric and 16 magnetic charges for the 15 + 1 Abelian vectors
of the N = 2 matter and gravity supermultiplets. The remaining charges from the decomposition (4.1) pertain to the graviphotons which are partners of the 6 remaining gravitino multiplets
6( 32 , 2(1), 12 ) in the N = 8 N = 2 reduction (1.1), which precisely have (6, 2) (6, 2) electric
and magnetic field strengths.
4.1. N = 2, 12 -BPS
For the N = 2, d = 4 supergravities listed in Table 1, the enhanced symmetry S 1 -BPS of
N = 2, d = 4 12 -BPS critical points of VBH,N =2 is given by [21,36]
S 1 -BPS = H0 HH ,
(4.2)
where H0 is the stabilizer of the N = 2, 12 -BPS-supporting BH charge orbit,5 and HH is the

1
H
stabilizer of G
HH . Furthermore, N = 2, 2 -BPS case has N = 2 quartic GV -invariant I4 > 0,
where I4 is nothing but a suitable truncation of the E7(7) -invariant J4 . Since the sign of the
U -duality group invariant (built out from the symplectic representation of the U -duality group)
does not change in the N = 8 N = 2 supersymmetry reduction, it is clear that the N = 2,
1
1
2 -BPS case comes from the reduction of the N = 8, 8 -BPS case.
Thus, S 1 -BPS must be included in the overall enhanced symmetry SU(6) SU(2)R of the
2
N = 8, 18 -BPS case:
S 1 -BPS SU(6) SU(2)R .
(4.3)
H
The very structure of the quaternionic Khler manifold G
HH yields that HH always include at
least one explicit factor SU(2), which is promoted to a global symmetry in the case nH = 0.
Thus, HH can always (for nH = 0) be rewritten as
HH
SU(2).
(4.4)
SU(2)
In general, the N = 2 R-symmetry group SU(2)R,N =2 is identified with the SU(2) factorized
in the r.h.s. of Eq. (4.4), which in the following treatment we will denote with the subscript H :
HH =
SU(2)R,N =2 = SU(2)H HH .
The identification determining the N = 2,
reads as follows (recall Eq. (3.8)):
(4.5)
1
2 -BPS
case as descendant of the N = 8,
Z12, 1 -BPS z1, 1 -BPS = ei/4 1 -BPS = Z 1 -BPS C0 .

8
1
8 -BPS
case
(4.6)
Therefore, at N = 2, 12 -BPS critical points of VBH,N =2 (which preserve 4 supersymmetry charges,

and are always stable [5], thus corresponding to attractor configurations), the N = 8 N = 2
kinematical decomposition (1.1) identifies SU(2)R on the r.h.s. of Eq. (4.3) with the N = 2
R-symmetry SU(2)H :
SU(2)R = SU(2)H .
(4.7)
5 Here and in the following treatment we will make use of the notation set up in [21]. H is defined (for n = 0) as
V
0
HV
[21].
H0 U (1)
73
Table 2
HH
(relevant at N = 2, 12 -BPS
The 12 -BPS supporting BH charge orbit O 1 -BPS , and the compact groups H0 and SU(2)
R
2
critical points) for the N = 2, d = 4 supergravities listed in Table 1
1 -BPS orbit
2
G
O 1 -BPS = HV
0
V
H0 U (1)
HH
SU(2)R =SU(2)H
SU(6)
HH , SU H (2) = SU(2)R = GH
SU(3) SU(3)
U (1)
SO (12)
SU(6)
SU(3,3)
SU(3)SU(3)
Sp(6,R)
SU(3)
J3H
J3C
J3R
SU(3)
SU(2)
stu
(SU(1,1))3
(U (1))2
(U (1))2
(SU(2))3
R
J3,M
SU(1, 1)
USp(6)
C
J3,M
SU(6)
Thus, Eq. (4.4) can be rewritten as

HH =
HH
SU(2)R ,
SU(2)R
(4.8)
which, by Eq. (4.3), implies that

H0
HH
SU(6).
SU(2)R
(4.9)
The corresponding data for all the N = 2, d = 4 supergravities which are consistent truncations
of the N = 8, d = 4 theory (listed in Table 1) are given in Table 2 (for the columns O 1 -BPS
2
and H0 refer to Tables 3 and 8 of [21]).
From Table 2 it is also evident that SU(2)R has necessarily to be chosen in HH , because
in all models H0 does not contain a factorized SU(2). Moreover, two orders of considerations
follow:
HH
is a proper subgroup of SU(6) in all models but the two limit models J3H
(i) H0 SU(2)
R
C (having n = 0, and thus H undefined and
(having nH = 0, and thus HH undefined) and J3,M
V
0
1
corresponding to a ReissnerNrdstrom extremal BH, only having 2 -BPS critical points).
For J3H , SU(2)R = SU(2)H is identified with the global symmetry SU(2) = GH due to
nH = 0.
C it holds that S
On the other hand, for J3,M
1
-BPS = HH = SU(6) SU(2)R , i.e. the enhanced
2
N = 2, 12 -BPS symmetry S 1 -BPS , the stabilizer of the quaternionic Khler manifold

2
GH
HH
and the
enhanced N = 8,
symmetry coincide.
(ii) Two models exist where an a priori arbitrariness in the identification of SU(2)H in HH
exists: J3R and stu.
However, in J3R such an arbitrariness is removed by the quantum numbers of the hypermultiplets scalars (which are always doublets of SU(2)H ); the right SU(2) to choose is the one
promoted to a global symmetry in the limit case nH = 0. On the other hand, in stu case the
arbitrariness of choice is removed by the noteworthy triality symmetry of the model.
1
8 -BPS
74
4.2. N = 2 non-BPS Z = 0
For the N = 2, d = 4 supergravities listed in Table 1, the overall symmetry Snon-BPS,Z=0 of
N = 2, d = 4 non-BPS Z = 0 critical points of VBH,N =2 is given by [21]
Snon-BPS,Z=0 = h HH ,
(4.10)
where h is the m.c.s. (factorized by U (1)) of the stabilizer H of the N = 2 non-BPS Z = 0supporting BH charge orbit [21]. Furthermore, N = 2 non-BPS Z = 0 case has N = 2 quartic
GV -invariant I4 > 0, as the N = 2, 12 -BPS case. Thus, it is clear that N = 2 non-BPS Z = 0 case
comes from the very same N = 8 N = 2 supersymmetry reduction giving raise to N = 2,
1
2 -BPS case. Thus, Snon-BPS,Z=0 must be included in the overall enhanced symmetry SU(6)
SU(2)R of the N = 8, 18 -BPS case:
Snon-BPS,Z=0 SU(6) SU(2)R .
(4.11)
The identification determining the N = 2 non-BPS Z = 0 case as descendant of the N = 8,

1
8 -BPS case reads as follows (recall that Znon-BPS,Z=0 = 0):
Z12, 1 -BPS z1, 1 -BPS = ei/4 1 -BPS = (Di Z)non-BPS,Z=0 = 0,
8
(4.12)
where i is one particular element of the set {1, . . . , nV }. In this sense, the key difference with
respect to the previously treated N = 2, 12 -BPS case is that the N = 2 central charge is interchanged with one N = 2 matter charge.
This leads to the fact that for N = 2 models under consideration which exhibit flat Hessian
directions at N = 2 non-BPS Z = 0 critical points of VBH,N =2 (namely J3H , J3C and J3R ) the
SU(2)R of the enhanced N = 8, 18 -BPS symmetry SU(2)R SU(6) is not identified with the
SU(2)R,N =2 (i.e. with (one of) the SU(2)(s) factorized in HH ) any more.
The stu model has h = SO(2). In such a model all goes the same way as for the previously
treated N = 2, 12 -BPS case, and consequently in stu model N = 2 non-BPS Z = 0 critical points
of VBH,N =2 are stable, i.e. there are no flat non-BPS Z = 0 Hessian directions at all. This can
be simply understood by noticing that in such an N = 2 framework triality symmetry puts nonBPS Z = 0 critical points on the very same footing of 12 -BPS critical points, which are always
stable and thus do not have any flat direction at all.
The corresponding data for all the maximal magic N = 2, d = 4 supergravities which are
consistent truncations of the N = 8, d = 4 theory (listed in Table 1) are given in Table 3 (for the
column h refer to Table 8 of [21]).
Let us consider two explicit examples, namely the models J3H and stu.
The model J3H has the highest number of vector multiplets (nV = 15) and no hypermultiplets
at all (nH = 0); thus, HH cannot be defined, and SU(2) = SU(2)H is promoted to a global
symmetry, which here coincides with GH itself. SU(2)R is identified with the factor SU(2) in
h = SU(4) SU(2), thus it holds that SU(4) GH = SU(4) SU(2)H SU(6). The 15, 15
and 20 of SU(6) decompose under SU(4) SU(2)H as follows:
15 = (4, 2) (6, 1) (1, 1);
15 = (4, 2) (6, 1) (1, 1);
20 = (4, 1) (4, 1) (6, 2).
(4.13)
75
Table 3
HH
(relevant at
The non-BPS Z = 0 supporting BH charge orbit Onon-BPS,Z=0 , and the compact groups h and SU(2)
H
N = 2 non-BPS Z = 0 critical points) for the N = 2, d = 4 supergravities listed in Table 1
non-BPS Z = 0 orbit
G
Onon-BPS,Z=0 = V
SO (12)
SU(4,2)
SU(3,3)
SU(2,1)SU(1,2)
Sp(6,R)
SU(2,1)
J3H
J3C
J3R
H )
h m.c.s.(
U (1)
HH
SU(2)H
SU(4) SU(2)R
HH , SU H (2) = GH
SU(2) SU(2) U (1)
U (1)
SU(2)
SU(2)
stu
(SU(1,1))3
(U (1))2
SO(2)
(SU(2))2 SU(2)R
R
J3,M
USp(6),
SU(2)H = SU(2)R
C
J3,M
SU(6),
SU(2)H = SU(2)R
Thus, by also recalling Eq. (3.14), one obtains that at N = 2 non-BPS Z = 0 critical points
the N = 8, 18 -BPS enhanced symmetry SU(6) SU(2)R decomposes under SU(4) SU(2)H
SU(2)R as follows:
m = 0:
(15, 1) (15, 1) = (4, 2, 1) (4, 2, 1) (6, 1, 1) (6, 1, 1)

(1, 1, 1) (1, 1, 1);
m = 0:
(20, 2) = (4, 1, 2) (4, 1, 2) (6, 2, 2).
(4.14)
As previously mentioned, in general the N = 2 vector multiplets and hypermultiplets scalar

degrees of freedom are respectively given by the singlets and doublets of SU(2)H . For the model
under consideration, all vector multiplets scalars are included in the N = 2, d = 4 spectrum,
whereas all hypermultiplets scalars are truncated away by dimensional reduction N = 8
N = 2. Thus, the representation decomposition (4.14) yields that at N = 2 non-BPS Z = 0
critical points the vector multiplets scalars and hypermultiplets scalars respectively sit in the
following representations of SU(4) SU(2)H SU(2)R :
14 m=0

(6, 1, 1) (6, 1, 1) (1, 1, 1) (1, 1, 1)
30 (real) vectors scalar degrees of freedom =
16 m=0
(all in the N =2, d=4 spectrum)

(4, 1, 2) (4, 1, 2) ;
16 m=0
m=0

24

40 (real) hypers scalar degrees of freedom = (4, 2, 1) (4, 2, 1) (6, 2, 2),
(all truncated away in the N =8N =2 reduction)
(4.15)
yielding a non-BPS Z = 0 mass splitting 14 m = 0/16 m = 0 of the vector multiplets scalar
degrees of freedom, matching the result obtained in [21].
The model stu is the one with the smallest number of vector multiplets (nV = 3) still exhibiting
non-BPS Z = 0 critical points. Without loss of generality (due to triality symmetry), one can
identify SU(2)R with the fourth factor SU(2) in HH = SO(4) SO(4) = (SU(2))4 , whereas the
N = 2 R-symmetry can be identified with the third factor SU(2) in HH . Thus, as yielded by
76
Table 3, the N = 2 non-BPS Z = 0 symmetry h HH can be rewritten as

stu:

2
h HH = SO(2) SU(2) SU(2)H SU(2)R .
(4.16)
Thus, it holds that SO(2) (SU(2))2 SU(2)H SU(6).

Thus, by also recalling Eq. (3.14), one obtains that at N = 2 non-BPS Z = 0 critical points the
N = 8, 18 -BPS enhanced symmetry SU(6) SU(2)R decomposes under (SU(2))2 SU(2)H
SU(2)R as follows:
m = 0:
(15, 1) (15, 1) = 6(1, 1, 1, 1) 2(2, 2, 1, 1) 2(2, 1, 2, 1) 2(1, 2, 2, 1);
m = 0:
(20, 2) = (2, 2, 2, 2) 2(1, 1, 2, 2) 2(1, 2, 1, 2) 2(2, 1, 1, 2).
(4.17)
Such a representation decomposition yields that at N = 2 non-BPS Z = 0 critical points the

vector multiplets scalars and hypermultiplets scalars respectively sit in the following representations of (SU(2))2 SU(2)H SU(2)R :
30 (real) vectors scalar degrees of freedom
(6 in the N =2, d=4 spectrum, 24 truncated away)
m=0

6(1, 1, 1, 1)
6 in the N =2, d=4 spectrum

m=0
m=0

2(2, 2, 1, 1) 2(1, 2, 1, 2) 2(2, 1, 1, 2) ;
24 truncated away
40 (real) hypers scalar degrees of freedom

(16 in the N =2, d=4 spectrum, 24 truncated away)
m=0

(2, 2, 2, 2)
16 in the N =2, d=4 spectrum

m=0
m=0

2(1, 1, 2, 2) 2(2, 1, 2, 1) 2(1, 2, 2, 1) ,
(4.18)
24 truncated away
yielding that the Hessian of VBH,N =2 has no flat directions at non-BPS Z = 0 critical points in
the stu model. As mentioned above, this can be traced back to the noteworthy triality symmetry
of the model under consideration, putting non-BPS Z = 0 critical points on the very same footing
of 12 -BPS critical points.
Thus, in this sense one can state that in the stu model the stability of 12 -BPS critical points
implies, by triality symmetry, the stability of non-BPS Z = 0 critical points. This can be quantitatively understood by considering the representation decomposition of SU(6) SU(2)R in
the 12 -BPS case. In such a case SU(2)R = SU(2)H , and SU(6) SU(2)R decomposes into
HH
SU(2)R = (U (1))2 (SU(2))3 SU(2)R (once again, the choice of SU(2)R as
H0 SU(2)
R
the fourth SU(2) does not imply any loss of generality, due to triality symmetry). It is thus easy to
realize that this amounts simply to interchange the third and fourth SU(2)s in the representation
decomposition (4.17).
77
Table 4
HH
(relevant at
The non-BPS Z = 0 supporting BH charge orbit Onon-BPS,Z=0 , and the compact groups h and SU(2)
H
N = 2 non-BPS Z = 0 critical points) for the N = 2, d = 4 supergravities listed in Table 1
non-BPS, Z = 0 orbit
G
Onon-BPS,Z=0 = V
SO (12)
SU (6)
SU(3,3)
SL(3,C)
Sp(6,R)
SL(3,R)
(SU(1,1))3
(SO(1,1))2
J3H
J3C
J3R
stu
h m.c.s.(H )
HH
SU(2)H
USp(6)
HH , SU H (2) = GH
SU(3)
U (1)
SO(3)
SU(2)
(SU(2))3
R
J3,M
SU(1, 1)
USp(6)
C
J3,M
SU(6)
5. N = 8 non-BPS critical points and N = 2 non-BPS Z = 0 critical points

For the N = 2, d = 4 supergravities listed in Table 1, the overall symmetry Snon-BPS,Z=0 of
N = 2, d = 4 non-BPS Z = 0 critical points of VBH,N =2 is given by [21]
Snon-BPS,Z=0 = h HH ,
(5.1)
where h is the m.c.s. of the stabilizer H of the N = 2 non-BPS Z = 0-supporting BH charge

orbit [21]. Furthermore, N = 2 non-BPS Z = 0 case has N = 2 quartic GV -invariant I4 < 0.
Thus, it is clear that N = 2 non-BPS Z = 0 case comes from the N = 8 N = 2 supersymmetry reduction given by Eq. (3.16). Thus, Snon-BPS,Z=0 must be included in the overall enhanced
symmetry USp(8) of the N = 8 non-BPS case:
Snon-BPS,Z=0 USp(8).
(5.2)
In order to determine the mass degeneracy pattern of the Hessian of VBH,N =2 at N = 2 nonBPS Z = 0 critical points, one will thus have to consider the decomposition of the representations
42 (m = 0), 27 (m = 0) and 1 (m = 0) of the enhanced N = 8 non-BPS symmetry USp(8) (recall
Eqs. (3.16) and (3.19)) into suitable representations of Snon-BPS,Z=0 . The embedding (5.2) is a
priori not unique, but only one embedding among the possible ones is consistent with the known
quantum numbers of the vector and hyper multiplets scalars in the consider models, and thus
consistent with the performed supersymmetry reduction N = 8 N = 2.
The corresponding data for all the N = 2, d = 4 supergravities which are consistent trunca refer
tions of the N = 8, d = 4 theory (listed in Table 1) are given in Table 4 (for the column h
to Table 8 of [21]).
In the following subsections we will analyze each model separately.
5.1. J3H
SO (12)
V
As given by Table 1, this model has (nV , nH ) = (15, 0), and G
HV = U (6) . HH cannot be
defined, and SU(2)H = GH is the global symmetry due to nH = 0. From Table 2 of [37] the
fundamental representation 56 of G = E7(7) decomposes along GV GH = SO (12) SU(2)H
as follows:
56 (32, 1) (12, 2),
(5.3)
78
yielding that the 32 real electric and magnetic charges {p 0 , p 1 , . . . , p 15 , q0 , q1 , . . . , q15 } of the
1 + 15 vectors of J3H lie in the SU(2)H -singlet real representation (32, 1) of SO (12) SU(2)H
(here and in what follows the index 0 pertains to the graviphoton). On the other hand, the
fundamental representation 8 of the enhanced N = 8 non-BPS symmetry USp(8) decomposes
along Snon-BPS,Z=0 = h SU(2)H = USp(6) SU(2)H as follows:
8 (6, 1) (1, 2).
(5.4)
The decomposition of the representations 42, 27 and 1 of USp(8) along Snon-BPS,Z=0 and its
interpretation in terms of the N = 2, d = 4 spectrum (and of the truncated scalar degrees of
freedom) reads as follows:
28 m=0 hypers scalar degrees of freedom truncated away

(14 , 2)
m = 0: 42
14 m=0 vectors scalar degrees of freedom in N =2, d=4 spectrum

(14, 1)
;

(6, 2)
27

m = 0:
(5.5)
(1,
1)
(14,
1)
;
1 m=0 vectors scalar degree of freedom in N =2, d=4 spectrum

1
(1, 1)
,
where 14 and 14 respectively stand for the two-fold and three-fold antisymmetric (traceless) of
USp(6).
It should be noticed that for J3H the embedding of Snon-BPS,Z=0 in the enhanced N = 8
non-BPS symmetry USp(8) is unique. Moreover, since J3H has the highest number nV = 15
of Abelian vector multiplets, all (would-be N = 2) vectors scalar degrees of freedom of the
starting N = 8 theory survive after the reduction N = 8 N = 2.
The N = 2 non-BPS Z = 0 mass degeneracy pattern of the vector multiplets scalar degrees
of freedom resulting from the decomposition (5.5) is nV + 1 = 16 m = 0/nV 1 = 14 m = 0,
thus confirming the Hessian splitting found in [10].
5.2. J3C
As given by Table 1, this model has (nV , nH ) = (9, 1), and
SU(2,1)
SU(2)H U (1) .
GV
HV
GH
HH
SU(3,3)
SU(3)SU(3)U (1)
From Table 2 of [37] the fundamental representation 56 of G = E7(7) decomposes

along GV GH = SU(3, 3) SU(2, 1) as follows:
56 (20, 1) (6, 3) (6, 3),
(5.6)
{p 0 , p 1 , . . . , p 9 , q
yielding that the 20 real electric and magnetic charges

0 , q1 , . . . , q9 } of the
1 + 9 vectors of J3C lie in the SU(2, 1)-singlet real representation (20, 1) of SU(3, 3) SU(2, 1).
On the other hand, the fundamental representation 8 of the enhanced N = 8 non-BPS symmetry
USp(8) decomposes along Snon-BPS,Z=0 = h HH = SU(3) SU(2)H U (1) as follows (here
and in what follows we disregard the quantum numbers of U (1), not essential for our purposes):
8 (3, 1) (3, 1) (1, 2).
(5.7)
79
4 m=0 hypers scalar degrees of freedom in N =2, d=4 spectrum

(1,
2)
(1,
2)
6 m=0 vectors scalar degrees of freedom truncated away

(3, 1) (3, 1)
m = 0: 42

2)
(6,
2)
(6,

(8, 1)
;
6
m
=
0
vectors
scalar
degrees
of
freedom
truncated
away

1)
(3,
1)
(3,

(8,
1)
27

m = 0:
(5.8)
2)
(3,
2)
(3,

(1,
1)
;

(1, 1)
.
1
It should be noticed that for J3C the embedding of Snon-BPS,Z=0 in the enhanced N = 8 nonBPS symmetry USp(8) is a priori not unique, but the only consistent with the N = 8 N = 2
reduction originating J3C is the following two-step one:
USp(8) USp(6) USp(2) SU(3) SU(2)H U (1).
(5.9)
Moreover, as evident from the decomposition (5.8), the N = 8 N = 2 reduction originating

J3C truncates away:
(1) 6 m = 0 and 6 m = 0 vectors scalar degrees of freedom, both sets sitting in the (3, 1)
(3, 1) of SU(3) SU(2)H ;
(2) 24 m = 0 and 12 m = 0 hypers scalar degrees of freedom, respectively sitting in the
(6, 2) (6, 2) and (3, 2) (3, 2) of SU(3) SU(2)H .
The resulting N = 2 J3C spectrum is composed by 4 m = 0 real hypers scalar degrees of
freedom (rearranging in 1 quaternionic hypermultiplet scalar), and by nV + 1 = 10 m = 0 and
nV 1 = 8 m = 0 real vectors scalar degrees of freedom, whose mass degeneracy pattern thus
confirms the Hessian splitting found in [10].
5.3. J3R
G2(2)
SU(2)SU(2)H
GV
HV
GH
HH
Sp(6,R)
U (3)
. From Table 2 of [37] the fundamental representation 56 of G = E7(7) decom-
80
poses along GV GH = Sp(6, R) G2(2) as follows:

56 (14 , 1) (6, 7),
(5.10)
where 14 is the three-fold antisymmetric (traceless) representation of Sp(6, R). The decomposition (5.10) yields that the 14 real electric and magnetic charges {p 0 , p 1 , . . . , p 6 , q0 , q1 , . . . , q6 }
of the 1 + 6 vectors of J3R lie in the G2(2) -singlet real representation (14 , 1) of Sp(6, R) G2(2) .
The symmetry group Snon-BPS,Z=0 of J3R reads
Snon-BPS,Z=0 = h HH = SO(3) SU(2) SU(2).
(5.11)
Thus, a priori Snon-BPS,Z=0 can be embedded in the enhanced N = 8 non-BPS symmetry USp(8)
in many ways, but the only consistent with the N = 8 N = 2 reduction originating J3R is the
following two-step one:
USp(8)
USp(6)
USp(2),

SO(3) SU(2) SU(2)H ,
HH = SU(2) SU(2)H ,
(5.12)
yielding that Snon-BPS,Z=0 can be rewritten as

Snon-BPS,Z=0 = SO(3) SU(2) SU(2)H .
(5.13)
Under the decomposition (5.12) one obtains

8 (6, 1) (1, 2) (3, 2, 1) (1, 1, 2),
(5.14)
and thence
42 (14, 1) (14 , 2) (5, 1, 1) (3, 3, 1) (1, 4, 2) (5, 2, 2);
27 (6, 2) (14, 1) (1, 1) (3, 2, 2) (5, 1, 1) (3, 3, 1) (1, 1, 1);
1 (1, 1, 1).
(5.15)
Such decompositions can be interpreted in terms of the N = 2, d = 4 spectrum (and of the

truncated scalar degrees of freedom) as follows:

(5, 1, 1)
9
m=0
vectors
scalar
degrees of freedom truncated away

(3, 3, 1)
m = 0 : 42

(1, 4, 2)

(5, 2, 2)
;
m = 0:
81

(5, 1, 1) (1, 1, 1)

27
(3, 2, 2)
9
m
=
0
vectors
scalar
degrees
of freedom truncated away

(3, 3, 1)
;
1
m
=
0
vectors
scalar
degree
of
freedom
in
N
=2,
d=4
spectrum

1
(1, 1, 1)
.
(5.16)
Such decompositions yield that the N = 8 N = 2 reduction originating J3R truncates away:
(1) 9 m = 0 and 9 m = 0 vectors scalar degrees of freedom, both sets sitting in the (3, 3, 1)
of SU(2) SU(2) SU(2)H ;
(2) 20 m = 0 and 12 m = 0 hypers scalar degrees of freedom, respectively sitting in the
(5, 2, 2) and (3, 2, 2) of SU(2) SU(2) SU(2)H .
The resulting N = 2 J3R spectrum is composed by 8 m = 0 real hypers scalar degrees of
confirms once again the Hessian splitting found in [10].
5.4. stu
SO(2,2)
SO(2)SO(2)
SO(4,4)
SO(4)SO(4) .
GV
HV
GH
HH
SU(1,1)
U (1)
From Eq. (182) of [35] the fundamental representation 56 of G =

E7(7) decomposes along GV GH = SU(1, 1) SO(2, 2) SO(4, 4) (SU(1, 1))3 SO(4, 4)
as follows (the three SU(1, 1) are actually indistinguishable due to triality symmetry):
56 (2, 2, 2, 1) (2, 1, 1, 8v ) (1, 2, 1, 8s ) (1, 1, 2, 8c ),
(5.17)
where 8v , 8s and 8c respectively are the vector, chiral spinorial and anti-chiral spinorial representations of SO(4, 4). The decomposition (5.17) yields that the 8 real electric and magnetic charges
{p 0 , p 1 , . . . , p 3 , q0 , q1 , . . . , q3 } of the 1 + 3 vectors of the stu model lie in the SO(4, 4)-singlet
real representation (2, 2, 2, 1) of (SU(1, 1))3 SO(4, 4). The symmetry group Snon-BPS,Z=0 of
the stu model reads

4
h =I
Snon-BPS,Z=0 = h HH stu= HH = SO(4) SO(4) SU(2) .
(5.18)
Thus, a priori Snon-BPS,Z=0 can be embedded in the enhanced N = 8 non-BPS symmetry USp(8)
in many ways, but the only consistent with the N = 8 N = 2 reduction originating the stu
model is the following two-step one:

4
USp(8) USp(4) USp(4) SO(4) SO(4) SU(2) .
(5.19)
We can choose the N = 2 R-symmetry SU(2)H to be the fourth one in Snon-BPS,Z=0 (as we will
see below, such an arbitrariness in the choice of the placement of the N = 2 R-symmetry inside
HH is actually removed by the triality symmetry of the stu model). Consequently, Snon-BPS,Z=0
can be rewritten as

3
Snon-BPS,Z=0 = SU(2) SU(2)H .
(5.20)
82
Thus, the fundamental representation 8 of the enhanced N = 8 non-BPS symmetry USp(8) decomposes along the chain of branchings (5.19) as follows:
8
USp(8)
(4, 1) (1, 4) (4s , 1) (1, 4s )

USp(4)USp(4)
SO(4)SO(4)
(2, 1, 1, 1) (1, 2, 1, 1) (1, 1, 2, 1) (1, 1, 1, 2),
(5.21)
SU(2)SU(2)SU(2)SU(2)H
where 4s is the spinorial of SO(4) (or, equivalently, the reduction of the fundamental of USp(4)
with respect to SO(4)).
Due to the chain of group inclusions (5.19) needed in the stu model in order to correctly embed
Snon-BPS,Z=0 into USp(8), the decomposition of the representations 42, 27 and 1 of USp(8) along
Snon-BPS,Z=0 should better be performed in two steps:
(i) Decomposition of USp(8) along USp(4) USp(4). It respectively yields (the prime distinguishes therepresentations of thetwo USp(4))
m = 0:
m = 0:
42 (4, 4 ) (5, 5 ) (1, 1 );

27 (4, 4 ) (5, 1 ) (1, 5 ) (1, 1 );
1 (1, 1 ).
(5.22)
(ii) Decomposition of USp(4) USp(4) along SO(4) SO(4). It will involve the representations 4s (previously introduced) and 4v (vector representation of SO(4) or, equivalently, reduction
of the antisymmetric traceless of USp(4) with respect to SO(4)). By exploiting the following decompositions of the representations (4, 4 ), (5, 5 ), (5, 1 ) and (1, 1 ) of USp(4) USp(4) along
SO(4) SO(4):

(4, 4 ) 4s , 4 s ;

(5, 5 ) 4v , 4 v (4v , 1 ) 1, 4 v (1, 1 );
(5, 1 ) (4v , 1 ) (1, 1 );
(1, 1 ) (1, 1 ),
(5.23)
one gets the following decompositions of representations 42, 27 and 1 of USp(8) along SO(4)
SO(4):

m = 0: 42 4s , 4 s 4v , 4 v (4v , 1 ) 1, 4 v 2(1, 1 );

27 (4s , 4 s ) (4v , 1 ) (1, 4 v ) 3(1, 1 );
m = 0:
(5.24)
1 (1, 1 ).
(iii) Further decomposition, performed by exploiting the group isomorphism SO(4)
SU(2) SU(2). Under the group isomorphism SO(4) (SU(2))2 , 4s and 4v respectively decompose as follows:
4s (2, 1) (1, 2);
4v (2, 2).
(5.25)
Thus, the decomposition of representations 42, 27 and 1 of USp(8) along (SU(2))4 = (SU(2))3
SU(2)H (embedded into USp(8) in the way given by the chain (5.19) of group inclusions), and
its interpretation in terms of the N = 2, d = 4 spectrum (and of the truncated scalar degrees of
83
freedom), reads as follows:


(2,
1,
2
,
1
)
(1,
2,
2
,
1
)
(2,
2,
1
,
1
)

(1, 2, 1 , 2 ) (2, 1, 1 , 2 ) (1, 1, 2 , 2 )

m = 0: 42

(2,
2, 2 , 2 )

2(1, 1, 1 , 1 )
;

(2,
1,
2
,
1
)
(1,
2,
2
,
1
)
(2,
2,
1
,
1
)

27 (1, 2, 1 , 2 ) (2, 1, 1 , 2 ) (1, 1, 2 , 2 )
m = 0:

3(1,
1,
1
,
1
)
;
1
m
=
0
vectors
scalar
degree
of
freedom
in
N
=2,
d=4
spectrum

1
(1, 1, 1 , 1 )
.
(5.26)
Such decompositions yield that the N = 8 N = 2 reduction originating the stu model
truncates away:
(1) 12 m = 0 and 12 m = 0 vectors scalar degrees of freedom, both sets sitting in the
(2, 1, 2 , 1 )(1, 2, 2 , 1 )(2, 2, 1 , 1 ) of (SU(2))3 SU(2)H (note the triality symmetry acting
on the first three quantum numbers);
(2) 12 m = 0 and 12 m = 0 hypers scalar degrees of freedom, both sets sitting in the
(1, 2, 1 , 2 )(2, 1, 1 , 2 )(1, 1, 2 , 2 ) of (SU(2))3 SU(2)H (note the triality symmetry acting
on the first three quantum numbers).
As it is seen, both the vectors and hypers scalar degrees of freedom truncated out receive
half of the contribution from the 42 (massless) of USp(8) and the other half of the contribution
from the 27 (massive) of USp(8). As it holds in general, the massive singlet representation 1 of
USp(8) always appears in the N = 2, d = 4 resulting spectrum.
The spectrum of the N = 2, d = 4 stu model determined by the decompositions (5.26) is
composed by 16 m = 0 real hypers scalar degrees of freedom (rearranging in 4 quaternionic hypermultiplet scalar), and by nV + 1 = 4 m = 0 and nV 1 = 2 m = 0 real vectors scalar degrees
of freedom, whose mass degeneracy pattern thus confirms once again the Hessian splitting found
in [10].
R
5.5. J3,M

F4(4)
USp(6)SU(2)H
GV
HV
GH
HH
SU(1,1)
U (1)
(recall that USp(2) SU(2)). From Table 2 of [37] the fundamental representation 56 of G = E7(7) decomposes along GV GH = SU(1, 1) F4(4) as follows:
56 (4, 1) (2, 26).
(5.27)
84
Such a decomposition yields that the 4 real electric and magnetic charges {p 0 , p 1 , q0 , q1 } of the
R lie in the F
1 + 1 vectors of J3,M
4(4) -singlet real representation (4, 1) of SU(1, 1) F4(4) . The
SU(1,1)
V
representation 4 of SU(1, 1) corresponds to spin s = 32 , and this identifies G
HV = U (1) as a
special Khler manifold (dimC = 1) with cubic holomorphic prepotential reading6 (in a suitable
system of special projective coordinates) F (t) = t 3 , C0 . The symmetry group Snon-BPS,Z=0
R is the same of the one of J H , and it reads (h
= I, as in the stu model)
of J3,M
3
Snon-BPS,Z=0 = h HH = HH = USp(6) SU(2)H .
(5.28)
R the embedding of S
As it holds also for J3H , in the model J3,M
non-BPS,Z=0 in the enhanced N = 8
non-BPS symmetry USp(8) is unique. The fundamental representation 8 of USp(8) decomposes
along USp(6) SU(2)H as follows:
8 (6, 1) (1, 2).
(5.29)

(14, 1)
m = 0: 42

(14 , 2)
;

(6,
2)

27
(14, 1)
m = 0:
(5.30)

(1, 1)
;

1
(1, 1)
,
where 14 and 14 respectively stand for the two-fold and three-fold antisymmetric (traceless) of
USp(6).
R truncates
Such decompositions yield that the N = 8 N = 2 reduction originating J3,M
away:
(1) 14 m = 0 and 14 m = 0 vectors scalar degrees of freedom, both sets sitting in the (14, 1)
of USp(6) SU(2)H ;
(2) 12 m = 0 hypers scalar degrees of freedom, sitting in the (6, 2) of USp(6) SU(2)H .
R spectrum is composed by 28 m = 0 real hypers scalar degrees of
The resulting N = 2 J3,M
6 For a discussion of (the N = 2, d = 4 attractor equations in the special Khler geometry of) SU(1,1) with cubic
U (1)
holomorphic prepotential, see e.g. [21,29] (and references therein) and [31].
85
confirms once again the Hessian splitting found in [10] (no flat directions of non-BPS Z = 0
R are all
Hessian, implying that the non-BPS Z = 0 critical points of VBH,N =2 in the model J3,M
stable).
C has
For what concerns the other mirror models, there is nothing more to say. Indeed, J3,M
nV = 0 and thus it corresponds to a ReissnerNrdstrom (extremal) BH with (graviphoton)
charges p 0 and q0 , only admitting 12 -BPS critical points for VBH,N =2 . Furthermore, as previH does not exist (at least as far d = 4 is concerned), and stu is self-mirror:
ously mentioned, J3,M
stu,M = stu.
6. Conclusion
In the present paper, in order to understand more in depth the nature of the non-BPS solutions
to attractor equations in N = 8, d = 4 supergravity, we considered the supersymmetry reduction
down to N = 2, d = 4 magic supergravities (and their mirror theories). The multiplets content
is given by nV vector supermultiplets, whose complex scalars span a special Khler manifold
of dimension nV , and by nH hypermultiplets, whose quaternionic scalars span a quaternionic
Khler manifold of dimension nH .
The mass spectrum of vector multiplets scalars (the only relevant for the Attractor Mechanism in ungauged supergravities) in N = 2 magic supergravities has been studied in [21]. By
taking into account also the hidden modes truncated away in the supersymmetry reduction
N = 8 N = 2, the splittings of the N = 2 spectra should reproduce the splittings of the full
spectra of the 70 real scalars of the parent N = 8 theory. We have shown how this works, and in
particular we reproduced the result of [10] about the mass splitting of the modes of the N = 2
non-BPS Z = 0 Hessian.
By the supersymmetry reduction N = 8 N = 2, the eventual instability of N = 2 non-BPS
Z = 0 solutions to attractor equations studied in [10] should reflect in a possible instability of
N = 8 non-BPS critical points of VBH in N = 8, d = 4 supergravity.
On the other hand, by assuming that supersymmetry determines the N = 8, 18 -BPS critical
points to be stable, it is possible to argue that the N = 2 non-BPS Z = 0 critical points of
VBH,N =2 should be stable (beside the N = 2, 12 -BPS critical points, whose stability is known
after [5]). Correspondingly, when covariantly differentiating VBH,N =2 beyond the second order,
the eventual flat directions of the non-BPS Z = 0 Hessian should suitably lift to directions
with strictly positive eigenvalues, or remain flat at all orders. Among the considered models,
only the N = 2, d = 4 stu supergravity (having (nV , nH ) = (3, 4), and thus self-mirror) exhibit
non-BPS, Z = 0 critical points stable already at the Hessian level. This can be understood by
noticing that in such an N = 2 framework triality symmetry puts non-BPS Z = 0 critical points
on the very same footing of 12 -BPS critical points, which are always stable [5] and thus do not
have any flat direction at all.
We conclude by saying that our analysis could be applied to non-BPS critical points of VBH in
2 < N < 8 (d = 4), extended supergravities, eventually comparing the N = 8 non-BPS spectrum
with spectra arising in 2 < N < 8 theories obtained by consistent supersymmetry reductions
(along the lines of [37]), as done in [36] for the N = 8, 18 -BPS spectrum. Ultimately, such a
procedure could be performed for the N = 1, d = 4 reduction of these theories, especially of the
N = 2 SK d-geometries [30].
86
Acknowledgements
The work of S.F. has been supported in part by European Community Human Potential Program under contract MRTN-CT-2004-005104 Constituents, fundamental forces and symmetries
of the universe and the contract MRTN-CT-2004-503369 The quest for unification: Theory
Confronts Experiments, in association with INFN Frascati National Laboratories and by D.O.E.
grant DE-FG03-91ER40662, Task C.
The work of A.M. has been supported by a Junior Grant of the Enrico Fermi Center, Rome,
in association with INFN Frascati National Laboratories, and in part by D.O.E. grant DE-FG0391ER40662, Task C.
A.M. would like to thank the Department of Physics and Astronomy, University of California
at Los Angeles, where this project was completed, for kind hospitality and stimulating environment.
We would like also to acknowledge Restaurant Lawrys-The Prime Rib in Beverly Hills,
for its inspiring atmosphere.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
S. Ferrara, R. Kallosh, A. Strominger, N = 2 extremal black holes, Phys. Rev. D 52 (1995) 5412, hep-th/9508072.
A. Strominger, Macroscopic entropy of N = 2 extremal black holes, Phys. Lett. B 383 (1996) 39, hep-th/9602111.
S. Ferrara, R. Kallosh, Supersymmetry and attractors, Phys. Rev. D 54 (1996) 1514, hep-th/9602136.
S. Ferrara, R. Kallosh, Universality of supersymmetric attractors, Phys. Rev. D 54 (1996) 1525, hep-th/9603090.
S. Ferrara, G.W. Gibbons, R. Kallosh, Black holes and critical points in moduli space, Nucl. Phys. B 500 (1997) 75,
hep-th/9702103.
A. Sen, Black hole entropy function and the attractor mechanism in higher derivative gravity, JHEP 0509 (2005)
038, hep-th/0506177.
K. Goldstein, N. Iizuka, R.P. Jena, S.P. Trivedi, Non-supersymmetric attractors, Phys. Rev. D 72 (2005) 124021,
hep-th/0507096.
A. Sen, Entropy function for heterotic black holes, JHEP 0603 (2006) 008, hep-th/0508042.
R. Kallosh, New attractors, JHEP 0512 (2005) 022, hep-th/0510024.
P.K. Tripathy, S.P. Trivedi, Non-supersymmetric attractors in string theory, JHEP 0603 (2006) 022, hep-th/0511117.
A. Giryavets, New attractors and area codes, JHEP 0603 (2006) 020, hep-th/0511215.
K. Goldstein, R.P. Jena, G. Mandal, S.P. Trivedi, A C-function for non-supersymmetric attractors, JHEP 0602 (2006)
053, hep-th/0512138.
M. Alishahiha, H. Ebrahim, Non-supersymmetric attractors and entropy function, JHEP 0603 (2006) 003, hep-th/
0601016.
R. Kallosh, N. Sivanandam, M. Soroush, The non-BPS black hole attractor equation, JHEP 0603 (2006) 060, hepth/0602005.
B. Chandrasekhar, S. Parvizi, A. Tavanfar, H. Yavartanoo, Non-supersymmetric attractors in R2 gravities,
JHEP 0608 (2006) 004, hep-th/0602022.
J.P. Hsu, A. Maloney, A. Tomasiello, Black hole attractors and pure spinors, JHEP 0609 (2006) 048, hep-th/
0602142.
S. Bellucci, S. Ferrara, A. Marrani, On some properties of the attractor equations, Phys. Lett. B 635 (2006) 172,
hep-th/0602161.
S. Bellucci, S. Ferrara, A. Marrani, Supersymmetric Mechanics. Vol. 2: The Attractor Mechanism and SpaceTime
Singularities, Lecture Notes in Physics, vol. 701, Springer-Verlag, Heidelberg, 2006.
S. Ferrara, R. Kallosh, On N = 8 attractors, Phys. Rev. D 73 (2006) 125005, hep-th/0603247.
M. Alishahiha, H. Ebrahim, New attractor, entropy function and black hole partition function, JHEP 0611 (2006)
017, hep-th/0605279.
S. Bellucci, S. Ferrara, M. Gnaydin, A. Marrani, Charge orbits of symmetric special geometries and attractors, Int.
J. Mod. Phys. A 21 (2006) 5043, hep-th/0606209.
D. Astefanesei, K. Goldstein, R.P. Jena, A. Sen, S.P. Trivedi, Rotating attractors, JHEP 0610 (2006) 058, hep-th/
0606244.
87
[23] R. Kallosh, N. Sivanandam, M. Soroush, Exact attractive non-BPS STU black holes, Phys. Rev. D 74 (2006) 065008,
hep-th/0606263.
[24] P. Kaura, A. Misra, On the existence of non-supersymmetric black hole attractors for two-parameter CalabiYaus
and attractor equations, hep-th/0607132.
[25] G.L. Cardoso, V. Grass, D. Lst, J. Perz, Extremal non-BPS black holes and entropy extremization, JHEP 0609
(2006) 078, hep-th/0607202.
[26] S. Bellucci, S. Ferrara, A. Marrani, A. Yeranyan, Mirror Fermat CalabiYau threefolds and LandauGinzburg black
hole attractors, hep-th/0608091.
[27] G.L. Cardoso, B. de Wit, S. Mahapatra, Black hole entropy functions and attractor equations, hep-th/0612225.
[28] R. DAuria, S. Ferrara, M. Trigiante, Critical points of the black-hole potential for homogeneous special geometries,
hep-th/0701090.
[29] S. Bellucci, S. Ferrara, A. Marrani, Attractor horizon geometries of extremal black holes, in: Contribution to the
Proceedings of the XVII SIGRAV Conference, Turin, Italy, 47 September 2006, hep-th/0702019.
[30] L. Andrianopoli, R. DAuria, S. Ferrara, M. Trigiante, Black hole attractors in N = 1 supergravity, hep-th/0703178.
[31] K. Saraikin, C. Vafa, Non-supersymmetric black holes and topological strings, hep-th/0703214.
[32] B. de Wit, F. Vanderseypen, A. Van Proeyen, Symmetry structures of special geometries, Nucl. Phys. B 400 (1993)
463, hep-th/9210068.
[33] S. Ferrara, M. Gnaydin, Orbits of exceptional groups, duality and bps states in string theory, Int. J. Mod. Phys.
A 13 (1998) 2075, hep-th/9708025.
[34] L. Andrianopoli, R. DAuria, S. Ferrara, M. Trigiante, Extremal black holes in supergravity, in: M. Gasperini,
J. Maharana (Eds.), String Theory and Fundamental Interactions, in: Lecture Notes in Physics, Springer-Verlag,
Berlin, 2007, hep-th/0611345.
[35] L. Andrianopoli, R. DAuria, S. Ferrara, P. Fr, M. Trigiante, E7(7) duality, BPS black hole evolution and fixed
scalars, Nucl. Phys. B 509 (1998) 463, hep-th/9707087.
[36] L. Andrianopoli, R. DAuria, S. Ferrara, U -invariants, black-hole entropy and fixed scalars, Phys. Lett. B 403
(1997) 12, hep-th/9703156.
[37] L. Andrianopoli, R. DAuria, S. Ferrara, Supersymmetry reduction of N -extended supergravities in four dimensions, JHEP 0203 (2002) 025, hep-th/0110277.
[38] S. Cecotti, S. Ferrara, L. Girardello, Geometry of type II superstrings and the moduli of superconformal field theories, Int. J. Mod. Phys. A 4 (1989) 2475.
[39] S. Ferrara, S. Sabharwal, Quaternionic manifolds for type II superstring vacua of CalabiYau spaces, Nucl. Phys.
B 332 (1990) 317.
[40] J.F. Luciani, Coupling of O(2) supergravity with several vector multiplets, Nucl. Phys. B 132 (1978) 325.
[41] L. Andrianopoli, R. DAuria, S. Ferrara, Flat symplectic bundles of N -extended supergravities, central charges and
black-hole entropy, Lectures given at the 5th Winter School on Mathematical Physics held at the Asia Pacific Center
for Theoretical Physics, Seul (Korea), February 1997, hep-th/9707203.
[42] E. Cremmer, B. Julia, The SO(8) supergravity, Nucl. Phys. B 159 (1979) 141.
[43] L. Andrianopoli, R. DAuria, S. Ferrara, U-duality and central charges in various dimensions revisited, Int. J. Mod.
Phys. A 13 (1998) 431, hep-th/9612105.
[44] S. Ferrara, C.A. Savoy, B. Zumino, General massive multiplets in extended supersymmetry, Phys. Lett. B 100
(1981) 393.
[45] E. Cartan, uvres compl, Editions du Centre National de la Recherche Scientifique, Paris, 1984.
[46] S. Ferrara, J.M. Maldacena, Branes, central charges and U-duality invariant BPS conditions, Class. Quantum
Grav. 15 (1998) 749, hep-th/9706097.
[47] M.J. Duff, J.T. Liu, J. Rahmfeld, Four-dimensional string/string/string triality, Nucl. Phys. B 459 (1996) 125, hep-th/
9508094.
[48] K. Behrndt, R. Kallosh, J. Rahmfeld, M. Shmakova, W.K. Wong, STU black holes and string triality, Phys. Rev.
D 54 (1996) 6293, hep-th/9608059.
[49] M. Gnaydin, G. Sierra, P.K. Townsend, Exceptional supergravity theories and the magic square, Phys. Lett. B 133
(1983) 72.
[50] M. Gnaydin, G. Sierra, P.K. Townsend, The geometry of N = 2 MaxwellEinstein supergravity and Jordan algebras, Nucl. Phys. B 242 (1984) 244.
[51] M. Gnaydin, G. Sierra, P.K. Townsend, Gauging the D = 5 MaxwellEinstein supergravity theories: More on
Jordan algebras, Nucl. Phys. B 253 (1985) 573.
[52] M. Gnaydin, G. Sierra, P.K. Townsend, More on d = 5 MaxwellEinstein supergravity: Symmetric space and
kinks, Class. Quantum Grav. 3 (1986) 763.
88
[53] P. Jordan, J. Von Neumann, E. Wigner, On an algebraic generalization of the quantum mechanical formalism, Ann.
Math. 35 (1934) 29.
[54] N. Jacobson, Ann. Math. Soc. Coll. Publ. 39 (1968).
[55] M. Gnaydin, Exceptional realizations of Lorentz group: Supersymmetries and leptons, Nuovo Cimento A 29
(1975) 467.
[56] M. Gnaydin, C. Piron, H. Ruegg, Moufang plane and octonionic quantum mechanics, Commun. Math. Phys. 61
(1978) 69.
[57] H. Freudenthal, Proc. Konink. Ned. Akad. Wetenschap A 62 (1959) 447.
[58] B.A. Rozenfeld, Dokl. Akad. Nauk SSSR 106 (1956) 600;
J. Tits, Mem. Acad. Roy. Belg. Sci. 29 (1955), fasc. 3.
[59] S. Helgason, Differential Geometry, Lie Groups and Symmetric Spaces, Academic Press, New York, 1978.
[60] R. Gilmore, Lie Groups, Lie Algebras, and Some of Their Applications, Dover Publications, 2006.
[61] M. Rios, Jordan algebras and extremal black holes, based on talk given at the 26th International Colloquium on
Group Theoretical Methods in Physics, hep-th/0703238.
[62] S. Ferrara, E.G. Gimon, R. Kallosh, Magic supergravities, N = 8 and black hole composites, Phys. Rev. D 74
(2006) 125018, hep-th/0606211.
Nuclear Physics B 788 [FS] (2008) 89119
Perturbative and nonperturbative correspondences

between compact and noncompact sigma-models
M. Niedermaier a,b,,1 , E. Seiler a,b , P. Weisz a,b
a Laboratoire de Mathmatiques et Physique Thorique, CNRS/UMR 6083, Universit de Tours,
Parc de Grandmont, 37200 Tours, France

b Max-Planck-Institut fr Physik, Fhringer Ring 6, 80805 Mnchen, Germany
Received 23 April 2007; accepted 31 May 2007

Available online 8 June 2007
Abstract
Compact (ferro- and antiferromagnetic) sigma-models and noncompact (hyperbolic) sigma-models are
compared in a lattice formulation in dimensions d 2. While the ferro- and antiferromagnetic models are
essentially equivalent, the qualitative difference to the noncompact models is highlighted. The perturbative
and the large N expansions are studied in both types of models and are argued to be asymptotic expansions
on a finite lattice. An exact correspondence between the expansion coefficients of the compact and the
noncompact models is established, for both expansions, valid to all orders on a finite lattice. The perturbative
one involves flipping the sign of the coupling and remains valid in the termwise infinite volume limit. The
large N correspondence concerns the functional dependence on the free propagator and holds directly only
in finite volume.
1. Introduction
Nonlinear sigma-models with maximally symmetric Riemannian target spaces naturally come
in dual pairs, one compact and the other noncompact. The generalized classical spin systems
associated with both the compact and the noncompact target spaces have a variety of applications,
see e.g. [13] for the less familiar noncompact models. Since the compact models are much better
* Corresponding author at: Laboratoire de Mathmatiques et Physique Thorique, CNRS/UMR 6083, Universit de
Tours, Parc de Grandmont, 37200 Tours, France.

E-mail address: max@phys.univ-tours.fr (M. Niedermaier).
1 Membre du CNRS.
doi:10.1016/j.nuclphysb.2007.05.028
90
M. Niedermaier et al. / Nuclear Physics B 788 [FS] (2008) 89119
understood and the target manifolds can be related to those of the noncompact models by analytic
continuation, it is natural to try to relate also the quantum (or statistical mechanics) properties of
the compact to that of the noncompact models. For definiteness we consider here as target-spaces
the N -dimensional sphere SN and its dual the hyperboloid HN , however the qualitative aspects
should be the same for other dual pairs.
The goal of this note is to present two examples of such relations for invariant correlation
functions. The first one concerns perturbation theory, the other the large N expansion. Provided
one has chosen a formulation where the expansions are valid asymptotic expansions, it turns out
that knowledge of the expansion coefficients in the compact model allows one to infer those in
the noncompact model. However it is crucial that the asymptotic expansions are known to exist
beforehand in both systems independently. The lattice formulation is especially suited to address
this and we shall see that on a finite lattice the relevant asymptotic expansions do exist; for the
perturbative one in Section 2 and for the large N expansion in a separate paper [4]. The perturbative correspondence simply involves a sign flip of the coefficients and in a formal expansion
(dimensional regularization and minimal subtraction) has been noted to low orders in [5,6] and
in the literature on Riemannian sigma models. Here we present a proof of the correspondence
to all loop orders, initially on a finite lattice; but the termwise infinite volume limit should exist
on account of the expected lattice counterpart of Davids theorem [7,8]. The perturbatively defined correlation functions, viewed as functions of the lattice points and of 1/, are thus likewise
related simply by flipping the sign of .
(s)
The large N correspondence is more subtle. In brief the sth order large N coefficient W,r
of
the invariant 2r-point functions on a finite lattice can in the compact (+) and in the noncompact
(s)
() model be expressed in terms of a single functional Xr [D]() of the leading order invariant
two-point function Dthe same functional for both systems. This large N correspondence is
computationally useful because the computations in the compact model, which do not require
gauge fixing, are much simpler. The detour over finite volume cannot be avoided as the correspondence is difficult to interpret directly in the infinite volume limit, see [9]. In addition is
important to appreciate that although the gap equations are related by flipping the sign of the
large N coupling = (N + 1)/, the leading order propagator D in the noncompact model
behaves very differently as a function of the lattice distance than D+ : while the latter shows
exponential decay in the thermodynamic limit, the former decreases only with a power law (for
d > 2) or increases logarithmically with the distance (for d = 2).
In a finite volume the gap equations for the compact as well as the noncompact model have the
same O(V ) number of solutions for the (mass)2 parameter; but whereas in the compact model
there is exactly one positive solution, which is the physical one, in addition to multiple negative
ones, the sign-flipped gap equation has only negative solutions of which only one is the right one
for defining D . The fundamental criterion by which the relevant solution is selected should be
the stability of the corresponding saddle point, but we find also an alternative characterization
selecting the physically relevant saddle point of the noncompact model. In the thermodynamic
limit, in both cases all the negative solutions of the gap equation disappear, therefore the large N
expanded correlation functions, viewed as functions of the lattice points, are not related in any
simple way, in particular not merely by flipping the sign of .
The derivations of the above results also highlight why one cannot expect useful correspondences to exist beyond asymptotic expansions. For example flipping the sign of the coupling in
the (exact) generating functionals for invariant correlation functions maps the ferromagnet onto
the compact antiferromagnet. In contrast, in the perturbative asymptotic expansion the sign flip
rather relates the ferromagnet to the noncompact model, while in the large N expansion the func-
91
tional dependence on the independently defined free propagators gets related. In the latter case
the relation to the compact model cannot be formulated directly in the thermodynamic limit but
one can take term by term the thermodynamic limit on both sides of the correspondence. This
somewhat tricky limit is studied in an accompanying paper [9] in the two-dimensional systems
for a number of physically interesting invariant quantities. Ultimately we expect the fundamental differences between the compact and the noncompact models to be rooted in two facts: the
presence of a infinite volume mass gap and the absence of long range order (for d 2) in the
compact models and the opposite characteristics in the noncompact models for all d 1.
The rest of the paper is organized as follows: In Section 2 we introduce generating functionals
for the invariant correlators in the three systems considered and establish the perturbative correspondence. In Section 3 dual formulations of the generating functions are introduced, suited for
the large N expansions. The large N correspondence is shown first from the generating functionals, then from the SchwingerDyson equations, and finally verified at low orders.
2. Invariant correlators and PT correspondence
Invariant correlators in the noncompact model have to be defined in terms of a gauge fixed
generating functional. A gauge fixing where one spin is kept fixed turns out to be advantageous.
To discuss the relation to the compact model we adopt the same gauge fixing there. This spin
model formulation is also convenient to discuss the relation between the exact generating functionals and for the proof of the perturbative (PT) correspondence.
2.1. Definitions
We begin by setting up the notation and the definitions for the invariant correlation functions considered and their generating functionals. We consider the O(N + 1) spherical and
the SO(1, N) hyperbolic sigma-models with standard lattice action, defined on a hypercubic
lattice Zd of volume V = || = Ld . The dynamical variables (spins) will be denoted
by nax , x , a = 0, . . . , N , in both cases, and periodic boundary conditions are assumed
throughout nx+L = nx . The constraint is n n = 1 in both cases, but with different dot
products; namely a b := a 0 b0 + a 1 b1 + + a N bN =: a c cd bd in the compact model, and
a b := a 0 b0 a 1 b1 a N bN =: a c cd bd in the noncompact model. We shall also use the

notation a = (a 1 , . . . , a N ) for vectors in RN , so that the bilinear forms read a b = a 0 b0 a b,
N
N+1
N
1,N
| n n = 1} is the N -sphere and H = {n R
|
in the two cases. Clearly S = {n R
n n = 1, n0 > 0} is the upper half of the two-sheeted N -dimensional hyperboloid. The invariance groups are O(N + 1) and SO0 (1, N), respectively.
Let us briefly note how SN and HN are related by symmetric space duality (see for instance
[10,11]). Recall that a symmetric space G/K has an involution associated with it such that
the Lie algebra g of G decomposes according to g = k m as a direct sum of vector spaces,
where k and m are even and odd under , respectively. Furthermore [k, k] k, [k, m] m and
[m, m] k. The dual Lie algebra g is then defined as g := k im and the corresponding dual
group G is the (simply connected) group whose Lie algebra is g . G contains a connected
subgroup K having k as its Lie algebra; the dual symmetric space (G/K) can then be defined
On g we have an invariant bilinear form B (in our case simply minus the Killing form)
as G /K.
which induces a dual bilinear form B on g . The bilinear forms B and B , restricted to m
and im, respectively, are positive definite in our case and define the metric of the tangent spaces
at the origins of the symmetric spaces G/K and (G/K) , respectively. By requiring invariance
92
under G or G , respectively, Riemannian metrics on both symmetric spaces are induced. For G =
0 (1, N) (the universal covering group of SO0 (1, N))
SO(N +1), K = SO(N ), this gives G = SO

and (G/K) = SO0 (1, N)/SO0 (N ) = SO0 (1, N)/SO0 (N ) or G/K = SN and (G/K) = HN .
The lattice actions for the two systems are
S =

(nx nx+ 1) =
nx (n)x ,
2 x
x,
(2.1)
where the upper sign refers to the compact model and the lower sign to the noncompact model. In
the compact model, with the conventions adopted in (2.3a), > 0 models ferromagnetic behavior
while < 0 models antiferromagnetic behavior. In the noncompact model only >
0 is allowed
and the action is unbounded from above, 0 S [n] < . The Laplacian is xy = [2x,y
x,y+ x,y ], as usual. We write
d+ (n) = dnN+1 (n n 1),

d (n) = 2dnN+1 (n n 1) n0 ,
(2.2)
for the invariant measure on SN and HN , respectively. Further (n, n ) is the invariant point
measure on SN , HN , and n = (1, 0, . . . , 0). Note that the measure d+ (n) is normalized while
HN has infinite volume.
In the compact model we consider two generating functionals, the usual one and a variant
with one spin frozen:

1
exp W [H ] = N
(2.3a)
d+ (nx ) exp S+ +
Hxy (nx ny 1) ,
2 x,y
x

1
d+ (nx ) + nx0 , n exp S+ +

Hxy (nx ny 1) ,
exp W+ [H ] = N+
2 x,y
x
(2.3b)
where Hxy 0 is a source field and the normalizations N , N+ are such that W [0] = 0. More
generally Wn [H ] will denote the generating functional with
the spin at site x0 fixed by a (nx0 , n)
insertion. In this notation one has W+ [H ] = Wn [H ] and d+ (n) exp Wn [H ] = exp W [H ]. In
fact
Wn [H ] = W [H ],
(2.4)
for all n SN . This can seen by performing a global rotation nx gnx =: n x , x , with
g O(N + 1) chosen such that gnx0 = n , say. The Boltzmann factor will then depend on n x ,
x = x0 , only and the d(n x0 ) = d(nx0 ) integration can be performed to give 1.
In the noncompact model only the fixed spin variant of the generating functional is well defined and we write

exp W [H ] = N
d (nx ) nx0 , n
x

1
Hxy (nx ny 1) ,
exp S +
2 x,y
where now Hxy < 0 sources give damping exponentials.
(2.5)
93
Partially connected 2r point functions are defined by

W [H ] =
1
W,r (x1 , y1 ; . . . ; xr , yr )Hx1 y1 Hxr yr ,
r!2r
r1
W,r (x1 , y1 ; . . . ; xr , yr ) := hx1 y1 hxr yr W [H ]|H =0 , hxy :=
.
Hxy
(2.6)
In particular W,1 (x, y) = nx ny 1, W,2 (x1 , y1 ; x2 , y2 ) := nx1 ny1 nx2 ny2 nx1 ny1
nx2 ny2 , where are the functional averages with respect to N1 eS . Note that
W,r (. . . ; x, x; . . .) = 0.
In the above we tacitly assumed that W [H ] and the correlation functions computed from it
do not depend on the site x0 of the frozen spin and are translation invariant. We show now that
this indeed the case. If we momentarily indicate the dependence on the site as Wx0 (and drop the
subscripts) one has trivially
Wx0 [a H ] = Wx0 +a [H ],
(a H )xy = Hx+a,y+a .
(2.7)
Thus, if Wx0 is independent of x0 it is also translation invariant. The Boltzmann factors (2.3b) and
(2.5) can be viewed as a function on the group via F (g0 , . . . , gs ) = f (g0 n , . . . , gs n ), where
we picked some ordering of the sites xi , i = 0, 1, . . . , s := V 1, identified nxi with gi n , and
wrote momentarily f for the Boltzmann factor. Then Wxi is of the form

d(nj ) ni , n f (n0 , . . . , ns ) = const Ui ,
j
Ui :=

dgj F (g0 , . . . , gi1 , e, gi+1 , . . . , gs ).
(2.8)
j =i
Using the invariance of F under gi h1 gi and the unimodularity and invariance of the measure
dg one verifies that Ui = U0 for all i.
A peculiarity of the fixed spin gauge is that an invariant 2r-spin correlator nx1 ny1 nx2
ny2 nxr nyr , can be re-interpreted as a noninvariant (2r 1)-spin correlator. This is manifest
when xi = x0 for some i; since nx0 = n the correlator then involves only 2r 1 fluctuating
spins. By the above argument any one site can play the role of x0 , so that picking such a reinterpretation amounts to picking a site where nxi = n . This feature holds both in the compact
and in the noncompact models, but initially only on a finite lattice. When the site x0 is kept
fixed in the interior of the lattice, these specific noninvariant 2r 1 point functions will have
a pointwise thermodynamic limit in both cases. However the fixed spin averages then do not
approach limits which can be interpreted in terms of averages without gauge fixing or with a
translation invariant gauge fixing. This changes if x0 is identified with a point on the boundary
and moves out to infinity as Zd . In this case, in dimensions d 2 an important difference
between the compact and noncompact models emerges: while in the compact model the Mermin
Wagner theorem assures that the noninvariant correlator is equal to its invariant average over
O(N + 1), in the noncompact model the non-amenability of SO(1, N) prevents this, and we find
either spontaneous symmetry breaking or divergence of the correlators, as discussed in [1].
The goal in the following will be to relate the perturbative and the large N expansions in the
compact and the noncompact models. For the perturbative expansions of the correlation func-
94
tions, we write
W,r r
n w,r .
(n)
(2.9)
n0
(s)
In a large N expansion := (N + 1)/ is kept fixed and the coefficient functions W,r in
W,r

1
r
(s)
W ,
(N + 1)s ,r
(N + 1)r1
(2.10)
s0
are sought.
2.2. Compact antiferromagnet and noncompact model
In the literature one finds statements, based on formal manipulations of functional integrals,
suggesting that compact and noncompact models are essentially related just by flipping the sign
of (in [2], however, the qualifying remark is added that such a relation is less clear at the
nonperturbative level). Since this simple flip rule is not completely correct and the relations
between the compact model, its analytic continuation to < 0 and the noncompact model are
somewhat subtle, let us try to clarify the situation.
The compact model defined in Eq. (2.1) describes for > 0 a ferromagnet, for < 0 an
antiferromagnet, so flipping the sign of turns a ferromagnet into an antiferromagnet, not the
noncompact model. It turns out that one can nevertheless relate the coefficients of the perturbation
expansion and the large N expansion of the compact ferromagnetic and the noncompact model
by a sign flip prescription, which does, however, not imply any simple relation between the
antiferromagnet and the noncompact model.
On the simple hypercubic lattices we are considering here, there is a transformation between
the cases > 0 and < 0, based on the fact that we can decompose into an even and an odd
sub-lattice
= + ,
where x + if and only if
x1 ++xd
x := (1)
d
i=1 xi
(2.11)
is even, otherwise x . Defining
(2.12)
we can define a map of the configurations by

nx (n x ) = x nx ,
(2.13)
mapping each configuration of the ferromagnet into one of the antiferromagnet with the same
(correctly normalized) Boltzmann weight. This yields the following identity relating the generating functionals of the antiferromagnetic and the ferromagnetic systems:
1

W+, [H ] = W+, H +
(2.14)
Hxy Hxy + const,
2 xy
:= H . To avoid
where we momentarily indicated the dependence on > 0 and set Hxy
x xy y
confusion, in the following will always be assumed to be positive. For the invariant two-point
function Eq. (2.14) gives
nx ny | = x y nx ny | ,
(2.15)
95
and similarly for the higher correlation functions. This implies that the physics of the antiferromagnet is essentially the same as for the ferromagnet, in particular they have the same mass gap,
phase structure etc.
The relation between the ferromagnetic O(N + 1) and the noncompact SO(1, N) systems
(the latter only exists in the ferromagnetic version) is, however, more subtle: starting from the
O(N + 1) model we re-parameterize the variables nx SN by introducing =: g 1 and nx =:
g x , so that

nx ny = n0x n0y + g x y = x y 1 g x2 1 g y2 + g x y ,
(2.16)
where the x are variables taking the values 1. To obtain W+ [H ] in this parameterization one
has to (i) integrate over all x subject to the constraint x2 < and (ii) sum over all x = 1
except x0 , which is 1. Explicitly

1/2

1,x0
d x (x0 ) || x2 1 g x2
exp W+ [H ] = N +
{x }

2

1
2
2
2
x 1 g x x+ 1 g x+ + g(x x+ )
exp
2g x,

1
exp
(2.17)
Hxy x y 1 g x2 1 g y2 + g x y 1 .
2 x,y
For + (ferromagnet) the term in which all x are equal dominates (for (antiferromagnet) instead the terms with alternating x dominate).
On the other hand, we may re-parameterize the noncompact model in a similar way by intro
ducing nx = g x , but this time

nx ny = n0x n0y g x y = 1 + g x2 1 + g y2 g x y .
(2.18)
Inserting this into the expression for W [H ] given in (2.3b), we now obtain

d x (x0 )(1 + g x0 )1/2
exp W [H ] = N
x

2

1
exp
1 + g x2 1 + g x+ g(x x+ )2
2g x,

1
2
2
exp
(2.19)
Hxy 1 + g x 1 + g y g x y 1 .
2 x,y
Comparing now Eqs. (2.17) and (2.19), we see that the partition function of the noncompact
model is obtained from the one of the compact one by (i) flipping the sign of , (ii) dropping all
terms except the one with all x = 1 and (iii) omitting the functions restricting the domain of
integration.
So it is clear that the analytic continuation of the compact system to negative (the antiferromagnet) is not equivalent to the noncompact SO(1, N) system, as one might infer from a
cursory reading of the literature (for instance [2]). This non-equivalence will be shown even more
manifestly below. Within the framework of the SchwingerDyson equations used in Section 3.5
96
a noteworthy consequence is that even the exact SD equations do not determine their solution
uniquely.
In the following sections we will see that nevertheless the perturbation expansions in a finite
volume of the ferromagnetic compact and the noncompact model are related simply by flipping
the sign of (but without the introduction of the sign factor x as in the antiferromagnet); furthermore we will derive a relation between the 1/N expansions of the two models.
2.3. Perturbative correspondence
We now consider the perturbative expansions (2.9) of the correlation functions. On a finite
lattice it turns out the perturbative series in the compact ferromagnetic and in the noncompact
models are asymptotic expansions which are simply related by a sign flip in the coupling,
; the perturbation expansion of the noncompact model is, however, not equal to that of the
antiferromagnet, as one might guess from this.
Concretely
(n)
(n)
w,r = ()r+n w+,r ,
r 1, n 0,
(2.20)
for the coefficients (2.9). The sign flip rule was stated explicitly by Hikami [6] (in a formal
continuum expansion based on dimensional regularization) but was presumably known to other
authors in special cases and at low orders, e.g. [5]. It has also been observed in the perturbative beta functional of Riemannian sigma models (using dimensional regularization and minimal
subtraction) and can readily be verified to all loop orders in this framework. But to the best of
our knowledge no general proof is available in print, so we present a simple proof here.
We set Z [H ] := exp W [H ] and define

1/2

PT
[H ]() :=
d x (x0 ) x2 1 1 x2
Z+
x

2

1
2
2
exp
1 1 x2 1 1 x+
+

)
x
x+
2 x,

1
exp
Hxy 1 1 x2 1 1 y2 + 1 x y 1 , (2.21)
2 x,y
and
PT
[H ]() :=
Z

1/2

d x (x0 ) x2 1 + 1 x2

2

1
2
2
1
2
1
exp
1 + x 1 + x+ (x x+ )
2 x,

1
exp
Hxy 1 + 1 x2 1 + 1 y2 1 x y 1 . (2.22)
2 x,y
PT [H ] is, up to a multiplicative constant, just the term of Z [H ] (Eq. (2.17)) in which = 1,
Z+
+
x
PT [H ] differs from Z [H ] in Eq. (2.19) by the presence of the functions and
for all x. Z
a multiplicative constant. We now state the
97
Result.

PT
[H ]() + O ,
Z [H ] = c ()Z
(2.23)
PT [H ]() have aswhere c () are normalization constants independent of H . Further the Z

ymptotic expansions of the form
PT
[H ]()
Z
an [H ]()n ,
(2.24)
n=0
with identical coefficients an [H ] in both cases.

The existence of the asymptotic expansions (2.24) is a standard result of Laplaces method
in asymptotic analysis (see for instance [12,13]). The relation (2.23) expresses the irrelevance of
boundaries away from the maximum as well as of the terms in which the x are not all equal. It
follows from two facts:
Fact 1. On a finite lattice in the compact ferromagnetic model all terms in Z+ [H ] except the one
with x = 1 are O( ).
Fact 2. On a finite lattice in the noncompact model for any x0 and 0 > 0 the expectation value

2
1
d (nx ) nx0 , n n2z 1
nz 1 :=
Z [H ]
x

1
Hxy (nx ny 1) ,
exp S +
(2.25)
2 x,y
is bounded by a exp(b) with a, b > 0.
Corollary. There are constants a , b > 0 such that for > 0 ,
PT

V 1
Z [H ]() N 2

1 a exp(b ).

Z [H ]()
(2.26)
Both facts have essentially the same origin: the fact that the energy in the omitted contributions is exponentially small. We give a detailed proof only of Fact 2 and add some comments on
the proof of Fact 1.
Proof of Fact 2. We do the integrations in the following way: first we integrate over all configurations having a fixed value s of the action and afterwards over s. Thus we can write

Z [H ]() = ds es H (s),
(2.27)
with some non-negative density function H . If we then can show that for any configuration
violating the field cutoff the action is larger than s0 (energy bound), it will follow that

2

Z [H ]() nx 1 ds es H (s) (s s0 ).
(2.28)
98
The right-hand side is easily seen to be exp( 2 s0 ) Z [H ](/2). By standard asymptotic

analysis (see for instance [12,13]) it follows that Z [H ]()/Z [H ](/2) goes to a finite limit
as , and therefore
Z [H ](/2)
a,
Z [H ]()
(2.29)
for 0 . So Eq. (2.28) implies Fact 2.

It remains to prove the energy bound. Since we are usingthe fixed spin gauge, we have
nx0 = n . Then by assumption there is a site x1 with n0x1 2. Choose a path P of length
|P | V from x0 to x1 and denote the part of the action corresponding to the path P by sP . We
claim that sP is minimized by moving in equal steps along the geodesic from nx0 = n to nx1 ,
which implies
s sP
c
c
,
|P | V
(2.30)
with some constant c.

To see the lower bound on sP it is sufficient to consider a path of three points, since the
minimization condition is local. Thus we only have to minimize e(n) := n0 n + n n1 over n
and show that the minimum is assumed for n being the midpoint of the geodesic in HN from n0
to n1 . Choosing coordinates such that n0 = n and n1 = (ch 1 , sh 1 , 0, . . . , 0) with 1 0 we
find

2
e(n) = n0 (1 + ch 1 ) n1 sh 1 n0 (1 + ch 1 ) n0 1 sh 1 .
(2.31)
Putting n0 = ch this becomes
1
,
2
using the convexity of ch function in the last step. It follows that
e(n) ch + ch(1 ) 2 ch
(2.32)

2
c
sP |P | ch 1 /|P | 1 1 ,
(2.33)
2|P | V
with c = (arccosh 2 )2 /2. The proof of Fact 2 is now complete with a as above and b =
c/V . 2
To show Fact 1 one proceeds essentially in the same way; only the energy bound is more
easily derived in a slightly different way: again it suffices to consider three spins n0 , n, n1 with
the energy
e(n) = n n0 + n n1 ,
(2.34)
direct minimization over n yields

n = (n0 + n1 ),
(2.35)
with fixed by the requirement n2 = 1. So n has to lie on the great circle connecting n0 and n1 .
As before we conclude from this that the minimal energy for a path is obtained by moving along
the shorter part of a great circle (geodesic) connecting n0 to n1 in equal steps.
The corollary as well as the main result are obvious consequences.
99
Thus we have learned that the perturbative expansions of the compact ferromagnetic model
and the noncompact model are related just by flipping the sign of , even though doing the same
sign flip to the full model leads to a completely different modelthe antiferromagnet.
The relation between the asymptotic expansions in 1/ for the compact and noncompact
models is given in (2.20). For the two point function this means for instance
(n)
(n)
w,1 = (1)n+1 w+,1 .
(2.36)
This should be contrasted with the relations between the perturbation expansions of the two-point
function between the ferromagnetic and antiferromagnetic compact models, which follows from
(2.13)
(n)
(n)
(x, y)|antiferro = x y w,1
(x, y)|ferro .
w,1
(2.37)
It is instructive to see what this means for the lowest order (tree graph) asymptotics of the two
point function:

1 (0)
nx ny = 1 + w+,1 (x, y) + O 2 ferromagnet,

1 (0)
nx ny = 1 + w+,1 (x, y) x y + O 2 antiferromagnet,

1 (0)
nx ny = 1 w+,1 (x, y) + O 2
noncompact model,
(2.38)
where
(0)
w+,1
(x, y) =
N cos p (x y) 1
,
V
Ep
(2.39)
p =0
with Ep as in Eq. (3.22) below.

Note that although W+ [H ] for the ferromagnet and the antiferromagnet are related by flipping
the sign of , this is no longer true for the asymptotic expansions, even in finite volume. Rather
the sign flipped perturbative expansion is that of the noncompact model.
3. Dual formulations and large N correspondence
The goal of this section is to establish a relation between the large N expansions of invariant
correlators in the compact and the noncompact sigma-models. In the compact model the large
N expansion is based on the familiar dual formulation, which here has to be modified due to the
fixed spin. In the noncompact model the duality transformation is ill defined but by performing
Gaussian integrations in horospherical coordinates, one can obtain a well-defined counterpart of
the dual formulation.
3.1. Noncompact generating functional via horospherical coordinates
Before turning to the expansions proper we present here an exact rewriting of W [H ] based
on a partial evaluation where the (dim HN )V integrations are reduced to V integrations. The point
of departure is the fact that the hyperboloid HN admits an alternative parameterization in terms
of so-called horospherical coordinates. These arise naturally from the Iwasawa decomposition of
SO0 (1, N). Here it suffices to note the relation to the hyperbolic spins
100
n0 = ch +
N1
1 2
ti e ,
2
n1 = sh +
i=1
n =e
i
N1
1 2
ti e ,
2
i=1
i = 2, . . . , N,
ti1 ,
(3.1)
and that (, t1 , . . . , tN1 ) RN defines a globally valid system of coordinates. It is convenient

to write t = (t1 , . . . , tN1 ) and t t = t1 t1 + tN1 tN1
. For the dot product of two spins
nx , ny HN this gives
1
nx ny = ch(x y ) + (tx ty )2 ex y ,
2
and for the measure in (2.2)
d (n) = e(N1) d dt1 dtN1 = e(N1) d d t.
(3.2)
(3.3)
The key advantage of horospherical coordinates is manifest from (3.1), (3.2): for a quadratic
action of the form S in (2.1) the integrations over the t variables are Gaussian and can be
performed without approximations. We refer to [4] for the derivation and only note the result

1
exp W [H ] = exp
Hxy N
dax
2 x,y
D(H ) x =x0
N +1
1 1

exp
ax + A x x .
Tr ln A
0 0
2
2
2
(3.4)
x =x0
Here
Axy = xy +
1
Hxy + xy ax = A xy + ax0 xx0 xy ,
(3.5)
and A xy is the matrix obtained from Axy by omitting the x0 th row and column. The domain
D(H ) is an algebraic variety described by

D(H ) = a [2d, ]V 1 | det A > 0, A positive semidefinite .
(3.6)
We also anticipate from [4] the following
Result. The correlation functions W,r admit an asymptotic expansion of the form (2.10), whose
(uniquely defined) expansion coefficients coincide with those defined by the Laplace expansion
of (3.4) where D(H ) has been replaced by RV 1 . In turn these coefficients coincide with those
of the formal large N expansion of

1
exp Wd [H ] = exp
Hxy N
dx exp (N + 1)S [, H ] ,
2 x,y
S [, H ] =
1
Tr ln A + i
2
x =x0
x =x0
1 1 1
A xx ,
0 0
2
(3.7)
where ax corresponds to 2ix . However

W [H ] = Wd [H ].
(3.8)
101
Heuristically the origin of the dual generating functional Wd [H ] can be understood by a dualization of the spatial spin components nx , x , and a formal contour deformation in field
space [9]. The formal nature of the latter is responsible for (3.8), although the large N expansion
coefficients of W [H ] are correctly reproduced.
3.2. Dual formulation: compact model
In the compact model the counterpart of (3.7) is obtained along the familiar lines: one first
implements the constraints nx nx = 1 via a Lagrange multiplier field and then performs the
Gaussian integrations. Of course in the compact model no gauge-fixing is required and the result
is well known. For the purposes of comparing with the noncompact model, however, we want
to perform the dualization here in the fixed spin gauge, i.e. for W+ [H ], in which case certain
modifications occur. The resulting dual generating functional W+d [H ] is an exact rewriting of the
original one: W [H ] = W+ [H ] = W+d [H ].
In the fixed spin gauge Gaussian integrals of the following form arise

1
dx (x0 ) exp
x Axy y +
Jx x
2 x,y
x
x

1
V 1
1
1/2
2
(det A)
exp
Jx A xy Jy ,
= (2)
(3.9a)
2 x,y

x20
V 1
1
1/2
2
dx exp
x Axy y = (2)
(det A)
exp
, (3.9b)
2 x,y
2(A1 )x0 x0
x =x0
for a real field x , x , and a symmetric invertible Ld Ld matrix A, for which we assume
that (A1 )x0 x0 = 0. Here A is the (Ld 1) (Ld 1) matrix arising from A by deleting its x0 th
row and column. The inverse of A can be expressed in terms of the inverse of A via

A 1

xy

(A1 )xx0 (A1 )yx0
= A1 xy
.
(A1 )x0 x0
(3.10)
This equation makes sense a priori for x, y = x0 , but we may trivially extend the matrix A 1
to an Ld Ld matrix by setting ([A 1 ]ext )x0 x = 0 in accordance with Eq. (3.10). Then A and
[A 1 ]ext are of course not inverse to each other

(A1 )x y
(3.11)
Axz A 1 ext zy = xy xx0 1 0 .
(A )x0 x0
z

To verify (3.9) one writes x,y x Axy y = x,y =x0 x Axy y +2x0 x =x0 Ax0 x x +x20 Ax0 x0 .
To get (3.9a) one first does the trivial x0 integration. To get (3.9b) one uses (3.11). The determinant of A is related to that of A by
det A =
det A
.
(A1 )x0 x0
(3.12)
Often a term in the x0 th matrix element on the diagonal of A has to be split off according to
Axy = A xy cxy x0 x . In this case the inverse of A is related to the inverse of A by
1

1 1

c
A xy = A 1 xy +
(3.13)
A
A
.
xx0
yx0
1 c(A 1 )x0 x0
102
1 1
In particular Ax0 x0 (A1 )1
x0 x0 = Ax0 x0 (A )x0 x0 and
1
1
= c +
,
1
1
(A )x0 x0
(A )x0 x0
(A1 )xx0
(A 1 )xx0
=
.
(A1 )x0 x0
(A 1 )x0 x0
(3.14)
With these preparations at hand the dualization of W+ [H ] is straightforward. It is instructive to

start from W [H ], single out one spin, nx0 , and to postpone its dualization. Inserting

dx
for x = x0 ,
n2x 1 = (N + 1)
(3.15)
exp i(N + 1)x n2x 1
2
the interchange in the order of integrations can be justified and the relevant Gaussian is of the
form (3.9b)

1N +1
dnx exp
nx Axy ny
2 x,y
x =x0

1 N + 1 1 1 2
(N+1)/2
= Const(det A)
exp
A x x nx0 ,
0 0
2
Hxy .
A xy = xy + 2ix (1 x0 x )xy
(3.16)
N +1
This gives for the dual generating functionals

exp W d [H ] = dnx0 n2x0 1 exp Wndx [H ],
0

1
Hxy N
dx exp (N + 1)Snx0 [, H ] ,
exp Wndx [H ] = exp
0
2 x,y
Snx0 [, H ] =
1
Tr ln A i
2

x =x0
x =x0
x +
1 2 1 1
n A xx .
0 0
2 x0
(3.17)
Wnd [H ] is the generating functional in the fixed spin gauge

Here we used the fact A = A;
x
0
and Snx0 [, H ] is its dual action. We shall also write S+ [, H ] for Sn [, H ] and W+d [H ] for
Wnd [H ]. In contrast to the noncompact model one has
W+d [H ] = W+ [H ],
W d [H ] = W [H ],
(3.18)
and also W+ [H ] = W [H ], by (2.4).

As a check on (3.17) one can verify that by dualizing also the last spin one recovers the
familiar expressions. Indeed,

2

(N + 1) 1 1 2
dnx0 nx0 1 exp
A x x nx0
0 0
2
(N1)/2

1
(N+1)/2
2
dx0 ei(N +1)x0 A 1 x x + 2ix0
.
=
(3.19)
0 0
N +1
Using now (3.13) for c = 2ix0 and (3.12) one obtains
exp W [H ] = N

x
103
N +1
dx exp
x ,
Tr ln A + i(N + 1)
2
x
Axy = xy + 2ix xy
Hxy ,
N +1
(3.20)
as required.
3.3. Finite volume mass gap and basic propagators
The saddle point expansions of Wd [H ] define to leading order the gap equations and the
invariant two-point functions. In both quantities the dependence on the site x0 of the fixed spin
drops out, see Section 3.6. The result for the leading order two-point function can be written in
the form
(0)
W,1 (x, y) = D (x y) 1 = nx ny |N= 1,
(3.21)
where D (x y) is the basic propagator. For the compact model it has the well-known structure
D+ (x) = D(x)|=+ , where
D(x) =
1 eipx
,
V p Ep +
(3.22)

with the sum over all p = 2
cos(p ).
L (n1 , . . . , nd ), ni = 0, 1, . . . , L 1, and Ep := 2d 2
Further + = + (, V ) is the dynamically generated finite volume mass2 term determined by
the gap equation D(0) = 1/. A subtlety is due to the fact that the equation for + (, V ) is an
algebraic equation of degree O(V ) and therefore has O(V ) solutions. Among those solutions
there is always a unique positive one, but it is not obvious that this is always the physically
relevant one: in dimension d 3 for weak coupling there is spontaneous symmetry breaking
and it has been found in [14,15] that at least in the determination of the constrained effective
potential one has to take a negative solution of the gap equation. Here we want to eliminate these
complications by always working in the strong coupling regime, which we will now specify.
In the infinite volume limit the gap equation becomes

1
dd p
,
1=
(3.23)
(2)d Ep +
and it has no negative solutions. It has a unique positive solution for (d), where (2) = 0
and (d) > 0 for d 3 ( (d) is known as the critical coupling of the spherical model). From
now on we will assume > (d) and always require + (, V ) > 0. In this case it is possible to
show that the finite volume corrections + (, V ) + (, ) are exponentially small.
In the noncompact model one has similarly D (x) = D(x)|= , where is a solution
of the gap equation D(0) = 1, see Eqs. (3.54), (3.56) in Section 3.6. Clearly analytically
continuing the multivalued algebraic function + (, V ) to negative values of and setting
(, V ) := + (, V ) will produce solutions of this gap equation. But again there are O(V )
such solutions, all of which are negative. To see this it is convenient to split off the zero momentum mode and to write

1
1
1
,
+ f (), f () = d
D(0) =
(3.24)
V
L
+ n
d
0 =n[0,L1]
104
where n = E2n/L , n = (n1 , . . . , nd ). Here f () is a meromorphic functions with simple poles

along the negative real axis. The first pole is at = 4 sin2 (/L), the others are at approximately
integer multiples thereof. Specifically

n = 4(/L)2 n21 + + n2d + O 1/L4 .
(3.25)
In terms of the function f the gap equations D(0) = 1 read
(3.26)
.
Ld
In the compact model (upper sign) is has a unique solution + (, L) > 0 for any given >
(d); all the other O(V ) solutions are negative. In the noncompact case there are only O(V )
negative solutions, one in each interval of the axis between consecutive poles of f .
To obtain an asymptotic expansion in 1/N one has to pick the solution corresponding to the
absolute maximum of the dual Boltzmann factor. But there is a simpler way to do this selection: if
the large N expansion is to yield an asymptotic expansion of the invariant two-point function, we
have to demand that D(x) 1 for all x. Since nx ny 1 in the noncompact model, this is
a necessary condition for the interpretation of D (x y) = D(x y)|= as the leading order
two-point function. This condition fixes the solution to lie in the interval (4 sin2 /L, 0).
Indeed, D(x) 1 implies

(1 cos q x) D(x) V , q = 0,
f (w) = 1
1
1
1
,
V 1 + /Eq
q = 0.
(3.27)
Since < 0 both factors on the left-hand side must be positive. Hence Eq + > 0 for all q = 0,
for which > 4 sin2 /L is a necessary and sufficient condition.
It is easy to sharpen this bound to
>
sin2
2d + 1
L
(3.28)
by noting that
0>
V
1
1
2d
=
+
.
2
Ep + 4 sin L +
p
(3.29)
From now on we write (, V ) (4 sin2 /L, 0) for the unique root of the D(0) = 1 gap
equation such that D (x) 1 for all x. Its large volume asymptotics comes out as
4
d = 2,
ln V (1 + O(1/ ln V )),
V (, V ) =
(3.30)
d2
( 1 + Cd )1 + O(V d ), d 3,
2 d d p 1
where Cd = 0 (2)
d Ep .
In summary, to leading order the invariant spin two-point functions in the compact ferromagnetic and in the noncompact model can be written in the form (3.21), where
D (x) := D(x)|= (,V ) ,
with (, V ) solution of D(0) = 1.
(3.31)
105
In both the compact model and the noncompact model there is a unique root of the gap equation
D(0) = 1 such that D (x) 1 for all x. From now on the symbols D (x) will always
refer to the unique propagators with this property. They satisfy
1 D (x) < 1
1 +
2
,
V
2
< D+ (x) 1.
V +
(3.32)
The dependent bounds follows from D(x) > 2/(V ) D(0). In particular 0 < < /V .
For V , D+ (x) approaches the massive free propagator of squared mass + (, ),
while D (x) becomes massless and naively appears to be ill-defined. If one inserts, however, for
the L-dependent solution of the gap equation D (0) = 1 one can rewrite D (x) in such
a way that the existence of the thermodynamic limit is manifest:
nx ny |N= = D (x y) = 1 +
1 cos k (x y)
,
V
Ek +
(3.33)
k =0
which for V has the well-defined limit

d d k 1 cos k (x y)
.
1+
(2)d
Ek
(3.34)
3.4. Large N correspondence

The results of Sections 2 and 3 now lead to the
Result. Let Wr,+ and Wr, denote the connected parts of the invariant 2r-point functions nx1
ny1 nxr nyr in the SO(N + 1) and in the SO(1, N) nonlinear sigma-model, respectively.
(With bare coupling , defined on a hypercubic lattice Zd of volume Ld with periodic
boundary conditions, and a suitable gauge fixing in the noncompact case.) Then the W,r have
well-defined large N asymptotic expansions of the form (2.10), with := (N + 1)/ fixed.
Further:
(a) The coefficient functions W,r are translation invariant. The lowest order two-point func(0)
tions have the form W,1 (x, y) = D (x, y) 1/, where D (x) = D(x)| . Here
D(x) is the free propagator of squared mass and (, V ) are the unique solutions of the
gap equations D(0) = 1 with the property D(x) 1, for all x.
(s)
(b) For all r 1, s 0, there exist unique functionals Xr [D]() of a free propagator D
(s)
(s)
(s)
such that W+,r = Xr [D+ ]() are the coefficients in the compact model and W,r =
(s)
(1)r Xr [D ]() are the coefficients in the noncompact model.
Proof. The existence of the large N asymptotic expansion has been shown in the compact
model in [16], for W [H ] (or its the counterpart with a linear source coupling). Since W+d [H ] =
W d [H ] = W [H ], this directly carries over to W+ [H ]. In the noncompact model the result has
been anticipated in Section 3.1; for the derivation we refer to the forthcoming paper [4]. The two
statements in part (a) follow from Sections 2.1 and 3.3, respectively. For (b) we use the result
stated in Section 3.1, namely that Wd [H ] (although not equivalent to W [H ] and not obtainable
106
from a well-defined duality transformation) generates the correct large N asymptotic expansion.
Given this, it suffices to note that W+d [H ] in (3.17) and Wd [H ] in (3.7) have the same structure
with actions that are related by the involution
S+ [, H ] S [, H ] for x x , .
W+d [H ]
(3.35)
Wd [H ]
Hence, if indeed the moments of

and
have asymptotic expansions of the
form (2.10), their coefficients are uniquely defined and must therefore also be related by the
involution (3.35).
For the sake of clarity let us add two remarks:
First, W+ [H ] and W [H ] do not coincide exactly, as follows from (3.8), (3.18) and the fact
that the boundary D(H ) in (3.6) is invariant under the involution (3.35). In physics terms this
amounts to the observation made before that the antiferromagnet and the noncompact model are
inequivalent.
Second, (b) does not state that the large N coefficients of the compact (ferromagnetic or
antiferromagnetic) model and the noncompact model are related by flipping the sign of . The
latter is certainly incorrect. Rather the dependence on D is same after flipping the sign of ,
but D+ and D are very different and independently defined functions of the lattice distance.
So, given the large N coefficients of W+,r , as a function of the lattice points, there would be no
simple way to construct the large N expansion of W,r . 2
3.5. SchwingerDyson equations
It is instructive to look at the previous result from the viewpoint of the SchwingerDyson
(SD) equations. Under the assumption that asymptotic expansions of W,r and W+,r are known
to exist, this also provides an alternative derivation of the result.
In the compact model the SchwingerDyson equations for the moments of W [H ] have, to our
knowledge, first been formulated by M. Lscher [17]. In the noncompact model the necessity to
gauge fix requires modifications. In the fixed spin gauge the modifications in the equations are
minimal, but they inevitably refer to the preferred point x0 . In the following we formulate the SD
equations in parallel for the gauge fixed functionals W [H ].
The result is:

z hzy W hzx W hzx hxy W (hzx W )(hxy W ) z=x

+
Hxz hzy W hzx W hxy W hxz hxy W (hzx W )(hxy W )
z =x
N(1 xy )(hxy W + 1) = 0.
(3.36)
Again the upper sign refers to the compact model and the lower one to the noncompact model. In
addition there is the following minor but essential modification: in the compact case (3.36) holds
for all x, y , since W+ [H ] coincides identically with the generating functional W [H ] defined
without gauge fixing, see (2.4). In contrast, in the noncompact case it holds for all x0 = x
and all y . Specializing to y = x0 thus results in an equation that is qualitatively different
from the others.

The derivation of (3.36) is based on the invariance of the product measure x d(nx ) under rotation of any one of the spins, nx1 , say, where x1 = x0 in the noncompact model. For an
infinitesimal rotation with the Lie algebra element t ab we write

xab1 ndx = x1 x ncx t ab c d .
(3.37)
For later use we pick an explicit basis and note the completeness relations:
ab d
so(N + 1):
t c = ca bd c b ad = (tab )c d ,
1 ab d
t c (tab )e f = ce f d c f e d ,
2
ab
ab d

so(1, N):
t c = ca bd c b ad = aa bb (ta b )c d ,
1 ab d
t c (tab )e f = ce f d c f e d ,
2
107
(3.38a)
(3.38b)
ab
where is defined by a c cd bd = a 0 b0 a 1 b1 a N bN . Denoting by H the source

dependent averages this gives in a first step

Hxz t ab nx nz O + xab O = 0,
xab S O +
(3.39)
H
for any local function O = O({n}) of the spins, and x = x0 in the noncompact case. Specializing
to O = nx t ab ny and summing over a, b, using the completeness relations (3.38) gives

Hxz [nz ny nz nx ny nx ]
z [nz ny nz nx ny nx ]|z=x +
z =x

N(1 xy )ny nx
= 0.
(3.40)
Replacing the H averages by functional derivatives of W [H ] results in (3.36).

Coupled bilinear equations for the Wr functions arise by differentiating (3.36) repeatedly and
setting H equal to zero. The equation obtained by s-fold differentiation involves all Wr functions
with r s + 2; as a consequence no finite subsystem of equations arises for any individual Wr .
Before turning to the large N expansion of the system (3.36) we wish to stress that even the
exact SD equations do not determine their solution uniquely. To see this, let us momentarily denote the functional equations (3.36) with the upper and the lower signs by (SD)+ and (SD) ,
respectively. Further we write W+, , W+, , and W, , > 0, for the generating functionals
of the ferromagnet, the antiferromagnet, and the noncompact model, respectively. Then, by construction, W+, solves (SD)+ while both W+, and W, solve (SD) . (Consistency requires
that substituting the rhs of (2.14) into the (SD) equation converts it back into a (SD)+ equation
for W+, evaluated on H . Using (f )x = x (f )x 4dx fx , this can be verified to be
the case.) Thus if the exact SD equations were to determine their solution uniquely, W+, and
W, would have to coincide. This however contradicts the discussion in Section 2.2, viz. the
antiferromagnet and the noncompact model are physically and mathematically inequivalent. The
upshot is that somehow initial or boundary conditions have to be imposed on (3.36) to specify a
solution uniquely. Since the exact equations couple all multipoint functions, this seems difficult
to do concretely.
In contrast, the large N ansatz (2.10) effectively converts the equations into ones which can be
solved recursively, and initial conditions can be specified. The structure of the large N expanded
SD equations is best explained by spelling out the first few (see [9]) from which one can read off
(s)
the recursion pattern for the Wr , r + s > 1, functions in Fig. 1.
To compute a given coefficient all quantities having arrows pointing towards it are needed.
The detailed form of the equations and the solutions is not essential for the following argument.
108

(0)
W1
(0)
(0)
W3
(1)
W2
W2
(1)
W1
Fig. 1. Recursion pattern for the solution of the large N expanded SD equations.
(0)
We only need: (i) the fact that the equation for W1 is autonomous and reads

(0)

(0)
(0)
(0)
z W1 (z, y) W1 (z, x) z=x W1 (x, y)z W1 (z, x)z=x
(0)
= 1 xy + W1 (x, y),
(3.41)
and (ii) the assumption that each recursion step in Fig. 1 has a unique solution. We first present
the solutions of (3.41) and then discuss the status of the assumption (ii).
A solution of (3.41) with the upper sign is
(0)
W1 (x, y) = D+ (x y) 1/ ferromagnet.
(3.42)
Two solutions of (3.41) with the lower sign are

(0)
W1 (x, y) = x y D+ (x y) 1/
(0)
W1 (x, y) = D (x
antiferromagnet,
y) 1/ noncompact model,
(3.43)
where for the verification of the first solution (f )x = x (f )x 4dx fx has been used.
(0)
For each solution W1 (x, x) = 0 amounts to the corresponding gap equation. As discussed in
Section 3.3 the gap equations have O(V ) solutions other than the physical ones entering D .
Each solution of the gap equation defines a different solution of (3.41) via (3.42), (3.43), and in
principle there could be others. The selection of a specific solutionhere D can be justified
either from the functional integral (in that it gives rise to valid asymptotic expansion) or from the
physics one seeks to describe.
The uniqueness of the recursion (ii), to all orders, is probably difficult to establish directly
from the SD equations. In a perturbative expansion of the Wr the problem becomes linear and
the fact that the perturbative expansions of the Wr is uniquely determined by the equations has
been pointed out by M. Lscher [17]. In other words a potential nonuniqueness is known to be
of order O( ), directly from the SD equations.
An indirect way to establish the uniqueness (ii) is by showing that the moments of the generating functionals W [H ] have well-defined asymptotic expansions of the form (2.10). The
expansion coefficients then must be unique, and since W [H ] are solutions of (3.36) the coef(s)
ficients W,r must be solutions of the expanded SD equations, and the unique solutions. This
also implies that the solutions must be translation invariant, despite the preferred role of the
x = x0 equation. As mentioned earlier, the fact that the moments of W+ [H ] have an asymptotic
expansion follows from [16]; for W [H ] this will be shown in [4].
Given the uniqueness of the recursion one can readily re-derive the result of Section 3.4. From
(3.36) and (2.10) it is clear that the involution
,
Wr(n) () ()r Wr(n) (),
(3.44)
109
will map the SchwingerDyson equations of the compact model onto those of the noncompact
model, to all orders of their large N expansions. To draw conclusions about the solutions, how(n)
ever, the uniqueness asserted in (ii) is crucial. It implies that a sequence of solutions Wr () in
(0)
the compact model, based on a specific choice of W1,+ (), has a unique counterpart in the non(0)
compact model based on the flipped initial solution W1, () of the lower sign SD equations
(3.41). Since the physical solutions of the compact ferromagnet and the noncompact model are
(s)
(s)
related in this way it follows that there exists unique functionals Xr [D]() such that W+,r =
(s)
(s)
(s)
(0)
Xr [D+ ]() and W,r = (1)r Xr [D ](), where D (x y) := (W,1 (x, y) + 1/) are
the physical solutions of (3.41). In summary, the previous discussion reproduces the parts (a), (b)
of the result in Section 3.4.
We add some remarks. First, the involution (3.44) is the counterpart of (3.35). Let us repeat
(n)
that the involutions (3.35) or (3.44) do not imply a simple relation between the Wr () and
(n)
(0)
(0)
Wr () as functions of the lattice points, because already W1 () and W1 () are based on
solutions of different gap equations. It is rather the functional dependence on D which is the
same.
The involution also does not map the large N coefficients of the compact model into those
of the antiferromagnet, defined by W+ [H ]| . The reason is that the large N saddle point
relevant for the antiferromagnet is not the one obtained by reversing the sign of in the gap
equation. Rather is given by the same equation as the saddle point for the ferromagnet, as is
manifest from (3.43). Generally the large N coefficients of the antiferromagnet are obtained
simply by introducing appropriate products of the alternating sign function x in those of the
ferromagnet, as explained in Section 2.2.
The special role of the SD equations (3.36) for y = x0 is related to the discussion in the
paragraph following Eq. (2.8). It may suffice to illustrate the point with the lowest order equation
(3.41). The solution of the y = x0 equation in (3.41) is

0
1
(0)
nx f.s. = n x + O
(3.45)
with n x := 1 + W1 (x0 , x).
N +1
The interpretation of this solution as the nonzero average of the n0x component follows from [1].
To higher orders the solutions of the y = x0 equations produce corrections to (3.45) or noninvariant three-point functions, etc.
3.6. Algorithm in fixed spin gauge
Here we illustrate the main results from Sections 3.1 (existence of the asymptotic expansion) and 3.4 (compactnoncompact correspondence) by an explicit computation of the two- and
four-point functions in the noncompact model to next-to-leading order. The starting point is the
generating functional (3.5), where according to the result described, D(H ) has been replaced
by RV 1 . Due to the gauge fixing a number of complications arise compared to the conventional large N computations; in particular it is not obvious how translation invariant correlation
functions are recovered.
Initially we keep the putative saddle point configuration x , x = x0 , generic and specialize to
x = only later. We thus write
ux
, ux R, x = x0 .
ax = x +
(3.46)
N +1
110
Although notationally cumbersome it is important to carefully distinguish hatted, tilde and

plain matrices. We set

Mxy := xy + x xy =: M xy + x0 xx0 xy ,
D xy := M 1 xy ,
Uxy := ux xy =: U xy + ux0 xx0 xy .
(3.47)
The matrices M and U are also defined for x, y = x0 . For M we assume x0 = + , which
U for
ensures its invertibility; in Uxy the variable ux0 is non-dynamical. Further we write M,
the matrices obtained from M, U , respectively, by deleting the x0 th row and column and put
D = M 1 . Then
1
1
H,
U +
A = M +
N
+1
N +1
1

1
1
A 1 = 1 +
DH
D,
D U +
N +1
N +1
1
1
H ,
A = M +
U +
N +1
N +1

1
1
D H .
ln A = ln M + ln 1 +
DU +
N +1
N +1
(3.48a)
(3.48b)
The required expansions are for square matrices a, b

1

a
1
b
1+
=1+
ql (a, b),
+
(N + 1)l/2
N +1 N +1
l1
q1 (a, b) = a, q2 (a, b) = a b, q3 (a, b) = a 3 + ab + ba,

2
q4 (a, b) = a 4 a 2 b aba ba 2 + b2 ,
(3.49a)

1
b
a
=
pl (a, b),
+
ln 1 +
(N + 1)l/2
N +1 N +1
l1
1
1
1
p1 (a, b) = a, p2 (a, b) = a 2 + b, p3 (a, b) = a 3 (ab + ba),
2
3
2
1 2
1 4 1 2
2
p4 (a, b) = a + ba + aba + a b b .
(3.49b)
4
3
2
Explicit formulas for the polynomials ql , pl , l 5, can easily be found but will not be needed.
With this notation one has

1
1
A x x = Dx0 x0 1 +
Dx x Ql ,
0 0
(N + 1)l/2 0 0
l1

)D
Ql := D x2
(3.50a)
ql (D U , DH
,
x x
0 x0
0 0
ln A = ln M +

l1
1
pl (D U , D H ),
(N + 1)l/2
(3.50b)
111
from which the expansion of S[a, H ], i.e.

1 1 1
1
1
ax
A xx,
S[a, H ] = Tr ln A +
0 0
2
2
2
S[a, H ] =

l0
x =x0
1
Sl [u, H ],
(N + 1)l/2
(3.51)
can readily be computed to any desired order. We present S1 , . . . , S4 first in the condensed form
arising from (3.49), (3.50)
1
1
1
ux + Q1 ,
S1 [u, H ] = Tr p1 (D U , D H ) +
2
2
2
x =x0

1
1
S2 [u, H ] = Tr p2 (D U , D H ) +
Q2 D x0 x0 Q21 ,
2
2

1
1
Q3 D x0 x0 2Q2 Q1 + D x20 x0 Q31 ,
S3 [u, H ] = Tr p3 (D U , D H ) +
2
2
1
S4 [u, H ] = Tr p4 (D U , D H )
2

1
Q4 D x0 x0 2Q3 Q1 + Q22 + 3D x20 x0 Q2 Q21 D x30 x0 Q41 .
+
(3.52)
2
Note that Sl is a homogeneous polynomial of order l in the ux s and H s, assigning ux degree 1
and Hxy degree 2.
For S1 and S2 this gives in a first step

1
D 2 ,
ux 1 + D xx D x2
S1 [u, H ] =
0 x0 x0 x
2
x =x0

D x20 x D x20 y
1 D x0 x D x0 y
1
1 2
ux uy Dxy +
Dx0 x0
S2 [u, H ] =
Dxy
4
2 D x0 x0
2
D x0 x0
D x20 x0 D x20 x0
x,y =x0

D x0 x D x0 y
1
+
(3.53)
Hxy D xy
.
2
D x0 x0 D x0 x0
x,y
Using (3.13), (3.14) one can now eliminate D xy in favor of Dxy := (M 1 )xy . This produces
an expression of the same form just with Dxy replacing D xy . Next we eliminate D xy in favor
of Dxy , using (3.10). Anticipating Dx0 x0 = 1/ (which can be shown to hold independent of
the gap equation) it follows that the saddle point equations S1 /ux = 0 are equivalent to
Dxx = 1,
x = x0 ,
(3.54)
which is the non-translation invariant precursor of the gap equation used in Section 3.3. Inserting
this into S2 gives
2
1
1
2
2
+
ux uy Dxy
2 Dxx
D
Hxy Dxy .
S2 [u, H ] =
(3.55)
yx
0
0
4
2 x,y
x,y =x0
So far translation invariance was not presupposed, i.e. we allowed for a nontrivial xdependence in the extremal configurations x , x = x0 . We anticipate now from [4] that the only
112
solution of (3.54) in the domain D(H = 0) is constant, i.e.

x = const,
x = x0 .
(3.56)
As seen in Section 3.3 then D(x) 1 fixes the constant to be = (, V ), i.e. the unique
solution of the translation invariant gap equation D(0) = 1 obeying ((4/(2d + 1))
sin2 /L, 0). As before we write D (x y) for the translation invariant propagator with gap .
Further, in Appendix A we prove that
S2 [u, H ]|x = 0.
(3.57)
This means that the saddle point is indeed a local minimum of the action (3.51). We conjecture
that it is in fact a global minimum of the action (3.51) in the domain D(H ).
We now return to the generating functional W [H ]. Substituting (3.55) one arrives at

1
exp W [H ] = exp
Hxy D (x y) + 1
2 x,y

1
dux exp +
ux uy D ,2 (x, y)
N
4
x =x0
x,y =x0

1 2
1
1
S4 + S3 + O
.
1+
(3.58)
N +1
2
(N + 1)2
Here
D ,2 (x, y) := D (x y)2 2 D (x x0 )2 D (y x0 )2 ,
and the terms occurring in the third line are generated by the expansion of

1
exp
S
[u,
H
]
,
l2 l
l3 (N + 1) 2
(3.59)
(3.60)
with the half-integer powers of 1/(N + 1) omitted. The conversion procedure exemplified for S2
1 1
is easily seen to generalize to all Sl , l 3. Indeed from (A 1 )1
x0 x0 = x0 + (A )x0 x0 it follows
that in an expansion

1 1
1 1
1
x1x 1
Q
A
=
D
+
,
A
(3.61)
1
x0 x0
x0 x0 l
0 0
l/2
(N + 1)
N +1
l2
all coefficients [(A 1 )1

x0 x0 ]l , l 2, have the following properties: (i) eliminating Dxy in favor of
1
Dxy = (M )xy via (3.13), the parameter drops out. (ii) The result of the replacement is an
expression of identically the same form, just with Dxy occurring for each instance of D xy . (iii) In
the result one can replace Dxy by D (x y) and D xy by D (x y) + D (x x0 )D (y x0 ).
The problem of computing W [H ] to order 1/(N + 1)p thus reduces to the algebraic problem
of computing the Sl [u, H ], l 2p + 2, with the indicated substitutions, and to performing Wick
contractions with the free propagator

(x, y),
(x, z)D ,2 (z, y) = x,y , x, y = x0 .
ux uy 0 = 2
(3.62)

z
In fact one has

(x, y) = (x y),

for x, y = x0 ,
(3.63)
113

where is the translation invariant free u-propagator defined by z (x z)D (z y)2 =
xy .
On the basis of the general results the H -moments computed with the algorithm described
above must be translation invariant and be related to that of the compact model by the involution
described. As a test of both the algorithm and the involution we now compute the two- and
the four-point function to sub-leading order. To this end explicit expressions for S3 and S4 are
needed.
It is convenient to parameterize them as follows
1
(0)
(1)
Tr p3 + Q3 + 2 Q3 ,
2
1
(0)
(1)
(2)
(0)
(1)
(2)
S4 = S4 + S4 + S4 = Tr p4 + Q4 + 2 Q4 + 3 Q4 .
2
(0)
(1)
S3 = S3 + S3 =
(n)
(3.64)
(n)
Here Sl , n = 0, 1, 2, are the pieces of order n in the source H . The contribution to Sl coming
(n)
(0)
(1)
from [(A 1 )1
x0 x0 ]l in (3.61) we denote by Ql , for n = 0, 1, 2, respectively. We need Q3 , Q3
(1)
(2)
and Q4 , Q4 . They come out as

(0)
Q3 =
ux uy uz D(x x0 )D(z x0 )D(x,

y)D(y,
z),
(3.65a)
Hxy uz D(x x0 )D(z x0 )D(y,

z),
(3.65b)
x,y,z =x0
Q(1)
3 =2
x,y;z =x0
(1)
Q4 =
Hxy uz uw D(z x0 )D(w,

x)
x,y;z,w =x0

w) + D(w x0 )D(z,
y) ,
2D(y x0 )D(z,

(2)
Hxy Hzw D(x x0 )D(w x0 )D(y,

z).
Q4 =
(3.65c)
(3.65d)
x,y,z,w
Here we omitted the subscript on the propagators and we shall continue to do so for the
remainder of this subsection.
(0)
(1)
(1)
(2)
For S3 , S3 , S4 , S4 one obtains
(0)
S3 =
1
6
(1)
S3 =
(1)
S4 =
ux uy uz D(x,
y)D(y,
z) D(z x) 2D(x x0 )D(z x0 ) ,
Hxy uz D(x,
z) D(z y) D(z x0 )D(y x0 ) ,
(2)
(3.66b)
x,y;z =x0
w) D(z y) 3D(z x0 )D(y x0 )

Hxy uz uw D(x,
w) D(z,
x,y;z,w =x0

y) D(w z) D(w x0 )D(z x0 ) ,
+ D(z,
S4 =
(3.66a)
x,y,z =x0

2
z) D(x w) D(x x0 )D(w x0 ) .

Hxy Hzw D(y,
4 x,y,z,w
(3.66c)
(3.66d)
114
The desired correlation functions can now be computed from

(2)
S4
1 (1) 2
+ S3
2

=
0
2
8
(0)
Hx1 ,y1 Hx2 ,y2 W2 (x1 , y1 ; x2 , y2 ),
x1 ,y1 ,x2 ,y2

(1)
=
Hxy W1 (x, y).
0
2 x,y
(1)
S4(1) + S3(0) S3
(3.67)
For W2(1) the x0 -dependent terms cancel out algebraically and one finds
(0)
W2 (x1 , y1 ; x2 , y2 ) = D(x1 x2 )D(y1 y2 ) + D(x1 y2 )D(y1 x2 )

D(x1 z)D(y1 z)(z w)D(w x2 )D(w y2 ).
2
z,w
(1)
S4 0 and
generates a large number of x0 -dependent terms

(x
z)D(z
x0 )2 = xx0 . The results are

z
The expansion of
which simplify after using
(1)
0
S4

xy
Hxy
(3.68)
(0) (1)
S3 S3 0
(z, w)D(z w)D(x w)D(y z)
z,w
Hxy D(x x0 )D(y x0 )
xy
(z, w)D(z w)D(z x0 )D(w x0 ),
(3.69a)
z,w
(0) (1)
0
S3 S 3
= q
Hxy
xy
D(z x)D(z y) + 3
Hxy D(x x0 )D(y x0 )
xy
(z, w)D(z w)D(z x0 )D(w x0 ).
(3.69b)
z,w
Here
q :=
(z)
D(z u)D(z v)(u v)D(u v).
(3.70)
u,v
Finally
(1)
W1 (x, y) = 2q
D(x w)D(y w)
+2
D(x z)D(y w)(z w)D(z w),
(3.71)
z,w
(1)
which obeys W1 (x, x) = 0. Recall that D(x) here is short for D (x) = D(x)|= . The results (3.71) and (3.68) are manifestly translation invariant and related to those of the compact
model (see [9,18]) by the involution described.
115
4. Conclusions
On the level of asymptotic expansions interesting and useful correspondences exist between
compact and noncompact nonlinear sigma-models. Here we established two such correspondences: in perturbation theory, where simply flipping the sign of allows to move between the
compact and noncompact models, and in the large N expansion, where one has to make a detour
through the finite volume in order to establish a correspondence that involves not just flipping
the sign of but also picking the physically appropriate solution of the gap equation.
Taking the sign flip prescription beyond perturbation theory leads from the compact model
not to the noncompact one, but rather to another compact model, the antiferromagnet.
In the literature [19] a sign flip rule for the large N coefficients in a continuum formulation has
been proposed. By Section 3.4 the rule is incorrect. In [2] the much weaker claim has been made
that there exists a scheme in which the coefficients of the large N beta function in the compact
and in the noncompact models are related by the sign flip . This statement seems to be in
agreement with our findings because it is based on re-organizing the perturbation expansion into
powers of 1/N , but it remains an open question whether the analytic continuation from positive
to negative employed in [2] to go from the compact to the noncompact model is justified in the
continuum.
Acknowledgements
We wish to thank A. Duncan and M. Lscher for discussions and correspondence.
Appendix A. Positivity of S2
Here we show that
S2 [u, H ]|x = 0,
(A.1)
which guarantees that the saddle point on which the large N expansion is based is indeed a local
minimum of the action (3.51).
We begin by noting that the source-dependent term separately obeys
1
Hxy D(x y) 0,
2 x,y
(A.2)
for all sources Hxy 0, as D(x y) 1. It therefore suffices to consider S2 [u, 0]. Next we
introduce

1
1
(p) :=
(A.3)
eipx D(x)2 =
,
V
(Ek + )(Ekp + )
x
k
and claim that in terms of it a sufficient condition for S2 [u, 0] 0 is

(p) < 0,
p = 0.
(A.4)
To see this we prepare

(0)
V
,
2
(A.5)
116
using once more 1 D(x). Then we rewrite S2 [u, 0] in Fourier space

1
u(p) Spq u(q),
4V 2 p,q

S00 = (0) V 2 (0) ,
S2 [u, 0] =
S0p = 2 v0 vp ,
vp := (p)eipx0 ,
Spq = pq V (p) + 2 vp vq ,
p, q = 0.
(A.6)
Subject to (A.4) and (A.5) the matrix S has the form of a positive rank one perturbation of a
positive diagonal matrix, which therefore must be itself positive.
It remains to show (A.4). We rewrite (p) in the form

2
1
+ J (p) , p = 0,
(p) =
Ep + 2
1 Ek + Epk Ep
,
J (p) =
(A.7)
V
(Ek + )(Epk + )
k
using the gap equation. Here J (p) manifestly has a finite thermodynamic limit. Further we claim
J (p) 0 for all p,
(A.8)
which implies (A.4).

The proof of (A.8) is based on a reorganization of the d-fold sum in (A.7) such that each
term is positive on account of the lemma below. We begin by noting that J (p) is completely
symmetric in all arguments p = (p1 , . . . , pd ), and that

J (2 p1 , p 1 ) = J (p),
(A.9)
p := (p1 , . . . , p1 , p+1 , . . . , pd ).
Since J (0) > 0 trivially, it suffices to restrict attention to momenta p = 0 with 0 p1
pd . To proceed we insert
Epk + Ek Ep = 8
cos
p
k
(k p)
sin
sin
2
2
2
(A.10)
into (A.7) to obtain

J (p) =

1 1
S(k |p),
d1
4 L
k , =
1
S(k |p) =
L
[sin2
k
k p
2
k
2
2 k p
sin
+
X
][sin
sin2 2
2
2
2
2 cos
p
2
sin
k
2
sin
+ Y ]
(A.11)
where we set
X :=

=
sin2
k
,
2
Y :=

=
The result (A.8) now follows from the
sin2
k p
,
2
2 sin
:= .
2
(A.12)
Lemma. For all L 2 and q [0, ], q

S(q, , x, y) :=
L1
1
L
[sin2
n=0
117
2
L Z,
q
n
2 cos q2 sin n
L sin( L 2 )
n
L
sin2
q
2
+ x][sin2 ( n
L 2 ) sin
+ y]
0,
(A.13)
for parameters x, y 0, x + y sin2 /L, and 0 < < /L.

Although not needed for the parameters in (A.12) we remark that (A.13) is in fact valid for all
x, y 0. In the application to (A.11) only the bound on needs to be verified. From (3.28) we
have however

4
sin
.
sin <
(A.14)
2
2d + 1
2L
Proof of the lemma. Trivially S(0, , x, y) > 0, S(2 q, , x, y) = S(q, , x, y). It thus suffices to consider q (0, ]. For L = 2 the inequality (A.13) is manifest, for L > 2 we make use
of the following summation formulas:
L1
1
L
sin2
n=0
and for q
2
L Z
1
n
2
a
2
2coth La
2
,
sh a
a = 0,
(A.15)
and a = b = 0:
L1
1
L
[sin2
n=0
+ sh2
1
q
2 b
+ sh2 a2 ][sin2 ( n
L 2 ) + sh 2 ]
L
sh 2 (a + b) sh(a + b)
2
Lb
ch(a + b) cos q
2 sh a sh b
n
L
sh La
2 sh

sh L2 (a b) sh(a b)
+
.
ch(a b) cos q
(A.16)
For a = b = 0 the result of the summation is obtained from the rhs of (A.16) by omitting the
second term in square brackets and specializing the remainder to a = b (i.e. no pole occurs for
a = b and q = 0). Both (A.15) and (A.16) can be obtained from
L1
a L1
1
1
=
,
2in
L
1 aL
n=0 e L a
a L = 1.
(A.17)
In the application to S we set sh2 a2 = x sin2 2 , sh2 b2 = y sin2 2 , and pick the roots
a, b > 0 for x, y > sin2 2 . In a first step one obtains (also using the 1d version of (A.10) backwards)
S=
2coth Lb
2(sh2 a2 + sh2 b2 + sin2 q2 )
2coth La
2
2
+
Lb
sh a
sh b
sh La
2 sh 2 sh a sh b

L
sh 2 (a + b) sh(a + b) sh L2 (a b) sh(a b)
+
.
ch(a + b) cos q
ch(a b) cos q
This can be rewritten in the form

L

sh L2 (a b) sh ab
sh 2 (a + b) sh a+b
cos2 q2
2
2
.
S=
Lb
2 q
2 q
ch a2 ch b2 sh La
sh2 a+b
sh2 ab
2 sh 2
2 + sin 2
2 + sin 2
(A.18)
(A.19)
118
Since the function

t (z) :=
z
sh Lz
2 sh 2
sh2
z
2
+ sin2
q
2
(A.20)
is positive and monotonically increasing on R+ it follows that for a, b > 0 and q (0, ] the rhs
of (A.19) is nonnegative.
It remains to consider the case where 0 < := ib < /L and a > 0. The expression (A.19)
can in this case be written as
S=
q
2
a
cos
2
2
4 cos2
L
coth La
2 s1 cot 2 s2
(ch a cos cos q)2 + sh2 a sin2

a
s1 = sh cos [ch a cos + 1 cos q],

2
2
a
s2 = ch sin [ch a cos 1 + cos q].

2
2
ch
(A.21)
Further
coth
L
La
s1 cot s2
2
2

a
La a
L
= (ch a cos ) ch cos

coth
th cot
tan
2
2
2
2
2
2

La a
a
+ (1 cos q) coth
sh cos + cot
sin ch .
2
2
2
2
2
2
(A.22)
The second term in (A.22) is manifestly positive. For the first term we note that cot L
2 tan 2
La
a
maps [0, /L] bijectively onto [1/L, 0], while a coth 2 th 2 maps [0, ) bijectively onto
[1/L, ). Hence
La a
th cot
tan 0,
2
2
2
2
which shows that S in (A.21) is nonnegative. This concludes the proof of the lemma.
coth
(A.23)
2
References
[1] A. Duncan, M. Niedermaier, E. Seiler, Vacuum orbit and spontaneous symmetry breaking in hyperbolic sigmamodels, Nucl. Phys. B 720 (2005) 235.
[2] J.J. Friess, S.S. Gubser, Non-linear sigma models with anti-de Sitter target spaces, Nucl. Phys. B 750 (2006) 111.
[3] T. Spencer, M.R. Zirnbauer, Spontaneous symmetry breaking of a hyperbolic sigma model in three dimensions,
Commun. Math. Phys. 252 (2004) 167187.
[4] M. Niedermaier, E. Seiler, in preparation.
[5] A. Houghton, A. Jevicki, R. Kenway, A. Pruisken, Noncompact sigma-models and the existence of a mobility edge
in disordered electron systems near two dimensions, Phys. Rev. Lett. 45 (1980) 394.
[6] S. Hikami, Three loop beta function of nonlinear sigma-models on symmetric spaces, Phys. Lett. B 98 (1981) 208.
[7] S. Elitzur, The applicability of perturbation expansion to two-dimensional Goldstone systems, Nucl. Phys. B 212
(1983) 501.
[8] F. David, Cancellation of infrared divergencies in the 2D nonlinear sigma-model, Commun. Math. Phys. 81 (1981)
149.
[9] A. Duncan, M. Niedermaier, P. Weisz, Noncompact sigma-modelsLarge N expansion and thermodynamic limit,
arXiv: 0607.2929 [hep-th].
[10] S. Kobayashi, K. Nomizu, Foundations of Differential Geometry, vol. II, Interscience, New York, 1969.
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
119
S. Helgason, Differential Geometry and Symmetric Spaces, Academic Press, New York, 1962.
E.T. Copson, Asymptotic Expansions, Cambridge Univ. Press, Cambridge, 1965.
J.D. Murray, Asymptotic Analysis, Oxford Univ. Press, Oxford, 1974.
H. Mukaida, Y. Shimada, Convex effective potential of O(N ) symmetric 4 theory for large N , Nucl. Phys. B 479
(1996) 663.
A. Duncan, E. Seiler, P. Weisz, unpublished notes, 2005.
A.J. Kupiainen, On the 1/n expansion, Commun. Math. Phys. 73 (1980) 273.
M. Lscher, SchwingerDyson equations in a finite volume, unpublished notes, 1980/81.
G. Cristofano, R. Musto, F. Nicodemi, R. Pettorino, F. Pezzella, 1/N contribution to physical quantities in the lattice
O(N ) sigma-model, Nucl. Phys. B 257 (1985) 505.
S.A. Brunini, M. Gomes, A.J. da Silva, Remarks on non-compact sigma-models, Phys. Rev. D 38 (1988) 706.
Statistical mechanics of vortices

from D-branes and T-duality
Minoru Eto a , Toshiaki Fujimori b, , Muneto Nitta c , Keisuke Ohashi b ,
Kazutoshi Ohta d , Norisuke Sakai b
a University of Tokyo, Institute of Physics, Komaba 3-8-1, Meguro-ku, Tokyo 153, Japan
b Department of Physics, Tokyo Institute of Technology, Tokyo 152-8551, Japan
c Department of Physics, Keio University, Hiyoshi, Yokohama, Kanagawa 223-8521, Japan
d Department of Physics, Graduate School of Science, Tohoku University, Sendai 980-8578, Japan
Received 9 April 2007; accepted 13 June 2007

Available online 29 June 2007
Abstract
We propose a novel and simple method to compute the partition function of statistical mechanics of
local and semi-local BPS vortices in the AbelianHiggs model and its non-Abelian extension on a torus.
We use a D-brane realization of the vortices and T-duality relation to domain walls. We use a special limit
where domain walls reduce to gas of hard (soft) one-dimensional rods for Abelian (non-Abelian) cases. In
the simpler cases of the AbelianHiggs model on a torus, our results agree with exact results which are
geometrically derived by an explicit integration over the moduli space of vortices. The equation of state for
U (N
) gauge theory deviates from van der Waals one, and the second virial coefficient is proportional to
1/ N, implying that non-Abelian vortices are softer than Abelian vortices. Vortices on a sphere are also
briefly discussed.
1. Introduction
Knowledge of moduli space structure of solitons is important in order to understand not only
their own dynamics but also non-perturbative effects in field theory. The moduli space structure of
E-mail addresses: meto@hep1.c.u-tokyo.ac.jp (M. Eto), fujimori@th.phys.titech.ac.jp (T. Fujimori),

nitta@phys-h.keio.ac.jp (M. Nitta), keisuke@th.phys.titech.ac.jp (K. Ohashi), kohta@phys.tohoku.ac.jp (K. Ohta),
nsakai@th.phys.titech.ac.jp (N. Sakai).
doi:10.1016/j.nuclphysb.2007.06.020
M. Eto et al. / Nuclear Physics B 788 [FS] (2008) 120136
121
the solitons signifies its topology, metrics and singularities. The volume of the moduli space also
possesses an important meaning and plays an essential role in understanding the non-perturbative
dynamics. Recently Nekrasov has shown that the prepotential of N = 2 supersymmetric gauge
theory in d = 3 + 1 can be obtained from some statistical partition function, whose free energy
measures the volume of the moduli space of YangMills instantons [1]. The instanton moduli
space is non-compact and of course the volume diverges, but a coefficient of the divergent part
gives perturbative and non-perturbative information on supersymmetric gauge theory. It is also
interesting that the partition function or the volume of the instantons is closely related to topological string amplitudes on suitable CalabiYau manifolds. Precisely speaking, the instantons,
which are used for the calculation of the volume, are not solutions of the equations of motion in
supersymmetric gauge theory. However, in the calculation of the partition function or the volume,
the so-called localization theorem says that the result does not depend on details of the moduli
space structure. This suggests that we do not need to know the exact solutions and metrics in
order to calculate the volume of the moduli space of some class of BPS solitons. However, the issue has not been settled except for YangMills instantons, since other examples and applications
have not been investigated yet. For instance, similar method should be applicable to calculate the
effective (super)potential of d = 1 + 1 supersymmetric gauge theory from the statistical partition
function associated with the volume of the vortex moduli space.
Another direct application of the volume of the moduli space is a classical statistical mechanics of the solitons. The Gibbs partition function of solitons at finite temperature T is given by
an integration over a phase space, which is a cotangent bundle T M of the moduli space M of
solitons:

1

d D x d D p eH (x ,p)/T
.
Z=
(1.1)
(2)D
T M
Here D represents the dimensions of the moduli space, and H (

x , p ) is a Hamiltonian of the
soliton system in terms of the moduli parameters. If we assume that the solitons are sufficiently
diluted and interactions can be ignored, we can regard the Hamiltonian as quadratic in momenta
H (
x , p ) = 12 g ij (x)pi pj and g ij is the (inverse of) metric on M. Then we can explicitly perform
the integration over the momenta (cotangent direction). The partition function becomes
D/2
D/2

T
T
Z=
(1.2)
d D x det gij (x) =
Vol(M),
2
2
M
which is proportional to the volume of the moduli space [2]. Many applications and calculations
from this point of view can be found [26] in the case of AbrikosovNielsenOlesen (ANO)
vortices [7], namely vortices in the AbelianHiggs model at the critical coupling (the BPS limit).
On the other hand, non-Abelian BPS vortices have been recently found [8,9] and extensively discussed in the (supersymmetric) non-AbelianHiggs model, which is the U (NC ) gauge
theory coupled with NF (> NC ) fundamental Higgs fields. Like (non-commutative) instantons,
single vortex in the NC = NF case is found to carry internal moduli CP NC 1 in addition to
position moduli. They can confine monopoles in the Higgs phase [10,11] and this fact is applied [12,13] to show coincidence of the BPS spectra of d = 3 + 1, N = 2 supersymmetric
gauge theory and d = 1 + 1, N = 2 supersymmetric CP n model. The moduli space of multiple
(k-)vortices has been conjectured in string theory [8] and then has been completely determined
in field theory [1416], which has turned out to be some resolution of k-symmetric product of
C CP NC 1 [14]. Although its explicit metric is still unknown except for a single (k = 1) vortex,
122
an integration formula of the Khler potential has been found [17]. The structure of the moduli
space of two (k = 2) vortices has been worked out [16,18] and is applied to non-Abelian duality
[19], and classical dynamics of non-Abelian vortex-strings has been clarified [20]. The extension
to a superconformal field theory [21] as well as the ChernSimonsHiggs theory [22] has been
discussed.
In this article, we propose a novel and simple derivation of the volume of the moduli space of
Abelian as well as non-Abelian (semi-)local vortices in order to apply to the statistical mechanics of vortex gas. We utilize a geometric and stringy (D-brane) interpretation of the Abelian and
non-Abelian BPS vortices in supersymmetric gauge theory [8] and use a T-dual relation between
the vortices and domain walls, which was proposed in [23,24]. This interpretation shows us a
schematic structure of the vortex moduli space without any detailed information on the exact
solutions and metrics. (See Fig. 17 of Ref. [24] for the moduli space of two non-Abelian local
vortices in U (2) gauge theory.) It is interesting to observe that the T-duality reduces the calculation of the volume into a simple problem of a gas of hard rods between a one-dimensional interval
in the Abelian case. The volume of the moduli space of the ANO vortices in the AbelianHiggs
model (with NC = NF = 1) has been already calculated by [24] (see also Section 7 in [5]) and
an analogy with the hard rod problem was pointed out there. However this relation has been mysterious for a long time. We can explain from a point of view of the T-duality in superstring theory
why the one-dimensional hard rod problem appears. The advantage of our method is that it can
be easily extended to the general cases, namely local and semi-local, Abelian and non-Abelian
vortices. We find that the T-duality enables us to reduce the problem of a gas of non-Abelian
local vortices to a gas of soft rods, in contrast to the hard rods in the
case of Abelian local vortices. We obtain that the second virial coefficient is proportional to 1/ N for non-Abelian local
vortices in U (N
) gauge theories at large N . This shows that the exclusion volume of a vortex
behaves as 1/ N as we increases N , getting closer to an ideal gas. Moreover, equation of state
for local vortices in U (N ) gauge theory deviates from van der Waals one, contrary to the case of
the AbelianHiggs model.
Comparing with the calculation in [25], our result in the Abelian case gives the precise
answer even though we take a special limit to the configuration and do not use the exact metrics.
So we expect that this is another example where the localization theorem effectively works. In
fact calculation of the volume of vortex moduli space using the localization theorem can be found
in the case of the AbelianHiggs model [6].
In Section 2 we give a brief review of brane configurations of vortices and domain walls.
In Section 3, we calculate the partition function of vortices in the AbelianHiggs model, which
agrees with the previous result. In Section 4 we give general formula of the partition function
of local/semi-local non-Abelian vortices which is new and shows the power of our method. We
perform explicit calculation in several cases. Section 5 is devoted to a discussion. An interesting
duality is observed and our method is extended to the case of a base manifold S 2 . Appendix A
gives some details of derivation of the virial coefficient.
2. Non-Abelian vortices and D-brane interpretation
We first start with (2 + 1)-dimensional U (NC ) gauge theory with NF ( NC ) massless Higgs
fields in the fundamental representation. The Lagrangian is given by

2

g 2
1
L = Tr 2 FMN F MN + DM H DM H
(2.1)
H H c1NC ,
2
2g
123
where M, N are the indices of spacetime (M, N = 0, 1, 2) and the spacetime metric is
chosen as MN = diag(+1, 1, 1). The field strength is defined by FMN i[DM , DN ] =
M WN N WM + i[WM , WN ], where WM is the U (NC ) gauge field. The scaler fields H rA
(r = 1, . . . , NC ), (A = 1, . . . , NF ) are expressed by elements of an NC NF matrix and the covariant derivative is defined by DM H M H + iWM H . The constants g and c (> 0) are the
gauge coupling constant and the FayetIliopoulos (FI) parameter respectively.
For static configurations, the energy is bounded from below as follows:

2
1
g2
+ Dz H (Dz H ) cF12
c1NC H H
E = d 2 x Tr 2 F12 +
2
g

c d 2 x F12 = 2ck,
(2.2)
2
1
where the integer k 2
d x F12 represents the number of vortices (vorticity). Since the
configurations of the BPS vortices saturate the inequality (2.2), they satisfy the following BPS
equations (the vortex equations):

g2
(2.3)
c1NC H H = 0.
2
If the fluctuation energy above the energy of the multi-vortex static configuration is less than that
of the mass gap between massless moduli fluctuations and massive fluctuations, the dynamics of
the system is well described by the geodesic motion on the moduli space Mk of the k vortices
(Mantons moduli/geodesic approximation [25]), so that the Lagrangian for multi-vortex system
is given by
Dz H = 0,
F12 +
i j ,
L = gi j (, )
(2.4)
where i , i (i = 1, . . . , D = dimC Mk ) are complex coordinates of the moduli space Mk and

is a Khler metric of Mk .
gi j (, )
In the cases of compact base manifolds, there is the maximum number for vortices allowed to
exist [26]. For general NC with a torus T 2 , this maximum number can be obtained by taking the
trace of the second equation of the BPS equation (2.3) and by integrating over T 2 as

g2c
g2
d 2 x Tr H H 0 A kBNC ,
A=
2k + NC
(2.5)
2
2
where A is the area of the torus and we have defined the Bradlow area
BNC
1 4
.
NC g 2 c
(2.6)
The limit to saturate this inequality is called the Bradlow limit. In particular, we can regard
B1 = g4
2 c as an effective area of the Abelian (ANO) vortex. Roughly speaking, the inequality
(2.5) implies that we can squeeze as many vortices as NC times the maximal number of the
Abelian (NC = 1) vortices for a given area A of the torus. This behavior can be understood intuitively as due to the freedom for vortices to avoid occupying the same points in their additional
internal moduli space CP NC 1 when they are squeezed too much in the actual configuration
space. This freedom allows the non-Abelian vortices to overlap in the configuration space more
easily compared to Abelian vortices. Therefore the effective area of the non-Abelian vortex BNC
becomes 1/NC < 1 times the area of the Abelian vortices B1 . The Bradlow area BNC indicates
124
(a)
(b)
(c)
Fig. 1. Duality between vortices and domain walls: NC = NF = 1, k = 2.

(a) Profile of energy density of 2-vortex
2
1) = 1
1
configuration on R S 1 with period 2 R, (b) profile of (x
2 R dx W2 , (c) profile of (x ) written on
1
R S with period 1/R.
the smallness of the effective area of vortices at the limiting high density of vortices. Note that
there exists another inequality in the case of non-Abelian gauge theories with NC > k. Since
all vortex solutions in this case can be embedded to a theory with a smaller number of colors
NC (= k ) = k, the vortex configuration for NC > k should satisfy the inequality for NC = k, that
is,
A
4
(= B1 )
g2c
(2.7)
stating that the area of the torus must be larger than that of one ANO vortex.
As we mentioned in Eq. (1.2), the classical partition function of k-vortex system is proportional to the volume of the moduli space Mk . To calculate this partition function (volume), we
utilize the duality relation between vortices and domain walls [23,24]. If we compactify x 2 direction on S 1 as x 2 x 2 + 2nR, a vortex configuration can be viewed as domain walls. The
profile of the kink solution of the domain walls is described by the eigenvalues of defined by

2R

1
1
,
dx 2 W2 x 1 , x 2
log P exp i
x =
(2.8)
2iR
0
where P stands for the path-ordered product. Note that under the gauge transformation U =
2
1 ) transforms as + n/R1NC and there is an identification
einx /R 1NC (n Z), (x
1 ) are interpreted as a function which takes
+ n/R1NC . Thus the eigenvalues of (x

1
value in S with radius 1/R. Fig. 1 depicts an example of the NC = NF = 1 case. The k-vortex
configuration (Fig. 1(a)) corresponds to the k-wall configuration (Fig. 1(b)) in the fundamental
1 ) with k windings (Fig. 1(c))
region, and it also can be interpreted as the configuration of (x
on the cylinder.
This duality can be regarded as a T-duality between the brane configurations of vortices [8]
and domain walls [28]. Our model can be embedded into the supersymmetric system with eight
supercharges and the associated model is realized by a combination of various kinds of branes in
type IIB superstring theory. The brane configurations are expressed in Table 1 and drawn in Fig. 2
schematically. The (2 + 1)-dimensional U (NC ) gauge theory coupled with NF massless hypermultiplets is realized on NC D3 branes in the HananyWitten setup [29] as in Fig. 2(a), since the
x 3 direction of the D3 brane worldvolume is a finite line segment. D1 branes correspond to vortices in the D3 brane worldvolume theory, since they are interpreted as codimension-two objects
on the D3 branes [8]. Fig. 2 shows the brane configuration T-dualized along the x 2 -direction.
125
Table 1
Brane configurations for vortices and domain walls: Branes are extended along directions denoted by , and are not
extended along directions denoted by . The symbol denotes the codimensions of the k D1-branes (D2 -brane) on the
worldvolume of the D3-branes (D2-branes) excluding the x 3 -direction which is a finite line segment
(a) Brane configuration for vortices
NC D3
NF D5
2 NS5
k D1
x0
x1
x2
x3
x4
x5
x6
x7
x8
x9
(b) Brane configuration for domain walls

NC D2
NF D4
2 NS5
k D2
x0
x1
x2
x3
x4
x5
x6
x7
x8
x9
(a)
(b)
Fig. 2. Brane configurations for vortices (a) and brane configuration for domain wall (b).
The theory without vortices is mapped to the world-volume theory on the NC D2 branes, which
is the (1 + 1)-dimensional U (NC ) gauge theory with NF (massive) hypermultiplets. The vortices
1)
are mapped to the kinky D2-branes representing domain walls [27,28]. The eigenvalues of (x
1
can be interpreted as the position of D2 branes in S .
3. A limit of the profile and the moduli integration: The AbelianHiggs model
If we use the T-dual relation via the brane configuration, the evaluation of the volume of the
vortex moduli space reduces to a calculation of the volume of the domain wall configurations
(kink profiles). However, it is difficult to solve the domain wall equations and to integrate over
their configuration space in general.
In order to proceed the evaluation, we demand an approximation which simplifies the profile
of the domain wall solution. Now, let us consider a special limit
g 2 c,
1
,
R
with d
2
g 2 cR
: fixed.
(3.1)
126
(a)
(b)
1 ) before and after taking the limit g 2 c, 1/R with d 2/g 2 cR finite.
Fig. 3. Profiles of (x
Although this is a rough approximation with respect to the kink profile, we will (amazingly) find
that we can nonetheless obtain the exact results. Note that using d defined above the inequalities
(2.5) and (2.7) are translated in the T-dualized picture to
2R
1
kd,
NC
2R d,
(3.2)
respectively. These inequalities can be easily shown even in this T-dualized picture as discussed
below. This fact is the first evidence for the usefulness of our T-dualized picture in this paper.
Here we describe a limit shape of the ANO vortices in the AbelianHiggs model (NC =
NF = 1) for simplicity. In this case, the moduli space of the multi-vortex system consists of the
position of each vortex. The positions of vortices in the x 2 -direction are translated into the phase
degrees of freedom of domain walls after the T-duality along x 2 -direction. If one vortex is located
1 ) is given by
at z = 0, the corresponding function (x
0
for x 1 < ,

1 1
d
d
d
x1 +
for < x 1 < ,
x =
(3.3)
Rd
2
2
2
1
for x 1 > .
R
2
From this expression, we find that the parameter d represents the effective thickness of the dual
1 ) before (Fig. 3(a)) and after (Fig. 3(b)) taking the
domain wall. Fig. 3 shows the profiles of (x
limit. We have previously shown both from field theory [30] and string theory (the s-rule) [28]
that these domain walls cannot pass through nor be compressed with each other. Therefore these
objects can be regarded as 1-codimension rigid bodies with length d. Using this fact, we can map
the multi-vortex system to the gas of hard rods, namely 1-dimensional gas of the rigid bodies.
Using the above limit, where the domain walls are regarded as the hard rods in 1-dimension,
let us now consider the vortices on a torus T 2 with periods 2R and 2R and dual domain wall
1 ) and Fig. 4(b) shows the correspondconfiguration space. Fig. 4(a) shows the profile of (x
ing system of hard rods. For the gas of identical hard rods with mass m on S 1 with period L,
the classical partition function can be easily calculated as

1 mT k
Zrods =
(3.4)
L(L kd)k1 .
k! 2
(a)
127
(b)
1 ) for vortices (a) and corresponding 1d gas of hard rods on S 1 (b).

Fig. 4. Profile of (x
In this case, the period is L = 2R and each rod has mass m = 2c which corresponds to
the mass of the vortex. There are additional phase degrees of freedom which correspond to the
positions of vortices in x 2 -direction. In the limit g 2 c, 1/R with d = g 22cR finite, these phase
degrees of freedom become independent and have the period 2R. In other words, the T-duality
maps the FI-parameter c to 2Rc. So we should replace c by 2Rc in the vortex picture. (Note
that the combination d = g 22cR is invariant under the T-duality.) Therefore the partition function
of k-vortex system on a torus T 2 is given by
N =NF =1
Zk,TC 2
1
(cT )k 2R (2R kd)k1 (2R)k
k!

1
4k k1
= (cT )k A A 2
,
k!
g c
(3.5)
where A = (2)2 RR is the area of T 2 . This result coincides with the exact partition function
which has already been known [35]. The inequality (2.5) with NC = 1 implies that the number of
2
2
vortices k must be less than g4c A. In the Bradlow limit k = g4c A, the partition function Eq. (3.5)
vanishes. In our interpretation of vortices as 1-dimensional rods, this maximum number can be
understood from the fact that sum of the length of rods cannot exceed the compact period of S 1 ,
namely g 22kcR = kd 2R from the first equation of (3.2).
We can derive the van der Waals equation of state for the vortex gas in the thermodynamic
limit A , k with keeping k/A fixed (see e.g. 7.15 in [5]),

4k
= kT .
P A 2
(3.6)
g c
Here P is the pressure of the vortex gas and the Boltzmann constant is unity (kB = 1) in our
notation. From this we find the pressure of the vortex gas diverges at the maximal number of
vortices determined by the area and couplings. Compared with the general van der Waals equation of state (P + ak 2 /A2 )(A kb) = kT we have a = 0 and b = 4/(g 2 c). The former can be
understood as the BPS property that there exists no potential energy between BPS vortices. The
latter implies the size (exclusion area) of the ANO vortex is 4/(g 2 c).
4. Local/semi-local non-Abelian vortices
So far we have seen that the partition function for multi-vortex system can be calculated by
using a system of hard rods in the case of the AbelianHiggs model (NC = NF = 1). We next
128
1 ) as the dualized configuration of k vortices. Kinks represent kNF domain walls.

Fig. 5. Eigenvalues of (x
extend this method to the more general cases of non-Abelian and Abelian gauge theories: local
non-Abelian vortices (NC = NF > 1) and semi-local (non-Abelian) vortices (NF > NC ). These
cases were not known previously, which proves the power of our method. To this end, we treat
vortices in a model with a twisted boundary condition

H x 1 , x 2 + 2R = H x 1 , x 2 e2iRM ,
(4.1)
where M = diag(1 /R, . . . , NF /R), 1 < < NF < 1 + 1 . We can reproduce the partition function with ordinary boundary condition by taking the limit A 0 after calculating the
partition function with the twisted boundary condition. In this case, k-vortex configuration corre 1 ) as shown in Fig. 5.
sponds to kNF domain walls represented as kinks of the eigenvalues of (x
By taking the limit (3.1), we can identify each domain wall with a 1-dimensional rod as before.
1 ) which correspond to domain walls have the same slope, which is
In this limit, all kinks of (x
1/(dR) = g 2 c/2. In this case there are NF types of rods whose masses are given by
mA = 2c(A+1 A ),
A = 1, . . . , NF 1,
mNF = 2c(1 + 1 NF ),
(4.2)
and the period of phases are given by

lA = 2R/(A+1 A ),
A = 1, . . . , NF 1,
lNF = 2R/(1 + 1 NF ).
(4.3)
Then we can calculate the partition function for the multi-vortex system in terms of the
1-dimensional hard rods. Note that we have to take into account the fact that there are k indistinguishable sets of rods corresponding to k indistinguishable vortices. Therefore the partition
function should be divided by k after integrating over the configuration space of distinguishable
rods. Then the partition function for multi-vortex system takes the form of
k
kNF
NF
1
T
NC ,NF
mA lA
dx1 dxkNF
Zk,T 2 =
2
k

=
T
2
kNF
A=1
(2c)kNF (2R)kNF
1
k

dy1 dykNF .
X
(4.4)
(a)
129
(b)
1 ) and slant configuration for NC = NF = 2, k = 2. (a) Configuration for

Fig. 6. The profile of eigenvalues of (x
NC = NF = 2, k = 2; (b) slant torus configuration.
(a)
(b)
Fig. 7. (a) Configuration of rods ( ) and particles () pierced by the rods for NC = NF = 2, k = 2 corresponding
Fig. 6. (b) Those for NC = NF = 3, k = 2. Each rod is divided by particles into the NC parts with different colors. Each
part of rods can overlap with parts of the other rods with different colors but not with parts with the same color.
Here the integration with respect to positions {xn } is taken over the configuration space of the
rods, which is denoted as X. In the last line, we have redefined the coordinates as yn = xn
A d if the nth rod sits between the Ath flavor and the (A + 1)th flavor, and the corresponding
domain of integration is denoted as X . The redefinition of the positions corresponds to consider
a slant torus (Fig. 6) where domain wall configurations are right-angled and perpendicular to the
x 1 -direction. Therefore we can associate the configurations of vortices with that of rods with
length d and particles with zero size pierced by the rods. Recalling the original profile of the
kink solution, we can see there exist some overlapping rules of the rods; a part of rods which
is divide by the particles can overlap with each other, but not be allowed for some of these.
So we also introduce colored rods in order to clarify the exclusion rules of rods. (See Fig. 7.)
Note that the above formula in the last line is completely independent of the parameters {A }.
Therefore, our result with the twisted boundary condition (4.1) (non-vanishing A ) is applicable
to the ordinary case with non-twisted boundary condition (A = 0).
The partition function (4.4) can be interpreted as the asymptotic form of partition function of
vortex gas on the rectangular torus in the limit Eq. (3.1). Let us assume that the partition function
is independent of details of the torus and depends only on the area of torus A. This is the case
for the partition function in the model with NC = NF = 1. Then we can calculate the partition
function for vortex gas on a surface which is topologically a rectangular torus with area A. In
addition, we can show that Eq. (4.4) gives the exact form of partition function as follows: The
limit Eq. (3.1) can be rewritten as
R=
R0
,
c = c0 ,
R0 , c0 , g: fixed.
(4.5)
130
From the explicit expression of the metric of moduli space [15] and dimensional analysis, we
can show that the partition function takes the form of
N ,NF
Zk,TC 2

= (cT )kNF AkNF f g 2 cA

= (c0 T )kNF A0 kNF f g 2 c0 A0 ,
A0 (2)2 R0 R ,
(4.6)
where f (g 2 cA) is an unknown function of dimensionless parameter g 2 cA. From this expression,
we find that the partition function is independent of . Therefore the limit gives the exact
partition function. In the discussion below, we assume that the partition function depends only on
the area of torus A. We will check that the partition function (4.4) indeed gives the exact result
for one vortex with NC = NF = N and NF > NC = 1.
We now perform the integration in (4.4) explicitly in some cases, all of which are new results.
4.1. Two (k = 2) local non-Abelian vortices with NC = NF = 2
Let us consider the domain of integration for NC = NF = 2 and k = 2 for example. The
domain can be easily seen from Figs. 6 and 7(a). In this case there are two rods with one particle
inside each of them. These rods can overlap with each other contrary to the case of NC = NF = 1.
However the edges of the rods cannot overlap with the particles inside the other rods. Therefore
the domain of integration is given by

0 < y1 < 2R , y1 < y2 < y1 + d, y2 < y3 , y1 + d < y4
X = (y1 , y2 , y3 , y4 )
.
y3 < y4 < y3 + d, y3 + d < y2 + 2R , y4 < y1 + 2R
(4.7)

Performing the integral in Eq. (4.4) over the domain (4.7), we immediately obtain the partition
function of the vortices (the volume of the domain wall (rods with particles) configuration space)

2

2 8
8
1
4 4
A
A
(cT
)
for 2 A,
2
2
2
3g c
g c
g c
NC =2,NF =2
Zk=2,T
=

2
2
4
8
1
16
4
(cT )4 A 2
for 2 A 2 .
A 2 A
6
g c
g c
g c
g c
(4.8)
Note that there exists a lower bound for the area A 4/(g 2 c) which corresponds to the Bradlow
limit (2.5). This is a new result which was not derived previously. Let us explain the physical
meaning of each factor restricting us to the first case. Remembering that each vortex carries an
internal orientation of CP 1 , we can understand that the factor (4/(g 2 c))2 corresponds to the
internal orientations (see Eq. (4.9) with N = 2 below) while the factor A does to the center of
mass. Then the remaining factor A 23 g8
2 c can be thought of as effective area of relative motion
moduli of two non-Abelian vortices. Comparing this result to that of two Abelian vortices in
Eq. (3.5) with k = 2, the extra factor 2/3 appears here, which implies that the effective area of
non-Abelian vortices is smaller than that of Abelian vortices. Interestingly, the factor 2/3 in the
exclusion area is different from the factor 1/2 in the Bradlow area B2 = 12 g4
2 c , contrary to the
Abelian case Eq. (3.6) in which the both factors coincide. We will see this in more detail around
(4.12), below.
(a)
131
(b)
1 ) on slant torus.
(i) Configurations of eigenvalues of (x
(a) NC = NF = 5, k = 1; (b) NC = 1, NF = 4, k = 2.
(a)
(b)
(ii) Configurations of rods and particles.

(a) NC = NF = 5, k = 1; (b) NC = 1, NF = 4, k = 2.
1 ) and corresponding rods and particles. (i) (a) There
Fig. 8. Examples of stand configurations on eigenvalues of (x
are 4 small kinks trapped in the region with length d. (i) (b) There are 3 small kinks trapped in each interval between
two regions with length d. (ii) (a) There are 4 particles trapped inside a rod. (ii) (b) There are 3 particles trapped in each
interval between 2 hard rods.
4.2. k local non-Abelian vortices with NC = NF = N

Let us next consider the single (k = 1) non-Abelian vortex. In this case the moduli space is
T 2 CP N1 . The corresponding configuration has a rod and N 1 particles trapped inside a
rod, see Fig. 8(ii) (a).
The partition function (4.4) is defined only in the region A 4/(g 2 c) since the number of
1 ) is NC and the period 2R should be larger than the length d of a rod:
eigenvalues of (x
namely this bound is the second inequality in Eq. (3.2),
N
1
T
NC =NF =N
Zk=1,T 2
=
(2c)N (2R)N 2R
d N1
2
(N 1)!

1
4 N1
=
.
(4.9)
(cT )N A 2
(N 1)!
g c
In this simple case, we can confirm that our result agrees with an explicit integration over the
exact metric on the non-Abelian vortex moduli space. Although the solution is not known, the
metric with the Khler class can be calculated [12,31], to give
ds 2 = 2c dz d z +
k 2
i j
8 i j (1 + |b | ) b b
dbi d b j ,
g2
(1 + |bk |2 )2
(4.10)
where bi are the inhomogeneous coordinates of orientational moduli CP N1 . Using this metric,
we can confirm that Eq. (4.9) agrees with the correct partition function.
132
For general number k of the non-Abelian vortices with NC = NF = N , we can calculate the
partition function (4.4) up to the next leading term in the expansion in terms of 1/A
N =N =N
Zk,TC 2 F

1
A
4 N1 k
= (cT )
k! (N 1)! g 2 c

k
4 2
1 DN (k 1) + O
.
A
g 2 cA
kN
(4.11)
We can show that the coefficient DN in Eq. (4.11) takes the form (the definition of DN is given
in Appendix A)
(2N 2)!!
DN
2 8 16 128
=
= 1, , , ,
,...
2
3 15 35 315
4/g c (2N 1)!!
(N = 1, 2, 3, 4, 5, . . .).
(4.12)
The first (leading) term in the partition function (4.11) represents the situation that all vortices
are separated, while the second (next leading) term implies that a pair of adjacent vortices is
overlapped. From the partition function (4.11), we can obtain the equation of state for dilute
vortex gas in the thermodynamic limit A , k with fixing k/A to enough small value
as
2

k
k
= kT .
P A 1 DN + O
(4.13)
A
A2
From this we see that the second virial coefficient DN represents the effective area of the nonAbelian vortices in dilute gas. We find that the third term O(k 2 /A2 ) does not vanish. This implies
that the equation of state deviates from van der Waals one, contrary to the Abelian case Eq. (3.5).
We find that DN < D1 = g4
2 c and DN behaves as
1
DN
2
4
N g2c
(4.14)
for large N . We see that the non-Abelian vortices can be closer to each other than the Abelian
vortices even though the area of individual vortex is g4
2 c for both non-Abelian and Abelian
vortices. Therefore we conclude that non-Abelian vortices are softer than Abelian vortices.
Clearly the inequality BN < DN holds with BN = N1 g4
2 c being the Bradlow area in (2.5). This
inequality implies that the effective area DN of a vortex in dilute vortex gas is larger than the
Bradlow area BN , which is the effective area in highly pressured gas. An intuitive understanding
of this inequality is as follows. It is known that the internal orientations of two non-Abelian
vortices are almost always aligned when the two vortices are approaching each other [20] (as
long as their speed is sufficiently slow). In other words, nearby non-Abelian local vortices behave
as if they are Abelian local vortices.
4.3. k semi-local vortices with NC = 1 and general NF
Finally we show that our method can be extended to the case of semi-local vortices with
NF > NC . Here we concentrate on the NC = 1 case which indicates essential features of semilocal vortices. The corresponding configuration has NF 1 particles trapped in each interval
between k hard rods, see Fig. 8(b). In this case, Eq. (4.4) reduces to
kNF
1
T
1
N =1,N
(2c)kNF (2R)kNF
Zk,TC 2 F =
2R (2R dk)kNF 1
2
k (kNF 1)!

1
1
4k kNF 1
= (cT )kNF
,
A A 2
k (kNF 1)!
g c
133
(4.15)
where 1/k factor is needed since k vortices cannot be distinguished. Substituting NF = 1 into
this result, the partition function of k ANO vortices in Eq. (3.5) is correctly recovered. From this
partition function, we can obtain the equation of state for the vortex gas in the thermodynamic
limit A , k with k/A finite as

4k
P A 2
(4.16)
= kNF T ,
g c
where P is the pressure. The factor NF in the right-hand side of Eq. (4.16) appears due to the
fact that the vortices have additional internal degrees of freedom in the case of NF > 1.
In this case, no field theoretical result has been known yet. Especially we can show that the
moduli space of k = 1 semi-local vortex (see Fig. 8(b)) is T 2 CP NF 1 by extending the moduli
matrix formalism [15] to T 2 . Second, as was done in [31], we can compute the moduli space
metric by using the integration formula of the Khler potential in [17], to give

k 2
i j
4 i j (1 + |b | ) b b
2
ds = 2c dz d z + 2c A 2
(4.17)
dbi d b j .
g c
(1 + |bk |2 )2
Here bi in the second term are the inhomogeneous coordinates of CP NF 1 and represent moduli
not localized around vortices. These moduli become non-normalizable on the plane C (A ).
However they contribute to the Khler potential since we are considering the vortices on the
compact manifold T 2 with the finite area A. From this metric we can confirm at least for k = 1
case that Eq. (4.15) gives the correct partition function.
5. Comments and discussions
We here comment on an important fact that there is a duality on the partition function (4.4),
without performing the difficult integration. That is, the partition function for k NC is invariant
under exchange

NC , g 2 N C , g 2 ,
with N C k + NF NC ,
4
4
A 2 .
2
g c
g c
(5.1)
One can confirm that this duality exists between the partition function (4.9) and the one of the
k = 1 case of (4.15). Also, the example of NC = NF = 2, k = 2 is self-dual. In fact, in the partition
function (4.8), the latter case (A/2 4/(g 2 c)) can be obtained by exchanging 4/(g 2 c) by
A 4/(g 2 c) in the former case (A/2 4/(g 2 c)). These observations are very analogous to
the duality relation between the non-commutative instantons and monopoles [32,33]. From string
theoretical point of view, there appears a slant torus due to the effect of the magnetic gauge field
or equivalently the B-field. The B-field is a source of the non-commutativity and associated
with the FI parameters in the effective theory on the instantons. In our case, the gauge coupling
plays a role of the non-commutative parameter and should relate to the modulus of the slant
134
torus. Better explanation on the above duality may be found in terms of the duality between the
non-commutative vortex and domain wall in the string theoretical picture.
Finally we would like to discuss vortices on a sphere S 2 . In this case we cannot use the above
T-dual argument naively since the sphere consists of non-trivial U (1) fibration. However we
here assume that the vortices on the sphere is equivalent to the vortices on the cylinder with
finite length in the computation of the volume of the moduli space; The cylinder is topologically
isomorphic to the sphere with two puncture points. When at least one vortex sits on a puncture
point (the north or south pole) of the sphere, the dimension of the corresponding subspace is less
than the dimension of the full moduli space. Therefore this subspace does not contribute to the
volume of the moduli space. In contrast to the torus case, we do not need to identify the both
sides in the x 1 -direction. Thus it is now sufficient to consider the gas of the hard rods in a finite
segment with length 2R . By using the system of the hard rods, we obtain the partition function
of vortices in the AbelianHiggs model on S 2 as

1
1
4k k
NC =NF =1
k

k
k
k
= (cT ) (2R kd) (2R) = (cT ) A 2
.
Zk,S 2
(5.2)
k!
k!
g c
This completely agrees with the results in [2,4,5]. This case also should be extendable to the
non-Abelian and/or semi-local vortices.
In conclusion, we have proposed a novel and simple method to compute the partition function
of vortices at finite temperature. Our result agrees with previously known cases (3.5) and (5.2)
of the local Abelian (ANO) vortices in the AbelianHiggs model on T 2 and S 2 , respectively.
Our method provides new results in more general cases of non-Abelian local vortices, (4.8)
and (4.11), and Abelian semi-local vortices, (4.15). In the two cases of k = 1, NC = NF = N
and k = 1, NC = 1 with general NF , we have confirmed that our results agree with explicit integrations over the exact moduli metrics (4.10) and (4.17), respectively. We have found that
non-Abelian vortices are reduced under T-duality to soft rods with particles inside them while
the Abelian vortices are to hard rods.
Our results will be applied to the thermal vortex gas in the early Universe or in superconductor. Extension to non-(or near-)critical coupling will be important to discuss more realistic
application. In particular the phase transition of vortex gas in the AbelianHiggs model was discussed previously [34,35]. Presence or absence of phase transitions in the non-Abelian and/or
semi-local vortices is very interesting to explore.
Acknowledgements
This work is supported in part by Grant-in-Aid for Scientific Research from the Ministry
of Education, Culture, Sports, Science and Technology, Japan Nos. 17540237 and 18204024
(NS). K. Ohta is supported in part by the 21st Century COE Program at Tohoku University Exploring New Science by Bridging Particle-Matter Hierarchy. The work of K. Ohashi
and M.E. is supported by Japan Society for the Promotion of Science under the Post-doctoral
Research Program. T.F. gratefully acknowledges support from a 21st Century COE Program
at Tokyo Tech Nanometer-Scale Quantum Physics by the Ministry of Education, Culture,
Sports, Science and Technology, and support from the Iwanami Fujukai Foundation. The authors thank the Yukawa Institute for Theoretical Physics at Kyoto University. Discussions during
the YITP workshop YITP-W-06-11 on String Theory and Quantum Field Theory were useful to complete this work. K. Ohta would like to thank S. Matsuura for useful discussions and
comments.
135
Appendix A. Virial expansion

In this appendix we give the definition of DN which appears as a coefficient of the next
leading term in the partition function (4.11). We consider k non-Abelian local vortices with
NC = NF = N on a torus T 2 . As we have mentioned, the system can be well described from
the picture of k soft rods with length d, each of them piercing N 1 particles therein, on S 1
in circumference L = 2R , as shown in Fig. 5. We can carry out an integration with respect
to only k parameters corresponding to positions of the rods and we find that the volume of the
configuration space can be rewritten as

1
VN,k
d kN y
k

k1

k
d
(Ld N1 )k
k(N1)
d
z max 1
MN,i , 0
,
=
(A.1)
k!
L
i=1
N,k
where the integration parameters z = zr,i (r = 1, . . . , N 1, i = 1, . . . , k) are dimensionless and

correspond to (N 1) relative positions of the particles inside each rod, and their integration
region is defined by N,k = {{zr,i } | 0 z1,i z2,i zN1,i 1, 1 i k}. Here we
have defined a dimensionless function, MN,i , for each rod (1 i k),
MN,i = max(z1,i , z2,i z1,i+1 , . . . , zN1,i zN2,i+1 , 1 zN1,i+1 ),
(A.2)
where we have defined zr,k+1 = zr,1 . All of them give the same contributions to the
integration.
Note that dimensionless parameter d/L appears in the integrand. We
find
kd

d
i MN,i in

this case of local vortices. Thus small kd/L guarantees that L d i MN,i in any point of the
integration region and that the integrand can be expanded with respect to d/L, since the maxfunction in Eq. (A.2) can be ignored.1 The DN appears as the coefficient of the next leading term
in this expansion of VN,k , namely in the virial expansion of soft rods in one dimension:

2
1 Ld N1 k
d
d
VN,k =
1 DN k(k 1) + O
,
k! (N 1)!
L
L2
(A.3)
where we have defined D N as an average value of MN,1 :

2
1

dzN1 dz N1 max(z1 z1 , . . . , zN1 zN1
, 0),
D N = + (N 1)!
N
0z1 zN1 1

0z1 zN1
1
where D N relates with DN by DN = D N g4

2 c = DN 2Rd. We can easily carry out this integration in each region where a definite order of the integration parameters such as z1 < z1 < z2 <
z2 < . It is convenient to divide the integration regions into several ones whose integral become identical. Therefore calculation for DN reduces to counting number of elements of the sets
and we have succeeded this counting and obtained the result as (4.12).
1 In the case of semi-local vortices, N > N > 1, we cannot take the similar operation since we cannot naively remove
F
C
the max-function from the integral representation (A.1) due to their size moduli integral.
136
References
[1] N.A. Nekrasov, Adv. Theor. Math. Phys. 7 (2004) 831, hep-th/0206161;
N.A. Nekrasov, hep-th/0306211.
[2] N.S. Manton, Nucl. Phys. B 400 (1993) 624.
[3] P.A. Shah, N.S. Manton, J. Math. Phys. 35 (1994) 1171, hep-th/9307165.
[4] N.S. Manton, S.M. Nasir, Commun. Math. Phys. 199 (1999) 591, hep-th/9807017;
S.M. Nasir, Phys. Lett. B 419 (1998) 253, hep-th/9807020.
[5] N.S. Manton, P. Sutcliffe, Topological Solitons, Cambridge Univ. Press, Cambridge, 2004.
[6] N.M. Romao, J. Phys. A 38 (2005) 9127, hep-th/0503014.
[7] A.A. Abrikosov, Sov. Phys. JETP 5 (1957) 1174, Zh. Eksp. Teor. Fiz. 32 (1957) 1442;
H.B. Nielsen, P. Olesen, Nucl. Phys. B 61 (1973) 45.
[8] A. Hanany, D. Tong, JHEP 0307 (2003) 037, hep-th/0306150.
[9] R. Auzzi, S. Bolognesi, J. Evslin, K. Konishi, A. Yung, Nucl. Phys. B 673 (2003) 187, hep-th/0307287.
[10] D. Tong, Phys. Rev. D 69 (2004) 065003, hep-th/0307302.
[11] R. Auzzi, S. Bolognesi, J. Evslin, K. Konishi, Nucl. Phys. B 686 (2004) 119, hep-th/0312233.
[12] M. Shifman, A. Yung, Phys. Rev. D 70 (2004) 045004, hep-th/0403149.
[13] A. Hanany, D. Tong, JHEP 0404 (2004) 066, hep-th/0403158.
[14] M. Eto, Y. Isozumi, M. Nitta, K. Ohashi, N. Sakai, Phys. Rev. Lett. 96 (2006) 161601, hep-th/0511088.
[15] M. Eto, Y. Isozumi, M. Nitta, K. Ohashi, N. Sakai, J. Phys. A 39 (2006) R315, hep-th/0602170.
[16] M. Eto, K. Konishi, G. Marmorini, M. Nitta, K. Ohashi, W. Vinci, N. Yokoi, Phys. Rev. D 74 (2006) 065021,
hep-th/0607070.
[17] M. Eto, Y. Isozumi, M. Nitta, K. Ohashi, N. Sakai, Phys. Rev. D 73 (2006) 125008, hep-th/0602289.
[18] K. Hashimoto, D. Tong, JCAP 0509 (2005) 004, hep-th/0506022;
R. Auzzi, M. Shifman, A. Yung, Phys. Rev. D 73 (2006) 105012, hep-th/0511150.
[19] M. Eto, L. Ferretti, K. Konishi, G. Marmorini, M. Nitta, K. Ohashi, W. Vinci, N. Yokoi, Nucl. Phys. B 780 (2007)
161, hep-th/0611313.
[20] M. Eto, K. Hashimoto, G. Marmorini, M. Nitta, K. Ohashi, W. Vinci, Phys. Rev. Lett. 98 (2007) 091602, hepth/0609214.
[21] D. Tong, JHEP 0612 (2006) 051, hep-th/0610214;
A. Ritz, hep-th/0612077.
[22] L.G. Aldrovandi, F.A. Schaposnik, hep-th/0702209.
[23] D. Tong, Phys. Rev. D 66 (2002) 025013, hep-th/0202012.
[24] M. Eto, T. Fujimori, Y. Isozumi, M. Nitta, K. Ohashi, K. Ohta, N. Sakai, Phys. Rev. D 73 (2006) 085008, hepth/0601181.
[25] N.S. Manton, Phys. Lett. B 110 (1982) 54.
[26] S.B. Bradlow, Commun. Math. Phys. 135 (1990) 1.
[27] N.D. Lambert, D. Tong, Nucl. Phys. B 569 (2000) 606, hep-th/9907098.
[28] M. Eto, Y. Isozumi, M. Nitta, K. Ohashi, K. Ohta, N. Sakai, Phys. Rev. D 71 (2005) 125006, hep-th/0412024.
[29] A. Hanany, E. Witten, Nucl. Phys. B 492 (1997) 152, hep-th/9611230.
[30] Y. Isozumi, M. Nitta, K. Ohashi, N. Sakai, Phys. Rev. Lett. 93 (2004) 161601, hep-th/0404198;
Y. Isozumi, M. Nitta, K. Ohashi, N. Sakai, Phys. Rev. D 71 (2005) 065018, hep-th/0405129;
Y. Isozumi, M. Nitta, K. Ohashi, N. Sakai, Phys. Rev. D 70 (2004) 125014, hep-th/0405194;
M. Eto, Y. Isozumi, M. Nitta, K. Ohashi, K. Ohta, N. Sakai, Y. Tachikawa, Phys. Rev. D 71 (2005) 105009, hepth/0503033.
[31] M. Eto, Y. Isozumi, M. Nitta, K. Ohashi, N. Sakai, Phys. Rev. D 72 (2005) 025011, hep-th/0412048.
[32] A. Hashimoto, K. Hashimoto, JHEP 9911 (1999) 005, hep-th/9909202;
K. Hashimoto, H. Hata, S. Moriyama, JHEP 9912 (1999) 021, hep-th/9910196;
K. Hashimoto, T. Hirayama, Nucl. Phys. B 587 (2000) 207, hep-th/0002090.
[33] D.J. Gross, N.A. Nekrasov, JHEP 0007 (2000) 034, hep-th/0005204.
[34] P.A. Shah, Nucl. Phys. B 438 (1995) 589, hep-th/9409145.
[35] K. Kajantie, M. Laine, T. Neuhaus, A. Rajantie, K. Rummukainen, Nucl. Phys. B 559 (1999) 395, hep-lat/9906028.
Conformal boundary loop models

Jesper Lykke Jacobsen a,b, , Hubert Saleur b,c
a LPTMS, Universit Paris-Sud, Btiment 100, Orsay 91405, France
b Service de Physique Thorique, CEA Saclay, Gif-sur-Yvette 91191, France
c Department of Physics and Astronomy, University of Southern California, Los Angeles, CA 90089, USA
Received 28 March 2007; accepted 29 June 2007

Abstract
We study a model of densely packed self-avoiding loops on the annulus, related to the TemperleyLieb
algebra with an extra idempotent boundary generator. Four different weights are given to the loops, depending on their homotopy class and whether they touch the outer rim of the annulus. When the weight of a
contractible bulk loop x q + q 1 (2, 2], this model is conformally invariant for any real weight of the
remaining three parameters. We classify the conformal boundary conditions and give exact expressions for
the corresponding boundary scaling dimensions. The amplitudes with which the sectors with any prescribed
number and types of non-contractible loops appear in the full partition function Z are computed rigorously.
Based on this, we write a number of identities involving Z which hold true for any finite size. When the
[r+1]
weight of a contractible boundary loop y takes certain discrete values, yr [r] q with r integer, other
q
identities involving the standard characters Kr,s of the Virasoro algebra are established. The connection
with Dirichlet and Neumann boundary conditions in the O(n) model is discussed in detail, and new scaling
dimensions are derived. When q is a root of unity and y = yr , exact connections with the Am type RSOS
model are made. These involve precise relations between the spectra of the loop and RSOS model transfer
matrices, valid in finite size. Finally, the results where y = yr are related to the theory of TemperleyLieb
cabling.
* Corresponding author at: LPTMS, Universit Paris-Sud, Btiment 100, Orsay 91405, France.
E-mail address: jesper.jacobsen@u-psud.fr (J.L. Jacobsen).

doi:10.1016/j.nuclphysb.2007.06.029
138
J.L. Jacobsen, H. Saleur / Nuclear Physics B 788 [FS] (2008) 137166
1. Introduction
Boundary conformal field theories (CFT) have lately played an increasingly important role
in statistical mechanics, condensed matter physics and string theory. In statistical mechanics,
they appear in most probabilistic applications of geometrical models (see, e.g., [1] for a recent
example), in particular through SLE [2]. In condensed matter, they contain all the information
about fixed points in theories which are gapless in the bulk, such as Kondo systems or edge
states in the fractional quantum Hall effect (see [3] for a review). In string theory, they provide
for instance microscopic techniques to study D-branes in curved backgrounds [4]. On top of this,
the study of boundary aspects is a crucial component of understanding and classifying CFTs at
large [5], and has lately played a crucial role in the solution of non-rational CFTs [6].
While progress in understanding conformal boundary conditions for rational CFTs has been
considerable, the situation is not so satisfactory for non-rational theories, which are however all
too frequent in statistical mechanics applications. A case in point concerns the loop or cluster
models, whichin one guise or anotherare hidden behind most simple models of interest,
such as Q-state Potts models, O(n) models, RSOS models, polymers, and percolation. We are
not aware of answers to most questions one might ask in this context, such as what are all
the conformal invariant boundary conditions, or what are the partition functions for the O(n)
models with Neumann boundary conditions, etc. The origin of this difficulty lies in our lack
of understanding of the bulk CFTs, which exhibit non-rational, logarithmic features, and for
which too little is known. Bulk exponents turned out to be tractable thanks to the Coulomb gas
technique, but this technique has not been generalized to the boundary case with sufficient control
yet (see [1] for recent progress in this direction).
We put forward in this paper a proposal for what we believe are all the conformal boundary
conditions of dense loop models. For each of those we determine the critical exponents, operator content and boundary partition functions, some of which have interesting probabilistic or
geometrical interpretations.
There seems to be much substance behind the results we uncover, and we hope to get back to
the question in more details in the near future. In the present paper, we only present the leading
argumentswhich are based on previously published but unexploited results, as well as algebraic
considerationstogether with intensive numerical checks and some combinatorial proofs.
To help the reader, we now give a quick summary of our results and notations. The boundary
loop model (BLM) to be studied is defined on a tilted square lattice (see Fig. 3), wrapped on an
annulus of width N strands and circumference M lattice spacings. Loops cover all the edges, and
interact in a specific way with the outer rim of the annulus, whereas they are simply reflected by
the inner rim (free boundary conditions). We denote by L the number of non-contractible loops
(note that L and N must have the same parity). Any loop has one of four weights (x, y, l or m,
see Fig. 8): l (respectively m) for a non-contractible loop never touching (respectively touching
at least once) the outer rim, and similarly x (respectively y) for contractible loops. We parametrize x = q + q 1 (2, 2] by q = ei/(p+1) (p real); the model is then critical with central
charge (2.7) for any real values of y, l, m and is endowed with the Uq (sl2 ) quantum group symmetry. We further parametrize y = y(r) as in (2.4). Our central claim is that for any real r, and
any L, there are two (distinct for L > 0) conformal boundary conditions: blobbed (respectively
unblobbed) in which the outermost non-contractible loop is required to (respectively required
not to) touch the outer rim of the annulus. (When L = 0 the two cases coincide.) The spectrum
generating functions in these two cases are (3.8), and the boundary conformal weights (critical
exponents) hr,rL are read off from (2.8). They combine to form the BLM partition function Z
139
through the amplitudes (3.11). When p 1 is integer, and when further r = 1, 2, . . . , p the BLM
model can be related to an RSOS model of the Ap type with specific boundary conditions (three
columns of fixed heights, see Fig. 12) through the rules (4.4). In the latter case, Z can be written
as a sum (3.18) over irreducible representations of the Virasoro algebra.
The paper is organized as follows. In Section 2 we review the algebraic framework used in our
study (the blob algebra) along with a few key results. In Section 3 we define the BLM, classify its
conformal boundary conditions, and give exact results for the associated critical exponents. Appendices A and B present a rigorous result on the amplitudes of the transfer matrix eigenvalues.
This is used in Section 3.4 to write a number of exact identitiesexact in finite sizerelating Z
to Virasoro characters Kr,s . In Section 3.5 we discuss the case of Neumann boundary conditions for the loop model, identifying in particular the Neumann to Dirichlet boundary condition
changing field. The relations to RSOS models are discussed in Section 4. Finally, in Section 5,
we comment on the relation between the blob algebra and the theory of TemperleyLieb cabling.
Our conclusionsand the prospects for (much) further workare given in Section 6.
2. The blob algebra
The TemperleyLieb (TL) algebra TN (x) on N strands is defined by the generators ei (i =
1, 2, . . . , N 1) acting on strands i and i + 1 and satisfying the well-known relations
ei ej = ej ei
for |i j | 2,
ei ei1 ei = ei ,
ei2 = xei .
(2.1)
In [7] this was generalized into the two parameter blob algebra BN (x, y), having an extra
generator b, and satisfying in addition the relations
b2 = b,
e1 be1 = ye1 ,
ei b = bei
for i = 2, 3, . . . , N 1.
(2.2)
For TN (x) it is well known how to interpret these algebraic relations graphically in terms
of the strands. The extra generator b marks the leftmost strand by adding a blob to it (see
Fig. 1). In this graphical representation any completed loop may be taken out and replaced by its
corresponding weight (see Fig. 2). A loop with no blob gets the usual weight of x, while a loop
with a blob gets a modified weight y. Note that several blobs on the same loop reduce to a single
blob. Obviously, only loops touching the left border can be blobbed.
This algebra has given rise to much work in recent years in the mathematical literature [8].
It has also been studied in the context of boundary conformal field theory [11] with results that
have some small overlap with ours. (In [11] this algebra is called the one boundary TL algebra,
a name we shall not adopt.) The blob algebra is in fact a quotient of the more general affine
Hecke algebra (like TL itself is a quotient of the ordinary Hecke algebra).
The representation theory of the two parameter algebra is richer than the representation theory
of the TL algebra. For a given x = q + q 1 , and assuming first that q = ei is not a root of unity,
exceptional cases occur whenever
y=
sin(r 1)
,
sin r
r integer.
(2.3)
140
Fig. 1. Graphical representation of the action of the generators of the blob algebra, here shown for a system of N = 4
strands.
Fig. 2. The word e1 e3 e6 e2 e6 be1 e3 = xye1 e3 e6 in the blob algebra B(x, y) on N = 7 strands.
[In the original paper [7] this corresponds to = r mod in the basic equation y = (q
q 1 e2i )/(1 e2i ).]
In the case
y=
sin(r + 1)
,
sin r
r integer,
the spectrum of the Hamiltonian

N1
ei
ab
H=
sin
(2.4)
(2.5)
i=1
(the normalization guarantees unit sound velocity) has been studied in the continuum most recently in [11] where it was found for any a > 0 to give rise to the generating function of scaled
gaps, in the sector with L non-contractible lines1 propagating:
Kr,r+L = Tr q L0 c/24 =
q hr,r+L q hr,rL
.
q c/24 P (q)
(2.6)
Here, as usual, L0 is a Virasoro generator,

c=1
6
p(p + 1)
is the central charge for =

hr,s =
(2.7)
p+1 ,
[(p + 1)r ps]2 1

4p(p + 1)
and
(2.8)

n
are the conformal weights of the Kac table. Moreover, P (q) =
n=1 (1 q ). We stress that in
(2.6) we have q = exp(2i ), where is the standard modular parameter. As this meaning of q
will be reserved for the argument of the spectrum generating functions, no confusion should arise
1 A non-contractible line is a strand propagating throughout the system. Thus, e acting on two non-contractible lines
i
at i and i + 1 is zero by definition. Fig. 2 has L = 1 non-contractible line running from the bottom to the top of the figure
(top and bottom are later wrapped onto an annulus). Note also that L and N must have the same parity.
141
with the other meaning of q as the quantum group deformation parameter q = ei appearing in
the parameterization of x.
The case r = 1 of (2.6) is the usual Hamiltonian for the loop model with free boundary conditions, since in this case y = 2 cos = x. In this case, the generating function of scaled gaps has
indeed been known for a very long time [12] to be given by K1,1+L . The case n > 1 as presented
in [11] appeared to be new, although it is in fact related to results in [12] (and has algebraic
connotations in terms of representation theory of the blob versus the Virasoro algebra). In this
paper, we shall discuss (2.6) further, interpret it in the language of the loop model, and correct it
whenever necessary.
We stress that the independence upon a is a truly remarkable phenomenon: it can be interpreted by saying that once the algebraic structure of the Hamiltonian is decided, the continuum
limit does not depend on the (boundary) details. Another way to view this independence is the following. The R matrix of the loop model with spectral parameter u, acting on strands i and i + 1,
sin(u)
can be written Ri (u) = 1 + f (u)ei , and f (u) = sin(
u) is easily found by solving the Yang
Baxter equations. Writing similarly the boundary matrix as B(u) = 1 + g(u)b, the Sklyanin (or
reflection) equations relate B(u) and R1 (u), giving rise to a solution for g(u) (see Eq. (40) in
[13]) that contains an arbitrary constant of separation . The arbitrariness of is analogous to
the a-independence discussed above.
3. Boundary loop model
We now want to study the blob algebra in the context of isotropic dense loop models, which are
described by a transfer matrix instead of a Hamiltonian. We recall that in the bulk, these models
are defined by dense coverings of the (tilted) square lattice with self avoiding and mutually
avoiding loops, each vertex allowing two possible configurations (see Fig. 3 and also Figs. 14,15
below), and each loop coming with a fugacity x. What happens at the boundary is the subject of
this paper.
The sum over TL generators in the Hamiltonian is replaced in that case by a product
T0 =
(N1)/2

(1 + e2i )
i=1
N/2

(1 + e2i1 ),
(3.1)
i=1
where denotes the integer part. By analogy we will supplement this by boundary contributions so the full transfer matrix reads
T = (1 + b)T0 .
(3.2)
In view of the a-independence of (2.6), we would expect the critical exponents associated
with T to be independent of . This independence is checked numerically below (see Fig. 4).
From a geometrical point of view the most natural choice is then = , so that after a trivial
rescaling T = bT0 . This can be interpreted as a lattice model for which every loop touching the
Fig. 3. Vertices of the boundary loop model on the tilted square lattice. Bulk vertices are in any of two different states
(corresponding to the generators 1 and ei ). Left boundary vertices are always blobbed (generator b), and right boundary
vertices are unblobbed (generator 1).
142
boundary gets a modified weight y instead of x. We will now study the conformal properties of
this model (which we call the boundary loop model), as a function of x and y.
3.1. Conformal boundary conditions and critical exponents
The results from [12] and more recently [11] suggest the following. Recall the parametrization
x = 2 cos ; for each y(r) we solve (2.4) to get r. This gives a real number r (0, ). The
leading eigenvalue in the loop model transfer matrix with L non-contractible lines should scale
with p integer and r = 1, 2, . . . , p

with conformal weight h(y) hr,r+L . For when = p+1
integer this is a rigorous consequence of [12] and the loop-RSOS correspondence, as will be
discussed in Section 4 below.
Fig. 4 serves the double purpose of checking this conjecture for L = 0 and p = 3 (the Ising
model), and establishing the independence of the exponents on the choice of in (3.2). The
agreement with the numerics is generally very good, and even in regions with strong corrections
to scaling (in particular y = 0) it should be noted that the finite-N effects have consistently
a trend and an amplitude compatible with the conjectured result in the thermodynamic limit. The
choice = appears to minimize the amplitude of the finite-size effects, and accordingly we
shall invariably adopt this choice for the subsequent numerical checks.
Another check, still for L = 0 but with a higher value p = 6 (the tricritical Potts model), is
shown in Fig. 5.
From the point of view of the blob algebra, the parametrization (2.4) has nothing special
compared with the other choice of sign in (2.3), which we rewrite here as
sin(r 1)
.
sin r
Of course by using symmetries of the sine function we can write as well
y(r ) =
(3.3)
sin(p + 2 r )
(3.4)
sin(p + 1 r )
so we see that the associated exponent reads as well hp+1r ,p+1r +L = hr 1,r L , by the
symmetry of (2.8).
y(r ) =
Fig. 4. (Color online.) Numerical check of the conjecture h(y) = hr,r as a function of y, here for L = 0 and p = 3
(the Ising model). The results were obtained from the finite-N corrections to the leading eigenvalue of the transfer
matrix (3.2), obtained by exact diagonalization techniques, for widths up to N = 20. The left (respectively right) panel
shows the choice = 1 (respectively = ), but the extrapolated results (N ) appear to be independent of . The
vertical lines represent the particular values (2.4).
143
Fig. 5. (Color online.) Another check of the conjecture h(y) = hr,r for L = 0, here for p = 6 (the tricritical Potts model).
We now have to make things a little more precise and technical. The sector with L noncontractible loops can in fact be considered from two points of view, depending on whether or
not the leftmost non-contractible loop is allowed to touch the left boundary. In Appendix A we
discuss in details the sector structure of the transfer matrix bT0 of (3.1) and show in particular
that each of its eigenvalues corresponds to a definite choice: either the leftmost contractible loop
is forced to touch the left boundary at least once, or it is forbidden from ever doing it. We shall
henceforth refer to these two cases as the blobbed (respectively the unblobbed) sector. Note that
in both sectors the contractible loops to the left of the leftmost non-contractible loop may of
course still touch the left boundary.
We claim that these two cases both correspond to conformal boundary conditions, which are
different. For entropic reasons, the largest eigenvalue in the blobbed sector is obviously greater
than the largest eigenvalue in the unblobbed sector. From the above arguments, the blobbed
sector therefore has the exponent hr,r+L indeed. The unblobbed sector meanwhile has a different
exponent hr,rL . This can be checked numerically, and is illustrated in Fig. 6. Note of course
that when L = 0 the two results actually coincide, as they should, since the two sectors are then
identical.
3.2. Relation to the Potts model
The boundary loop model is closely related to the Q = x 2 state Potts model at coupling (inverse temperature) J . Indeed, ignoring first boundary effects, the Potts model partition function
can be written as [9]

B
Z=
(3.5)
eJ 1 QC = QV /2
Q/2 .
clusters
loops
The first sum is over bond percolation clusters consisting of B bonds and C connected components. The second sum is over loops on the medial lattice that separate the clusters and their duals,
with being the number of loops and V the number of Potts spins. We have here supposed that
the model is defined on a square lattice and stands at its critical temperature, eJ 1 = Q1/2 . The
equivalence between the cluster and loop formulations is obtained by applying the Euler relation.
144
Fig. 6. (Color online.) Exponents h(y) in the presence of L = 2 non-contractible loops, here for p = 3 (the Ising model).
The lower (respectively upper) curves show the numerical results for the (constrained) boundary loop model. They agree
with the conjecture h(y) = hr,r+L (respectively h(y) = hr,rL ) as claimed.
Note that the local configurations of the loops correspond precisely to the first two vertices of
Fig. 3.
In the sector L = 0 we have claimed that the exponent of the boundary loop model is hr,r in
the parameterization (2.4). We now wish to check that this claim is consistent with known results
on the Potts model. To that end, consider the Q-state Potts model in the same annular geometry
as the boundary loop model. Denote by P1 and P2 two points on the left boundary, and let
boundary spins on the interval P1 P2 be constrained to take a subset of Qs states (with Qs Q).
In particular, when Qs = 1, this boundary condition corresponds to the Potts spins being fixed
on the interval P1 P2 and free on the remainder of the boundary. The modified partition function
reads

(P1 P2 )

B C Qs C(P1 P2 )
J
V /2
/2 Qs
e 1 Q
=Q
Q
,
Z(P1 P2 ) =
(3.6)
Q
Q
clusters
loops
where we have used the Euler relation as before. The number of clusters (respectively loops)
that touch P1 P2 is denoted C(P1 P2 ) (respectively (P1 P2 )), and obviously we have C(P1 P2 ) =
(P1 P2 ). We stress that C and still denote the total number of clusters and loops. Now, (3.6) is
a special case of the boundary loop model with the correspondence between weights
Q = x2,
Qs = xy.
(3.7)
In particular, for Qs = 1 we have r = p 1 in (2.4). The corresponding special case of

our general claim is therefore that the operator that changes the Potts model boundary conditions from free to fixed is p1,p1 = 1,2 . This indeed coincides with a well-known result of
Cardy [10].
Another verification is furnished by Q = 3, Qs = 2. The claim is then that the operator that
1
changes the boundary conditions from free to mixed is 2,2 with conformal weight h2,2 = 40
.
This is again a well-known result.
145
3.3. Spectrum generating functions

Further numerical study shows that the spectrum generating functions for the blobbed and
unblobbed sectors of the boundary loop model are simply the characters of generic irreducible
representations of the Virasoro algebra, i.e., respectively,
q hr,r+L c/24
,
P (q)
q hr,rL c/24
Unblobbed sector: ZL (r) =
.
P (q)
Blobbed sector:
ZL (r) =
(3.8)
This can be related with the structure of the basis of the blob algebra and the absence of truncation
of the Bratelli diagram [7]. The precise finite-size definition of ZL and ZL in terms of the transfer
matrix blocks TL and TL (defined in Appendix A) is given in (B.1)(B.2); note that we have set
j = L/2 in Appendix B.
The leading behavior in these expressions defines the exponents hr,rL and has already been
checked in Figs. 46. The coefficients of the terms up to level 6 in the development
1/P (q) = 1 + q + 2q 2 + 3q 3 + 5q 4 + 7q 5 + 11q 6 +
(3.9)
have been verified by computing the first 32 eigenvalues of T , for sizes up to N = 24, and looking
for integer gaps in the spectrum of critical exponents. For definiteness we have concentrated
on the case p = 3 and L = 2. Independent computations were made in the blobbed and the
unblobbed sectors. Moreover, the verification was made both for y = x and for a generic value
of y, and in either case the absence of singular vectors up to and including level 6 was ascertained.
It is important to stress that the spectrum generating functions are not given by (2.6) for the
BLM. The full loop transfer matrix actually contains more information, but can be truncated
when r is integer.
3.4. Partition function identities
We are now interested in the situation where the non-contractible loops wrap around the annulus. The question then arises of which weight should be given to these loops (while our results
for the exponents have some overlap with [11], the following has never appeared before). We
will give in general the weight l to non-contractible loops that do not touch the boundary (i.e.,
the outer rim of the annulus), and weight m to those that do (there is at most one). We now claim
that the full partition functions are given by adding sectors ZL (r) and ZL (r) with the following
amplitudes. We parametrize l, m in terms of two numbers , :
l = 2 cosh ,
sinh( + )
.
m=
sinh
(3.10)
Then the amplitudes are

sinh(L + )
,
sinh
sinh(L )
.
DL =
sinh()
DL =
(3.11)
146
Fig. 7. Pictorial representation of the recursion relation (3.13).
A complete proof of this statement is presented in Appendix B. Alternatively, we can argue

that, because one can get from a sector L to a sector L + 2 by adding two non-contractible
loops at the right of the annulus (here seen as a periodic strip), which can or not get contracted
with the first L ones, the generic amplitudes have to obey a recursion relation (which has a deep
algebraic nature, see [14]). Considering first the amplitudes DL0 for the simpler problem with
transfer matrix T0 (i.e., without the boundary generator b) one has
0
0
+ DL2
DL0 D20 = DL0 + DL+2
(3.12)
with initial values D00 = 1 and D20 = l 2 1. A pictorial rendering of this relation is shown in
Fig. 7. The solution reads explicitly DL0 = UL (l/2), where UL is the Lth Chebyshev polynomial
of the second kind. Turning now to the boundary loop model we must have similarly
DL D20 = DL + DL+2 + DL2 ,
+ DL2
.
DL D20 = DL + DL+2
(3.13)
The initial values can then be determined to be D0 = D0 = 1, D2 = l 2 lm 1 and D2 = lm 1,

from which the general formulas (3.11) follow. In addition to the actual proof of Appendix B,
we have checked (3.11) by formal manipulations of transfer matrices up to size N = 6, along the
lines of [15].
The amplitudes can then be used to write down the general partition functions in the case N
even (we set L = 2j )

sinh(2j + ) q hr,r+2j
sinh(2j ) q hr,r2j
c/24
Z=q
(3.14)
.
sinh
P (q)
sinh
P (q)
j =0
j =1
This partition function corresponds to the most general case represented in Fig. 8: contractible
loops in the bulk get the weight x, those touching the boundary a weight y, non-contractible
loops not touching the boundary a weight l and the others a weight m.
The simplest situation occurs when
sinh(r + 1)
,
(3.15)
sinh r
that is, the parameter m has the same formal algebraic relationship with l that y does with x.
Then = rthe same r as in (2.4)and one can write

sinh(2j + r) q hr,r+2j
sinh(2j r) q hr,r2j
c/24
Z=q
(3.16)
.
sinh r
P (q)
sinh r
P (q)
m=
j =0
j =1
147
Fig. 8. (Color online.) The most general model studied in this paper. We distinguish four types of self-avoiding loops on
the annulus: contractible loops touching (blue color in the figure) or not touching (green) the outer rim of the annulus,
and similarly for non-contractible loops (red or purple).
If moreover one is in a degenerate case where r is integer, one can reorganize the sum (3.16) into
Z=
r1

sinh(2j + r)
sinh(2j r) q hr,r2j
Kr,r+2j q c/24
.
sinh r
sinh r
P (q)
j =0
(3.17)
j =1
By pairing up terms with j and r j in the second sum, this can in turn be rewritten as
Z=

j =[r/2]
sinh(2j + r)
Kr,r+2j ,
sinh r
(3.18)
where denotes the integer part. When r is even, the contribution from j = r/2 actually
disappears. We thus get a sum over irreducible representations of the Virasoro algebra.
We claim that the subtractions occurring in the partition function do occur in finite size as
well. These subtractions involve the conformal weights hr,r+2j and hr,r2j and correspond
respectively to the blobbed sector with L = 2j non-contractible loops and the unblobbed sector
with L = 2j + 2r. In other words, we claim there are level coincidences in finite size between
the blobbed and unblobbed sectors when r is an integer: this in fact follows from the theory of
representations of the blob algebra.
To be more precise, one can make sense of expressions such as (3.18) in finite size by replacing the definition (2.6) of the characters Kr,s by traces of transfer matrix blocks, as outlined
in Appendix B. Care should then be taken that the annulus is wide enough to accommodate the
prescribed number of non-contractible lines, which amounts to replacing the upper limit in the
summation (3.18) by N/2. A numerical check of the level coincidences is shown in Table 1, for
the case N = 6, p = 6 and r = 1 (i.e., y = x). We find indeed that the level spectrum of the
, for all allowed
transfer matrix (see Appendix A) block T2j +2r is a proper subset of that of T2j
values of j (i.e., j = 0, 1, . . . , N/2 2r 1). We have checked this statement for several other
values of N , p and r, including on examples involving many more levels.
148
Table 1
Complete set of levels for the various transfer matrix blocks (cf. Appendix A) in the case of N = 6 strands, for p = 6
(the tricritical Potts model) and r = 1 (i.e., y = x). The left column shows N 1 log (rounded to 12 digits), where
1 is the transfer matrix eigenvalue, and the remaining columns show the multiplicities of within the various blocks.
A blank entry denotes zero multiplicity. The level coincidences mentioned in the text are clearly observed. (There is one
extra coincidence of the level 0.085859268861, between T2 and T4 .)
f
T0
0.338946565198
0.295466694605
0.247750936031
0.238365774752
0.210311165453
0.201367085486
0.195052854163
0.184121456296
0.171718537721
0.168641596700
0.149149353298
0.134931152674
0.125859871619
0.110023422061
0.097996769822
0.085859268861
0.069318933352
0.067354247176
0.048891555720
0.041781216253
0.036787385047
0.027592998174
0.012286435222
0.003076941021
0.000000000000
T2
T4
T6
T2
T4
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
Although we have restricted to N even, similar results can be written for N odd, with the
difference that L is odd (and thus never vanishes). The generating function then reads

sinh[(2j + 1) + ] q hr,r+2j +1
sinh[(2j + 1) ] q hr,r2j 1
c/24
Z=q
.
sinh
P (q)
sinh
P (q)
j =0
j =0
(3.19)
3.5. Neumann boundary conditions
If we go back to the derivation of the partition function for the O(n) model [16] we see that
Dirichlet boundary conditions for the O(n) variable will translate into having a loop extremity
on every point of the boundary, and the open lines thus obtained get a weight one (since the
O(n) variable is fixed to say S = (1, 0, . . . , 0) on the boundary). Let us assume this carries over
to the fully packed case and our geometry (see Fig. 9), where we thus demand that every point
on the boundary looks like the top diagram on Fig. 10, which we call a fork. We also require that
open loops thus formed all carry a weight unity. It is then easy to see that this is equivalent to
marking loops touching the boundary with a blob having parameter y = 1. In turn, with the usual
149
Fig. 9. (Color online.) Configuration in which the two boundaries of the annulus (here shown as a periodic strip) have
respectively Dirichlet and Neumann boundary conditions. RSOS faces are shown in black, and the loops that separate
clusters at constant height are green. The open lines going from the left to the left boundary carry unit weight.
Fig. 10. (Color online.) Two possible interactions between the loops and the left boundary: fork (top image) and blob
(bottom image).
parametrization for x we get the associated conformal weight to be

hTwist = hp/2,p/2 =
p2 4
16p(p + 1)
(3.20)
since y = 1 in (2.4) when r = p/2. This corresponds to a twist operator, or the dimension of the
boundary field changing boundary conditions from Neumann to Dirichlet.
We note that this dimension of the twist operator agrees with the one proposed in [17] after
some reinterpretation of the results. The formula given in this reference is slightly different
hK
Twist = h p+1 , p+1 =
2
(p + 1)2 4
16p(p + 1)
(3.21)
but turns out to hold for the dilute phase of the O(n) model [16] only [18]. For the dense case, it
has to be replaced by (3.20).
We can now write the partition function with Dirichlet boundary conditions on one rim of the
annulus and Neumann on the other, simply by setting r = p2 in our formulas. An interesting limit
150
Fig. 11. (Color online.) Exponents h(y) with L = 0 non-contractible loops, in the limit p . The exact result is a step
1 (shown as a big cross).
function, passing through the value h(1) = 16
to consider is then p , m = 1, l = 2 where we find from (3.14)

Z = q 1/24

q (2j +1) /16
P (q)
(3.22)
j =
which is the usual DirichletNeumann partition function for the free boson. The presence of the
two kinds of terms hr,rL in (3.14) is crucial to recover this limit.
Note that in the limit p , the exponent h(y) becomes a step function:
(L1)2
for y < 1,
4
2
h(y) = (2L1) for y = 1,
(3.23)
16
L2
for y > 1.
4
The numerical check, shown in Fig. 11, is compatible with this behavior, although the finite-size
corrections are of course large near the step.
Note that in general blob and fork do not coincide: this is true only when the associated
weights are equal to unity. One can play the game of defining a fork algebra such that the top
diagram in Fig. 10 defines the fork operator f . Required relations to give to every open loop
a weight z are then obviously
f 2 = zf,
e1 f e1 = ze1 .
(3.24)
But, by a rescaling f f/z we get the same relations as the blob relations with y = 1!
4. Relation to RSOS models
We now want to tackle the boundary loop model by another route. Recall that in the numerical
studies of Saleur and Bauer [12] it was found that for Ap RSOS models [with central charge
(2.7)] the annulus partition function is exactly the character da when the following boundary
151
Fig. 12. RSOS model in an annular geometry, for N = 6. Heights on the rightmost and the two leftmost layers are fixed
as shown. Time flows vertically, and there are periodic boundary conditions in the time direction.
conditions are imposed: all heights on the right boundary of the annulus are fixed to a (Dirichlet
boundary conditions) while on the left boundary, the heights on the boundary are fixed to b and
those in the layer next to the boundary are fixed to c, with d = min(b, c).2 Note that without the
constraint on the next-to-leftmost heights, each height in that layer may take either of the values
b + 1 and b 1. This is illustrated in Fig. 12.
Let us now see how this choice of boundary conditions translates in the loop model. To do so,
we first note that if the loop model is defined on N strands, a time slice of the RSOS model
is defined by N + 1 heights. A state of the model is a collection of those heights, denoted
|l1 , l2 , . . . , lN+1 , and we have the RSOS constraint li = 1, 2, . . . , p with |li+1 li | = 1. Fig. 12
corresponds to a case of even N . We will restrict to this case to start. We will also restrict to the
case where c = b + 1 so b = d, c = d + 1; note that then d a and N have the same parity.
We next recall briefly how the TemperleyLieb generators act in the RSOS representation.
Taking the time to flow downwards in Fig. 13, the generator ei acts as
ei |l1 , . . . , li1 , li , li+1 , . . . , lN+1
(Sli Sl )1/2
i
|l1 , . . . , li1 , li , li+1 , . . . , lN+1 ,
= (li1 , li+1 )
Sli1

(4.1)
li
where the Sl are the components of the PerronFrobenius eigenvector of the adjacency matrix of
Ap , and read explicitly
Sl = [l]q =
l
sin( p+1
)
sin( p+1
)
(4.2)
where we have introduced the q-deformed numbers [l]q (q l q l )/(q q 1 ).

The transfer matrix of the RSOS model has the usual form (3.1), but now in terms of the
ei defined by (4.1). The graphical expansion of the partition function is obtained by taking, for
each elementary face transfer matrix 1 + ei , either the identityin which case a vertical bar is
2 In order to respect the use of a, b and c in Ref. [12], note that throughout Section 4, a is not the parameter of (2.5),
b is not the blob generator in (2.2), and c is not the central charge (2.7).
152
Fig. 13. Labelling of RSOS heights around a lattice face.
drawn diagonally across the face, indicating that li = li for this termor the TemperleyLieb
generator ej , in which case a horizontal bar is drawn, indicating that li1 = li+1 . One gets in this
way clusters of constant heights, which can be separated by non-intersecting loops drawn on the
dual lattice.
The analysis of the weights needs some modification as compared to Pasquiers original treatment [19] of the RSOS model on a torus, but some of the key elements can be taken over. We first
consider the case of l1 = l2N +1 d; the heights l2 are also fixed and we shall assume l2 = d + 1
(the case l2 = d 1 being similar). We further assume that there is at least one cluster connecting
the left and right boundary of the annulus (in the limit where the aspect ratio = M/N
this is almost surely the case). There is then L = 0 non-contractible lines. Since all clusters connecting the two boundaries have the same height d, we might just as well identify them. In the
graphical expansion this can be represented by adding extra vertical bars on the two boundaries.
With this identification it follows that any loop is at the junction between exactly two distinct
clusters. Now orient every loop in the clockwise direction. When traversing any loop along this
direction, the cluster to its right is said to be surrounded by the loop, whereas the cluster to its left
is said to surround the loop. Let us now represent each distinct cluster by a node, and each loop
by a directed link going from the cluster that it surrounds towards the cluster that it is surrounded
by. This then defines a directed rooted tree, where every link is oriented towards the root, the root
being the unique cluster connecting the two boundaries.
The weight appearing in (4.1) is then distributed on individual loop turns as follows: Consider
a loop that surrounds a cluster at height l and that is surrounded by a cluster at height k. When
making a right (respectively left) turn whilst being tangential to a horizontal bar, the loop picks
up a factor (Sl /Sk )1/2 (respectively (Sl /Sk )1/2 ). Turns being tangential to a vertical bar do not
carry any weight. The complete loop, being clockwise, makes four more right than left turns, but
only two of those excess turns carry any weight, so the complete weight is Sl /Sk .
The directed tree is now undone by summing over the heights of its nodes. We start at a leave
node. For any fixed height k of the node adjacent to the leaf, the height of the leaf node can
be l = k 1. Summing over l gives a weight3 (Sk+1 + Sk1 )/Sk = 2 cos( p+1

)note that this is
independent of kwhich can be attributed to the loop represented by the link directed away from
the leaf. We then remove the leaf and its outgoing link, and proceed iteratively, moving always
from the leaves and towards the root. A special case occurs when summing over the heights of
those nodes adjacent to the root whose outgoing link corresponds to a loop that touches the left
boundary. Indeed, the heights of those nodes have been fixed to d + 1, and we obtain then the
weight Sd+1 /Sd = [d + 1]q /[d]q . Finally, when the whole tree has been undone, only the root
node remains, but since its height is fixed it contributes no additional weight.
3 This is true even when l = 1 or l = p, since S = S
0
p+1 = 0 from (4.2).
153
Fig. 14. (Color online.) Configuration of the RSOS clusters (in red color) for a case where a = d. Periodic boundary
conditions are imposed in the time (vertical) direction.
Consider now the loop model with weights x = 2 cos( p+1

) and y = [d + 1]q /[d]q i.e.,
setting r = d in (2.4)and no non-contractible lines, L = 0. Its configurations are in one-toone correspondence with those of the RSOS model treated above and the weights x, y are the
same. The spectrum of Tloop therefore contains the complete spectrum of TRSOS , and we have
verified this numerically. In particular the leading eigenvalues of these transfer matrices coincide.
However, Tloop also contains eigenvalues not present in TRSOS . This was to be expected, since the
loop model contains non-local information not present in the RSOS model (allowing in particular
its definition for non-integer p).
An example of the loop-RSOS equivalence is illustrated in Fig. 14, still for the case d = a.
Denoting by l1 , . . . , l5 the heights of the five clusters4 (labelled from top to bottom) we find the
overall weight (for simplicity we denote [l]q l):
(l2 l5 )1/2 (l1 l3 )1/2 (l1 l4 )1/2 (l3 l4 )1/2 (l2 l5 )1/2 l4 l3 l4
=
.
l1
l2
l2
l2
l4
l5 l2 l2
(4.3)
Note that l4 = l1 by the periodic boundary conditions in the time direction. Each of the two
factors on the right-hand side is associated with a loop. Since l3 can take values d 1 while
l2 = l5 = d and l4 = l1 = d + 1, the overall weight is 2 cos( p+1

) SSd+1
= xy as claimed.
d
While this construction clearly works when there are no non-contractible lines running
through the annulus (i.e., a = d), things are more delicate when there are. Fig. 15 shows for
instance that a configuration with two non-contractible lines does not necessarily contribute to
d = 3, a = 1. A little thought suggests to simply eliminate non-contractible loops touching the
boundary in this case.
We have indeed checked numerically that when d a is negative, the eigenvalues of the RSOS
model are all found in the loop transfer matrix for the blobbed sector, and the leading eigenvalues
coincide. Meanwhile if d a is positive, this is true provided one considers instead the unblobbed
sector for the loop model. In both cases one needs to have L = d a and y = y(d).
This corresponds simply to accepting configurations such as the one on the right of Fig. 15,
but not the one on the left.
4 Note that in this example the subscripts are used differently than in (4.1).
154
Fig. 15. (Color online.) The choice a = 3, b = 4, c = 1 is not compatible with the configuration of loops shown on the
left panel (even though there are two non-contractible loops). It is however compatible with the configuration shown on
the right panel.
Of course the RSOS model contains more situations: one can decide to have c = b 1 instead,
or to have N odd. The complete set of rules is as follows:
d a > 0 and has same parity as N:
d a 0 and has same parity as N:
d a 1 and has opposite parity of N:
d a < 1 and has opposite parity of N:
L = |d a|,
L = |d a|,
L = |d + 1 a|,
L = |d + 1 a|,
y = y(d),
y = y(d),
y = y(p d),
y = y(p d),
unblobbed,
blobbed,
blobbed,
unblobbed.
(4.4)
The first two cases coincide with what we discussed if N is even and generalize it easily if N
is odd. In these two cases, the left boundary necessarily sees b = d + 1, c = d, and the value of y
follows from the analysis of the loop model.
The last two cases meanwhile require that b = d, c = d + 1. The new value of y giving
the correct weights to the boundary loops is thus y = [d]q /[d + 1]q , which coincides with y =
[p + 1 d]q /[p d]q .
Note that the rules (4.4) are compatible with the global symmetry li p li of the RSOS
model configurations and weights (4.2). On the level of the sectors this reads (d, a) (p
d, p + 1, a) and is nothing but the usual symmetry of the Kac table (2.8).
An extensive numerical check of (4.4) is given in Table 2. We show all levels observed for
the RSOS model with N = 6 and p = 6, along with the sector label (d, a), and we give the
()
corresponding multiplicities of these levels in the loop model, where each sector TL (r) is characterized by the number of non-contractible lines L, the blobbing of the leftmost string ( denotes
the blobbed sector), as well as r which determines the boundary weight y(r) through (2.4). Remarkably, all of the RSOS levels are also observed in the loop model, with the sector given
precisely by the rules (4.4). This observation extends to other values of N and p (with any parities).
Note that the dimension of the loop (respectively RSOS) model transfer matrix does not (respectively does) depend on r. This apparent paradox is resolved by the fact that a given loop
model sector contains in general extra eigenvalues (not shown in Table 2) which are not present
in the RSOS model. It should also be noted that the dominant eigenvalue in a given RSOS sector
is always observed to be dominant as well in the corresponding loop model sector (4.4).
155
Table 2
Complete set of levels for the various sectors (d, a) of the RSOS model, here for size N = 6 and p = 6 (the tricritical
Potts model), cf. Table 1. The left column shows N 1 log (rounded to 12 digits), where 1 is the transfer matrix
()
eigenvalue. The remaining columns show the corresponding multiplicities of within the various sectors TL of the
loop model (the asterisk denotes the blobbed sector, see Appendix A). A blank entry denotes zero multiplicity. The
boundary weight y = y(r) is given by (2.4), with r indicated in the top of the table
f
0.338946565198
0.324025479536
0.316673567023
0.310464485251
0.302390277235
0.295466694605
0.291196112101
0.289510241627
0.288254559167
0.241054964628
0.238365774752
0.237748281915
0.234731074254
0.230275094022
0.230197803440
0.226587692821
0.223700268691
0.220405235066
0.212046452393
0.210311165453
0.201367085486
0.200352797500
0.194920765586
0.191048760878
0.187939781596
0.186511467791
0.186351229891
0.184916564626
0.184121456296
0.183007717844
0.182528404723
0.179523463630
0.178529556360
0.176928229253
0.172266060806
0.171718537721
0.170284900502
0.168641596700
0.167982529467
0.163038059540
0.155628452273
0.153242699723
0.149149353298
0.147179401602
0.140243221820
0.139719837688
(d, a)
T0
(1, 1)
(2, 2)
(3, 3)
(2, 3)
(1, 2)
(1, 3)
(2, 4)
(3, 2)
(2, 1)
(2, 2)
(1, 3)
(3, 3)
(2, 3)
(2, 4)
(1, 2)
(3, 2)
(2, 1)
(3, 1)
(2, 5)
(1, 1)
(1, 3)
(1, 4)
(2, 2)
(2, 4)
(2, 2)
(3, 2)
(3, 3)
(3, 3)
(1, 3)
(2, 1)
(2, 3)
(1, 2)
(2, 3)
(2, 4)
(3, 2)
(1, 1)
(3, 1)
(1, 5)
(2, 6)
(2, 5)
(2, 2)
(1, 4)
(1, 3)
(3, 3)
(2, 4)
(2, 3)
T2
2
T4
2
T2
2
T4
4
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
(continued on next page)
156
Table 2 (continued)
f
0.135157811358
0.134931152674
0.132209035214
0.115345804299
0.110023422061
0.109559975348
0.105095398499
0.104419222242
0.099484992148
0.097996769822
0.097573977039
0.095502814093
0.092050608093
0.088269162793
0.086753609983
0.085859268861
0.085433490626
0.085246090691
0.085077485734
0.082112734265
0.080981552646
0.080291708506
0.067354247176
0.066058643675
0.064397781477
0.061125554776
0.059945596252
0.056741093622
0.054049770890
0.052888923290
0.044940510316
0.041781216253
0.037564796511
0.036787385047
0.036787385047
0.033076470754
0.032439276523
0.030419378593
0.028358275153
0.027839432948
0.027592998174
0.020938336515
0.015656249492
0.012286435222
0.012266183325
0.007688713211
0.005272945987
0.003076941021
(d, a)
T0
(3, 2)
(1, 5)
(2, 6)
(2, 2)
(1, 1)
(3, 3)
(3, 1)
(2, 3)
(2, 5)
(1, 3)
(1, 2)
(2, 2)
(1, 4)
(3, 3)
(2, 4)
(1, 5)
(2, 4)
(3, 2)
(2, 1)
(2, 3)
(3, 2)
(2, 6)
(1, 3)
(3, 3)
(2, 3)
(1, 2)
(2, 4)
(3, 2)
(2, 1)
(2, 5)
(1, 4)
(1, 3)
(2, 2)
(1, 5)
(1, 6)
(2, 4)
(3, 3)
(2, 6)
(3, 2)
(2, 3)
(1, 1)
(3, 1)
(2, 2)
(1, 3)
(2, 5)
(3, 3)
(2, 4)
(1, 5)
T2
2
T4
2
T2
2
T4
4
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
157
Finally, we should mention a couple of fine details. In general the levels are not observed
to be degenerate. One exception visible in Table 2 is that the level f = 0.036787385047 is
present in two distinct sectors, but since this is so both on the RSOS and the loop side, definite
labels respecting (4.4) can be assigned as shown in the table. Another exception is the level
f = 0.085859268861 which apart from its assignment to T4 (r = 1) by (4.4) appears also in
two more loop sectors, T2 (r = 1) and T4 (r = 1). This type of degeneracies can be explained by
quantum group arguments.
Recall now that the character of the irreducible representation of the Virasoro algebra with
highest weight hda can be written as [20]
da =
Kd,a+2(p+1)n
n=0
Kd,2n(p+1)a .
(4.5)
n=1
We wish to recover this character from the knowledge of the loop model partition function. This
involves as usual an infinite series of additions and subtractions of sectors which is made possible
in finite size by the quantum group symmetry. We suppose we are in the situation with heights
b = d, c = d + 1. To go from d on the left side to a on the right side we need to sandwich between
the left and the right the adjacency matrix of the Ap diagrams so that a random walk on this
diagram, from boundary to boundary, hopping from non-contractible cluster to non-contractible
cluster, takes one from b to a. All the steps are identical with those in [12] where partition
functions with boundary conditions fixed at the leftmost and rightmost sides were computed.
Using the eigenvectors of the adjacency matrix as in Eqs. (4.15)(4.19) of [12], the required
expression is thus (we have set H = p + 1)

H 1
2
pd
ip
pa
sin
sin
Z =
, = d .
H
H
H
H
(4.6)
p=1
Here Z is calculated with running over ip

H and for each the value of mthe weight of
non-contractible loopsfollows from the same argument as in the case of the leading exponent
ip
r = 1 in the bulk case, implying = d so m = sinh(d+1)
sinh d , = H . Using the expression (3.18)
we see that the functions Kd,d+2j (r = d) get a combinatorial factor

H 1
p
p
2 1
cos(2j + d a)
cos(2j + d + a)
.
H
2
H
H
(4.7)
p=1
Since d a and 2j have the same parity, this select the conditions
2j + d a = 2n1 H,
2j + d + a = 2n2 H
(4.8)
for the first and second term respectively. Remembering that j runs only from [r/2] to infinity,
and that when the foregoing conditions are satisfied one gets a factor H2 from the sum cancelling
the prefactor in (4.6), it follows that
Z=

n=0
Kd,a+2nH

n=1
which is the result we wanted.
Kd,a+2nH
(4.9)
158
5. More algebraic considerations

There are many ways to think of the blob algebra. We would like now to think of it within the
theory of cabling, i.e., tensor products of spin 1/2 representations of Uq (sl2 ). Consider therefore
r + 1 representations of spin 1/2, and suppose we wish to project them on the maximally qsymmetric representation. The object doing this is the q-symmetrizer, whose expression is well
known by induction to be [21]
Sr+1 (e1 , . . . , er ) = Sr tr Sr er Sr
(5.1)
with the boundary condition S1 = 1. Here the tr are numbers given by

tr =
sin r
sin(r + 1)
(5.2)
and the ei are TL generators which obey (2.1) as before and act on the tensor product of the
ith and (i + 1)th spin 1/2 representation. These generators will not be identified with the ones
used in the previous sections however (see below). The first values are well known: S1 = 1,
e1
S2 = 1 2 cos
, etc.
Let us now add to our system r 1 ghost strings as in Fig. 16 (labelled 2 r, . . . , 0) on
which generators e2r , . . . , e0 act, and let us symmetrize on these ghost strings and the first one
of our system. We then define br through
br = Sr (e2r , . . . , e0 ).
(5.3)
We have then
b1 = 1,
e0
,
2 cos
b3 = 1 t2 (e1 + e0 ) + t1 t2 (e1 e0 + e0 e1 )
b2 = 1
(5.4)
and so on. By construction, the br s are projectors

br2 = br .
(5.5)
It is easy to show that they satisfy moreover

e1 b r e 1 =
sin(r + 1)
e1 Pr1 ,
sin r
(5.6)
where we denote by Pr1 the symmetrizer on the ghost strings only, Pr1 = Sr1 (e2r , . . . , e1 ).
Of course, [ei , Pr1 ] = 0 for all i 1. We can thus consider a modified version of the TL algebra where instead of the generators ei we have the modified generators ei Pr1 ei . They
obviously satisfy the bulk TL relations (2.1), since Pr1 is a projector. Moreover we now have
the relations (2.2) of the blob algebra, with b = br and (2.4). We have thus made the link with
the foregoing discussion and shown that for r an integer, the boundary conditions corresponding
to the value (2.4) can be obtained by adding ghost strings on the left boundary and symmetrizing
them with the first real string in the system.
This should not come as a surprise: it is easy to show using Pasquiers 6j calculations [22]
that the insertion of br through the ghost strings construction translates into RSOS language by
having heights increasing linearly from the left-hand side of the ghost strings up to the first height
159
Fig. 16. (Color online.) Ghost strings 2 r, . . . , 0, represented by thin red lines. They are symmetrized among themselves
and with the first real line (the symmetrization is shown symbolically by horizontal green bars).
from l = 1 to l = r, and then because br symmetrizes on r strings and thus on the first real string
as well, from l = r to l = r + 1. This proves that the transfer matrix bT0 describes exactly the
mixed RSOS boundary conditions.
Note that one does not have to introduce the ghost strings. Algebraically, the same would be
obtained by replacing these r 1 strings by a spin r1
2 (in which case, Pr1 = 1 identically). The
operator br then amounts to projecting the product of this representation and the first spin 1/2
representation onto spin r/2.
In this form, the identification was already mentioned in [12].
The formulas obtained in this section should match the ones which appeared in a recent paper
by Pearce et al. [23]; the derivation there is based on boundary integrability, and does not refer
to the blob or boundary TemperleyLieb algebra, as far as we can see.
6. Conclusion
In conclusion, it is important to stress that we have found a continuum of conformal boundary
conditions for the dense loop models. How to incorporate them into a consistent conformal field
theory remains an open problem. Note that none of these boundary conditions involves the number of times loops touch the boundary, which would correspond somehow to modified spinspin
couplings on the boundary. Rather, giving a different weight y to loops touching the boundary
can be interpreted in the O(n) model most easily as restricting the degrees of freedom on the
boundary to take values in a sub manifold of dimension y. This is not without reminding us of
results in the WZW cases [24].
The results described in this paper point to many further directions. Among those are:
geometric applications of the generating functions on the annulus;

derivation of the Bethe ansatz equations and of the spectrum of scaled gaps;
extensions to the dilute case;
extension to the double boundary case;
study of boundary conditions in c = 1 theories.
We hope to report on these soon. To conclude, we now give one example of application. We
consider the case x = 0, c = 2 and the partition function with x = l = 0, y = m = 1 (i.e., loops
touching the boundary get a weight one, whether contractible or not, the others are not allowed).
160
After some simple manipulations, (3.14) can be written as

Z=
q 1/24
2
(1)j q (4j 1) /32 .
P (q)
(6.1)
j =
This is easy to interpret geometrically as the partition function of a gas of dense loops, all contractible, which are all constrained to touch the left boundary. The loops can also be replaced
by trees. Meanwhile, this partition function can also be interpreted in the symplectic fermion
theory [25].
Note added in proof
After the completion of this work, I. Kostov has studied (in the preprint hep-th/0703221) the
coupling of our boundary loop model to two-dimensional quantum gravity. His results corroborate those presented here.
Acknowledgements
H. Saleur thanks I. Kostov and V. Schomerus for many interesting discussions, and for communication of their unpublished results [18,25].
Appendix A. Transfer matrix structure
In this appendix we discuss the construction and structure of the transfer matrix of the boundary loop model, corresponding to taking the limit in (3.2).
We recall the most general case of the boundary loop model, cf. Fig. 8. The model is defined
on an annulus with the exterior boundary being distinguished. Each loop touching at least once
this boundary gets a weight y if it is contractible (i.e., homotopic to a point), and m if it is not.
Each loop that never touches the distinguished boundary gets a weight x if it is contractible, and
l if it is not. For an annulus of width N strands (we assume N even) and circumference M, this
defines a partition function ZN,M (x, y, l, m) which can be expressed in terms of the Mth power
of the transfer matrix T = bT0 , where T is given in terms of TL generators by (3.1), and the
generators b, ei obey the blob algebra BN (x, y) defined by (2.1), (2.2).
For simplicity we henceforth represent the annulus as a rectangle of width N and height M,
with periodic boundary conditions identifying its top and bottom sides. The distinguished boundary is taken to be the left side. The transfer matrix then acts on states which can be depicted
graphically as fully-packed non-crossing link patterns within a slab bordered by two horizontal
rows, each of N points. A link joining the top and the bottom row is called a string, and any other
link is called an arc. The action of a word in BN (x, y) on a state is obtained by adjoining the
word to the top row of the slab (i.e., time propagates upwards). The 20 possible states for N = 4
are represented in Fig. 17.
Note that links touching a point on the left boundary are necessarily blobbed (shown by a
circle in Fig. 17). Only links up to and including the leftmost string can be blobbed. The states
can be ordered as follows: First we sort the states according to a decreasing number of strings L.
For fixed L > 0, we place first the states in which the leftmost string is unblobbed. And finally
we group together states with fixed L and fixed blobbing of the leftmost string, according to the
link/arc configuration of the lower row of the slab. This gives the order of the rows of Fig. 17.
161
Fig. 17. Ordering of the states of the transfer matrix T , here for width N = 4 strands.
Fig. 18. Reduced states for N = 4.
The ordering of states within each row is according to the link/arc configuration of the upper row
of the slab.
With this ordering of the states, T has a blockwise lower triagonal structure, with each block
corresponding to a group of states as defined above. The reason is that acting by ei can annihilate two strings (if their positions on the top side of the slab are i and i + 1) but cannot
create any strings. Likewise, acting by b can blob the leftmost string, but it cannot subsequently be unblobbed. The triagonal structure implies that the eigenvalues of T are the union
of eigenvalues of the blocks on its diagonal. Moreover, blocks differing only by the configuration of the bottom row are identical. For the purpose of studying only the spectrum of T
the bottom row can therefore be completely forgotten, leading to a much smaller transfer matrix.
Indeed let us define a reduced state as a non-crossing link pattern on N points. A (full) state
can be turned into a pair of reduced states by cutting each of its strings and pulling apart the
upper and lower parts. For convenience, a cut string will still be called a string with respect to
the reduced state. For N = 4 there are 8 reduced states, shown in Fig. 18.
162
The blocks on the diagonal of T are denoted TL and TL , where L is the number of strings, and
the presence (respectively absence) of the asterisk () indicates that the leftmost link is blobbed
(respectively unblobbed). Note that the blocks TL and TL can be constructed in terms of the
reduced states. The numerical studies of the spectral properties of T reported in this paper were
done by diagonalizing these blocks in the basis of reduced states. Their dimensions read (see
Appendix B for proofs of closely related statements)

N 1
dim TL =
for L = 0, 2, . . . , N 2,
(N L 2)/2

N 1
(A.1)
for L = 2, 4, . . . , N.
dim TL =
(N L)/2
With the terminology being fixed, the annulus partition function can now be written as [15]
ZN,M (x, y, l, m) = u|T M |v.
(A.2)
Here, the right vector |v is the unit vector corresponding to the state with N strings. The left
vector u| is obtained by identifying the top and bottom rows for each state; counting the number
of loops of each type (contractible or not, blobbed or not) gives the corresponding weight as a
monomial in x, y, l, and m. For instance, with N = 4 and the ordering of the states shown in
Fig. 17, we have
|u = (1, 0, 0, . . . , 0),

|v = l 3 m, l 2 y, lmx, lm, y, . . . , xy .
(A.3)
Appendix B. Exact eigenvalue amplitudes

The goal of this appendix is to provide a rigorous combinatorial proof of the amplitude formulae (3.11) by generalizing the working of [26]. The discussion assumes knowledge of the transfer
matrix blocks TL and TL defined in Appendix A.
Following [26], we introduce the characters
M
Kk = Tr(T2k )M ,
(B.1)
Kk = Tr T2k
,
where we stress that the trace is over reduced states. Also, let Zj (respectively Zj ) be the annulus
partition function constrained to have exactly 2j unblobbed non-contractible loops (respectively
2j 1 unblobbed and 1 blobbed non-contractible loops). In other words, Zj (respectively Zj )
consists of the terms in the full partition function ZN,M (x, y, l, m) whose l, m dependence is l 2j
(respectively l 2j 1 m). The goal is to search for a decomposition of the form
Zj =
N/2

(j )
(j )
Dk l 2k Kk + Dk l 2k1 mKk ,
k=j
Zj =
N/2

(j )
l Kk + Dk l 2k1 mKk ,
(j ) 2k
(B.2)
(j )
, and Dk are coefficients to be determined. In the notation of (3.11)
Dk
k=j
(j )
(j )
where Dk , Dk , Dk
(j )
163
Fig. 19. Construction of invariant restricted states. (a) A configuration contributing to Zj with N = 12 and j = 2, here
depicted as a state. (b) Application on the bottom of the reduced state corresponding to the top half of (a). (c) After
removal of arcs one has simply 2j links.
we have then
D2j
j

(i)
(i)
Dj l 2i + Dj l 2i1 m ,
i=0
D2j =
j

(i)
(i) 2i1
Dj l 2i + Dj

m
(B.3)
i=0
(j )
(j )
(j )
with D0 = D0 = Dj = 0.
Rather than solving directly for the decomposition of Zj in terms of Kk , the idea [26] is now
to turn the problem upside down and look for the decomposition of Kk in terms of Zj :
Kk =
N/21

j =k
Kk =
N/21

j =k
(k) Zj
Ej 2j
Ej(k)
N/2

j =k+1
Zj
(k)
Ej 2j 1 ,
l
m
N/2
Zj (k) Zj
+
E
j 2j 1 .
l 2j
l
m
(B.4)
j =k
The determination of the coefficients E can be turned into a combinatorial counting problem
as follows. First, recall that Kk and Kk were defined as traces over restricted states (in contradistinction to the partition function which, as we have seen in (A.2), is more complicated than just
a trace). We must now determine how many times each Zj and Zj occurs within a given trace.
Consider therefore some configuration C on the annulus that contributes to (say) Zj (i.e., has 2j
non-contractible unblobbed loops). An example with j = 2 and N = 12 is shown in Fig. 19a. It
is convenient to not represent the contractible loops within the configuration, i.e., to depict it as a
state. This configuration will contribute to the trace only over such restricted states S that are left
invariant by the action of the configuration. Therefore, S must contain the same arcs as does C
in its top row (see Fig. 19b). It suffices therefore to determine the parts of S which connect onto
the starting points of the 2j non-contractible loops (see Fig. 19c). Since the goal is to determine
the contribution to (say) Kk , precisely 2k strings and j k arcs must be used.
164
With this in mind, the coefficients E are then determined as the following counting problems.
In all cases, construct a reduced state on 2j strands, using 2k strings and j k arcs. Further:
(k)
For Ej , all strings are unblobbed, but the exterior arcs to the left of the first string may be
blobbed.
(k)
For Ej , the problem is the same, except that the leftmost string must be blobbed. But
(k)
evidently this leads to the same counting, and so Ej
(k)
= Ej .
(k)
For Ej , the leftmost strand becomes blobbed (since we are considering a contribution from
Zj ), and so must connect onto a blobbed object (arc or string). But as the strings are unblobbed (since we are considering a contribution from Kk ), it follows that the leftmost object
is a blobbed arc.
(k)
For Ej , the leftmost object (arc or string) as well as the leftmost string must be blobbed.
These counting problems are easily solved using generating function techniques. As a
warmup, consider the counting of restricted states made up of only arcs. Associate to each pair
of sites an activity z. A state is either empty, or has a leftmost arc which divides the space into
two parts (inside the arc and to its right) each of which can accommodate an independent arc
state. The generating function f (z) therefore satisfies f (z) = 1 + z[f (z)]2 with regular solution
f (z) =
1 4z
=
Cn z n .
2z
(B.5)
n=0
(2n)!
The coefficients are the celebrated Catalan numbers Cn = n!(n+1)!
.
Consider next states made up of only arcs, but in which exterior arcs may (but need not) be
blobbed. Call the generating function g(z). If the state is non-empty, the leftmost arc is necessarily exterior. Inside it are f (z) states, and to its right g(z) states. Thus, g(z) = 1 + 2zf (z)g(z),
or

1
2n n
g(z) =
(B.6)
z .
=
1 4z n=0 n
(k)
We can now attack the case of Ej . Since there are 2k strings (all unblobbed), all of which
divide the space into independent parts, the generating function reads

2k
hk (z) = g(z) zk f (z) =
n=k
(k)
(k)
and we infer that Ej = Ej

For
reads
(k)
Ej
2j
j k
2n
zn
nk
it follows from the above observations that the generating function with 2k strings

2n 1
zn
ik (z) = zf (z)hk (z) =
nk1
n=k+1
(k)
(B.7)
and so Ej =
2j 1
j k1
(B.8)
165
(k)
Finally, for Ej we must distinguish between the cases where the leftmost object is an arc
or a string. This gives the generating function

2k

2n 1 n
jk (z) = 1 + zf (z)g(z) zk f (z) =
(B.9)
z
nk
(k)
Ej
n=k
2j 1
whence
= j k .
To summarize, we have shown that
N/21
N/2
2j
Zj

Zj
2j 1
,
Kk =
+
j k l 2j
j k 1 l 2j 1 m
j =k
Kk =
N/21

j =k
j =k+1
N/2
Zj
2j Zj 2j 1
.
+
j k l 2j
j k l 2j 1 m
(B.10)
j =k
It is easily checked that this system of equations is invertible, and after some straightforward
manipulations the Zj and Zj can be isolated. In the notation of (B.2) the solution reads

j +k
j +k1
(j )
(j )
Dk = (1)j +k
,
Dk = (1)j +k1
,
2k
2k 1

j +k1
j +k1
(j )
(j )
,
Dk = (1)j +k
.
Dk = (1)j +k
(B.11)
2k
2k 1
Applying (B.3) then finally leads to (3.11).
References
[1] J. Cardy, The O(n) model on the annulus, J. Stat. Phys. 125 (2006) 1, cond-mat/0604043.
[2] M. Bauer, D. Bernard, SLE, CFT and zig-zag probabilities, in: Proceedings of the NATO Conference Conformal
Invariance and Random Spatial Processes, Edinburg, 2003, math-ph/0401019.
[3] H. Saleur, Lectures on non-perturbative field theory and quantum impurity problems, in: Topological Aspects of
Low Dimensional Systems, in: Les Houches Lectures, vol. 69, Springer, 1999, p. 473, cond-mat/9812110.
[4] V. Schomerus, Class. Quantum Grav. 19 (2002) 57815847, hep-th/0209241.
[5] V.B. Petkova, J.B. Zuber, Conformal boundary conditions and what they teach us, hep-th/0103007.
[6] V. Schomerus, Phys. Rep. 431 (2006) 3986, hep-th/0509155.
[7] P. Martin, H. Saleur, Lett. Math. Phys. 30 (1994) 189206.
[8] P.P. Martin, D. Woodcock, J. Algebra 225 (2000) 957, math.RT/0205263.
[9] R.J. Baxter, S.B. Kelland, F.Y. Wu, J. Phys. A: Math. Gen. 9 (1976) 397.
[10] J.L. Cardy, J. Phys. A 25 (1992) L201.
[11] A. Nichols, V. Rittenberg, J. de Gier, J. Stat. Mech. 0503 (2005) P003, cond-mat/0411512;
A. Nichols, J. Stat. Mech. 0601 (2006) P003, hep-th/0509069;
A. Nichols, J. Stat. Mech. 0602 (2006) L004, hep-th/0512273.
[12] H. Saleur, M. Bauer, Nucl. Phys. B 320 (1989) 591624.
[13] A. Doikou, P.P. Martin, J. Phys. A 36 (2003) 22032226, hep-th/0206076.
[14] N. Read, H. Saleur, Nucl. Phys. B 613 (2001) 409, hep-th/0106124.
[15] J.L. Jacobsen, J. Salas, J. Stat. Phys. 122 (2006) 705760, cond-mat/0407444.
[16] B. Nienhuis, Phys. Rev. Lett. 49 (1982) 10621065.
[17] I. Kostov, B. Ponsot, D. Serban, Nucl. Phys. B 683 (2000) 309362, hep-th/0307189.
[18] I. Kostov, private communication.
[19] V. Pasquier, J. Phys. A 20 (1987) L1229.
[20] A. Rocha-Caridi, in: S. Lepowski, S. Mandelstam, I.M. Singer (Eds.), Vertex Operators in Mathematics and Physics,
in: MSRI Publications, vol. 3, Springer, New York, 1985, p. 451.
166
[21] L.H. Kaufman, S.H. Links, Temperley Lieb Recoupling Theory and Invariants of Three Manifolds, Princeton Univ.
Press, 1994.
[22] V. Pasquier, Commun. Math. Phys. 118 (1988) 355.
[23] J. Rasmussen, P. Pearce, J.-B. Zuber, Logarithmic minimal models, J. Stat. Mech. 0611 (2006) P017, hepth/0607232.
[24] A.Y. Alekseev, V. Schomerus, Phys. Rev. D 60 (1999) 061901, hep-th/9812193.
[25] T. Creutzig, T. Quella, V. Schomerus, New boundary conditions for the c = 2 system, unpublished.
[26] J.-F. Richard, J.L. Jacobsen, Nucl. Phys. B 750 (2006) 250264, math-ph/0605016.
Form factors in finite volume I:

Form factor bootstrap and truncated conformal space
B. Pozsgay a , G. Takcs b,
a Institute for Theoretical Physics, Etvs University, Budapest, Hungary
b HAS Research Group for Theoretical Physics, H-1117 Budapest, Pzmny Pter stny 1/A, Hungary
Received 13 June 2007; accepted 29 June 2007
Abstract
We describe the volume dependence of matrix elements of local fields to all orders in inverse powers
of the volume (i.e., only neglecting contributions that decay exponentially with volume). Using the scaling
LeeYang model and the Ising model in a magnetic field as testing ground, we compare them to matrix
elements extracted in finite volume using truncated conformal space approach to exact form factors obtained
using the bootstrap method. We obtain solid confirmation for the form factor bootstrap, which is different
from all previously available tests in that it is a non-perturbative and direct comparison of exact form factors
to multi-particle matrix elements of local operators, computed from the Hamiltonian formulation of the
quantum field theory. We also demonstrate that combining form factor bootstrap and truncated conformal
space is an effective method for evaluating finite volume form factors in integrable field theories over the
whole range in volume.
1. Introduction
The matrix elements of local operators (form factors) are central objects in quantum field theory. In two-dimensional integrable quantum field theory the S matrix can be obtained exactly in
the framework of factorized scattering (see [1,2] for reviews), and using the scattering amplitudes
as input it is possible to obtain a set of axioms satisfied by the form factors [3], which provides
the basis for the so-called form factor bootstrap (see [4] for a review).
DOI of companion paper: 10.1016/j.nuclphysb.2007.07.008.
E-mail addresses: pozsi@bolyai.elte.hu (B. Pozsgay), takacs@elte.hu (G. Takcs).

doi:10.1016/j.nuclphysb.2007.06.027
168
B. Pozsgay, G. Takcs / Nuclear Physics B 788 [FS] (2008) 167208
Although the connection with the Lagrangian formulation of quantum field theory is rather
indirect in the bootstrap approach, it is thought that the general solution of the form factor axioms
determines the complete local operator algebra of the theory [5]. This expectation was confirmed
in many cases by explicit comparison of the space of solutions to the spectrum of local operators
as described by the ultraviolet limiting conformal field theory [69]; mathematical foundation is
provided by the local commutativity theorem stating that operators specified by solutions of the
form factor bootstrap are mutually local [4]. Another important piece of information comes from
correlation functions: using form factors, a spectral representation for the correlation functions
can be built which provides a large distance expansion [10,11], while the Lagrangian or perturbed
conformal field theory formulation allows one to obtain a short-distance expansion, which can
then be compared provided there is an overlap between their regimes of validity [11]. Other
evidence for the correspondence between the field theory and the solutions of the form factor
bootstrap results from evaluating sum rules like Zamolodchikovs c-theorem [12,13] or the theorem [14], both of which can be used to express conformal data as spectral sums in terms of
form factors. Direct comparisons with multi-particle matrix elements are not so readily available,
except for perturbative or 1/N calculations in some simple cases [3].
Therefore, part of the motivation of this paper is to provide non-perturbative evaluation of
form factors from the Hamiltonian formulation, which then allows for a direct comparison with
solutions of the form factor axioms. Another goal is to have a better understanding of finite
size effects in the case of matrix elements of local operators, and to contribute to the investigation of finite volume [15] (and also finite temperature [16]) form factors and correlation
functions.
Based on what we learned from our previous investigation of decay rates in finite volume [17],
in this paper we determine form factors using a formulation of quantum field theory in finite volume. In two spacetime dimensions this is most efficiently done using the truncated conformal
space approach (TCSA) developed by Yurov and Zamolodchikov [18], but one could have also
made use of, for example, lattice field theory: the only important point is to have a method
which can be used to determine energy levels and matrix elements of local operators as functions of the volume. We give a relation between finite and infinite volume multi-particle form
factors, which is a natural extension of the results by Lellouch and Lscher for two-particle decay matrix elements [19]. It is used to determine form factors in two important examples of
integrable quantum field theory: the scaling LeeYang model and the Ising model in a magnetic
field. We show that these results agree very well with the predictions of the form factor bootstrap. We only treat matrix elements of local fields between multi-particle states for which there
are no disconnected pieces (which appear whenever there are particles with coincident rapidities
in the left and right multi-particle states); the treatment of disconnected pieces, together with
a number of theoretical arguments and ramifications, are postponed to a subsequent publication [20].
The organization of the paper is as follows. In Section 2, after a brief review of necessary facts
concerning the form factor bootstrap we give the general description of form factors (without
disconnected pieces) to all orders in 1/L based on an analysis of two-point correlation functions.
Section 3 describes the two models we chose for demonstration, and also specifies the method of
evaluating form factors from truncated conformal space. Numerical results on elementary form
factors (vacuummany-particle matrix elements) are given in Section 4, while the general case
is treated in Section 5. We give our conclusions in Section 6.
169
2. Form factors in finite volume

2.1. Form factor bootstrap
Here we give a very brief summary of the axioms of the form factor bootstrap, in order to
set up notations and to provide background for later arguments; the interested reader is referred
to Smirnovs review [4] for more details. Let us suppose for simplicity that the spectrum of
the model consists of particles Ai , i = 1, . . . , N , with masses mi , which are assumed to be
strictly non-degenerate, i.e., mi = mj for any i = j (and therefore also self-conjugate). Because
of integrability, multi-particle scattering amplitudes factorize into the product of pairwise twoparticle scatterings, which are purely elastic (in other words: diagonal). This means that any
two-particle scattering amplitude is a pure phase, which we denote by Sij ( ) where is the
relative rapidity of the incoming particles Ai and Aj . Incoming and outgoing asymptotic states
can be distinguished by ordering of the rapidities:

|1 , . . . , n in
i1 ...in : 1 > 2 > > n ,
|1 , . . . , n i1 ...in =
|1 , . . . , n out
i1 ...in : 1 < 2 < < n ,
and states which only differ in the order of rapidities are related by
|1 , . . . , k , k+1 , . . . , n i1 ...ik ik+1 ...in
= Sik ik+1 (k k+1 )|1 , . . . , k+1 , k , . . . , n i1 ...ik+1 ik ...in
from which the S matrix of any multi-particle scattering process can be obtained. The normalization of these states is specified by the following inner product for the one-particle states:
j
| i = ij 2( ).
The form factors of a local operator O(t, x) are defined as

O
(1 , . . . , m |1 , . . . , n )j1 ...jm ;i1 ...in = j1 ...jm 1 , . . . , m |O(0, 0)|1 , . . . , n i1 ...in .
Fmn
(2.1)
With the help of the crossing relations

O
(1 , . . . , m |1 , . . . , n )j1 ...jm ;i1 ...in
Fmn
O

= Fm1n+1
(1 , . . . , m1
|m + i, 1 , . . . , n )j1 ...jm1 ;jm i1 ...in

n
k1

+
Sil ik (l k )
2jm ik (m k )
k=1
l=1
O

Fm1n1
(1 , . . . , m1
|1 , . . . , k1 , k+1 . . . , n )j1 ...jm1 ;jm i1 ...ik1 ik+1 ...in
all form factors can be expressed in terms of the elementary form factors
FnO (1 , . . . , n )i1 ...in = 0|O(0, 0)|1 , . . . , n i1 ...in
which satisfy the following axioms:

(2.2)
170
I. Exchange:
FnO (1 , . . . , k , k+1 , . . . , n )i1 ...ik ik+1 ...in
= Sik ik+1 (k k+1 )FnO (1 , . . . , k+1 , k , . . . , n )i1 ...ik+1 ik ...in .
(2.3)
II. Cyclic permutation:

FnO (1 + 2i, 2 , . . . , n ) = FnO (2 , . . . , n , 1 ).
(2.4)
III. Kinematical singularity:

O
i Res Fn+2
( + i, , 1 , . . . , n )ij i1 ...in
=

n

= 1 i j
Si ik ( k ) FnO (1 , . . . , n )i1 ...in .
(2.5)
k=1
IV. Dynamical singularity:

j
O
i Res Fn+2
+ i u ij k /2, i u ik /2, 1 , . . . , n ij i
=
1 ...in
O
= ijk Fn+1
(, 1 , . . . , n )k i1 ...in
(2.6)
whenever k occurs as the bound state of the particles i and j , corresponding to a bound state pole
of the S matrix of the form
i(ijk )2

Sij iukij
,
iukij
(2.7)
where ijk is the on-shell three-particle coupling and ukij is the so-called fusion angle. The fusion
angles satisfy
m2k = m2i + m2j + 2mi mj cos ukij ,
j
2 = ukij + uik + uij k

and we also used the notation u kij = ukij . Axioms IIV are supplemented by the assumption
of maximum analyticity (i.e., that the form factors are meromorphic functions which only have
the singularities prescribed by the axioms) and possible further conditions expressing properties
of the particular operator whose form factors are sought.
2.2. Finite size corrections for form factors
Let us consider the spectral representation of the Euclidean two-point function
n

dk

O(x)O
(0, 0) =
FnO (1 , 2 , . . . , n )i1 ...in
2
n=0 i1 ...in k=1

n

+
O
Fn (1 , 2 , . . . , n )i1 ...in exp r
mik cosh k ,
k=1
(2.8)
171
where

FnO (1 , 2 , . . . , n )+
i1 ...in = i1 ...in 1 , . . . , n |O (0, 0)|0

= FnO (1 + i, 2 + i, . . . , n + i)i1 ...in

(which is just the complex conjugate of FnO for unitary theories) and r = 2 + x 2 is the length
of the Euclidean separation vector x = (, x).
In finite volume L, the space of states can still be labeled by multi-particle states but the
momenta (and therefore the rapidities) are quantized. Denoting the quantum numbers I1 , . . . , In
the two-point function of the same local operator can be written as

O(, 0)O (0, 0) L

=

0|O(0, 0)|{I1 , I2 , . . . , In }i1 ...in ,L
n=0 i1 ...in I1 ...In

i1 ...in {I1 , I2 , . . . , In }|O (0, 0)|0L exp
n

mik cosh k ,
(2.9)
k=1
where we supposed that the finite volume multi-particle states |{I1 , I2 , . . . , In }i1 ...in ,L are orthonormal and for simplicity restricted the formula to separation in Euclidean time only. The
index L signals that the matrix element is evaluated in finite volume L. Using the finite volume
expansion developed by Lscher in [21] one can easily see that

O(, 0)O (0) O(, 0)O (0) L O eL ,
(2.10)
where is some characteristic mass scale.
Note that Lschers finite volume expansion is derived using covariant perturbation theory,
but the final expansion is obtained after resummation in the coupling constant, and is expected
to hold non-perturbatively. The finite volume corrections are then expressed in terms of the exact
one-particle-irreducible vertex functions of the infinite volume theory. According to Lschers
classification of finite volume Feynman graphs, the difference between the finite and infinite
volume correlation function is given by contributions from graphs of nontrivial gauge class,
i.e., graphs in which some propagator has a nonzero winding number around the cylinder. Such
graphs always carry an exponential suppression factor in L, whose exponent can be determined
by analyzing the singularities of the propagators and vertex functions entering the expressions. In
a massive theory, all such singularities lie away from the real axis of the Mandelstam variables,
and the one with the smallest imaginary part determines . It turns out that the value of is
determined by the exact mass spectrum of the particles and also the bound state fusions between
them [21,22]. Therefore it is universal, which means that it is independent of the correlation
function considered. In general m1 where m1 is the lightest particle mass (the mass gap
of the theory), because there are always corrections in which the lightest particle loops around
the finite volume L, and so the mass shell pole of the corresponding exact propagator is always
present. Contributions from such particle loops to the vacuum expectation value are evaluated
in Section 4.1 (for a graphical representation see Fig. 4.2), while an example of a finite volume
correction corresponding to a bound state fusion (a so-called -term) is discussed in Section 4.1.2
(Fig. 4.7).
To relate the finite and infinite volume form factors a further step is necessary, because the
integrals in the spectral representation (2.8) must also be discretized. Let us consider this problem
172
first for the case of free particles:

n

n

dk
dpk
f (1 , . . . , n ) =
f (p1 , . . . , pn ),
2
2k
k=1
k=1
where
pk = mik sinh k ,
k = mik cosh k
are the momenta and energies of the particles. In finite volume

2Ik
L
and it is well known (as a consequence of the Poisson summation formula, cf. [23]) that
pk =

I1 ,...,In
2I1
2In
g
,...,
L
L
L
2
n
n

dpk g(p1 , . . . , pn ) + O LN
(2.11)
k=1
provided the function g and its first N derivatives are integrable. Recalling that form factors are
analytic functions for real momenta, in our case this is true for derivatives of any order, due to the
exponential suppression factor in the spectral integrals, provided the form factors grow at most
polynomially in the momentum, i.e.,

Fn (1 + , 2 + , . . . , n + ) ex|| as | | .
This is true if we only consider operators which have a power-like short distance singularity in
their two-point functions [24]:
1
.
r 2
Such operators are called scaling fields and are generally assumed to form a closed algebra under
the operator product expansion.
Therefore the discrete sum differs from the continuum integral only by terms decaying faster
than any power in 1/L, i.e., by terms exponentially suppressed in L. Taking into account that
(2.8) is valid for any pair O, O of scaling fields, we obtain
0|O(x)O(0)|0
0|O(0, 0)|{I1 , . . . , In }i1 ...in ,L =
1
i1 ...in (1 , . . . , n )
(0)

FnO (1 , . . . , n )i1 ...in + O e L ,
(2.12)
where
sinh k =
2Ik
mik L
and
i1 ...in (1 , . . . , n ) =
(0)
n

mik L cosh k .
(2.13)
k=1
(0)
n is nothing else than the Jacobi determinant corresponding to changing from the variables

2Ik to the rapidities k . The term O(e L ) signifies that our considerations are valid to all
173
orders in 1/L, although our argument does not tell us the value of : to do that, we would need
more information about the correction term in the discretization (2.11).
In the case of interacting particles a more careful analysis is necessary because the quantization rules are different from the free case. In a two-dimensional integrable quantum field theory,
general multi-particle levels are determined by the BetheYang equations

Qk (1 , . . . , n )i1 ...in = mik L sinh k +
(2.14)
ik il (k l ) = 2Ik , k = 1, . . . , n,
l=k
which are valid to any order in 1/L, where

ij ( ) = i log Sij ( )
(2.15)
are the two-particle scattering phase-shifts. Therefore the proper generalization of (2.12) is
0|O(0, 0)|{I1 , . . . , In }i1 ...in ,L =
1
i1 ...in (1 , . . . , n )

FnO (1 , . . . , n )i1 ...in + O e L ,
(2.16)
where
i1 ...in (1 , . . . , n ) = det J (n) (1 , . . . , n )i1 ...in ,
Qk (1 , . . . , n )i1 ...in
(n)
Jkl (1 , . . . , n )i1 ...in =
,
l
k, l = 1, . . . , n,
(2.17)
and k are the solutions of the BetheYang equations (2.14) corresponding to the state with the
specified quantum numbers I1 , . . . , In at the given volume L.
Similar arguments were previously used to obtain the finite size dependence of kaon decay
matrix elements by Lin et al. [25]; they also give a more detailed analysis of the discretization
in (2.11). In principle it could be that < if the discretization introduces larger errors than
the finite size correction of the two-point functions. However, it is possible to give an argument
that actually = . Recall that the Poisson formula gives the discrete sum in terms of a Fourier
transform: the leading term is the Fourier transform of the summand evaluated at wave number 0
(i.e., the integral) and the corrections are determined by the decay of the Fourier transform at
large wave numbers. The function we need to consider is
h(p1 , . . . , pn )

=
i1 ...in (1 , . . . , n )1 FnO (1 , 2 , . . . , n )i1 ...in FnO (1 , 2 , . . . , n )+
i1 ...in
i1 ...in
exp r
n

mik cosh k ,
(2.18)
k=1
where pk = mik sinh k are the momentum variables. Due to the analyticity of the form factors
for real rapidities, this function is analytic for physical (real) momenta, and together with all of
its derivatives decays more rapidly than any power at infinity. Therefore its Fourier transform
taken in the momentum variables has the same asymptotic property, i.e., it (and its derivatives)
decay more rapidly than any power:
1 , . . . , n ) e ||
h(
174

for large . As a result, discretization introduces an error of order e L . The asymptotic exponent of the Fourier transform can be generally determined by shifting the contour of the
integral transform and is given by the position of the singularity closest to the real momentum
domain (this is essentially the procedure that Lscher uses in [21]). Singularities of the form
factors are given by the same analytic structure as that of the amplitudes which determine the
exponent in Eq. (2.10) (see the discussion there). Thus we expect that = and, as a result,
= .
This argument is just an intuitive reasoning, although it can be made a little more precise.
First of all, we must examine whether the determinant i1 ...in (1 , . . . , n ) can have any zeros. It
can always be written in the form
n

i1 ...in (1 , . . . , n ) =
mik L cosh k 1 + O(1/L) .
k=1
The leading factor can only be zero when k = i

2 for some k, which corresponds to pk = imk ,
giving = mk in case this is the closest singularity. That gives a correction
emk L
which is the same as the contribution by an on-shell propagator wound around the finite volume,
and such corrections are already included in the eL term of (2.10). Another possibility is that
some phase-shift function ( ) in the O(1/L) terms contributes a large term, which balances the
1/L prefactor. For that its argument must be close to a singularity, and then according to Eq. (2.7)
we can write
( ) log( u) O(L),
where u is the position of the singularity in the phase-shift.1 This requires that the singularity
is approached exponentially close (as a function of the volume L), but the positions of all these
singularities are again determined by singularities of the vertex functions, so this gives no new
possibilities for the exponent .
A further issue that can be easily checked is whether the Fourier integral is convergent for
large momenta; the function (2.18) is cut off at the infinities by the factor
n

exp(mik r cosh k )
k=1
which can only go wrong if for some k

e cosh k < 0
but that requires
m k >
2
which is already farther from the real momentum domain then the position of the on-shell propagator singularity.
1 Strictly speaking, this argument only works in an integrable field theory where we know the form (2.17) of the
determinant, and also know the singularity structure of the phase-shift : the two-particle S matrix amplitude as a function
of the rapidity can only have poles (possibly of higher order, but that does not affect the main line of reasoning here).
175
Obtaining a sound mathematical proof requires applying Lschers finite volume expansion
directly to form factors which is out of the scope of the present work. The corrections computed
in Section 4.1 can be considered as the simplest examples, but we intend to take this line of
investigation further in the future [26].
We also remark that there are no finite volume states for which the quantum numbers of any
two of the particles are identical. The reason is that
Sii (0) = 1
(with the exception of free bosonic theories) and so the wave function corresponding to the
appropriate solution of the BetheYang equations (2.14) vanishes. We can express this in terms
of form factors as follows:
0|O(0, 0)|{I1 , I2 , . . . , In }i1 ...in ,L = 0
whenever Ik = Il and ik = il for some k and l. Using this convention we can assume that the summation in (2.9) runs over all possible values of the quantum numbers without exclusions. Note
that even in this case the relation (2.16) can be maintained since due to the exchange axiom (2.3)
FnO (1 , . . . , n )i1 ...in = 0
whenever k = l and ik = il for some k and l.
It is also worthwhile to mention that there is no preferred way to order the rapidities on the
circle, since there are no genuine asymptotic in/out particle configurations. This means that in
relation (2.16) there is no preferred way to order the rapidities inside the infinite volume form
factor function FnO . Different orderings are related by S-matrix factors according to the exchange
axiom (2.3), which are indeed phases. Such phases do not contribute to correlation functions
(cf. the spectral representation (2.8)), nor to any physically meaningful quantity derived from
them. In Section 4.2.1 we show that relations like (2.16) must always be understood to hold only
up to physically irrelevant phase factors.
The quantity i1 ...in (1 , . . . , n ) is nothing else than the density of states in rapidity space. It is
also worthwhile to mention that relation (2.16) can be interpreted as an expression for the finite
volume multi-particle state in terms of the corresponding infinite volume state as follows

1
{I1 , . . . , In }
|1 , . . . , n i1 ...in .
=
(2.19)
i1 ...in ,L
i1 ...in (1 , . . . , n )
This relation between the density and the normalization of states is a straightforward application
of the ideas put forward by Saleur in [27]. Using the crossing formula (2.2), Eq. (2.19) allows us
to construct the general form factor functions (2.1) in finite volume as follows:

j1 ...jm {I1 , . . . , Im } O(0, 0) {I1 , . . . , In } i ...i ,L
1
O (
Fm+n
m
+ i, . . . , 1 + i, 1 , . . . , n )jm ...j1 i1 ...in
i1 ...in (1 , . . . , n )j1 ...jm (1 , . . . , m )

+ O eL
(2.20)
provided that there are no rapidities that are common between the left and the right states, i.e.,
the sets {1 , . . . , n } and {1 , . . . , m } are disjoint. The latter condition is necessary to eliminate
disconnected pieces.
We stress that Eqs. (2.16), (2.20) are exact to all orders of powers in 1/L; we refer to the
corrections non-analytic in 1/L (eventually decaying exponentially as indicated) as residual finite
size effects, following the terminology introduced in [17].
176
3. Form factors from truncated conformal space

3.1. Scaling LeeYang model
3.1.1. Truncated conformal space approach for scaling LeeYang model
We use the truncated conformal space approach (TCSA) developed by Yurov and Zamolodchikov in [18]. The ultraviolet conformal field theory has central charge c = 22/5 and a unique
nontrivial primary field with scaling weights = = 1/5. The cylinder of circumference L
can be mapped unto the complex plane using
2
2
( ix),
z = exp
( + ix).
L
L
The field is normalized so that it has the following operator product expansion:
z = exp
(3.1)
(z, z )(0, 0) = C(zz)1/5 (0, 0) + (zz)2/5 I + ,
(3.2)
where I is the identity operator and the only nontrivial structure constant is
C = 1.911312699 i.
The Hilbert space of the conformal model is given by

Vh V h ,
HLY =
h=0,1/5
where Vh (V h ) denotes the irreducible representation of the left (right) Virasoro algebra with
highest weight h.
The Hamiltonian of scaling LeeYang model takes the following form in the perturbed conformal field theory framework:
L
H
SLY
= H0LY
+ i
dx (0, x),
(3.3)
where
H0LY

2
c
=
L0 + L 0
L
12
is the conformal Hamiltonian. When > 0 the theory above has a single particle in its spectrum
with mass m that can be related to the coupling constant as [28]
= 0.09704845636 m12/5
and the bulk energy density is given by
3 2
B=
m .
12
The S-matrix reads [29]
SLY ( ) =
sinh + i sin 2
3
sinh i sin 2
3
(3.4)
(3.5)
(3.6)
177
and the particle occurs as a bound state of itself at = 2i/3 with the three-particle coupling
given by
2 = 2 3,
where the negative sign is due to the non-unitarity of the model. In this model we define the
phase-shift via the relation
SLY ( ) = ei()
(3.7)
so that (0) = 0. This means a redefinition of Bethe quantum numbers Ik in the BetheYang
equations (2.17) such they become half-integers for states composed of an even number of particles; it also means that in the large volume limit, particle momenta become
2Ik
.
L
Due to translational invariance of the Hamiltonian (3.3), the conformal Hilbert space H can be
split into sectors characterized by the eigenvalues of the total spatial momentum
m sinh k =
2
(L0 L 0 )
L
the operator L0 L 0 generates Lorentz transformations and its eigenvalue is called Lorentz spin.
For a numerical evaluation of the spectrum, the Hilbert space is truncated by imposing a cut in
the conformal energy. The truncated conformal space corresponding to a given truncation and
fixed value s of the Lorentz spin reads

HTCS (s, ecut ) = | H (L0 L 0 )| = s|,

c
| = e|: e ecut .
L0 + L 0
12
On this subspace, the dimensionless Hamiltonian matrix can be written as

2
c
l 22
(s)1 (s)
G
B
L0 + L 0
+i
,
hij =
(3.8)
l
12
(2)12
where energy is measured in units of the particle mass m, l = mL is the dimensionless volume
parameter,
P=
(s)
Gij = i|j
(3.9)
is the conformal inner product matrix and

(s)
Bij = i|(z, z )|j |z=z=1
(3.10)
is the matrix element of the operator at the point z = z = 1 on the complex plane between vectors |i, |j from HTCS (s, ecut ). The natural basis provided by the action of Virasoro generators
is not orthonormal and therefore G(s)1 must be inserted to transform the left vectors to the dual
basis. The Hilbert space and the matrix elements are constructed using an algorithm developed
by Kausch et al. and first used in [30].
Diagonalizing the matrix hij we obtain the energy levels as functions of the volume, with
energy and length measured in units of m. The maximum value of the cutoff ecut we used was 30,
in which case the Hilbert space contains around one thousand vectors, slightly depending on the
spin.
178
3.1.2. Exact form factors of the primary field

Form factors of the trace of the stress-energy tensor were computed by Al.B. Zamolodchikov in [11], and using the relation
= i(1 )
(3.11)
we can rewrite them in terms of . They have the form

Fn (1 , . . . , n ) = Hn Qn (x1 , . . . , xn )
n
n

f (i j )
x i + xj
(3.12)
i=1 j =i+1
with the notations

cosh 1
v(i )v(i + ),
cosh + 1/2

t
t
sinh t
2 sinh 3 sinh 6 it
v( ) = exp 2 dt
,
e
t sinh2 t
f ( ) =
xi = e ,
Hn =
31/4
21/2 v(0)
n
.
The exact vacuum expectation value of the field is

= 1.239394325 i m2/5
which can be readily obtained using (3.4), (3.11) and also the known vacuum expectation value
of [11]
m2
= .
4 3
The functions Qn are symmetric polynomials in the variables xi . Defining the elementary symmetric polynomials of n variables by the relations
n

(x + xi ) =
n
i=1
(n)
x ni i (x1 , . . . , xn ),
(n)
= 0 for i > n,
i=0
they can be constructed as

Q1 = 1,
(2)
(3) (3)
Q2 = 1 ,
(n)
Qn = 1(n) n1
Pn ,
Pn = det M(n)
Q3 = 1 2 ,
n > 3,
(n)
(n)
where Mij = 3i2j +1 ,
i, j = 1, . . . , n 3.
Note that the one-particle form factor is independent of the rapidity:

F1 = 1.0376434349 im2/5 .
(3.13)
179
3.2. Ising model with magnetic perturbation

The critical Ising model is the described by the conformal field theory with c = 1/2 and has
two nontrivial primary fields: the spin operator with = = 1/16 and the energy density
with = = 1/2. The magnetic perturbation
L
H
= H0I
+h
dx (0, x)
0
is massive (and its physics does not depend on the sign of the external magnetic field h). The
spectrum and the exact S matrix is described by the famous E8 factorized scattering theory [31],
which contains eight particles Ai , i = 1, . . . , 8, with mass ratios given by
m2 = 2m1 cos ,
5
m3 = 2m1 cos ,
30
7
m4 = 2m2 cos
,
30
2
m5 = 2m2 cos
,
15
m6 = 2m2 cos ,
30
m7 = 2m4 cos ,
5
m8 = 2m5 cos
5
and the mass gap relation is [32]
m1 = (4.40490857 . . .)|h|8/15
or
15/8
h = h m1
h = 0.06203236 . . . .
(3.14)
The bulk energy density is given by

B = 0.06172858982 m21 .
We also quote the scattering phase shift of two A1 particles:

1
sinh + i sin x
1
2
S11 ( ) =
,
{x} =
.
15 3 5
sinh i sin x
(3.15)
(3.16)
All other amplitudes Sab are determined by the S matrix bootstrap [31]; the only one we need
later is that of the A1 A2 scattering, which takes the form

1
4
2
7
.
S12 ( ) =
5 15 5 15
To have an unambiguous definition of the quantum numbers Ii entering the BetheYang equations (2.14), it is convenient to define phase shift functions ab which are continuous and odd
180
functions of the rapidity difference ; we achieve this using the following convention:
Sab ( ) = Sab (0)eiab () ,
where ab is uniquely specified by continuity and the branch choice
ab (0) = 0
and it is an odd function of due to the following property of the scattering amplitude:
Sab ( )Sab ( ) = 1
(which also implies Sab (0) = 1). The above redefinition of the phase shift compared to the
original one in Eq. (2.15) contains as a special case the LeeYang definition (3.7) and also entails
appropriate redefinition of quantum numbers depending on the sign of Sab (0).
3.2.1. Truncated fermionic space approach for the Ising model
The conformal Ising model can be represented as the theory of a massless Majorana fermion
with the action

1
+ ).
AIsing =
d 2 z (
2
On the conformal plane the model has two sectors, with the mode expansions

r1/2 NeveuSchwarz (NS) sector,
rZ+ 12 br z
(z) =
r1/2
Ramond (R) sector.
rZ br z
and similarly for the antiholomorphic field . The Hilbert space is the direct sum of a certain
projection of the NS and R sectors, with the Virasoro content

Vh V h .
HIsing =
1
h=0, 12 , 16
(s)
The spin field connects the NS and R sectors, and its matrix elements Bij in the sector with
a given conformal spin s (cf. Eq. (3.10)) can be most conveniently computed in the fermionic
basis using the work of Yurov and Zamolodchikov [33], who called this method the truncated
fermionic space approach. The fermionic basis can easily be chosen orthonormal, and thus in
this case the metrics G(s) on the spin subspaces (cf. Eq. (3.9)) are all given by unit matrices of
appropriate dimension. Apart from the choice of basis all the calculation proceeds very similarly
to the case of the LeeYang model. Energy and volume are measured in units of the lowest
particle mass m = m1 and using relation (3.14) one can write the dimensionless Hamiltonian in
the form (3.8). The highest cutoff we use is ecut = 30, in which case the Hilbert space contains
around three thousand vectors (slightly depending on the value of the spin chosen).
We remark that the energy density operator can be represented in the fermionic language as
=
which makes the evaluation of its matrix elements in the fermionic basis extremely simple.
181
3.2.2. Form factors of the energy density operator

The form factors of the operator in the E8 model were first calculated in [34] and their
determination was carried further in [35]. The exact vacuum expectation value of the field is
given by [36]
= h |h|8/15 ,
h = 2.00314 . . . ,
or in terms of the mass scale m = m1

= 0.45475 m.
(3.17)
The form factors are not known for the general n-particle case in a closed form, i.e., no formula
similar to that in (3.12) exists. They can be evaluated by solving the appropriate polynomial
recursion relations derived from the form factor axioms. We do not present explicit formulae
here; instead we refer to the above papers. For practical calculations we used the results computed
by Delfino, Grinza and Mussardo, which can be downloaded from the Web in Mathematica
format [37].
Our interest in the Ising model is motivated by the fact that this is the simplest model in which
form factors of an operator different from the perturbing one are known, and also its spectrum
and bootstrap structure is rather complex, both of which stands in contrast with the much simpler
case of scaling LeeYang model.
3.3. Evaluating matrix elements of a local operator O in TCSA
3.3.1. Identification of multi-particle states
Diagonalizing the TCSA Hamiltonian (3.8) yields a set of eigenvalues and eigenvectors at
each value of the volume, but it is not immediately obvious how to select the same state at
different values of the volume. Therefore in order to calculate form factors it is necessary to
identify the states with the corresponding many-particle interpretation.
Finding the vacuum state is rather simple since it is the lowest lying state in the spin-0 sector
and its energy is given by
E0 (L) = BL + ,
where the ellipsis indicates residual finite size effects decaying exponentially fast with volume L
and B is the bulk energy density which in the models we consider is exactly known (3.5), (3.15).
One-particle states can be found using that their energies can be expressed as

2s 2
(s)
+ m2i +
Ei (L) = BL +
L
again up to residual finite size effects where s is the spin of the sector considered and i is the
species label (every sector contains a single one-particle state for each species).
Higher multi-particle states can be identified by comparing the measured eigenvalues to the
levels predicted by the BetheYang equations. Fixing species labels i1 , . . . , in and momentum
quantum numbers I1 , . . . , In , Eq. (2.14) can be solved to give the rapidities 1 , . . . , n of the
particles as function of the dimensionless volume parameter l = mL. Then the energy of the
multi-particle state in question is
(I ...I )
n
1
Ei1 ...i
(L) = BL +
n
n

k=1
mik cosh k +
182
Fig. 3.1. The first 13 states in the finite volume spectrum of scaling LeeYang model. We plot the energy in units
of m (with the bulk subtracted): e(l) = (E(L) BL)/m, against the dimensionless volume variable l = mL. n-particle
states are labeled by |I1 , . . . , In , where the Ik are the momentum quantum numbers. The state labeled |2, 1, 3 is
actually two-fold degenerate because of the presence of |2, 1, 3 (up to a splitting which vanishes as el , cf. the
discussion in Section 4.3). The dots are the TCSA results and the continuous lines are the predictions of the BetheYang
equations (2.14). The points not belonging to any of the BetheYang lines drawn are two- and three-particle states which
are only partly contained in the first 13 levels due to line crossings, whose presence is a consequence of the integrability
of the model.
which can be compared to the spectrum.

For each state there exists a range of the volume, called the scaling region, where L is large
enough so that the omitted residual finite size effects can be safely neglected and small enough so
that the truncation errors are also negligible. More precisely, the scaling region for any quantity
depending on the volume can be defined as the volume range in which the residual finite size
corrections and the truncation errors are of the same order of magnitude; since both sources of
error show a dependence on the state and the particular quantity considered (as well as on the
value of the cutoff), so does the exact position of the scaling region itself.
In the scaling region, we can use a comparison between the BetheYang predictions and the
numerical energy levels to sort the states and label them by multi-particle quantum numbers. An
example is shown in Fig. 3.1, where we plot the first few states in the spin-0 sector of the scaling
LeeYang model and their identification in terms of multi-particle states is given. In this case, the
agreement with the predicted bulk energy density and the BetheYang levels in the scaling region
is better than one part in 104 for every state shown (with the TCSA cutoff taken at ecut = 30).
3.3.2. Evaluation of matrix elements
Suppose that we computed two Hamiltonian eigenvectors as functions of the volume L (labeled by their quantum numbers in the BetheYang description (2.14), omitting the particle
183
species labels for brevity):

{I1 , . . . , In } =
i (I1 , . . . , In ; L)|i,
L
i

{I , . . . , I } =
j (I1 , . . . , Ik ; L)|j
k L
1
j
in the sector with spin s and spin s , respectively. Let the inner products of these vectors with
themselves be given by
N=
(s)
i (I1 , . . . , In ; L)Gij j (I1 , . . . , In ; L),
i,j
N =
(s )
i (I1 , . . . , Ik ; L)Gij j (I1 , . . . , Ik ; L).
i,j
It is important that the components of the left eigenvector are not complex conjugated. In the
Ising model we work in a basis where all matrix and vector components are naturally real. In the
LeeYang model, the TCSA eigenvectors are chosen so that all of their components i are either
purely real or purely imaginary depending on whether the basis vector |i is an element of the
h = h = 0 or the h = h = 1/5 component in the Hilbert space. It is well known that the Lee
Yang model is non-unitary, which is reflected in the presence of complex structure constants as
indicated in (3.2). This particular convention for the structure constants forces upon us the above
inner product, because it is exactly the one under which TCSA eigenvectors corresponding to
different eigenvalues are orthogonal. We remark that by redefining the structure constants and
the conformal inner product it is also possible to use a manifestly real representation for the Lee
Yang TCSA (up to some truncation effects that lead to complex eigenvalues in the vicinity of
level crossings [18]). Note that the above conventions mean that the phases of the eigenvectors
are fixed up to a sign.
Let us consider a spinless primary field O with scaling weights O = O , which can be
described as the matrix
(s ,s)
Oij
= i|O(z, z )|j |z=z=1 ,
|i HTCS (s , ecut ), |j HTCS (s, ecut ),
between the two truncated conformal space sectors. Then the matrix element of O can be computed as

m2O {I1 , . . . , Ik }O(0, 0){I1 , . . . , In } L

1
2 2O 1
(s ,s)
=
j (I1 , . . . , Ik ; L)Oj l l (I1 , . . . , In ; L),

mL
N N
(3.18)
j,l
where the volume dependent prefactor comes from the transformation of the primary field O
under the exponential map (3.1) and we wrote the equation in a dimensionless form using the
mass scale m. The above procedure is a generalization of the one used by Guida and Magnoli to
evaluate vacuum expectation values in [38]; it was extended to one-particle form factors in the
context of the tricritical Ising model by Fioravanti et al. in [39].
184
Fig. 4.1. The vacuum expectation value of in finite volume. The dashed line shows the exact infinite volume value,
while the continuous line corresponds to Eq. (4.1).
4. Numerical results for elementary form factors

4.1. Vacuum expectation values and one-particle form factors
4.1.1. Scaling LeeYang model
Before the one-particle form factor we discuss the vacuum expectation value. Let us define
the dimensionless function
(l) = im2/5 0||0L ,
where the finite volume expectation value is evaluated from TCSA using (3.18). We performed
measurement of as a function of both the cutoff ecut = 21 . . . 30 and the volume l = 1 . . . 30
and then extrapolated the cutoff dependence fitting a function
12/5
(l, ecut ) = (l) + A(l)ecut
(where the exponent was chosen by verifying that it provides an optimal fit to the data). The
data corresponding to odd and even values of the cutoff must be extrapolated separately [17],
therefore one gets two estimates for the result, but they only differ by a very small amount (of
order 105 at l = 30 and even less for smaller volumes). The theoretical prediction for (l) is

(l) = 1.239394325 + O el .
The numerical result (after extrapolation) is shown in Fig. 4.1 from which it is clear that there is a
long scaling region. Estimating the infinite volume value from the flattest part of the extrapolated
curve (at l around 12) we obtain the following measured value
(l = ) = 1.23938 . . . ,
where the numerical errors from TCSA are estimated to affect only the last displayed digit, which
corresponds to an agreement within one part in 105 .
185
Fig. 4.2. Graphical representation of Eq. (4.1).

Table 4.1
Comparison of Eq. (4.1) to TCSA data
l
2
3
4
5
6
7
8
9
10
11
12
(l) (predicted)
(l) (TCSA)
1.048250
1.184515
1.222334
1.233867
1.237558
1.238774
1.239182
1.239321
1.239369
1.239385
1.239391
1.112518
1.195345
1.224545
1.234396
1.237698
1.238811
1.239189
1.239317
1.239360
1.239373
1.239375
There is also a way to compute the leading exponential correction, which was derived by
Delfino [40]:
1
L = +
(4.1)
F2 (i, 0)ii K0 (mi r) + ,
where

K0 (x) =
d cosh ex cosh
is the modified Bessel-function, and the summation is over the particle species i (there is only
a single term in the scaling LeeYang model). This agrees very well with the numerical data, as
demonstrated in Table 4.1 and also in Fig. 4.1. Using Lschers finite-volume perturbation theory
introduced in [21], the correction term can be interpreted as the sum of Feynman diagrams where
there is exactly one propagator that winds around the cylinder, and therefore Eq. (4.1) can be
represented graphically as shown in Fig. 4.2.
To measure the one-particle form factor we use the correspondence (2.16) between the finite
and infinite volume form factors to define the dimensionless function

1/4
f1s (l) = im2/5 l 2 + (2s)2
0||{s}L ,
186
Fig. 4.3. One-particle form factor from sectors with spin s = 0, 1, 2. The continuous line shows the exact infinite volume
prediction.
where |{s}L is the finite volume one-particle state with quantum number I = s, i.e., from the
spin-s sector. The theoretical prediction for this quantity is

f1s (l) = 1.0376434349 + O el .
(4.2)
The numerical results (after extrapolation in the cutoff) are shown in Fig. 4.3. The scaling region
gives the following estimates for the infinite volume limit:
f10 (l = ) = 1.037654 . . . ,
f11 (l = ) = 1.037650 . . . ,
f12 (l = ) = 1.037659 . . .
which show good agreement with Eq. (4.2) (the relative deviation is again around 105 , as for
the vacuum expectation value).
4.1.2. Ising model in magnetic field
For the Ising model, we again start with checking the dimensionless vacuum expectation value
for which, using Eq. (3.17) we have the prediction

1
L = 0.45475 + O el ,
m
where m = m1 is the mass of the lightest particle and l = mL as before. The TCSA data are
shown in Fig. 4.4. Note that there is substantial dependence on the cutoff ecut and also that
extrapolation in ecut is really required to achieve good agreement with the infinite volume limit.
Reading off the plateau value from the extrapolated data gives the estimate
(l) =
1
= 0.4544 . . .
m
for the infinite volume vacuum expectation value, which has 8 104 relative deviation from the
exact result. Our first numerical comparison thus already tells us that we can expect much larger
187
Fig. 4.4. Measuring the vacuum expectation value of in the Ising model.
Fig. 4.5. The volume dependence of the vacuum expectation value of in the Ising model, showing the extrapolated
value and the prediction from Eq. (4.1), normalized by the value in the infinite volume limit.
truncation errors than in the LeeYang case. It is also clear from Fig. 4.4 that in order to attain
suitable precision in the Ising model extrapolation in the cutoff is very important.
Defining the function
= L /
(l)
we can calculate the leading exponential correction using Eq. (4.1) and the exact two-particle
form factors from [37]. It only makes sense to include particles i = 1, 2, 3 since the contribution
of the fourth particle is subleading with respect to two-particle terms from the lightest particle
due to m4 > 2m1 . The result is shown in Fig. 4.5; we do not give the data in numerical tables, but
we mention that the relative deviation between the predicted and measured value is better than
103 in the range 5 < l < 10.
From now on we normalize all form factors of the operator by the infinite volume vacuum
expectation value (3.17), i.e., we consider form factors of the operator
= /
(4.3)
188
which conforms with the conventions used in [35,37]. We define the dimensionless one-particle
form factor functions as

1/4
mi l 2
s
2
fi (l) =
+ (2s)
0| |{s}i,L .
m1
In the plots of Fig. 4.6 we show how these functions measured from TCSA compare to predictions from the exact form factors for particles i = 1, 2, 3 and spins s = 0, 1, 2, 3.
It is evident that the scaling region sets in much later than for the LeeYang model; therefore
for the Ising model we do not plot data for low values of the volume (all plots start from l
10 . . . 15). This also means that truncation errors in the scaling region are also much larger than
in the scaling LeeYang model; we generally found errors larger by an order of magnitude after
extrapolation in the cutoff. We remark that extrapolation improves the precision by an order of
magnitude compared to the raw data at the highest value of the cutoff.
Note the rather large finite size correction in the case of A3 . This can be explained rather
simply as the presence of a so-called -term. We can again apply Lschers finite-volume perturbation theory, which we use in the form given by Klassen and Melzer in [22] for finite volume
mass corrections. The generalization to one-particle matrix elements is straightforward, and for
a static particle it gives the diagram depicted in Fig. 4.7, whose contribution has the volume
dependence

m2
e311 L , 311 = m21 3 = 0.10453 m1 ,
4
i.e., we can expect a contribution suppressed only by e0.1l . A numerical fit of the l-dependence
in the s = 0 case is perfectly consistent with this expectation. As a result, no scaling region
can be found, because truncation errors are too large in the volume range where the exponential
correction is suitably small. We do not elaborate on this issue further here; we only mention that
starting from this point there are other interesting observations that can be made, and we plan to
return to them in a separate publication [26].
4.2. Two-particle form factors
Following the ideas in the previous subsection, we can again define a dimensionless function
for each two-particle state as follows:
f2 (l)I1 I2 = im2/5 0||{I1 , I2 }L ,
l = mL.
Relation (2.16) gives the following prediction in terms of the exact form factors:
f2 (l)I1 I2 =
im2/5
11 (1 (l), 2 (l))

F2 1 (l), 2 (l) + O el ,
where 1 (l), 2 (l) solve the BetheYang equations

l sinh 1 + (1 2 ) = 2I1 ,
l sinh 2 + (2 1 ) = 2I2
(4.4)
189
Fig. 4.6. One-particle form factors measured from TCSA (dots) compared to the infinite-volume prediction from exact
form factors. All numerical data have been extrapolated to ecut = and s denotes the Lorentz spin of the state considered. The relative deviation in the scaling region is around 103 for A1 and A2 ; there is no scaling region for A3 (see
the discussion in the main text).
190
Fig. 4.7. Leading finite size correction (a so-called -term) to the one-particle form factor of A3 , which results from
the process of splitting up into two copies of A1 which then wind around the cylinder once before recombining into A3
again.
and the density of states is given by

11 (1 , 2 ) = l 2 cosh 1 cosh 2 + l cosh 1 (2 1 ) + l cosh 2 (1 2 )
the phase shift is defined according to Eq. (3.7) and
( ) =
d( )
.
d
There is a further issue to take into account: the relative phases of the multi-particle states are
a matter of convention and the choice made in Section 3.3.2 for the TCSA eigenvectors may
differ from the convention adapted in the form factor bootstrap. Therefore in the numerical work
we compare the absolute values of the functions f2 (l) computed from TCSA with those predicted from the exact form factors. Note that this issue is present for any non-diagonal matrix
element, and was in fact tacitly dealt with in the case of one-particle matrix elements treated in
Section 4.1.1.
The prediction (4.4) for the finite volume two-particle form factors is compared with spin-0
states graphically in Fig. 4.8 and numerically in Table 4.2, while the spin-1 and spin-2 case
is presented in Fig. 4.9 and in Table 4.3. These contain no more than a representative sample
of our data: we evaluated similar matrix elements for a large number of two-particle states for
values of the volume parameter l running from 1 to 30. The behaviour of the relative deviation
is consistent with the presence of a correction of el type up to l 9 . . . 10 (i.e., the logarithm
of the deviation is very close to being a linear function of l), and after l 16 . . . 18 it starts to
increase due to truncation errors. This is demonstrated in Fig. 4.10 using the data presented in
Table 4.3 for spin-1 and spin-2 states,2 but it is equally valid for all the other states we examined.
In the intermediate region l 10 . . . 16 the two sources of numerical deviation are of the same
order, and so that range can be considered as the optimal scaling region: according to the data in
the tables agreement there is typically around 104 (relative deviation). It is also apparent that
2 Note that the dependence of the logarithm of the deviation on the volume is not exactly linear because the residual
finite size correction can also contain a factor of some power of l, and so it is expected that a log l contribution is also
present in the data plotted in Fig. 4.10.
191
Fig. 4.8. Two-particle form factors in the spin-0 sector. Dots correspond to TCSA data, while the lines show the corresponding form factor prediction.
Table 4.2
Two-particle form factors |f2 (l)| in the spin-0 sector
l
2
4
6
8
10
12
14
16
18
20
I1 = 1/2, I2 = 1/2
I1 = 3/2, I2 = 3/2
I1 = 5/2, I2 = 5/2
TCSA
FF
TCSA
FF
TCSA
FF
0.102780
0.085174
0.056828
0.036058
0.023168
0.015468
0.010801
0.007869
0.005950
0.004643
0.120117
0.086763
0.056769
0.035985
0.023146
0.015463
0.010800
0.007867
0.005945
0.004634
0.058158
0.058468
0.050750
0.042123
0.034252
0.027606
0.022228
0.017976
0.014652
0.012061
0.066173
0.059355
0.050805
0.042117
0.034248
0.027604
0.022225
0.017972
0.014645
0.012050
0.039816
0.042072
0.039349
0.035608
0.031665
0.027830
0.024271
0.021074
0.018268
0.015844
0.045118
0.042729
0.039419
0.035614
0.031664
0.027828
0.024267
0.021068
0.018258
0.015827
scaling behaviour starts at quite low values of the volume (around l 4 the relative deviation is
already down to around 1%).
It can be verified by explicit evaluation that in the scaling region the BetheYang density of
states () given in (2.17) differs by corrections of relative magnitude 101 102 (analytically:
of order 1/ l) from the free density of states ( 0 ) in (2.13), and therefore without using the proper
interacting density of states it is impossible to obtain the precision agreement we demonstrated.
In fact the observed 104 relative deviation corresponds to corrections of order l 4 at l = 10, but
it is of the order of estimated truncation errors.3
3 Truncation errors can be estimated by examining the dependence of the extracted data on the cutoff e , as well as
cut
by comparing TCSA energy levels to the BetheYang predictions.
192
Fig. 4.9. Two-particle form factors in the spin-1 and spin-2 sectors. Dots correspond to TCSA data, while the lines show
the corresponding form factor prediction.
Table 4.3
Two-particle form factors |f2 (l)| in the spin-1 and spin-2 sectors
l
2
4
6
8
10
12
14
16
18
20
I1 = 3/2, I2 = 1/2
I1 = 5/2, I2 = 3/2
I1 = 5/2, I2 = 1/2
I1 = 7/2, I2 = 3/2
TCSA
FF
TCSA
FF
TCSA
FF
TCSA
FF
0.077674
0.072104
0.056316
0.042051
0.031146
0.023247
0.017619
0.013604
0.010717
0.008658
0.089849
0.073571
0.056444
0.042054
0.031144
0.023245
0.017616
0.013599
0.010702
0.008580
0.048170
0.049790
0.045031
0.039191
0.033469
0.028281
0.023780
0.019982
0.016831
0.014249
0.054711
0.050566
0.045100
0.039193
0.033467
0.028279
0.023777
0.019977
0.016822
0.014227
0.064623
0.062533
0.051828
0.041370
0.032757
0.026005
0.020802
0.016808
0.013735
0.011357
0.074763
0.063932
0.052009
0.041394
0.032759
0.026004
0.020799
0.016802
0.013724
0.011337
0.042031
0.044034
0.040659
0.036284
0.031850
0.027687
0.023941
0.020659
0.017835
0.015432
0.047672
0.044716
0.040724
0.036287
0.031849
0.027684
0.023936
0.020652
0.017824
0.015413
These results are very strong evidence for the main statement in (4.4) (and thus also (2.16)),
namely, that all 1/L corrections are accounted by the proper interacting state density factor and
that all further finite size corrections are just residual finite size effects decaying exponentially
in L. In Section 4.3 we show that data from higher multi-particle form factors fully support the
above conclusions drawn from the two-particle form factors.
4.2.2. Ising model in magnetic field
In this case, there is some further subtlety to be solved before proceeding to the numerical comparison. Namely, there are spin-0 states which are parity reflections of each other, but
are degenerate according to the BetheYang equations. An example is the state |{1, 1}12 in
Fig. 4.11(b), which is degenerate with |{1, 1}12 to all orders in 1/L. In general the degeneracy
193
Fig. 4.10. Estimating the error term in (4.4) using the data in Table 4.3. The various plot symbols correspond to the same
states as specified in Fig. 4.9.
of these states is lifted by residual finite size effects (more precisely by quantum mechanical
tunnelinga detailed discussion of this mechanism was given in the framework of the k-folded
sine-Gordon model in [41]). Since the finite volume spectrum is parity symmetric, the TCSA
eigenvectors correspond to the states

{1, 1} = 1 {1, 1}
{1, 1} 12,L .
12,L
12,L
2
Because the Hilbert space inner product is positive definite, the TCSA eigenvectors |{1, 1}
12
can be chosen orthonormal and the problem can be resolved by calculating the form factor matrix
element using the two-particle state vectors

+
1
{1, 1} 12,L {1, 1} 12,L .
2
Because the Ising spectrum is much more complicated than that of the scaling LeeYang model
(and truncation errors are larger as well), we only identified two-particle states containing two
copies of A1 , or an A1 and an A2 . The numerical results are plotted in Figs. 4.11(a) and 4.11(b),
respectively. The finite volume form factor functions of the operator (4.3) are defined as

f11 (l)I1 I2 = 11 1 (l), 2 (l) 0| |{I1 , I2 }11 ,
where
l sinh 1 + 11 (1 2 ) = 2I1 ,
l sinh 2 + 11 (2 1 ) = 2I2 ,
11 (1 , 2 ) = l 2 cosh 1 cosh 2 + l cosh 1 11 (2 1 ) + l cosh 2 11 (1 2 ),
d11 ( )
11 ( ) =
d
194
Fig. 4.11. Two-particle form factors in the Ising model. Dots correspond to TCSA data, while the lines show the corresponding form factor prediction.
and
f12 (l)I1 I2 =
11 (1 , 2 )0| |{I1 , I2 }12
with
l sinh 1 + 12 (1 2 ) = 2I1 ,
m2
l sinh 2 + 12 (2 1 ) = 2I2 ,
m1
m2 2
m2
12 (1 , 2 ) =
l cosh 1 cosh 2 + l cosh 1 12 (2 1 ) +
l cosh 2 12 (1 2 ),
m1
m1
d12 ( )
12 ( ) =
d
195
and are compared against the form factor functions

F2 1 (l), 2 (l) 11
and

F2 1 (l), 2 (l) 12
respectively.
Although (as we already noted) truncation errors in the Ising model are much larger than in the
LeeYang case, extrapolation in the cutoff improves them by an order of magnitude compared to
the evaluation at the highest cutoff (in our case 30). After extrapolation, deviations in the scaling
region become less than 1% (with a minimum of around 103 in the A1 A1 , and 104 in the A1 A2
case), and even better for states with nonzero total spin. As noted in the previous subsection this
means that the numerics is really sensitive to the dependence of the particle rapidities and state
density factors on the interaction between the particles; generally the truncation errors in the
extrapolated data are about two orders of magnitude smaller than the interaction corrections.
It is a general tendency that the agreement is better in the sectors with nonzero spin, and the
scaling region starts at smaller values of the volume. This is easy to understand for the energy levels, since for low-lying states nonzero spin generally means higher particle momenta. The higher
the momenta of the particles, the more the BetheYang contributions dominate over the residual
finite size effects. This is consistent with the results of Rummukainen and Gottlieb in [42] where
it was found that resonance phase shifts can be more readily extracted from sectors with nonzero
momentum; our data show that this observation carries over to general matrix elements as well.
4.3. Many-particle form factors
We also performed numerical evaluation of three- and four-particle form factors in the scaling
LeeYang model; some of the results are presented in Figs. 4.12 and 4.13, respectively. For
the sake of brevity we refrain from presenting explicit numerical tables; we only mention that
the agreement between the numerical TCSA data and the prediction from the exact form factor
solution is always better than 103 in the scaling region. For better visibility we plotted the
functions

2/5
k (1 , . . . , k )0||{I1 , . . . , Ik }L , l = mL,

fk (l)I1 ...Ik = im
for which relation (2.16) gives:

fk (l)I1 ...Ik = im2/5 Fk (1 , . . . , k ) + O el .
(4.5)
Due to the fact that in the LeeYang model there is only a single particle species, we introduced
the simplified notation n for the n-particle Jacobi determinant.
The complication noted in Section 4.2.2 for the Ising state |{1, 1}12 is present in the Lee
Yang model as well. The BetheYang equations give degenerate energy values for the states
|{I1 , . . . , Ik }L and |{Ik , . . . , I1 }L (as noted before, the degeneracy is lifted by quantum mechanical tunneling). For states with nonzero spin this causes no problem, because these two states
are in sectors of different spin (their spins differ by a sign) and similarly there is no difficulty
when the two quantum number sets are identical, i.e.,
{I1 , . . . , Ik } = {I1 , . . . , Ik }
196
Fig. 4.12. Three-particle form factors in the spin-0 sector. Dots correspond to TCSA data, while the lines show the
corresponding form factor prediction.
Fig. 4.13. Four-particle form factors in the spin-0 sector. Dots correspond to TCSA data, while the lines show the corresponding form factor prediction.
since
then there is a single state. However, there are states in the zero spin sector (i.e., with
k Ik = 0) for which
{I1 , . . . , Ik } = {I1 , . . . , Ik }.
We use two such pairs of states in our data here: the three-particle states |{3, 1, 2}L ,
|{2, 1, 3}L and the four-particle states |{7/2, 1/2, 3/2, 5/2}L , |{5/2, 3/2, 1/2, 7/2}L .
Again, the members of such pairs are related to each other by spatial reflection, which is a symmetry of the exact finite-volume Hamiltonian and therefore (supposing that the eigenvectors are
197
orthonormal) the finite volume eigenstates correspond to

{I1 , . . . , Ik } = 1 {I1 , . . . , Ik } {Ik , . . . , I1 }
L
L
L
2
and this must be taken into account when evaluating the form factor matrix elements. In the
LeeYang case, however, the inner product is not positive definite (and some nonzero vectors
may have zero length, although this does not happen for TCSA eigenvectors, because they are
orthogonal to each other and the inner product is non-degenerate), but there is a simple procedure
that can be used in the general case. Suppose the two TCSA eigenvectors corresponding to such
a pair are v1 and v2 . Then we can define their inner product matrix as
gij = vi G(0) vj
using the TCSA inner product (3.9). The appropriate basis vectors of this two-dimensional subspace, which can be identified with |{I1 , . . . , Ik }L and |{Ik , . . . , I1 }L , can be found by
solving the two-dimensional generalized eigenvalue problem
g w = P w
for the vector (w1 , w2 ) describing orientation in the subspace, with

0 1
P=
.
1 0
This procedure has the effect of rotating from the basis of parity eigenvectors to basis vectors
which are taken into each other by spatial reflection.
4.3.2. Ising model in a magnetic field
As we already noted, it is much harder to identify4 higher states in the Ising model due to the
complexity of the spectrum, and so we only performed an analysis of states containing three A1
particles. We define

f111 (l)I1 I2 I3 = 111 1 (l), 2 (l), 3 (l) 0| |{I1 , I2 , I3 }111 ,
where i (l) are the solutions of the three-particle BetheYang equations in (dimensionless) volume l and 111 is the appropriate 3-particle determinant. The results of the comparison can be
seen in Fig. 4.14. The numerical precision indicated for two-particle form factors at the end of
Section 4.2.2, as well as the remarks made there on the spin dependence apply here as well; we
only wish to emphasize that for A1 A1 A1 states with nonzero total spin the agreement between
the extrapolated TCSA data and the form factor prediction in the optimal part of the scaling
region is within 2 104 .
5. General form factors without disconnected pieces
Let us consider a matrix element of the form

j1 ...jm {I1 , . . . , Im } O(0, 0) {I1 , . . . , In } i ...i
1
n ,L
4 To identify A A A states it is necessary to use at least e

cut = 22 or 24 and even then the agreement with the Bethe
1 1 1
Yang prediction is still only within 20%, but the identification can be made for the first few A1 A1 A1 states using data
up to ecut = 30. Truncation errors are substantially decreased by extrapolation to ecut = .
198
Fig. 4.14. Three-particle form factors in the Ising model. Dots correspond to TCSA data, while the lines show the
corresponding form factor prediction.
Disconnected pieces are known to appear when there is at least one particle in the state on the left
which occurs in the state on the right with exactly the same rapidity. The rapidities of particles
as a function of the volume are determined by the BetheYang equations (2.14)

Qk (1 , . . . , n )i1 ...in = mik L sinh k +
ik il (k l ) = 2Ik , k = 1, . . . , n,
l=k
and
Qk (1 , . . . , m )j1 ...jm = mjk L sinh k +
jk jl (k l ) = 2Ik ,
k = 1, . . . , m.
l=k
Due to the presence of the scattering terms containing the phase shift functions , equality of two
quantum numbers Ik and Il does not mean that the two rapidities themselves are equal in finite
volume L. It is easy to see that there are only two cases when exact equality of some rapidities
can occur:
(1) The two states are identical, i.e., n = m and
{j1 . . . jm } = {i1 . . . in },
{I1 , . . . , Im } = {I1 , . . . , In }
in which case all the rapidities are pairwise equal, or
(2) Both states are parity symmetric states in the spin zero sector, i.e.,
{I1 , . . . , In } {In , . . . , I1 },
{I1 , . . . , Im } {Im , . . . , I1 }
and the particle species labels are also compatible with the symmetry, i.e., in+1k = ik and
jm+1k = jk . Furthermore, both states must contain one (or possibly more, in a theory with
more than one species) particle of quantum number 0, whose rapidity is then exactly 0 for
any value of the volume L due to the symmetric assignment of quantum numbers.
199
Discussion of such matrix elements raises many interesting theoretical considerations and is
postponed to the followup paper [20]; here we only concentrate on matrix elements for which
there are no disconnected contributions.
In this model there is a single particle species, so we can introduce the following notations:

I ,...,I
fkn (l)I11 ,...,Ikn = im2/5 {I1 , . . . , Ik }(0, 0){I1 , . . . , In } L
and also

I ,...,I
fkn (l)I11 ,...,Ikn

= im2/5 k (1 , . . . , k ) n (1 , . . . , n ) {I1 , . . . , Ik }(0, 0){I1 , . . . , In } L
for which relation (2.20) yields

I ,...,I
( + i, . . . , + i, , . . . , )

Fk+n
1
n
1
k
+ O el ,
n (1 , . . . , n )k (1 , . . . , m )

= im2/5 Fk+n
(k + i, . . . , 1 + i, 1 , . . . , n ) + O el .
fkn (l)I11 ,...,Ikn = im2/5

I ,...,I
fkn (l)I11 ,...,Ikn
(5.1)
For the plots we chose to display f or f depending on which one gives a better visual picture.
The numerical results shown here are just a fraction of the ones we actually obtained, but all
of them show an agreement with precision 104 103 in the scaling region (the volume range
corresponding to the scaling region typically varies depending on the matrix element considered
due to variation in the residual finite size corrections and truncation effects).
The simplest cases involve one- and two-particle states: the one-particleone-particle data in
Fig. 5.1 actually test the two-particle form factor F2 , while the one-particletwo-particle plot
Fig. 5.2 corresponds to F3 (we obtained similar results on F4 using matrix elements f22 ). Note
that in contrast to the comparisons performed in Sections 4.2 and 4.3, these cases involve the
form factor solutions (3.12) at complex values of the rapidities. In general, all tests performed
with TCSA can test form factors at rapidity arguments with imaginary parts 0 or , which are the
only parts of the complex rapidity plane where form factors eventually correspond to physical
matrix elements.
One-particlethree-particle and one-particlefour-particle matrix elements f13 and f14 contribute another piece of useful information. We recall that there are pairs of parity-related states
in the spin-0 factors which we cannot distinguish in terms of their elementary form factors. In
Section 4.3 we showed the example of the three-particle states

{3, 1, 2}
and {2, 1, 3} L
L
and the four-particle states

{7/2, 1/2, 3/2, 5/2}

L
and {5/2, 3/2, 1/2, 7/2} L .
In fact it is only true that they cannot be distinguished if the left state is parity-invariant. However,
using a one-particle state of nonzero spin on the left it is possible to distinguish and appropriately
label the two states, as shown in Figs. 5.3 and 5.4. This can also be done using matrix elements
200
Fig. 5.1. One-particleone-particle form factors in LeeYang model. Dots correspond to TCSA data, while the lines show
Fig. 5.2. One-particletwo-particle form factors in LeeYang model. Dots correspond to TCSA data, while the lines show
with two-particle states of nonzero spin: the two-particlethree-particle case f23 is shown in
Fig. 5.5 (similar results were obtained for f24 ). Examining the data in detail shows that the
identifications provided using different states on the left are all consistent with each other.
201
Fig. 5.3. One-particlethree-particle form factors in LeeYang model. Dots correspond to TCSA data, while the lines
show the corresponding form factor prediction.
Fig. 5.4. One-particlefour-particle form factors in LeeYang model. Dots correspond to TCSA data, while the lines
It is also interesting to note that the f14 (Fig. 5.4) and f23 data (Fig. 5.5) provide a test for the
five-particle form factor solutions F5 . This is important since it is progressively harder to identify
many-particle states in the TCSA spectrum for two reasons. First, the spectrum itself becomes
more and more dense as we look for higher levels; second, the truncation errors grow as well.
Both of these make the identification of the energy levels by comparison with the predictions
of the BetheYang equations more difficult; in the LeeYang case we stopped at four-particle
levels. However, using general matrix elements and the relations (5.1) we can even get data for
202
Fig. 5.5. Two-particlethree-particle form factors in LeeYang model. Dots correspond to TCSA data, while the lines
Fig. 5.6. Three-particlethree-particle and four-particlefour-particle form factors in LeeYang model. Dots correspond
to TCSA data, while the lines show the corresponding form factor prediction.
form factors up to 8 particles, a sample of which is shown in Fig. 5.6 (f33 and f44 , corresponding
to 6 and 8 particle form factors) and Fig. 5.7 (f34 which corresponds to 7 particle form factors).
5.2. Ising model in magnetic field
In the case of the Ising model, we define the functions

fj1 ...jm ;i1 ...in (l) = i1 ...in (1 , . . . , n )j1 ...jm (1 , . . . , m )

j1 ...jm {I1 , . . . , Im } {I1 , . . . , In } i ...i

1
n ,L
203
Fig. 5.7. Three-particlefour-particle form factors in LeeYang model. Dots correspond to TCSA data, while the lines
which are compared against form factors
Fm+n
(m + i, . . . , 1 + i, 1 , . . . , n )jm ...j1 i1 ...in ,
where i and j denote the rapidities obtained as solutions of the appropriate BetheYang equations at the given value of the volume. We chose states for which the necessary form factor
solution was already known (and given in [37]), i.e., we did not construct new form factor solutions ourselves.
One-particleone-particle form factors are shown in Fig. 5.8; these provide another numerical
test for the two-particle form factors examined previously in Section 4.2.2. One-particletwoparticle form factors, besides testing again the three-particle form factor A1 A1 A1 (Fig. 5.9 (a))
also provide information on A1 A1 A2 (Fig. 5.9(b)).
Finally, one-particlethree-particle and two-particletwo-particle matrix elements can be
compared to the A1 A1 A1 A1 form factor, which again shows that by considering general matrix elements we can go substantially higher in the form factor tree than using only elementary
form factors (Fig. 5.10).
We remark that the cusps on the horizontal axis in the form factor plots correspond to zeros
where the form factors change sign; they are artifacts introduced by taking the absolute value of
the matrix elements. The pattern of numerical deviations between TCSA data and exact form factor predictions is fully consistent with the discussion in the closing paragraphs of Sections 4.2.2
and 4.3.2. The deviations in the scaling region are around 1% on average, with agreement of the
order of 103 in the optimal range.
6. Conclusions
In this work we gave an expression for the matrix elements of local operators between general
multi-particle states in finite volume which is valid to any order in the expansion in the inverse of
the volume 1/L. It was shown in Section 2.2 using a very general argument that all the remaining
volume dependence is non-analytic in 1/L (it is given by residual finite size effects vanishing
exponentially with increasing volume). It is also clear that the derivation itself does not depend
on integrability, neither it is restricted to (1 + 1)-dimensional field theories and therefore relations
204
Fig. 5.8. One-particleone-particle form factors in the Ising model. Dots correspond to TCSA data, while the lines show
(2.16) and (2.20) can be extended to general quantum field theories (substituting the rapidities
with appropriate kinematical parametrization and the n with the proper state densities), with the
only condition that their spectrum of excitations must possess a mass gap.
(1 + 1)-dimensional integrable field theories are special in the respect that multi-particle states
in finite volume can be described using the BetheYang equations (2.14) and so the n-particle
state density n can be obtained in the general closed form (2.17). Another important feature is
that there are exact results for matrix elements of local operators in infinite volume which can
be obtained from the form factor bootstrap briefly reviewed in Section 2.1. Therefore they are
ideal toy models to test ideas about finite size corrections. Such an approach is also interesting
due to a fundamental property of the bootstrap, namely that it is only indirectly related to the
actual Lagrangian (or Hamiltonian) field theory. As we discussed in the introduction, testing the
conjectured form factors against field theory usually involves calculating two-point functions
using spectral representations, or sum rules derived from such expansions; however, direct non-
205
Fig. 5.9. One-particletwo-particle form factors in the Ising model. Dots correspond to TCSA data, while the lines show
perturbative comparison of the actual form factors to matrix elements computed from the field
theory have been very restricted so far.
Using TCSA we were able to give an extensive and direct numerical comparison between
bootstrap results for form factors and matrix elements evaluated non-perturbatively. One of the
advantages is that we can compare matrix elements directly, without using any proxy (such as
a two-point function or a sum rule); the other is the very high precision of the comparison and
also that it is possible to test form factors of many particles which have never been tested using
spectral sums, mostly due to the fact that usually their contribution to spectral expansions is
extremely small, and evaluating it also involves calculating multidimensional integrals to very
high precision, which a numerically difficult task. The second problem is actually related to the
first, since due to the smallness of the contribution from higher particle terms all the lower ones
must be evaluated to sufficiently high precision. Our approach, in contrast, makes it possible to
have a test of entire one-dimensional sections of the form factor functions using the volume as
206
Fig. 5.10. One-particlethree-particle and two-particletwo-particle form factors in the Ising model. Dots correspond to
TCSA data, while the lines show the corresponding form factor prediction.
a parameter, and the number of available sections only depends on our ability to identify multiparticle states in finite volume.
Our results can also be viewed in the context of finite volume form factors [15] (which is also
related to the problem of finite temperature form factors; for a review on the latter see [16] and
references therein). The relations (2.16), (2.20) give finite volume form factors expressed with
their infinite volume counterparts to all orders in 1/L (where L denotes the volume), i.e., up to
exponentially decaying terms in L. This gives finite volume form factors in large volume with
very high precision. On the other hand, what we determine numerically in TCSA are actually
the finite volume form factors themselves, which is an approach that primarily works in small
enough volume due to the truncation errors. In the LeeYang case, the combination of the two
approaches gives the finite volume form factors involving up to four particles with better than
103 relative precision, as demonstrated by the excellent agreement in the scaling region of
TCSA where their domains of validity overlap. For the Ising model the numerical precision is
207
not as good, but with some care a precision of around 103 can be achieved for most of the
matrix elements considered in this paper.
An open question which is not discussed in this paper is the case of matrix elements with
disconnected pieces. Results on such matrix elements are already available, but we postpone
them to a followup paper [20] where we also plan to discuss many theoretical issues related to
crossing and disconnected contributions in finite volume.
Acknowledgements
We wish to thank Z. Bajnok and L. Palla for useful discussions. This research was partially
supported by the Hungarian research funds OTKA T043582, K60040 and TS044839. G.T. was
also supported by a Bolyai Jnos research scholarship.
References
[1] A.B. Zamolodchikov, Al.B. Zamolodchikov, Ann. Phys. 120 (1979) 253291.
[2] G. Mussardo, Phys. Rep. 218 (1992) 215379.
[3] M. Karowski, P. Weisz, Nucl. Phys. B 139 (1978) 455;
B. Berg, M. Karowski, P. Weisz, Phys. Rev. D 19 (1979) 2477;
M. Karowski, Phys. Rep. 49 (1979) 229.
[4] F.A. Smirnov, Form-factors in completely integrable models of quantum field theory, Adv. Ser. Math. Phys. 14
(1992) 1208.
[5] J.L. Cardy, G. Mussardo, Nucl. Phys. B 340 (1990) 387402.
[6] A. Koubek, G. Mussardo, Phys. Lett. B 311 (1993) 193201, hep-th/9306044.
[7] A. Koubek, Nucl. Phys. B 428 (1994) 655680, hep-th/9405014.
[9] F.A. Smirnov, Nucl. Phys. B 453 (1995) 807824, hep-th/9501059.
[10] V.P. Yurov, Al.B. Zamolodchikov, Int. J. Mod. Phys. A 6 (1991) 34193440.
[11] Al.B. Zamolodchikov, Nucl. Phys. B 348 (1991) 619641.
[12] A.B. Zamolodchikov, Pisma Zh. Eksp. Teor. Fiz. 43 (1986) 565, JETP Lett. 43 (1986) 730.
[13] J.L. Cardy, Phys. Rev. Lett. 60 (1988) 2709;
A. Cappelli, D. Friedan, J.I. Latorre, Nucl. Phys. B 352 (1991) 616670;
D.Z. Freedman, J.I. Latorre, X. Vilasis, Mod. Phys. Lett. A 6 (1991) 531542.
[14] G. Delfino, P. Simonetti, J.L. Cardy, Phys. Lett. B 387 (1996) 327333, hep-th/9607046.
[15] F.A. Smirnov, hep-th/9802132;
V.E. Korepin, N.A. Slavnov, Int. J. Mod. Phys. B 13 (1999) 29332942, math-ph/9812026;
G. Mussardo, V. Riva, G. Sotkov, Nucl. Phys. B 670 (2003) 464578, hep-th/0307125.
[16] B. Doyon, SIGMA 3 (2007) 011, hep-th/0611066.
[17] G. Takcs, B. Pozsgay, Nucl. Phys. B 748 (2006) 485523, hep-th/0604022.
[20] B. Pozsgay, G. Takcs, Form factors in finite volume II: Disconnected terms and finite temperature correlators,
Nucl. Phys. B 788 (2008) 209, arXiv: 0706.3605 [hep-th].
[22] T.R. Klassen, E. Melzer, Nucl. Phys. B 362 (1991) 329388.
[24] G. Delfino, G. Mussardo, Nucl. Phys. B 455 (1995) 724758, hep-th/9507010.
[25] C.J.D. Lin, G. Martinelli, C.T. Sachrajda, M. Testa, Nucl. Phys. B 619 (2001) 467498, hep-lat/0104006.
[26] B. Pozsgay, in preparation.
[27] H. Saleur, Nucl. Phys. B 567 (2000) 602610, hep-th/9909019.
[29] J.L. Cardy, G. Mussardo, Phys. Lett. B 225 (1989) 275278.
[30] H. Kausch, G. Takcs, G. Watts, Nucl. Phys. B 489 (1997) 557579, hep-th/9605104.
208
[31] A.B. Zamolodchikov, Adv. Stud. Pure Math. 19 (1989) 641;

A.B. Zamolodchikov, Int. J. Mod. Phys. A 3 (1988) 743.
[32] V.A. Fateev, Phys. Lett. B 324 (1994) 4551.
[34] G. Delfino, P. Simonetti, Phys. Lett. B 383 (1996) 450456, hep-th/9605065.
[35] G. Delfino, P. Grinza, G. Mussardo, Nucl. Phys. B 737 (2006) 291303, hep-th/0507133.
[36] V.A. Fateev, S. Lukyanov, A.B. Zamolodchikov, Al.B. Zamolodchikov, Nucl. Phys. B 516 (1998) 652674, hepth/9709034.
[37] http://people.sissa.it/~delfino/isingff.html.
[38] R. Guida, N. Magnoli, Phys. Lett. B 411 (1997) 127133, hep-th/9706017.
[39] D. Fioravanti, G. Mussardo, P. Simon, Phys. Rev. E 63 (2001) 016103, cond-mat/0008216.
[40] G. Delfino, J. Phys. A 34 (2001) L161L168, hep-th/0101180.
[41] Z. Bajnok, L. Palla, G. Takcs, F. Wgner, Nucl. Phys. B 587 (2000) 585618, hep-th/0004181.
[42] K. Rummukainen, S.A. Gottlieb, Nucl. Phys. B 450 (1995) 397436, hep-lat/9503028.
Form factors in finite volume II:

Disconnected terms and finite temperature correlators
B. Pozsgay a , G. Takcs b,
a Institute for Theoretical Physics, Etvs University, Budapest, Hungary
b HAS Research Group for Theoretical Physics, H-1117 Budapest, Pzmny Pter stny 1/A, Hungary
Received 28 June 2007; accepted 11 July 2007
Abstract
Continuing the investigation started in a previous work, we consider form factors of integrable quantum
field theories in finite volume, extending our investigation to matrix elements with disconnected pieces.
Numerical verification of our results is provided by truncated conformal space approach. Such matrix elements are important in computing finite temperature correlation functions, and we give a new method for
generating a low temperature expansion, which we test for the one-point function up to third order.
1. Introduction
The matrix elements of local operators, the so-called form factors are central objects in
quantum field theory. In two-dimensional integrable quantum field theory, the S matrix can be
obtained exactly in the framework of factorized scattering (see [1,2] for reviews). Using the scattering amplitudes as input, it is possible to obtain a set of axioms [3] which provides the basis
for the form factor bootstrap (see [4] for a review).
Although in the bootstrap approach the connection with the Lagrangian formulation of quantum field theory is rather indirect, it is thought that the general solution of the form factor axioms
determines the complete local operator algebra of the theory [5], which was confirmed in many
cases by explicit comparison of the space of solutions to the spectrum of local operators [69].
DOI of companion paper: 10.1016/j.nuclphysb.2007.06.027.
E-mail addresses: pozsi@bolyai.elte.hu (B. Pozsgay), takacs@elte.hu (G. Takcs).

doi:10.1016/j.nuclphysb.2007.07.008
210
Another important piece of information comes from correlation functions: using form factors,
a spectral representation for the correlation functions can be built which provides a large distance expansion [10,11], while the Lagrangian or perturbed conformal field theory formulation
allows one to obtain a short-distance expansion, which can then be compared provided there is an
overlap between their regimes of validity [11]. Other evidence for the correspondence between
the field theory and the solutions of the form factor bootstrap results from evaluating sum rules
like Zamolodchikovs c-theorem [12,13] or the -theorem [14], both of which can be used to
express conformal data as spectral sums in terms of form-factors. Direct comparisons with multiparticle matrix elements are not so readily available, except for perturbative or 1/N calculations
in some simple cases [3]. One of our aims is to provide non-perturbative evaluation of form factors from the Hamiltonian formulation, which then allows for a direct comparison with solutions
of the form factor axioms.
Based on what we learned from our previous investigation of decay rates in finite volume [15],
in our previous paper [16] we determined form factors using a formulation of the field theory in
finite volume. We used the truncated conformal space approach (TCSA) developed by Yurov
and Al.B. Zamolodchikov [17] as a basis for numerical comparison to non-perturbative Hamiltonian formulation of quantum field theory, and also its fermionic version in the case of the
Ising model [18]. We were able to give an extensive and direct numerical comparison between
bootstrap results for form factors and matrix elements evaluated non-perturbatively. One of the
advantages is that we can compare matrix elements directly, without using any proxy (such as a
two-point function or a sum rule); the other is the very high precision of the comparison and also
that it is possible to test form factors of many particles which have never been tested using spectral sums. Our approach, in contrast, makes it possible to test entire one-dimensional sections of
the form factor functions using the volume as a parameter, and the number of available sections
only depends on our ability to identify multi-particle states in finite volume. Part of the motivation of this work is to complete the non-perturbative evaluation of form factors by extending our
results to matrix elements with disconnected pieces.
Another motivation is provided by the fact that such matrix elements are relevant for the calculation of finite temperature correlators. Finite temperature correlation functions have attracted
quite a lot of interest recently [1927]. Leclair and Mussardo proposed an expansion for the onepoint and two-point functions in terms of form factors dressed by appropriate occupation number
factors containing the pseudo-energy function from the thermodynamical Bethe Ansatz [20]. It
was shown by Saleur [21] that their proposal for the two-point function is incorrect; on the other
hand, he gave a proof of the LeclairMussardo formula for one-point functions provided the operator considered is the density of some local conserved charge. His proof is based on a conjecture
concerning the expression of diagonal finite volume matrix elements in terms of connected form
factors. In view of the evidence it is now generally accepted that the conjecture made by Leclair
and Mussardo for the one-point functions is correct; in contrast, the case of two-point functions
(and also higher ones) is not yet fully understood (see the introductory part of Section 7 for more
details). Here we investigate how finite temperature one-point functions can be expanded systematically using finite volume L as a regulator and make a proposal which is expected to be
valid for multi-point correlators as well.
Our exposition is structured as follows. In Section 2, after recalling the form factor bootstrap
axioms, we present a brief review of the approach developed in our earlier paper [16] (to which
we refer the interested reader for more details), and then we state our main result which is the
description of all matrix elements containing disconnected contributions. In Section 3 we briefly
recall the two models used for numerical comparison, which are the scaling LeeYang model and
211
the Ising model in a magnetic field. We omit the description of the method for obtaining matrix
elements from truncated conformal space, and instead we refer the interested reader to [16] where
all the necessary details can be found.
As we showed in [16], there are essentially two types of matrix elements with disconnected
contributions. Section 4 is devoted to the first type, which is the case of diagonal matrix elements;
we present a general formula for them in terms of the symmetric evaluation of the diagonal form
factor and test it against truncated conformal space. In Section 5 we analyze diagonal matrix
elements in terms of connected form factor amplitudes, and we show that our results are fully
consistent with the above-mentioned conjecture made by Saleur in [21]. In Section 6 we discuss
the second type of matrix elements with disconnected contributions, namely those with particles
of exactly zero momentum in the finite volume states. Adding the results presented in Sections 4
and 6 to those obtained in [16], we achieve a complete description of all multi-particle matrix
elements of a general local operator to all orders in 1/L. Section 7 is devoted to finite temperature correlation functions: we propose a systematic method for deriving a low-temperature
expansion, which is applied to one-point functions and tested by comparing the results to the
LeclairMussardo expansion [20]. We also briefly discuss the extension of our method to the
evaluation of two-point functions. Section 8 is reserved for the conclusions.
2. Form factors in finite volume: A brief review
2.1. Form factor bootstrap
Here we give a very brief summary of the axioms of the form factor bootstrap, because we
need them in the sequel; for more details we refer to Smirnovs review [4]. Let us suppose for
simplicity that the theory has particles Ai , i = 1, . . . , N , with masses mi which are strictly nondegenerate, i.e., mi = mj for any i = j (and therefore the particles are also self-conjugate).
Because of integrability, multi-particle scattering amplitudes factorize into the product of pairwise two-particle scatterings, which are purely elastic (in other words: diagonal). This means
that any two-particle scattering amplitude is a pure phase, which we denote by Sij ( ) where
is the relative rapidity of the incoming particles Ai and Aj . Incoming and outgoing asymptotic
states can be distinguished by the ordering of the rapidities:

|1 , . . . , n in
i1 ...in : 1 > 2 > > n ,
|1 , . . . , n i1 ...in =
|1 , . . . , n out
i1 ...in : 1 < 2 < < n ,
and states which only differ in the order of rapidities are related by
|1 , . . . , k , k+1 , . . . , n i1 ...ik ik+1 ...in
= Sik ik+1 (k k+1 )|1 , . . . , k+1 , k , . . . , n i1 ...ik+1 ik ...in .
The normalization of these states is specified by giving the following inner product among oneparticle state:
j
| i = ij 2( ).
For a local operator O(t, x) the form factors are defined as

O
(m , . . . , 1 |1 , . . . , n )j1 ...jm ;i1 ...in = j1 ...jm 1 , . . . , m |O(0, 0)|1 , . . . , n i1 ...in . (2.1)
Fmn
With the help of the crossing relations
212

O
Fmn
(1 , . . . , m |1 , . . . , n )j1 ...jm ;i1 ...in
O

= Fm1n+1
(1 , . . . , m1
|m + i, 1 , . . . , n )j1 ...jm1 ;jm i1 ...in
n
2jm ik (m k )
k=1
k1
Sil ik (l k )
l=1
O

(1 , . . . , m1
|1 , . . . , k1 , k+1 . . . , n )j1 ...jm1 ;jm i1 ...ik1 ik+1 ...in
Fm1n1
(2.2)
all form factors can be expressed in terms of the elementary form factors
FnO (1 , . . . , n )i1 ...in = 0|O(0, 0)|1 , . . . , n i1 ...in
(2.3)
which satisfy the following axioms:

(I) Exchange:
FnO (1 , . . . , k , k+1 , . . . , n )i1 ...ik ik+1 ...in
= Sik ik+1 (k k+1 )FnO (1 , . . . , k+1 , k , . . . , n )i1 ...ik+1 ik ...in .
(2.4)
(II) Cyclic permutation:

FnO (1 + 2i, 2 , . . . , n ) = FnO (2 , . . . , n , 1 ).
(2.5)
(III) Kinematical singularity

O
( + i, , 1 , . . . , n )ij i1 ...in
i Res Fn+2
=

n

= 1 ij
Siik ( k ) FnO (1 , . . . , n )i1 ...in .
(2.6)
k=1
(IV) Dynamical singularity

j
O
+ i u ij k /2, i u ik /2, 1 , . . . , n ij i
i Res Fn+2
=
O
= ijk Fn+1
(, 1 , . . . , n )ki1 ...in
1 ...in
(2.7)
whenever k occurs as the bound state of the particles i and j , corresponding to a bound state pole
of the S matrix of the form
i(ijk )2

k
Sij iuij
,
iukij
(2.8)
where ijk is the on-shell three-particle coupling and ukij is the so-called fusion angle. The fusion
angles satisfy
m2k = m2i + m2j + 2mi mj cos ukij ,
j
2 = ukij + uik + uij k

and we also used the notation u kij = ukij . The axioms (I)(IV) are supplemented by the
assumption of maximum analyticity (i.e., that the form factors are meromorphic functions which
only have the singularities prescribed by the axioms) and possible further conditions expressing
properties of the particular operator whose form factors are sought.
213
We remark that with the exception of free bosonic theories, all known exact S matrices satisfy
Sii (0) = 1
and therefore the elementary form factors (2.3) have an exclusion property: they vanish whenever
the rapidities of two particles belonging to the same species coincide.
2.2. Finite volume matrix elements to all orders in 1/L
Following our conventions in [16], the finite volume multi-particle states can be denoted

{I1 , . . . , In }
,
i ...i ,L
n
where the Ik are momentum quantum numbers and ik are particle species labels. We order the
momentum quantum numbers in a monotonically decreasing sequence: In I1 , which is
just a matter of convention. The corresponding energy levels are determined by the BetheYang
equations

Qk (1 , . . . , n ) = mik L sinh k +
(2.9)
ik il (k l ) = 2Ik , k = 1, . . . , n,
l=k
which must be solved with respect to the particle rapidities k , where

ij ( ) = i log Sij ( )
are the two-particle scattering phase-shifts and the energy (with respect to the finite volume
vacuum state) can be computed as
n
mik cosh k .
k=1
The density of n-particle states can be calculated as

i1 ...in (1 , . . . , n ) = det J (n) ,
(n)
Jkl =
Qk (1 , . . . , n )
,
l
k, l = 1, . . . , n.
(2.10)
We are interested in matrix elements of local operators between finite volume multi-particle
states:

j1 ...jm {I1 , . . . , Im } O(0, 0) {I1 , . . . , In } i ...i ,L
1
which can be obtained numerically using truncated conformal space (for details see [16, Section 3.3]). On the other hand, using our previous results (Eq. (2.16) of [16]), the finite volume
behaviour of local matrix elements can also be given as

j1 ...jm {I1 , . . . , Im } O(0, 0) {I1 , . . . , In } i ...i ,L
1
O (
Fm+n
m
+ i, . . . , 1 + i, 1 , . . . , n )jm ...j1 i1 ...in
i1 ...in (1 , . . . , n )j1 ...jm (1 , . . . , m )

+ O e L
(2.11)
and k (k ) are the solutions of the BetheYang equations (2.9) corresponding to the state with
the specified quantum numbers I1 , . . . , In (I1 , . . . , In ) at the given volume L. The above relation is valid provided there are no disconnected terms, i.e., the left and the right states do
214
not contain particles with the same species and rapidity: the sets {(i1 , 1 ), . . . , (in , n )} and
{(j1 , 1 ), . . . , (jm , m )} are disjoint.
We recall from [16] that Eqs. (2.9), (2.11) are exact to all orders of powers in 1/L; we refer to
the corrections non-analytic in 1/L (eventually, as indicated, decaying exponentially) as residual
finite size effects, following the terminology introduced in [15].
2.3. Disconnected contributions
Let us consider a matrix element of the form

j1 ...jm {I1 , . . . , Im } O(0, 0) {I1 , . . . , In } i ...i
1
n ,L
Disconnected terms appear when there is at least one particle in the state on the left which occurs
in the state on the right with exactly the same rapidity. The rapidities of particles as a function of
the volume are determined by the BetheYang equations (2.9)

ik il (k l ) = 2Ik , k = 1, . . . , n,
Qk (1 , . . . , n ) = mik L sinh k +
l=k
and
Qk (1 , . . . , m ) = mjk L sinh k +
jk jl (k l ) = 2Ik ,
k = 1, . . . , m.
l=k
Due to the presence of the interaction terms containing the phase shift functions , equality of
two quantum numbers Ik and Il does not mean that the two rapidities themselves are equal in
finite volume L. It is easy to see that in the presence of nontrivial scattering there are only two
cases when exact equality of the rapidities can occur:
(1) The two states are identical, i.e., n = m and
{j1 . . . jm } = {i1 . . . in },
{I1 , . . . , Im } = {I1 , . . . , In }.
In Section 4 we show that the corresponding diagonal matrix element can be written as a
sum over all bipartite divisions of the set of the n particles involved (including the trivial
ones when A is the empty set or the complete set {1, . . . , n})

i1 ...in {I1 . . . In } O {I1 . . . In } i1 ...in ,L

1
=
F(A)L {1, . . . , n} \ A L + O eL ,
({1, . . . , n})L
A{1,2,...,n}
where |A| denotes the cardinal number (number of elements) of the set A

{k1 , . . . , kr } L = ik1 ...ikr (k1 , . . . , kr )
is the r-particle BetheYang Jacobi determinant (2.10) involving only the r-element subset
1 k1 < < kr n of the n particles, and

s
F {k1 , . . . , kr } L = F2r
(k1 , . . . , kr )ik1 ...ikr ,
F2ls (1 , . . . , l )i1 ...il = lim F2lO (l + i + , . . . , 1 + i + , 1 , . . . , l )i1 ...il il ...i1
0
is the so-called symmetric evaluation of diagonal multi-particle matrix elements.
215
(2) Both states are parity symmetric states in the spin zero sector, i.e.,
{I1 , . . . , In } {In , . . . , I1 },
{I1 , . . . , Im } {Im , . . . , I1 }
and the particle species labels are also compatible with the symmetry, i.e., in+1r = ir and
jm+1r = jr . Furthermore, both states must contain one (or possibly more, in a theory with
more than one species) particle of quantum number 0, whose rapidity is then exactly 0 for any
value of the volume L due to the symmetric assignment of quantum numbers. In Section 5
we state the following conjecture
f2k+1,2l+1

= {I1 , . . . , Ik , 0, Ik , . . . , I1 } {I1 , . . . , Il , 0, Il , . . . , I1 } L

1
=
2k+1 (1 , . . . , k , 0, k , . . . , 1 )2l+1 (1 , . . . , l , 0, l , . . . , 1 )

Fk,l (1 , . . . , k |1 , . . . , l ) + mLF2k+2l (i + 1 , . . . , i + k ,

i k , . . . , i 1 , 1 , . . . , l , l , . . . , 1 ) + O eL ,
where n is a shorthand notation for the n-particle BetheYang density (2.10) and equality
is understood up to phase conventions (cf. Section 5) and
Fk,l (1 , . . . , k |1 , . . . , l )
O
(i + 1 + , . . . , i + k + , i k + , . . . , i 1 + ,
= lim F2k+2l+2
0
i + , 0, 1 , . . . , l , l , . . . , 1 )
is defined by assigning the same shift to all rapidities entering the left (or equivalently the
right) state and taking the limit 0. For the sake of simplicity we assumed above that
there is a single particle species with mass m, but the prescription can be easily extended to
theories with more than one particle species; an example is shown in Section 7.2.
3. Exact form factors
The Hamiltonian of scaling LeeYang model takes the following form in the perturbed conformal field theory framework:
L
H
SLY
= H0LY
+ i
dx (0, x),
0
where
H0LY

2
c
=
L0 + L 0
L
12
is the conformal Hamiltonian and is the only nontrivial primary field, which has conformal
weights = = 1/5. When > 0 the theory above has a single particle in its spectrum with
216
mass m that can be related to the coupling constant as [28]

= 0.09704845636 m12/5
and the bulk energy density is given by
3 2
B=
m .
12
The S-matrix reads [29]
SLY ( ) =
(3.1)
sinh + i sin 2
3
(3.2)
sinh i sin 2
3
and the particle occurs as a bound state of itself at = 2i/3 with the three-particle coupling
given by
2 = 2 3,
where the negative sign is due to the nonunitarity of the model. In this model we define the
phase-shift via the relation
SLY ( ) = ei()
so that (0) = 0. This means a redefinition of Bethe quantum numbers Ik in the BetheYang
equations (2.10) such they become half-integers for states composed of an even number of particles; it also means that in the large volume limit, particle momenta become
2Ik
.
L
Form factors of the trace of the stress-energy tensor were computed by Al.B. Zamolodchikov
in [11], and using the relation
m sinh k =
= i(1 )
we can rewrite them in terms of . They have the form
Fn (1 , . . . , n ) = Hn Qn (x1 , . . . , xn )
n
n

f (i j )
x i + xj
i=1 j =i+1
with the notations

cosh 1
v(i )v(i + ),
cosh + 1/2

t
t
sinh t
2 sinh 3 sinh 6 it
v( ) = exp 2 dt
,
e
t sinh2 t
0

n
31/4
Hn =
xi = ei ,
21/2 v(0)
f ( ) =
and the exact vacuum expectation value of the field is

= 1.239394325 im2/5 .
(3.3)
217
The functions Qn are symmetric polynomials in the variables xi . Defining the elementary symmetric polynomials of n variables by the relations
n
n

(n)
(x + xi ) =
x ni i (x1 , . . . , xn ),
i=1
(n)
= 0 for i > n,
i=0
they can be constructed as

(2)
Q1 = 1,
(3) (3)
Q2 = 1 ,
(n) (n)
Qn = 1 n1 Pn ,
Pn = det M(n)
Q3 = 1 2 ,
n > 3,
(n)
(n)
where Mij = 3i2j +1 , i, j = 1, . . . , n 3.
3.2. Ising model with magnetic perturbation

The critical Ising model is the described by the conformal field theory with c = 1/2 and has
two nontrivial primary fields: the spin operator with = = 1/16 and the energy density
with = = 1/2. The magnetic perturbation, defined using the Hamiltonian (where H0I
denotes the Hamiltonian of the c = 1/2 conformal field theory)
L
H
= H0I
+h
dx (0, x)
0
is massive (and its physics does not depend on the sign of the external magnetic field h). The
spectrum and the exact S matrix is described by the famous E8 factorized scattering theory [30],
which contains eight particles Ai , i = 1, . . . , 8, with known mass ratios, and the mass gap relation
is [31]
m1 = (4.40490857 . . .)|h|8/15
or
15/8
h = h m1
h = 0.06203236 . . . .
(3.4)
The bulk energy density is given by

B = 0.06172858982 m2 .
(3.5)
We also quote the scattering phase shift of two A1 particles for = 0, which has the form

1
sinh + i sin x
1
2
S11 ( ) =
(3.6)
,
{x} =
.
15 3 5
sinh i sin x
All the other amplitudes Sab are determined by the S matrix bootstrap [30]; we only quote the
A1 A2 scattering amplitude

1
4
2
7
S12 ( ) =
5 15 5 15
because it enters some matrix elements examined later. In this model we define the phase-shifts
by the relations (for detailed explanation cf. [16])
S11 ( ) = ei11 ()
and
S12 ( ) = ei12 ()
218
so that again 11 (0) = 12 (0) = 0. The form factors of the operator in the E8 model were
first calculated in [32] and their determination was carried further in [33]. The exact vacuum
expectation value of the field is given by [34]
= h |h|8/15 ,
h = 2.00314 . . . ,
or in terms of the mass scale m = m1

= 0.45475 m.
For practical evaluation of form factors we used the results computed by Delfino, Grinza and
Mussardo, which can be downloaded from the Web in Mathematica format [35]. They use
the following normalized operator:

=

and so all data we plot in the sequel are understood with the same normalization.
4. Diagonal matrix elements
4.1. Form factor perturbation theory and disconnected contributions
In the framework of conformal perturbation theory, we consider a model with the action

A(, ) = ACFT dt dx (t, x) dt dx (t, x)
(4.1)
such that in the absence of the coupling , the model defined by the action A(, = 0) is integrable. The two perturbing fields are taken as scaling fields of the ultraviolet limiting conformal
field theory, with left/right conformal weights h = h < 1 and h = h < 1, i.e., they are
relevant and have zero conformal spin, resulting in a Lorentz-invariant field theory.
The integrable limit A(, = 0) is supposed to define a massive spectrum, with the scale
set by the dimensionful coupling . The exact spectrum in this case consists of some massive
particles, forming a factorized scattering theory with known S matrix amplitudes, and characterized by a mass scale M (which we take as the mass of the fundamental particle generating the
bootstrap), which is related to the coupling via the mass gap relation
= M 22h ,
where is a (non-perturbative) dimensionless constant.
Switching on a second independent coupling in general spoils integrability, deforms the
mass spectrum and the S matrix, and in particular allows decay of the particles which are stable
at the integrable point. One way to approach the dynamics of the model is the form factor perturbation theory proposed in [36]. Let us denote the form factors of the operator in the = 0
theory by
Fn (1 , . . . , n )i1 ...in = 0| (0, 0)|1 . . . n =0
i1 ...in .
Using perturbation theory to first order in , the following quantities can be calculated [36]:
(1) The vacuum energy density is shifted by an amount
Evac = 0| |0=0 .
(4.2)
219
2 gets a correction
(2) The mass (squared) matrix Mab
2
Mab
= 2F2 (i, 0)a b ma ,mb
(4.3)
(where the bar denotes the antiparticle) supposing that the original mass matrix was diagonal
2 = m2 .
and of the form Mab
a ab
(3) The scattering amplitude for the four particle process a + b c + d is modified by
cd
(, ) = i
Sab
F4 (i, + i, 0, )cdab
,
ma mb sinh
= a b .
(4.4)
It is important to stress that the form factor amplitude in the above expression must be defined
as the so-called symmetric evaluation
lim F4 (i + , + i + , 0, )cdab
(see Eq. (4.9) below). It is also necessary to keep in mind that Eq. (4.4) gives the variation
of the scattering phase when the center-of-mass energy (or, the Mandelstam variable s) is
kept fixed [36]. Therefore, in terms of rapidity variables, this variation corresponds to the
following:
cd (, = 0)
cd (, )
Sab
Sab
cd

+
Sab (, ) =

=0
where
=
ma ma + ma ma + (mb ma + ma mb ) cosh
ma mb sinh
is the shift of the rapidity variable induced by the mass corrections given by Eq. (4.3).
It is also possible to calculate the (partial) decay width of particles [33], but we do not need it
here.
We can use the above results to calculate diagonal matrix elements involving one particle. For
simplicity we present the derivation for a theory with a single particle species. Let us start with
the one-particle case. The variation of the energy of a stationary one-particle state with respect
to the vacuum (i.e., the finite volume particle mass) can be expressed as the difference between
the first order perturbative results for the one-particle and vacuum states in volume L:

m(L) = L {0} {0} L 0| |0L .

(4.5)
On the other hand, using Lschers results [37] it only differs from the infinite volume mass in
terms exponentially falling with L. Using Eq. (4.3)

F (i, 0) + O eL .
m
Similarly, the vacuum expectation value receives only corrections falling off exponentially
with L. Therefore we obtain

1
F (i, 0) + mL0| |0 +
{0} {0} L =
mL
with the ellipsis denoting residual finite size corrections. Note that the factor mL is just the
one-particle BetheYang Jacobian 1 ( ) = mL cosh evaluated for a stationary particle = 0.
m(L) =
220
Fig. 4.1. Diagonal 1-particle matrix elements in the scaling LeeYang model. The discrete points correspond to the TCSA
data, while the continuous line corresponds to the prediction from exact form factors.
We can extend the above result to moving particles in the following way. Up to residual finite
size corrections, the one-particle energy is given by

E(L) = m2 + p 2
with
2s
,
L
where s is the Lorentz spin (which is identical to the particle momentum quantum number).
Therefore
p=
EE = mm
whereas perturbation theory gives:

E = L {s} {s} L 0| |0L
and so we obtain

{s} {s} L =
1
1 ( )

F (i, 0) + 1 ( )0| |0 + ,
(4.6)
where

2s
sinh =
1 ( ) = m2 L2 + 4 2 s 2 .
mL
Fig. 4.1 shows the comparison of Eq. (4.6) to numerical data obtained from LeeYang TCSA:
the matching is spectacular, especially in the so-called scaling region (the volume range where
residual finite size corrections are of the order of truncation errors, cf. [16]) where the relative
deviation is less than 104 . Here and in all following plots we use the dimensionless volume
221
Fig. 4.2. Diagonal 1-particle matrix elements in the Ising model. The discrete points correspond to the TCSA data, while
the continuous line corresponds to the prediction from exact form factors.
parameter l = mL, and the matrix elements are also measured in units of m (cf. [16] for details). Diagonal one-particle matrix elements for the Ising model are shown in Fig. 4.2, where
we similarly use natural units given by the mass m = m1 of the lightest particle A1 , just as in all
subsequent plots related to the Ising model.
One can use a similar argument to evaluate diagonal two-particle matrix elements in finite
volume. Let us assume that the theory considered has diagonal scattering as in Section 2.1. The
two-particle BetheYang equations remain valid even in a non-integrable theory as long as the
total energy of the two-particle state remains under the inelastic threshold [38], and therefore the
energy levels can be calculated from
mi1 L sinh 1 + (1 2 ) = 2I1 ,
mi2 L sinh 2 + (2 1 ) = 2I2
222
and (up to residual finite size corrections)

E2 (L) = E2pt (L) E0 (L) = mi1 cosh 1 + mi2 cosh 2 ,
where i1 and i2 label the particle species. After a somewhat tedious, but elementary calculation
the variation of this energy difference with respect to can be determined, using (4.3) and (4.4):
E2 (L) =

L
F4 (2 + i, 1 + i, 1 , 2 )i2 i1 i1 i2
i1 i2 (1 , 2 )

+ mi1 L cosh 1 F2 (i, 0)i2 i2 + mi2 L cosh 2 F (i, 0)i1 i1 ,
where all quantities (such as BetheYang rapidities i , masses mi and the two-particle state
density 2 ) are in terms of the = 0 theory. This result expresses the fact that there are two
sources for the variation of two-particle energy levels: one is the mass shift of the individual
particles, and the second is due to the variation in the interaction. On the other hand, in analogy
with (4.5) we have

E2 (L) = L i i {I1 , I2 } {I1 , I2 } i i ,L 0| |0L

1 2
1 2
and so we obtain the following relation:

i1 i2 {I1 , I2 } {I1 , I2 } i1 i2 ,L

1
F4 (2 + i, 1 + i, 1 , 2 )i2 i1 i1 i2 + mi1 L cosh 1 F2 (i, 0)i2 i2
=
i1 i2 (1 , 2 )

+ mi2 L cosh 2 F2 (i, 0)i1 i1 + 0| |0 + ,
(4.7)
where the ellipsis again indicate residual finite size effects. The above argument is a generalization of the derivation of the mini-Hamiltonian coefficient C in Appendix C of [15]. This formula
is tested against numerical data in the LeeYang model in Fig. 4.3, and the agreement is as precise as it was for the one-particle case. Similar results can be found in the Ising case; they are
shown in Fig. 4.4.
4.2. Generalization to higher number of particles
Let us now introduce some more convenient notations. Given a state

{I1 . . . In }
i ...i
1
we denote

{k1 , . . . , kr } L = ik1 ...ikr (k1 , . . . , kr ),
(4.8)
where l , l = 1, . . . , n, are the solutions of the n-particle BetheYang equations (2.9) at volume L
with quantum numbers I1 , . . . , In and ({k1 , . . . , kr }, L) is the r-particle BetheYang Jacobi
determinant (2.10) involving only the r-element subset 1 k1 < < kr n of the n particles,
evaluated with rapidities k1 , . . . , kr . Let us further denote

s
F {k1 , . . . , kr } L = F2r
(k1 , . . . , kr )ik1 ...ikr ,
where
s
F2n
(1 , . . . , n )i1 ...in = lim F2n
(n + i + , . . . , 1 + i + , 1 , . . . , n )i1 ...in in ...i1
0
(4.9)
223
Fig. 4.4. Diagonal 2-particle matrix elements in the Ising model. The discrete points correspond to the TCSA data, while
the continuous line corresponds to the prediction from exact form factors.
is the so-called symmetric evaluation of diagonal n-particle matrix elements, which we analyze
more closely in the next subsection. Note that the exclusion property mentioned at the end of
Section 2.1 carries over to the symmetric evaluation too: (4.9) vanishes whenever the rapidities
of two particles of the same species coincide.
Based on the above results, we conjecture that the general rule for a diagonal matrix element
takes the form of a sum over all bipartite divisions of the set of the n particles involved (including
the trivial ones when A is the empty set or the complete set {1, . . . , n}):

i1 ...in {I1 . . . In } {I1 . . . In } i1 ...in ,L

1
=
F(A)L {1, . . . , n} \ A L + O eL .
(4.10)
({1, . . . , n})L
A{1,2,...n}
224
This rule can be tested against matrix elements with n = 3 and n = 4 in the LeeYang model,
which are displayed in Figs. 4.5 and 4.6, respectively. The agreement is excellent as before, with
the relative deviation in the scaling region being of the order of 104 .
225
5. Diagonal matrix elements in terms of connected form factors

In this section we discuss diagonal matrix elements in terms of connected form factors, and
prove that a conjecture made by Saleur in [21] exactly coincides with our Eq. (4.10). To simplify
notations we omit the particle species labels; they can be restored easily if needed.
5.1. Relation between connected and symmetric matrix elements
The purpose of this discussion is to give a treatment of the ambiguity inherent in diagonal
matrix elements. Due to the existence of kinematical poles (2.6) the expression
F2n (1 + i, 2 + i, . . . , n + i, n , . . . , 2 , 1 )
which is relevant for diagonal multi-particle matrix elements, is not well defined. Let us consider
the regularized version
F2n (1 + i + 1 , 2 + i + 2 , . . . , n + i + n , n , . . . , 2 , 1 ).
It was first observed in [36] that the singular parts of this expression drop when taking the limits
i 0 simultaneously; however, the end result depends on the direction of the limit, i.e., on the
ratio of the i parameters. The terms that are relevant in the limit can be written in the following
general form:
F2n (1 + i + 1 , 2 + i + 2 , . . . , n + i + n , n , . . . , 2 , 1 )
n
n
n
n

1
ai1 i2 ...in (1 , . . . , n ) i1 i2 . . . in + ,
=
i
i=1
i1 =1 i2 =1
(5.1)
in =1
where ai1 i2 ...in is a completely symmetric tensor of rank n and the ellipsis denote terms that
vanish when taking i 0 simultaneously.
In our previous considerations we used the symmetric limit, which is defined by taking all i
equal:
s
F2n
(1 , 2 , . . . , n ) = lim F2n (1 + i + , 2 + i + , . . . , n + i + , n , . . . , 2 , 1 ).
0
It is symmetric in all the variables 1 , . . . , n . There is another evaluation with this symmetry
property, namely the so-called connected form factor, which is defined as the i independent part
of Eq. (5.1), i.e., the part which does not diverge whenever any of the i is taken to zero:
c
F2n
(1 , 2 , . . . , n ) = n!a12...n
(5.2)
where the appearance of the factor n! is simply due to the permutations of the i .
5.1.1. The relation for n 3
We now spell out the relation between the symmetric and connected evaluations for n = 1, 2
and 3.
The n = 1 case is simple, since the two-particle form factor F2 (1 , 2 ) has no singularities at
1 = 2 + i and therefore
F2s ( ) = F2c ( ) = F2 (i, 0).
It is independent of the rapidities and will be denoted F2c in the sequel.
(5.3)
226
For n = 2 we need to consider

F4 (1 + i + 1 , 2 + i + 2 , 2 , 1 )
a11 12 + 2a12 1 2 + a22 22

1 2
(5.4)
which gives
F4s (1 , 2 ) = a11 + 2a12 + a22 ,
F4c (1 , 2 ) = 2a12 .
The terms a11 and a22 can be expressed using the two-particle form factor. Taking an infinitesimal, but fixed 2 = 0
Res F4 (1 + i + 1 , 2 + i + 2 , 2 , 1 ) = a22 2
1 =0
whereas according to (2.7)

Res F4 (1 + i + 1 , 2 + i + 2 , 2 , 1 )

= i 1 S(1 2 )S(1 2 i 2 ) F2 (2 + i + 2 , 2 ).
1 =0
To first order in 2

S(1 2 i 2 ) = S(2 1 + 2 ) = S(2 1 ) 1 + i(2 1 ) 2 + ,
where
d
log S( )
d
is the derivative of the two-particle phase shift defined before. Therefore we obtain
( ) = i
a22 = (2 1 )F2c
and similarly
a11 = (1 2 )F2c
and so
F4s (1 , 2 ) = F4c (1 , 2 ) + 2(1 2 )F2 (i, 0).
(5.5)
In the case of the trace of the energymomentum tensor the following expressions are
known [24]
F2 = 2m2 ,
F4,s
= 8m (1 2 ) cosh
2

2

1 2
,
2
F4,c = 4m2 (1 2 ) cosh(1 2 )

and they are in agreement with (5.5).
For n = 3, a procedure similar to the above gives the following relation:

F6s (1 , 2 , 3 ) = F6c (1 , 2 , 3 ) + F4c (1 , 2 ) (1 3 ) + (2 3 ) + permutations

+ 3F2c (1 2 )(1 3 ) + permutations ,
(5.6)
where we omitted terms that only differ by permutation of the particles.
227
5.1.2. Relation between the connected and symmetric evaluation in the general case
Our goal is to compute the general expression
F2n (1 , . . . , n | 1 , . . . , n )
= F2n (1 + i + 1 , 2 + i + 2 , . . . , n + i + n , n , . . . , 2 , 1 ).
(5.7)
Let us take n vertices labeled by the numbers 1, 2, . . . , n and let G be the set of the directed
graphs Gi with the following properties:
Gi is tree-like.
For each vertex there is at most one outgoing edge.
For an edge going from i to j we use the notation Eij .
Theorem 1. Eq. (5.7) can be evaluated as a sum over all graphs in G, where the contribution of
a graph Gi is given by the following two rules:
Let Ai = {a1 , a2 , . . . , am } be the set of vertices from which there are no outgoing edges in
Gi . The form factor associated to Gi is
c
F2m
(a1 , a2 , . . . , am ).
(5.8)
For each edge Ej k the form factor above has to be multiplied by

j
(j k ).
k
Note that since cannot contain cycles, the product of the i / j factors will never be trivial
(except for the empty graph with no edges).
Proof. The proof goes by induction in n. For n = 1 we have
F2s (1 ) = F2c (1 ) = F2 (i, 0).
This is in accordance with the theorem, because for n = 1 there is only the trivial graph which
contains no edges and a single node.
Now assume that the theorem is true for n 1 and let us take the case of n particles. Consider
the residue of the matrix element (5.7) at n = 0 while keeping all the i finite
R = Res F2n (1 . . . n | 1 . . . n ).
n =0
According to the theorem the graphs contributing to this residue are exactly those for which the
vertex n has an outgoing edge and no incoming edges. Let Rj be sum of the diagrams where the
outgoing edge is Enj for some j = 1, . . . , n 1, and so
R=
n1
Rj .
j =1
The form factors appearing in Rj do not depend on n . Therefore we get exactly the diagrams
that are needed to evaluate F2(n1) (1 . . . n1 | 1 . . . n1 ), apart from the proportionality factor
228
associated to the link Enj and so

j
Rj = (j n )F2(n1) (1 . . . n1 | 1 . . . n1 )
n
and summing over j gives

R = 1 (1 n ) + 2 (2 n ) + + n1 (n1 n )
F2(n1) (1 n1 | 1 n1 ).
(5.9)
In order to prove the theorem, we only need to show that the residue indeed takes this form. On
the other hand, the kinematical residue axiom (2.6) gives

n1

S(n j )S(n j i j ) F2(n1) (1 . . . n1 | 1 . . . n1 )
R=i 1
j =1
which is exactly the same as Eq. (5.9) when expanded to first order in j .
We thus checked that the theorem gives the correct result for the terms that include a 1/ n
singularity. Using symmetry in the rapidity variables this is true for all the terms that include
at least one 1/ i for an arbitrary i. There is only one diagram that cannot be generated by the
inductive procedure, namely the empty graph. However, there are no singularities (1/ i factors)
c ( , . . . , ) by definition. 2
associated to it, and it gives F2n
1
n
We now illustrate how the theorem works. For n = 2, there are only three graphs, depicted in
Fig. 5.1. Applying the rules yields

1 2
c
F4 (1 , 2 | 1 , 2 ) = F4 (1 , 2 ) + (1 2 )
+
Fc
2 1 2
which gives back (5.5) upon putting 1 = 2 . For n = 3 there are 4 different kinds of graphs, the
representatives of which are shown in Fig. 5.2; all other graphs can be obtained by permuting the
node labels 1, 2, 3. The contributions of these graphs are
(a):
(b):
F6c (1 , 2 , 3 ),
2
(1 2 )F4c (2 , 3 ),
1
(a)
(b)
(c)
Fig. 5.1. The graphs relevant for n = 2.
(a)
(b)
(c)
(d)
Fig. 5.2. The graphs relevant for n = 3.
(c):
(d):
229
2 3
3
(1 2 )(2 3 )F2c = (1 2 )(2 3 )F2c ,
1 2
1
2 2
c
(1 2 )(3 2 )F2 .
1 3
Adding up all the contributions and putting 1 = 2 = 3 we recover Eq. (5.6).

5.2. Consistency with Saleurs proposal
Saleur proposed an expression for diagonal matrix elements in terms of connected form factors in [21], which is partially based on earlier work by Balog [39] and also on the determinant
formula for normalization of states in the framework of algebraic Bethe Ansatz, derived by
Gaudin, and also by Korepin (see [40] and references therein). To describe it, we must extend the
normalization of finite volume states defined in [16] to the case when the particle rapidities form
a proper subset of some multi-particle BetheYang solution.
According to [16], the normalization of a finite volume state is given by

{I1 , . . . , In } =
L
1
n (1 , . . . , n )
|1 , . . . , n
in terms of the infinite volume state with rapidities 1 , . . . , n , which are the solutions of the
BetheYang equations (2.9) for the given quantum numbers I1 , . . . , In at volume L (we again
omit the particle species labels, and also denote the n-particle determinant by n ). Let us take a
subset of particle indices A {1, . . . , n} and define the corresponding sub-determinant by
(n)
n (1 , . . . , n |A) = det JA ,
where JA(n) is the sub-matrix of the matrix J (n) defined in Eq. (2.10) which is given by choosing
the elements whose indices belong to A. The full matrix can be written explicitly as
J (n)
E L + + +
1
12
1n
12
=
..
.
1n
12
E2 L + 21 + 23 + + 2n
..
.
2n
...
...
..
.
...
1n
2n
,
..
.
En L + 1n + + n1,n
where the following abbreviations were used: Ei = mi cosh i , ij = j i = (i j ). Note that

n depends on all the rapidities, not just those which correspond to elements of A. It is obvious
that
n (1 , . . . , n ) n (1 , . . . , n |{1, . . . , n}).
Saleur proposed the definition

{k }kA {k }kA L = n (1 , . . . , n |A),
(5.10)
where

{k }kA
L
is a partial state which contains only the particles with index in A, but with rapidities that
solve the BetheYang equations for the full n-particle state. Note that this is not a proper state in
230
the sense that it is not an eigenstate of the Hamiltonian since the particle rapidities do not solve
the BetheYang equations relevant for a state consisting of |A| particles (where |A| denotes the
cardinal numberi.e., number of elementsof the set A). The idea behind this proposal is that
the density of these partial states in rapidity space depends on the presence of the other particles
which are not included, and indeed it is easy to see that it is given by n (1 , . . . , n |A).
In terms of the above definitions, Saleurs conjecture for the diagonal matrix element is

i1 ...in
{I1 . . . In } {I1 . . . In } i
1 ...in ,L

1
c
{k }kA (
1 , . . . , n |A) + O eL
F2|A|
n (1 , . . . , n ) A{1,2,...,n}
(5.11)
which is just the standard representation of the full matrix element as the sum of all the connected
contributions provided we accept Eq. (5.10). The full amplitude is obtained by summing over all
possible bipartite divisions of the particles, where the division is into particles that are connected
to the local operator, giving the connected form factor F c and into those that simply go directly
from the initial to the final state which contribute the norm of the corresponding partial multiparticle state.
Using the results of Section 5.1, it is easy to check explicitly (which we did up to n = 3) that
our rule for the diagonal matrix elements as given in Eq. (4.10) is equivalent to Eq. (5.11). We
now give a complete proof for the general case.
Theorem 2.

c
s
{k }kA (
1 , . . . , n |A) =
{k }kA {k }kN\A ,
F2|A|
F2|A|
AN
(5.12)
AN
where we denoted N = {1, 2, . . . , n}.

Proof. The two sides of Eq. (5.12) differ in two ways:
The form factors on the right-hand side are evaluated according to the symmetric prescription, and in addition to the connected part also they contain extra terms, which are
proportional to connected form factors with fewer particles.
The densities on the left-hand side are not determinants of the form (2.10) written down in
terms of the particles contained in N \ A: they contain additional terms due to the presence
of the particles in A as well.
Here we show that Eq. (5.12) is merely a reorganization of these terms.
For simplicity consider first the term on the left-hand side which corresponds to A = {m + 1,
m + 2, . . . , n}, i.e.,
c
(m+1 , . . . , n )(
1 , . . . , n |A).
F2m
We expand in terms of the physical multi-particle densities . In order to accomplish this, it is

useful to rewrite the sub-matrix JNn \A as
J (n) |N\A = J m (1 , . . . , m )
n
i=m+1 1i
231
n
i=m+1 2i
..
n
i=m+1 mi
Jm
where
is the m-particle Jacobian matrix which does not contain any terms depending on the
particles in A. The determinant of JNn \A can be written as a sum over the subsets of N \ A. For
a general subset B N \ A let us use the notation B = {b1 , b2 , . . . , b|B| }. We can then write

n

|B|

(n)
=
b ,c
N \ (A B)
,
(
1 , . . . , n |A) = det J
(5.13)
i
N\A
i=1
ci =m+1
where (N \ (A B)) is the -density (2.10) written down with the particles in N \ (A B).
Applying a suitable permutation of variables we can generalize Eq. (5.13) to an arbitrary
subset A N :
|B|

(
1 , . . . , n |A) = det J (n) |N\A =
(5.14)
N \ (A B)
bi ,ci ,
B
i=1
where the second summation goes over all the sets C = {c1 , c2 , . . . , c|B| } with |C| = |B| and
ci A. The left-hand side of Eq. (5.12) can thus be written as

c
(5.15)
{k }kA (
1 , . . . , n |A) =
F2|A|
N \ (A B)
F(A,B,C) ,
AN
A,BN
AB=
where
|B|

c
{k }kA
bi ,ci .
F(A,B,C) = F2|A|
i=1
We now show that there is a one-to-one correspondence between all the terms in (5.15) and
s are expanded according
those on the right-hand side of (5.12) if the symmetric evaluations F2k
to Theorem 1. To each triplet (A, B, C) let us assign the graph G(A,B,C) defined as follows:
The vertices of the graph are the elements of the set A B.
There are exactly |B| edges in the graph, which start at bi and end at ci with i = 1, . . . , |B|.
s
({k }kAB ) is nothing else than F(A,B,C) which
The contribution of G(A,B,C) to F2(|A|+|B|)
can be proved by applying the rules of Theorem 1. Note that all the possible diagrams with
at most n vertices are contained in the above list of the G(A,B,C) , because a general graph G
satisfying the conditions in Theorem 1 can be characterized by writing down the set of vertices
with and without outgoing edges (in this case B and A) and the endpoints of the edges (in this
case C).
It is easy to see that the factors (N \ (A B)) multiplying the F(A,B,C) in (5.15) are also
s
the correct ones: they are just the density factors multiplying F2(|A|+|B|)
({k }kAB ) on the righthand side of (5.12). 2
232
6. Zero-momentum particles
In the scaling LeeYang model, with a single type of particle, there can only be a single particle of zero momentum in a multi-particle state due to the exclusion principle. For the momentum
to be exactly zero in finite volume it is necessary that the all other particles should come with
quantum numbers in pairs of opposite sign, which means that the state must have 2n + 1 particles
in a configuration

{I1 , . . . , In , 0, In , . . . , I1 } .
L
Therefore we consider matrix elements of the form

{I1 , . . . , Ik , 0, Ik , . . . , I1 } {I1 , . . . , Il , 0, Il , . . . , I1 } L

(with k = 0 or l = 0 corresponding to a state containing a single stationary particle). We also
suppose that the two sets {I1 , . . . , Ik } and {I1 , . . . , Il } are not identical, otherwise we have the
case of diagonal matrix elements treated in Section 4.
We need to examine form factors of the form
F2k+2l+2 (i + 1 , . . . , i + k , i k , . . . , i 1 , i + , 0, 1 , . . . , l , l , . . . , 1 ),
where the particular ordering of the rapidities was chosen to ensure that no additional S matrix
factors appear in the disconnected terms of the crossing relation (2.2). Using the singularity
axiom (2.6), plus unitarity and crossing symmetry of the S-matrix it is easy to see that the residue
of the above function at = 0 vanishes, and so it has a finite limit as 0. However, this limit
depends on direction just as in the case of the diagonal matrix elements considered in Section 4.
Therefore we must specify the way it is taken, and just as previously we use a prescription that
is maximally symmetric in all variables: we choose to shift all rapidities entering the left-hand
state with the same amount to define
Fk,l (1 , . . . , k |1 , . . . , l )
= lim F2k+2l+2 (i + 1 + , . . . , i + k + , i k + , . . . , i 1 + ,
0
i + , 0, 1 , . . . , l , l , . . . , 1 ).
(6.1)
Using the above definition, by analogy to (4.10) we conjecture that

f2k+1,2l+1 = {I1 , . . . , Ik , 0, Ik , . . . , I1 } {I1 , . . . , Il , 0, Il , . . . , I1 } L

1
=
2k+1 (1 , . . . , k , 0, k , . . . , 1 )2l+1 (1 , . . . , l , 0, l , . . . , 1 )

Fk,l (1 , . . . , k |1 , . . . , l ) + mLF2k+2l (i + 1 , . . . , i + k ,

i k , . . . , i 1 , 1 , . . . , l , l , . . . , 1 ) + O eL ,
(6.2)
where denote the solutions of the appropriate BetheYang equations at volume L, n is a
shorthand notation for the n-particle BetheYang density (2.10) and equality is understood up to
phase factors. We recall from our previous work [16] that relative phases of multi-particle states
are in general fixed differently in the form factor bootstrap and TCSA. Also note that reordering
233
particles gives phase factors on the right-hand side according to the exchange axiom (2.4). This
issue is obviously absent in the case of diagonal matrix elements treated in Sections 4 and 5,
since any such phase factor cancels out between the state and its conjugate. Such phases do not
affect correlation functions, or as a consequence, any physically relevant quantities since they
can all be expressed in terms of correlators.
There is some argument that can be given in support of Eq. (6.2). Note that the zeromomentum particle occurs in both the left and right states, which actually makes it unclear how
to define a density similar to in (5.10). Such a density would take into account the interaction
with the other particles. However, the nonzero rapidities entering of the two states are different
and therefore there is no straightforward way to apply Saleurs recipe (5.11) here. Using the
maximally symmetric definition (6.1) the shift can be equally put on the right-hand side rapidities as well, and therefore we expect that the density factor multiplying the term F2k+2l in
(6.2) would be the one-particle state density in which none of the other rapidities appear, which
is exactly mL for a stationary particle. This is a natural guess from Eq. (4.10) which states that
when diagonal matrix elements are expressed using the symmetric evaluation, only densities of
the type appear.
Another argument can be formulated using the observation that Eq. (6.2) is only valid if Fk,l
is defined as in (6.1); all other possible ways to take the limit can be related in a simple way
to this definition and so the rule (6.2) can be rewritten appropriately. Let us consider two other
natural choices
+
(1 , . . . , k |1 , . . . , l ) = lim F2k+2l+2 (i + 1 , . . . , i + k , i k , . . . ,
Fk,l
0
i 1 , i, , 1 , . . . , l , l , . . . , 1 ),
(1 , . . . , k |1 , . . . , l ) = lim F2k+2l+2 (i + 1 , . . . , i + k , i k , . . . ,

Fk,l
0
i 1 , i + , 0, 1 , . . . , l , l , . . . , 1 )
in which the shift is put only on the zero-momentum particle on the right/left, respectively. Using
the kinematical residue axiom (2.6), F can be related to F via
+
Fk,l (1 , . . . , k |1 , . . . , l ) = Fk,l
(1 , . . . , k |1 , . . . , l
+2
l
(i )F2k+2l (i + 1 , . . . , i + k , i k , . . . ,
i=1
i 1 , 1 , . . . , l , l , . . . , 1 ),
(1 , . . . , k |1 , . . . , l )
Fk,l (1 , . . . , k |1 , . . . , l ) = Fk,l
k
(i )F2k+2l (i + 1 , . . . , i + k , i k , . . . ,
i=1
i 1 , 1 , . . . , l , l , . . . , 1 ).
With the help of the above relations Eq. (6.2) can also be rewritten in terms of F . The way F
and therefore also Eq. (6.2) are expressed in terms of F shows a remarkable and natural symmetry under the exchange of the left and right state (and correspondingly F + with F ), which
provides a further support to our conjecture.
The above two arguments cannot be considered as a proof; we do not have a proper derivation
of relation (6.2) at the moment. On the other hand, as we now show it agrees very well with
234
Fig. 6.1. 1-particle3-particle matrix elements in the scaling LeeYang model. The discrete points correspond to the
TCSA data, while the continuous line corresponds to the prediction from exact form factors.
numerical data which would be impossible if there were some additional terms present; such
terms, as shown in our previous work [16] would contribute corrections of order 1/ l in terms of
the dimensionless volume parameter l = mL.
Data for the case of 13 and 33 matrix elements are shown in Figs. 6.1 and 6.2, respectively.
In order to strengthen the support for Eq. (6.2) we must find 5-particle states. This is not easy
because they are high up in the spectrum, and identification using the process of matching against
BetheYang predictions (as described in [16]) becomes ambiguous. We could identify the first
235
Fig. 6.3. Identifying the 5-particle state using form factors. The discrete points correspond to the TCSA data, while the
continuous line corresponds to the prediction from exact form factors.
5-particle state by combining the BetheYang matching with predictions for matrix elements
with no disconnected pieces given by Eq. (2.11), as shown in Fig. 6.3. Some care must be taken
in choosing the other state because many choices give matrix elements that are too small to be
measured reliably in TCSA: since vector components and TCSA matrices are mostly of order 1
or slightly less, getting a result of order 104 or smaller involves a lot of cancellation between
a large number of individual contributions, which inevitably leads to the result being dominated
236
by truncation errors. Despite these difficulties, combining BetheYang level matching with form
factor evaluation we could identify the first five-particle level up to l = 20.
The simplest matrix element involving a five-particle state and zero-momentum disconnected
pieces is the 15 one, but the prediction of Eq. (6.2) turns out to be too small to be usefully
compared to TCSA. However, it is possible to find 35 matrix elements that are sufficiently large,
and the data shown in Fig. 6.4 confirm our conjecture with a relative precision of somewhat better
than 103 in the scaling region.
We close by noting that since the agreement is better than one part in 103 in the scaling region,
which is typically found in the range of volume l 10 . . . 20, and also this precision holds for
quite a large number of independent matrix elements, the presence of additional terms in
Eq. (6.2) can be confidently excluded.
6.2. Ising model in magnetic field
In Fig. 6.5 we show how the prediction (6.2) describes a 13 matrix element in the Ising
model; since all particles in this example are of species A1 , the formula carries over without
essential modifications.
However, due to the fact that the Ising model has more than one particle species, it is possible
to have more than one stationary particles in the same state. Our TCSA data allow us to locate
one such state, with a stationary A1 and A2 particle, and extending our previous considerations
we have the prediction

f1,12 =1 {0} {0, 0} 12 =
m1 L m2 L

lim F3 (i + , 0, 0)112 + m1 LF1 (0)2 ,
where F1 (0)2 is the one-particle form factor corresponding to A2 . This is compared to TCSA
data in Fig. 6.6 and a convincing agreement is found.
237
Fig. 6.5. A1 A1 A1 A1 matrix element in Ising model with a zero-momentum particle.
Fig. 6.6. A1 A1 A2 matrix element in Ising model with zero-momentum particle.
Note that in both of Figs. 6.5 and 6.6 there is a point which obviously deviates from the
prediction. This is a purely technical issue, and is due to the presence of a line crossing close
to this particular value of the volume which makes the cutoff dependence more complicated and
so slightly upsets the extrapolation in the cutoff. We also remark that we cannot check further
matrix elements at the moment, because the appropriate form factor solutions have not yet been
computed.
7. Finite temperature correlators
In this section we show how a systematical low-temperature expansion for correlation functions can be developed using the results presented so far. Finite temperature correlation functions
have attracted quite a lot of interest recently. Leclair and Mussardo proposed an expansion for
the one-point and two-point functions in terms of form factors dressed by appropriate occupation number factors containing the TBA pseudo-energy function [20], based on a quasi-particle
description motivated by the thermodynamic Bethe Ansatz. As discussed in the introduction,
their proposal for the two-point function was shown to be incorrect by Saleur [21]; on the other
hand, he also gave a proof of the LeclairMussardo formula for one-point functions based on
the conjecture formulated in Eq. (5.11), provided the operator considered is the density of some
local conserved charge. Since we proved that our formula (4.10) for diagonal matrix elements is
238
equivalent to Saleurs conjecture, our results in Section 4 can be considered as a very convincing
numerical evidence for the correctness of his argument.
Another proposal for finite-temperature one-point functions was made by Delfino [23], who
attempted to express them in terms of free-particle occupation numbers and the symmetric evaluation of diagonal matrix elements. It was shown by Mussardo that this proposal is not correct
using a counter-example where it disagreed with the LeclairMussardo expansion [24].
Furthermore, Castro-Alvaredo and Fring also argued [25] that two-point functions cannot be
obtained by a simple dressing procedure analogous to the LeclairMussardo expansion for onepoint functions. They argued that one needs a more drastic change in the form factor program.
All these issues are connected to the problem of finding a proper definition of the disconnected
pieces. From the crossing relation (2.2), these are infinite for the form factors defined in infinite
volume, and subtraction of such infinities must be made with care in order to obtain the correct
finite pieces. Because of the above difficulties there is also a development in the direction of
finite temperature form factors (for a review cf. [41]); with further development, this other line
of thought can also give a very useful formulation of finite temperature correlation functions.
Here we use the idea that putting the system into a finite volume L provides a regularization
for the form factors, which can even be considered physical since in the real world there are
no infinite systems.1 Our expressions for the finite volume form factors are valid up to exponential corrections in the volume, which makes it clear that performing the calculation in finite
volume and then taking the limit L we should recover the proper finite temperature correlation function. Here we present the computation for the case of the one-point function up to
the first three nontrivial orders; the calculation gets complicated for higher orders, but the recipe
is straightforward. On general theoretical grounds, it is quite clear that our approach should also
apply to the two-point function, or indeed to any multi-point correlator, but in order to keep the
exposition short we do not go into these details here and leave them to future investigations.
7.1. LeclairMussardo series expanded
The finite temperature expectation value of a local operator O is defined by
OR =
Tr(eRH O)
,
Tr(eRH )
where R = 1/T is the temperature dependent extension of the Euclidean time direction used
in thermal quantum field theory and H is the Hamiltonian. To keep the exposition simple we
assume that the spectrum contains a single massive particle of mass m. Leclair and Mussardo
proposed the following expression for the low temperature (T m, or equivalently mR 1)
expansion of the above one-point function:

n
(i )

1
e
1
di
F c (1 , . . . , n ),
OR =
n! (2)n
1 + e (i ) 2n
n=0
(7.1)
i=1
1 There is actually a little subtlety here, since we impose periodic boundary conditions which are also nonphysical, but
we make use of the old intuition that nothing can actually depend very much on the choice of the boundary condition if
the system is very large and has a finite correlation length (i.e., a mass gap).
239
c is the connected diagonal form factor defined in Eq. (5.2) and ( ) is the pseudowhere F2n
energy function, which is the solution of the thermodynamic Bethe Ansatz equation

d

( ) log 1 + e ( ) .
( ) = mR cosh( )
(7.2)
2
The solution of this equation can be found by successive iteration, which results in

d
1
d

mR cosh
( ) = mR cosh( )
+
( )e
( )e2mR cosh
2
2
2

d d

+
(7.3)
( )( )emR cosh emR cosh + O e3mR .
2 2
Using this expression, it is easy to derive the following expansion from (7.1)

d c mR cosh
R
O = O +
e2mR cosh
F e
2 2

d1 d2 c
1
F4 (1 , 2 ) + 2(1 2 )F2c emR cosh 1 emR cosh 2
+
2
2 2
3mR
,
+O e
(7.4)
where O denotes the zero-temperature vacuum expectation value. The above result can also be
written in terms of the symmetric evaluation (4.9) as

d s mR cosh
R
e2mR cosh
O = O +
F2 e
2

d1 d2 s
1
F4 (1 , 2 )emR(cosh 1 +cosh 2 ) + O e3mR ,
+
(7.5)
2
2 2
where we used relations (5.3) and (5.5).
For completeness we also quote Delfinos proposal:

n

1 1
emR cosh i
R
s
OD =
(7.6)
di
(1 , . . . , n )
F2n
n! (2)n
1 + emR cosh i
n=0
i=1
which gives the following result when expanded to second order:

d s mR cosh
R
OD = O +
e2mR cosh
F2 e
2

d1 d2 s
1
F4 (1 , 2 )emR(cosh 1 +cosh 2 ) + O e3mR .
+
(7.7)
2
2 2
Note that the two formulae coincide with each other to this order, which was already noted
in [23]. However, this is not the case in the next order. Obtaining the third order correction
from the LeclairMussardo expansion is a somewhat lengthy, but elementary computation, which
results in

d1 d2 d3 s
1
F (1 , 2 , 3 )emR(cosh 1 +cosh 2 +cosh 3 )
6
2 2 2 6

d1 d2 s
d1 s 3mR cosh 1
mR(cosh 1 +2 cosh 2 )
+
F (1 , 2 )e
F e
2 2 4
2 2

d1 d2 s
1
(7.8)
F (1 2 )emR(cosh 1 +2 cosh 2 ) ,
2
2 2 2
240
where we used Eqs. (5.3), (5.5), (5.6) to express the result in terms of the symmetric evaluation.
On the other hand, expanding (7.6) results in

1
d1 d2 d3 s
6
2 2 2 6

d1 d2 s
d1 s 3mR cosh 1
.
(7.9)
F4 (1 , 2 )emR(cosh 1 +2 cosh 2 ) +
F e
2 2
2 2
It can be seen that the two proposals differ at this order (the last term of (7.8) is missing
from (7.9)), which was already noted by Mussardo using a toy model in [24], but our computation here is model independent and shows the general form of the discrepancy. We also need
the third order correction explicitly so that we can compare it to the result of the computation
performed in the next section.
7.2. Low-temperature expansion for one-point functions
We now evaluate the finite temperature expectations value in a finite, but large volume L:
OR
L=
TrL (eRHL O)
,
TrL (eRHL )
(7.10)
where HL is the finite volume Hamiltonian, and TrL means that the trace is now taken over the
finite volume Hilbert space. For later convenience we introduce a new notation:

|1 , . . . , n L = {I1 , . . . , In } L ,
where 1 , . . . , n solve the BetheYang equations for n particles with quantum numbers
I1 , . . . , In at the given volume L. We can develop the low temperature expansion of (7.10) in
powers of emR using

TrL eRHL O

(1)
emR cosh (1) O (1) L
= OL +
(1)
1 mR(cosh (2) +cosh (2) ) (2) (2) (2) (2)
1
2
1 , 2 O 1 , 2 L
e
2 (2) (2)
1 ,2

1
+
6
(3)
(3)
(3)
emR(cosh 1
(3)
(3)
+cosh 2 +cosh 3 )
1(3) , 2(3) , 3(3) O 1(3) , 2(3) , 3(3) L
(3)
1 ,2 ,3
+O e

4mR
(7.11)
and

1 mR(cosh( (2) )+cosh( (2) ))
(1)
1
2
TrL eRHL = 1 +
emR cosh( ) +
e
2
(1)
(2) (2)
1
+
6

(3) (3) (3)
1 ,2 ,3
1 ,2
(3)
(3)
(3)
mR(cosh 1 +cosh 2 +cosh 3 )

+ O e4mR .
(7.12)
241
The denominator of (7.10) can then be easily expanded:

1
(1)
emR cosh +
=1
RH
L)
TrL (e
(1)

emR cosh
(1)
2
(1)

3
1 mR(cosh (2) +cosh (2) )
mR cosh (1)
1
2
e
2 (2) (2)
(1)
1 ,2

emR cosh
(1)
1
6

(1)
(2)
(2)
emR(cosh 1
(2)
+cosh 2 )
(2)
1 ,2

(3)
emR(cosh 1
(3)
(3)
+cosh 2 +cosh 3 )

+ O e4mR .
(7.13)
(3) (3) (3)

1 ,2 ,3
The primes in the multi-particle sums serve as a reminder that there exist only states for which
all quantum numbers are distinct. Since we assumed that there is a single particle species, this
means that terms in which any two of the rapidities coincide are excluded. All n-particle terms
in (7.11) and (7.12) have a 1/n! prefactor which takes into account that different ordering of the
same rapidities give the same state; as the expansion contains only diagonal matrix elements,
phases resulting from reordering the particles cancel. The upper indices of the rapidity variables
indicate the number of particles in the original finite volume states; this is going to be handy
when replacing the discrete sums with integrals since it keeps track of which multi-particle state
density is relevant.
We also need an extension of the finite volume matrix elements to rapidities that are not necessarily solutions of the appropriate BetheYang equations. The required analytic continuation
is simply given by Eq. (4.10)
1 , . . . , n |O|1 , . . . , n L

1
=
n (1 , . . . , n )L

L
s
{i }iA n|A| {i }i A
,
F2|A|
/ L+O e
(7.14)
A{1,2,...,n}
where we made explicit the volume dependence of the n-particle density factors. The last term
serves as a reminder that this prescription only defines the form factor to all orders in 1/L
(i.e., up to residual finite size corrections), but this is sufficient to perform the computations
in the sequel.
Using the leading behaviour of the n-particle state density, contributions from the n-particle
sector scale as Ln , and for the series expansions (7.11), (7.12) and (7.13) it is necessary that
mL emR . However if mR is big enough there remains a large interval
1 mL emR ,
where the expansions are expected to be valid. After substituting these expansions into (7.10) we
will find order by order that the leading term of the net result is O(L0 ), and the corrections scale
as negative powers of L. Therefore in (7.10) we can continue analytically to large L and take the
L limit.
242
7.2.1. Corrections of order emR

Substituting the appropriate terms from (7.13) and (7.11) into (7.10) gives the result

(1) (1) (1)
O L OL + O e2mR .
emR cosh
OR
L = OL +
(1)
Taking the L limit one can replace the summation with an integral over the states in the
rapidity space:

d
1 ( )
2
i
and using (4.6) we can write

1 ( ) |O|L OL = F2s + O eL
so we obtain
(7.15)

d s mR cosh
+ O e2mR
F2 e
2
which coincides with Eq. (7.5) to this order.
O = O +
R
7.2.2. Corrections of order e2mR

Substituting again the appropriate terms from (7.13) and (7.11) into (7.10) gives the result

(1) (1) (1)
O L OL
emR cosh
OR
L = OL +

(1)
(1)
mR cosh 1

(1)
1
(1)
mR cosh 2
(1) (1)
2 O 2 L
OL
(1)
2

1 mR(cosh (2) +cosh (2) ) (2) (2) (2) (2)
1
2
1 , 2 O 1 , 2 L OL
e
2 (2) (2)
1 ,2

+ O e3mR .
The O(e2mR ) terms can be rearranged as follows. We add and subtract a term to remove the
constraint from the two-particle sum:

1 mR(cosh (2) +cosh (2) ) (2) (2) (2) (2)
1
2
1 , 2 O 1 , 2 L OL
e
+
2 (2) (2)
1 ,2
1
2
(2)

(2) (2) (2) (2)
1 , 1 O 1 , 1 L OL
(2)
1
2
(2)
e2mR cosh 1
1 =2
(1)
1
(2)
(2)
(1)
emR(cosh 1
(1)
+cosh 2 )

(1) (1)
(1) (1)
1 O 1 L + 2 O 2 L 2OL .
(1)
The 1 = 2 terms correspond to insertion of some spurious two-particle states with equal
Bethe quantum numbers for the two particles (I1 = I2 ). The two-particle BetheYang equations
in this case degenerates to the one-particle case (as discussed before, the matrix elements can be
243
defined for these states without any problems since we have the analytic formula (7.14) valid
to any order in 1/L). This also means that the density relevant to the diagonal two-particle sum
is 1 and so for large L we can substitute the sums with the following integrals

d1,2
d
1 (1,2 ),
1 ( ),
2
2
(1)
(2)
(2)
1,2
(2)
(2)
1 =2
d1 d2
2 (1. , 2 ).
2 2
1 ,2
Let us express the finite volume matrix elements in terms of form factors using (4.6) and (4.7):

(2) (2) (2) (2)
2 (1 , 2 ) 1 , 2 O 1 , 2 L OL

1 (1 )1 (2 ) 1 |O|1 L + 2 |O|2 L 2OL = F4s (1 , 2 ) + O eL .
Combining the above relation with (7.15), we also have
, |O|, L OL =

21 ( ) s
F2 + O eL ,
2 (, )
where we used that F4s (, ) = 0, which is just the exclusion property mention after Eq. (4.9).
Note that

1 ( )2
= 1 + O L1
2 (, )
and therefore in the limit L we obtain

d1 d2 s
d 2mR cosh s 1
F2 +
e
F (1 , 2 )emR(cosh 1 +cosh 2 )
2
2
2 2 4
which is equal to the relevant contributions in the LeclairMussardo expansion (7.5).
7.2.3. Corrections of order e3mR
This calculation is rather long, and so it is relegated to Appendix A. The net result is

1
d1 d2 d3 s
6
2 2 2 6

d1 d2 s
d1 s 3mR cosh 1
mR(cosh 1 +2 cosh 2 )
+
F4 (1 , 2 )e
F e
2 2
2 2

1
d1 d2 s
F (1 2 )emR(cosh 1 +2 cosh 2 )
2
2 2 2
(7.16)
which agrees exactly with Eq. (7.8).

7.3. Remarks
There are a few remarks which we wish to make. First, we see that the proposals by Leclair
and Mussardo and by Delfino differ at the order e3mR . The reason for this difference can be
understood in the formalism developed here. Namely, the expansions (7.11) and (7.13) both
contain positive powers of L. On physical grounds, they are expected to cancel completely order
by order in the emR expansion. However, the state densities depend on the interaction as well.
244
This dependence is of order L1 , and it actually characterizes the ambiguity in the definition
of the diagonal matrix element resulting from the resolution of the singularity (see Eq. (5.1)).
Naively it drops out in the L limit, but actually some of these terms is multiplied by a
positive L power from (7.13). In our derivation we evaluated every relevant contribution to all
orders in 1/L (i.e., we only neglected residual finite size corrections). As a result, we could take
the limit L properly and get the correct finite part of the resulting expression.
Taking this line of thought further, note that the leading term of every multi-particle density
(whether it is degenerate in the sense defined in Appendix A, or not) is always a product of
Ei L factors where i runs over the number of particles and Ei is their energy. Therefore density
terms whose leading behaviour is L0 do not contribute explicit factors. As far as there are
only contributions of this type, the expansion of the one-point function, when written in terms of
F s is just the same as in a free field theory. Indeed in the free field limit the LeclairMussardo
expansion and the Delfino proposal are identical, since the pseudo-energy function is just ( ) =
c F s (more generally, due to the absence of kinematical singularities the
mR cosh and F2n
2n
i 0 limit of (5.1) is independent of the direction).
To have terms that depend explicitly on the interaction we need density contributions that
naively scale as a positive power of L. When combining all such terms at a given order, the
leading term must drop out, and the final result can only have a behaviour L0 at large L. It is
clear from our calculation detail above and in Appendix A that the first order at which such an
anomalous contribution arises is that of e3mR . Up to that order every individual term is finite
as L . However, at third order there appear some anomalous density terms, namely those
collected in (A.7), which individually grow linearly in L. As required by general principles, the
linear contribution cancels between them and so the L limit is well defined. However,
the subleading terms always contain dependence on , and indeed they all vanish for a free
theory (when = 0), therefore it is only such terms that can contribute explicit dependence
in the expansion. As a result, there remains an anomalous term which is just (1 times) the
derivative of the phase shift, and leads to the correction (A.8), which is exactly the term absent
in Delfinos expression.
Strictly speaking, the above discussion is only valid if the expansion is written in terms of
s ; rewriting it in terms of the connected form factors F c obviously
the symmetric evaluation F2n
2n
introduces further dependence. As shown in the above argument, the real difference between
the free and the interacting case can be properly observed when the expansion is written in terms
s , therefore it seems a more natural choice than using the connected form factors, as the
of F2n
behaviour specific to interacting theories can be seen much more clearly.
Another important point is that our results give an independent support for the Leclair
Mussardo expansion. It is known that it coincides precisely with the exact TBA result for the
trace of the energymomentum tensor [20], and Saleur presented an argument for its validity
when the operator considered is the density of a local conserved charge [21]. These arguments
work to all orders, but only for a restricted set of local operators. On the other hand, our calculation above is model independent, and although we only worked it out to order e3mR , we
expect that it coincides with the LeclairMussardo expansion to all orders. For a complete proof
we need a better understanding of its structure, which is out of the scope of the present work.
Furthermore, our method has a straightforward extension to higher point correlation functions.
For example, a two-point correlation function
R TrL (eRHL O1 (x)O2 (0))

O1 (x)O2 (0) L =
TrL (eRHL )
245
can be expanded inserting two complete sets of states

RE (L)
n
e
n|O(x)|mL m|O(0)|nL .
TrL eRHL O1 (x)O2 (0) =
(7.17)
m,n
Since we now have a complete description of finite volume matrix elements to all orders in 1/L,
the above expression can be evaluated along the lines presented in Section 7.2, provided that
the intermediate state sums are properly truncated. We leave the explicit evaluation of expansion (7.17) to further investigations.
Finally note that besides giving a systematic expansion in powers of emR , our method also
gives the L dependence to all orders in 1/L (i.e., up to residual finite size effects), therefore it
can also be used to study finite size corrections of correlators in the low temperature regime.
8. Conclusions
In this work we completed the description of finite volume matrix elements of local operators
by considering those with disconnected pieces. There are two types of such matrix elements,
namely (1) diagonal ones and (2) ones involving parity-invariant zero-spin states with zeromomentum particles. Our description is valid to any order in 1/L, i.e., up to residual finite size
corrections decaying exponentially with the volume L. The precise statements were formulated
in Section 2.3 and we then gave extensive numerical evidence for them. We also formulated and
proved a general theorem relating the different possible evaluations of diagonal matrix elements,
and showed that our results coincide with the proposal made by Saleur [21].
We then showed how to perform an expansion for finite temperature correlation functions,
using the fact that finite volume acts as a regulator for the otherwise infinite disconnected pieces.
The case we considered explicitly was that of one-point functions at finite temperature. We evaluated the first few orders in the low temperature expansion and showed that they coincide with the
result conjectured by Leclair and Mussardo [20], but are different from Delfinos proposal [23]
at third order. Some important aspects of this expansion were already discussed in Section 7.3,
which we do not repeat here.
There is a number of interesting issues remaining. Our approach gives the finite volume form
factors up to residual finite size effects, but combined with truncated conformal space one can
achieve a precision of order 104 in the scaling LeeYang model, and 103 in the Ising model
with magnetic field. It would be interesting to see how these results can be related to other approaches to finite volume form factors (see [42]) and whether the picture can be completed to
give some sort of exact description in the case of integrable field theories. It also seems worthwhile to formulate a higher dimensional generalization of these results extending the approach
of Lellouch and Lscher [43], which is expected to be relevant for lattice field theory.
Another open issue is to give a more concise formulation of the finite temperature expansion
discussed in Section 7 that would make possible a partial resummation to recover the Leclair
Mussardo expression (7.1) which involves dressed form factors.
It is even more interesting to write down the expansion for two-point correlators following the
ideas outlined in Section 7.3; a better method of organizing the contributions could be of great
help here as well. Results for the two-point function can be compared, e.g., to evaluation of correlation functions from truncated conformal space, and can also be used in further development
of the finite temperature form factor program [41].
246
Acknowledgements
We wish to thank Z. Bajnok and L. Palla for useful discussions. This research was partially
supported by the Hungarian research funds OTKA T043582, K60040 and TS044839. G.T. was
also supported by a Bolyai Jnos research scholarship.
Appendix A. e3mR corrections to the finite temperature one-point function
In order to shorten the presentation, we introduce some further convenient notations:
Ei = m cosh i ,
1 , . . . , n |O|1 , . . . , n L = 1 . . . n|O|1 . . . nL ,
n (1 , . . . , n ) = (1 . . . n).
Summations will be shortened to

,
1 ...n

1 ...n
1...n

1...n
Given these notations, we now multiply (7.11) with (7.13) and collect the third order correction
terms:

1 R(E1 +E2 +E3 )
123|O|123L OL
e
6
123

R(E +E )
RE1 1
2
3 23|O|23 O
e
e
L
L
2
1
23

1 R(E1 +E2 )
+
eRE1
eRE2
e
2
1
2
12

eRE3 3|O|3L OL .

3
To keep trace of the state densities, we avoid combining rapidity sums. Now we replace the
constrained summations by free sums with the diagonal contributions subtracted:

=
,
12

12
1=2

=
+
+
,
+2
123
123
1=2,3
2=3,1
1=3,2
1=2=3
where the diagonal contributions are labeled to show which diagonal it sums over, but otherwise
the given sum is free, e.g.,

1=2,3

(3)
(3)
(3)
(3)
(3)
shows a summation over all triplets 1 , 2 , 3 where 1 = 2

also be equal with the other two). We also make use of the notation
247
(3)
and 3
runs free (it can
s
(1 , . . . , n )
F (12 . . . n) = F2n
so the necessary matrix elements can be written in the form

(123) 123|O|123L OL = F (123) + (1)F (23) + + (12)F (3) + ,

(122) 122|O|122L OL = 2(2)F (12) + 2(12)F (3) + (22)F (1),

(111) 111|O|111L OL = 3(111)F (1),

(12) 12|O|12L OL = F (12) + (1)F (2) + (2)F (1),

(11) 11|O|11L OL = 2(1)F (1),

(1) 1|O|1L OL = F (1),
(A.1)
where we used that F and are entirely symmetric in all their arguments, and the ellipsis in
the first line denote two plus two terms of the same form, but with different partitioning of the
rapidities, which can be obtained by cyclic permutation from those displayed. We also used the
exclusion property mentioned after Eq. (4.9).
We can now proceed by collecting terms according to the number of free rapidity variables.
The terms containing threefold summation are
1

1 R(E1 +E2 +E3 )
123|O|123L OL
23|O|23L OL
e
6
2
123
1 2,3

1
3|O|3L OL .
+
2
1
1,2
Replacing the sums with integrals

d1
(1),
2
1

d1 d2
(12),
2 2
1,2

d1 d2 d3
(123)
2 2 2
1,2,3
and using (A.1) we get

1
d1 d2 d3 R(E1 +E2 +E3 )
F (123) + 3(1)F (23) + 3(12)F (3)
e
6
2 2 2

d1 d2 d3 R(E1 +E2 +E3 )
1
(1)F (23) + 2(1)(2)F (3)
e
2
2 2 2

d1 d2 d3 R(E1 +E2 +E3 )
1
+
e
(1)(2)F (3) (12)F (3) ,
2 2 2
2
where we reshuffled some of the integration variables. Note that all terms cancel except the one
containing F (123) and writing it back to its usual form we obtain

d1 d2 d3 s
1
F (1 , 2 , 3 )emR(cosh 1 +cosh 2 +cosh 3 ) .
(A.2)
6
2 2 2 6
248
It is also easy to deal with terms containing a single integral. The only term of this form is

1 R(E1 +E2 +E3 )
123|O|123L OL .
e
3
1=2=3
(3)
(3)
(3)
When all rapidities 1 , 2 , 3

the one-particle case
are equal, the three-particle BetheYang equations reduce to
(3)
mL sinh 1 = 2I1 .
Therefore the relevant state density is that of the one-particle state:

d1 3RE1
1
(11)
d1 3RE1
(1) 111|O|111L OL =
(1)
e
e
F (1)
3
2
2
(111)

d1 3mR cosh 1 s
F2 ,
e
2
(A.3)
where we used that

(1)
(11)
1
(111)
when L .
The calculation of double integral terms is much more involved. We need to consider

1
+
+
eR(E1 +E2 +E3 ) 123|O|123L OL

6
1=2,3
1=3,2
2=3,1

1 R(E1 +E2 +E3 )
+
23|O|23L OL
e
2
1 2=3
1
2

eR(E1 +E2 +E3 ) 3|O|3L OL .
(A.4)
1=2 3
We need the density of partially degenerate three-particle states. The relevant BetheYang equations are
mL sinh 1 + (1 2 ) = 2I1 ,
mL sinh 2 + 2(2 1 ) = 2I2 ,
where we supposed that the first and the third particles are degenerate (i.e., I3 = I1 ), and used
a convention for the phase-shift and the quantum numbers where (0) = 0. The density of these
degenerate states is then given by

LE1 + (1 2 )
(1 2 )
(13,
2) = det
,
LE2 + 2(1 2 )
2(1 2 )
where we used that ( ) = ( ). Using the above result and substituting integrals for the sums,
we can rewrite Eq. (A.4) in the form
249
2)
d1 d2 R(2E1 +E2 ) (13,
e
2(1)F (12) + 2(12)F (1) + (11)F (2) +
2 2
(112)

d1 d2 R(E1 +2E2 )
2(2)
1
(1)(2)
e
F (2)
+
2
2 2
(22)

d1 d3 R(2E1 +E3 )
1
1
+
(1)(3)
e
F (3),
2
2 2
(3)
where the ellipsis denote two terms that can be obtained by cyclical permutation of the indices
1, 2, 3 from the one that is explicitly displayed, and these three contributions can be shown to be
equal to each other by relabeling the integration variables:

d1 d2 R(2E1 +E2 ) (13,
1
2)
e
2(1)F (12) + 2(12)F (1) + (11)F (2)
2
2 2
(112)

d1 d2 R(E1 +2E2 )
2(2)
1
(1)(2)
e
F (2)
+
2
2 2
(22)

d1 d3 R(2E1 +E3 )
1
1
+
(A.5)
(1)(3)
e
F (3).
2
2 2
(3)
We first evaluate the terms containing F (23) which results in

d1 d2 s
F (1 , 2 )emR(cosh 1 +2 cosh 2 )
2 2 4
(A.6)
using that

(13,
2)
(1) = 1 + O L1 .
(112)
We can now treat the terms containing the amplitude F (1) = F (2) = F (3) = F2s . Exchanging
the variables 1 2 in the second line and redefining 3 2 in the third line of Eq. (A.5)
results in

2(1)2 (2)
F2s
d1 d2 R(2E1 +E2 )
(13,
2)
e
2(12) + (11) +
+ (1) .
2
2 2
(112)
(11)
The combination of the various densities in this expression requires special care. From the large
L asymptotics
(i) Ei L,
(ij ) Ei Ej L2 ,
(ij k) Ei Ej Ek L3 ,
(13,
2) E1 E2 L2
it naively scales with L. However, it can be easily verified that the coefficient of the leading term,
which is linear in L, is exactly zero. Without this, the large L limit would not make sense, so
this is rather reassuring. We can then calculate the subleading term, which requires tedious but
elementary manipulations. The end result turns out to be extremely simple

2(1)2 (2)
(13,
2)
2(12) + (11) +
+ (1) = (1 2 ) + O L1
(112)
(11)
so the contribution in the L limit turns out to be just

1
d1 d2 s
F (1 2 )emR(2 cosh 1 +cosh 2 ) .

2
2 2 2
Summing up the contributions (A.2), (A.3), (A.6) and (A.8) we indeed obtain (7.16).
(A.7)
(A.8)
250
References
[1] A.B. Zamolodchikov, Al.B. Zamolodchikov, Ann. Phys. 120 (1979) 253291.
[2] G. Mussardo, Phys. Rep. 218 (1992) 215379.
[3] M. Karowski, P. Weisz, Nucl. Phys. B 139 (1978) 455;
B. Berg, M. Karowski, P. Weisz, Phys. Rev. D 19 (1979) 2477;
M. Karowski, Phys. Rep. 49 (1979) 229.
[4] F.A. Smirnov, Form-factors in completely integrable models of quantum field theory, Adv. Ser. Math. Phys. 14
(1992) 1208.
[5] J.L. Cardy, G. Mussardo, Nucl. Phys. B 340 (1990) 387402.
[6] A. Koubek, G. Mussardo, Phys. Lett. B 311 (1993) 193201, hep-th/9306044.
[9] F.A. Smirnov, Nucl. Phys. B 453 (1995) 807824, hep-th/9501059.
[12] A.B. Zamolodchikov, Pisma Zh. Eksp. Teor. Fiz. 43 (1986) 565, JETP Lett. 43 (1986) 730.
[13] J.L. Cardy, Phys. Rev. Lett. 60 (1988) 2709;
A. Cappelli, D. Friedan, J.I. Latorre, Nucl. Phys. B 352 (1991) 616670;
D.Z. Freedman, J.I. Latorre, X. Vilasis, Mod. Phys. Lett. A 6 (1991) 531542.
[14] G. Delfino, P. Simonetti, J.L. Cardy, Phys. Lett. B 387 (1996) 327333, hep-th/9607046.
[15] B. Pozsgay, G. Takcs, Nucl. Phys. B 748 (2006) 485523, hep-th/0604022.
[16] B. Pozsgay, G. Takcs, Form factors in finite volume I: Form factor bootstrap and truncated conformal space, arXiv:
0706.1445 [hep-th].
[19] A. Leclair, F. Lesage, S. Sachdev, H. Saleur, Nucl. Phys. B 482 (1996) 579612, cond-mat/9606104.
[20] A. Leclair, G. Mussardo, Nucl. Phys. B 552 (1999) 624642, hep-th/9902075.
[21] H. Saleur, Nucl. Phys. B 567 (2000) 602610, hep-th/9909019.
[22] S.L. Lukyanov, Nucl. Phys. B 612 (2001) 391412, hep-th/0005027.
[23] G. Delfino, J. Phys. A 34 (2001) L161L168, hep-th/0101180.
[24] G. Mussardo, J. Phys. A 34 (2001) 73997410, hep-th/0103214.
[25] O.A. Castro-Alvaredo, A. Fring, Nucl. Phys. B 636 (2002) 611631, hep-th/0203130.
[26] F.H.L. Essler, R.M. Konik, Applications of massive integrable quantum field theories to problems in condensed
matter physics in: M. Shifman et al. (Ed.), From Fields to Strings, vol. 1, World Scientific, Singapore, 2005, pp. 684
830, cond-mat/0412421;
R.M. Konik, Phys. Rev. B 68 (2003) 104435, cond-mat/0105284.
[27] B.L. Altshuler, R.M. Konik, A.M. Tsvelik, Nucl. Phys. B 739 (2006) 311327, cond-mat/0508618.
[29] J.L. Cardy, G. Mussardo, Phys. Lett. B 225 (1989) 275278.
[30] A.B. Zamolodchikov, Adv. Stud. Pure Math. 19 (1989) 641;
A.B. Zamolodchikov, Int. J. Mod. Phys. A 3 (1988) 743.
[31] V.A. Fateev, Phys. Lett. B 324 (1994) 4551.
[32] G. Delfino, P. Simonetti, Phys. Lett. B 383 (1996) 450456, hep-th/9605065.
[33] G. Delfino, P. Grinza, G. Mussardo, Nucl. Phys. B 737 (2006) 291303, hep-th/0507133.
[34] V.A. Fateev, S. Lukyanov, A.B. Zamolodchikov, Al.B. Zamolodchikov, Nucl. Phys. B 516 (1998) 652674, hepth/9709034.
[35] http://people.sissa.it/~delfino/isingff.html.
[36] G. Delfino, G. Mussardo, P. Simonetti, Nucl. Phys. B 473 (1996) 469508, hep-th/9603011.
[39] J. Balog, Nucl. Phys. B 419 (1994) 480506.
[40] V.E. Korepin, N.M. Bogoliubov, A.G. Izergin, Quantum Inverse Scattering Method and Correlation Functions,
Cambridge Univ. Press, 1993.
[41] B. Doyon, SIGMA 3 (2007) 011, hep-th/0611066.

[42] F.A. Smirnov, hep-th/9802132;
V.E. Korepin, N.A. Slavnov, Int. J. Mod. Phys. B 13 (1999) 29332942, math-ph/9812026;
G. Mussardo, V. Riva, G. Sotkov, Nucl. Phys. B 670 (2003) 464578, hep-th/0307125.
251

Nucl - Phys.B v.788

Diunggah oleh

Informasi Dokumen

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Nucl - Phys.B v.788

Diunggah oleh

Hak Cipta:

Format Tersedia

Nuclear Physics B 788 (2008) 120

The -resonance in a finite volume

D-53115 Bonn, Germany

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

i0|T (x) (0)|0 =

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

These projectors obey the following relations

2.3. The poles of the propagator

Fig. 2. The self-energy of the at O( 3 ) in the SSE.

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

as well as the mesonbaryon loop functions

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

where the Poisson formula

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

The final result for the tadpoles is given by

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

The second expression corresponds to the subtraction term

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

and the function r( 2 ) is regular at 2 = E (1) .

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

V. Bernard et al. / Nuclear Physics B 788 (2008) 120

Nuclear Physics B 788 (2008) 2146

The soft-energy region in the radiative decay

P.D. Ruiz-Femena / Nuclear Physics B 788 (2008) 2146

P.D. Ruiz-Femena / Nuclear Physics B 788 (2008) 2146

P.D. Ruiz-Femena / Nuclear Physics B 788 (2008) 2146

where we have used the center of mass variables (r |x|)

P.D. Ruiz-Femena / Nuclear Physics B 788 (2008) 2146

as Gf m exp(r)/r. Therefore the Coulomb interaction can be treated perturbatively in the

P.D. Ruiz-Femena / Nuclear Physics B 788 (2008) 2146

i0 = 0) can be calculated easily:

P.D. Ruiz-Femena / Nuclear Physics B 788 (2008) 2146

0 (x)xi /ar, we can rewrite the electric amplitude as

The general structure of D ij (k) reads

The term proportional to D()

and the total electric amplitude is then equal to

P.D. Ruiz-Femena / Nuclear Physics B 788 (2008) 2146

where the leading term is given by

not considered there.

P.D. Ruiz-Femena / Nuclear Physics B 788 (2008) 2146

P.D. Ruiz-Femena / Nuclear Physics B 788 (2008) 2146

dr rG0 (0, r; )er/a sin

where the magnetic amplitude in the dipole approximation,

P.D. Ruiz-Femena / Nuclear Physics B 788 (2008) 2146

and the NRQED spectrum in this region, up to NLO, thus reads:

[11] has been done by Voloshin [5], showing a good agreement.

P.D. Ruiz-Femena / Nuclear Physics B 788 (2008) 2146

P.D. Ruiz-Femena / Nuclear Physics B 788 (2008) 2146

4. Threshold expansion of the QED 1-loop diagrams

P.D. Ruiz-Femena / Nuclear Physics B 788 (2008) 2146

P.D. Ruiz-Femena / Nuclear Physics B 788 (2008) 2146

is always hard, when q  m we can expand

P.D. Ruiz-Femena / Nuclear Physics B 788 (2008) 2146

i0|T (x) (0)|0 =

is always hard, when q m we can expand