Correlation functions in the QCD vacuum

Edward V. Shuryak
Department of Physics, State University of New York, Stony Brook, New York 11794
Correlation functions are one of the key tools used to study the structure of the QCD vacuum. They are
constructed out of the fundamental fields and can be calculated using quantum-field-theory methods, such
as lattice gauge theory. One can obtain many of these functions using the rich phenomenology of hadron
physics. They are also the object of study in various quark models of hadronic structure. This review begins with available phenomenological information about the correlation functions, with their most important properties emphasized. These are then compared with predictions of various theoretical approaches,
including lattice numerical simulations, the operator product expansion, and the interacting instanton approximation.

CONTENTS
I. Introduction
A. Preface
B. Why the correlation functions?
C. Different types of correlation functions
D. General relations and inequalities
II. Phenomenology of Mesonic Correlation Functions
A. Vector currents and correlators
B. Vector 1=1 (or p) channel
C. (o and <p channels
D. Strange vector (or K*) channel
E. Axial 1=1 (or A l) channel
F. Pseudoscalar correlation functions for the SUP) octet
(the TT,K,7] channels)

G. The SU(3) singlet correlation functions: axial, pseudoscalar, and gluonic ones
H. General properties of the scalar correlators
III. Theory of Mesonic Correlation Functions
A. Potential models and heavy quarkonia
B. Operator product expansion and QCD sum rules
C. Interacting instanton approximation
IV. Other Correlation Functions
A. Light-and-heavy mesons
B. Does the constituent quark model make sense?
C. Diquarks and baryons containing a heavy quark
D. Ordinary baryons. Why is the nucleon so light?
V. Correlation Functions at Nonzero Temperatures and/or
Densities
A. Melting the QCD vacuum
B. The low-temperature and low-density limits
C. Quark propagation in the quark-gluon plasma at high
temperatures
D. Lattice data
1. Screening masses
2. Binding of qq pairs propagating in spatial direction
3. Baryonic number susceptibility
E. Sum rules based on the operator product expansion at
finite temperatures and densities
1. Modifications of the operator product expansion
2. Operator expectation values
3. Some results
VI. Summary and Discussion
A. Summary of phenomenological observations
B. What new experiments are needed?
C. Further lattice studies
D. Theoretical problems
Notes added in proof
Acknowledgments
References
Reviews of Modern Physics, Vol. 65, No. 1, January 1993

I. INTRODUCTION
1
1
2
3
5
6
6
7

A. Preface

8
10
11
13

14
16
17
17
19
22
24
24
26
28
28
31
31
33
34
35
35
36
37
37
38
38
39
40
40
41
42
42
43
44
44

It is widely recognized that one of the central problems

of strong-interaction physics is in understanding the
structure of the ground state of quantum chromodynamics, the QCD vacuum. This "vacuum" is composed of
gauge and quark fields interacting in a complicated nonperturbative way. A general introduction to this subject
and references to original papers can be found in many
reviews, including Shuryak (1984, 1988a) and Shifman
(1992).
This review deals with a part of the problem, that related to correlation functions. Our purpose is to collect
the available phenomenological information and compare
it with predictions of various theoretical approaches. We
do not discuss the theoretical ideas in depth, and in many
places we ignore the related technicalities. Instead, we
proceed from a qualitative discussion directly to the
specific results. We concentrate on the main phenomena
to be explained by the theory and comment to what extent this goal has been achieved.
An important motivation for surveying the field is that
theorists working on strong-interaction physics are divided into several poorly interacting communities, according to the method they follow. The main theoretical approaches are (1) lattice gauge theory (LGT); (2) QCD
sum rules, based on the operator product expansion
(OPE); (3) interacting instanton approximation (IIA); and
(4) quark potential models for hadronic spectroscopy and
reactions.
We hope that this review can bridge the theoretical
gaps and bring these isolated communities together. Our
main target is to make better connections between lattice
calculations and the other theoretical analyses. The lattice community, having the most powerful methods
based on first principles, has great opportunities for making better contact with the phenomenology of stronginteraction physics.
As a common denominator for our discussion, we have
chosen the point-to-point correlation function in the
coordinate representation, with which it becomes possible to compare phenomenological information and
Copyright 1993 The American Physical Society

Edward V. Shuryak: Correlation functions in the QCD vacuum

theoretical predictions. As we shall see, it is quite

straightforward to "translate back" the main results of
the O P E and instanton frameworks to the coordinate
representation of the correlators.
Unfortunately, most lattice studies of correlation functions use either sources in the form of three-dimensional
"walls" or even more complicated nonlocal sources.
However, the point-to-point correlators are the basic objects; they carry more information and can be measured
by simple modification of the existing techniques. Apart
from the qualitatively new short-range phenomena to be
learned, one may obtain more accurate and reliable comparisons between lattice and empirical correlators. This
contrasts with the traditional technique of only studying
the asymptotic large-distance behavior of the correlators
related to the masses of the lightest hadrons.
It is hoped that some part of this review may be of interest to experimentalists working in different branches
of strong-interaction physics. We comment often on the
mutual consistency of various sets of data, on most desirable new experiments, on the main source of experimental uncertainties, etc. A brief summary of the experimental side of the problem can be found in Sec. VLB.
The paper is organized as follows. In Sec. II we discuss the available phenomenological information about
mesonic correlation functions, outlining a set of major
facts to be explained by the theory. Section III is devoted to theoretical predictions, which we consider only
briefly on a matter-of-fact level, concentrating on the results and their relation to experiment. In Sec. IV we discuss some other correlation functions, including lightand-heavy mesons and baryons, as well as ordinary
baryons, pointing out several other important observations related to quark properties and interactions. There
also we consider experimental information and theoretical ideas together. Section V is devoted to correlation
functions at nonzero temperature a n d / o r density. The
hadrons are expected to "melt" at some critical temperature into free quarks and gluons, and one can study this
phenomenon using the correlation functions. Our conclusions and suggestions are summarized in Sec. VI.
The reader who would like to see from the start which
operators and correlation functions are considered in this
review is invited to look at Table I.

B. Why the correlation functions?

Here we define what we mean by the correlation functions, discuss their asymptotic behavior at small and
large distances, and then try to explain why they play
such an important role in studies of the Q C D vacuum.
Below we deal with two types of operators: mesonic
ones of the type

and baryonic ones,

0^x)

= Mjfk^Jk

Rev. Mod. Phys., Vol. 65, No. 1, January 1993

d.2)

Here the color indices i>j,k are explicitly shown; we shall

omit them below. Other indices like spin and flavor are
not here specified, but shall be later. As all color indices
are properly contracted and all quark fields are taken at
the same point JC, these operators are manifestly gauge invariant.
The correlation function is defined as the vacuum expectation value (VEV) of the product 1 of two operators
taken at two points x and y:
K(x-y) = (0\O(x)O(y)\0) .

(1.3)

The first comment is that the vacuum is homogeneous;

so one of the points can be the origin of the coordinate
system, say, y = 0 . A second comment is that we assume
throughout this paper that the separation between the
points (x-y) is spacelike. The reason is we prefer to deal
with virtual propagation of quarks or hadrons from one
point to another, to have simple decaying functions instead of functions having a complicated oscillatory behavior.
There is an old question that one inevitably asks at this
point: why is there a correlation between fields outside
the light cone? It was essentially answered by Feynman:
particles can propagate along any path going from x to y.
Depending on the reference frame, an observer can consider the path t6 be a sequence of spontaneous paircreation and annihilation events. This correlation does
not contradict causality, because one cannot use it for
signal transfer. See textbooks on quantum field theory
for more.
One can look at the pairs of points of the correlator in
two ways. Either they are two points in space, taken at
the same instant of time and separated by the spatial distance x, or they are two events separated by some interval in imaginary or Euclidean time: ix0iy0 = r. Below
we use both interpretations, depending on which is more
convenient at the moment. We hope the reader will not
be confused by our using the symbols x and r interchangeably.
Let us now discuss the behavior of the correlation
functions at small and large distances. At small x
(remember, y =0) the asymptotic freedom of Q C D tells
us that quarks and gluons propagate freely, up to small
and calculable radiative corrections. Therefore K(x) in
the mesonic (baryonic) case is essentially the square (or
cube) of the free-quark propagator, S(x)=
(q(x)q(O)),
depending on whether a mesonic or baryonic correlator
is under consideration. From dimensional arguments,
the quark propagator S(x) is seen to scale as S(x)~x
~3,

Actually, it is the time-ordered product that is usually denoted by T: it is always implied below. We do not go into details
here, but only mention that such ^-ordering just corresponds to
using the standard path integrals and Feynman propagators for
particles propagating from x to y.

Edward V. Shuryak: Correlation functions in the QCD vacuum

ignoring small quark masses. 2 So for the mesonic or
baryonic correlators, there follows a small x limit
K(x)~ x ~6 (or ~x ~ 9 ) from the simple dimensional arguments 3 alone.
If quarks are allowed to propagate to larger distances,
they start interacting more strongly with vacuum fields.
If corrections are not too large, one can take these effects
into account using the operator product expansion (OPE)
formalism (see Sec. III.B). At intermediate distances,
description of the correlation functions becomes, in general, very complicated, and one may only evaluate them
by using either lattice numerical simulations or some vacuum models (e.g., the instanton model described in Sec.
III.C).
At large distances one can again understand the behavior of the correlation functions, using now completely
different kinds of arguments. Instead of thinking in
terms of fundamental fields, one may just use the formal
relation for the time evolution of an operator
O(t) = eiHTO(0)e~iHt,
where i f is a Hamiltonian, and
then insert a complete set of physical intermediate states
between the two operators:
K(t) = J,\(0\O(0)\n)\2e~iEnt

(1.4)

Now one can analytically continue the correlation function into the Euclidean time r=it and get a sum over decreasing exponents. 4
Physically, application of such relations in Q C D means
that one consider propagation of physical excitations or
hadrons between our two points, leading to the prediction that K(x)~exp(
mx) for large x, where m is the
mass of the lightest particle with the corresponding quantum numbers. This is essentially the idea of Yukawa, to
relate the range of the nuclear forces with the pion mass.
It is now easy to understand why the correlation functions are so important in nonperturbative Q C D and hadronic physics. The reason is that the same function can
be considered on two different levels: (1) in terms of the
fundamental Q C D fields, quark and gluons, or (2) in
terms of the physical intermediate states, using the vast
hadronic phenomenology of masses, coupling constants,
form factors, etc.
Moreover, there is a third approach to the correlation
functions. There are useful models originating from the

The coefficient is also easy to find by solving the Dirac equation for free massless particles: S(x)=z(iytJjdti){ l/4ir2x2).
3
Of course, QCD does have a dimensional parameter AQCD,
which eventually fixes the scale of all dimensional quantities.
However, in perturbation theory it only comes in via the radiative corrections. Therefore, at small x, those produce corrections to our estimates above containing as(x)~ l/ln(;cA).
4
A reader who does not like Euclidean time can repeat this exercise for spatially separated points and sum over virtual momenta of the intermediate states. The result is the same, due to
the four-dimensional symmetry of the Euclidean space-time.
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

original quark model of the '60s based on "constituent"

quarks and their effective interactions. It is instructive to
explain what we want to learn from the correlation functions in this language: it is the interquark effective interaction.
Application of these models to hadronic spectroscopy
reminds one of nuclear physics in its early days, when
only limited information about the nuclear forces was
known. Besides knowledge of the bound states, such as
the deuteron, one had only qualitative information that
the potential was attractive and of short range.
Indeed, potential-type quark models are successfully
applied to the evaluation of hadronic parameters. This is
discussed in detail by Godfrey and Isgur (1985) for
mesons and by Capstick and Isgur (1986) for baryons.
One obtains the average characteristics of the few lowest
hadronic states in each channel, and the theory is sensitive mainly to interquark interaction averaged over the
size of these states. The hadronic phenomenology
demonstrates the existence of flavor- and spinindependent confining forces, complemented by some
short-range spin-spin interaction.
However, we lack detailed knowledge of how the interquark interaction depends on distance and momenta.
Returning to the analogy with nuclear physics, we comment that only the extensive studies of nucleon-nucleon
scattering eventually showed all the details of nuclear
forces with their complicated spin-isospin structure.
Although qq or qq scattering is experimentally impossible to study, due to confinement, a set of various mesonic
correlation functions K(x) plays essentially the same role
as that played in nuclear physics by the scattering phase
shifts. These correlation functions are discussed below.
Roughly speaking, we shall describe virtual qq or qq
scattering, using wave packets of variable size instead of
physical hadrons.
C. Different types of correlation functions
The correlation functions in Euclidean space-time
K{x) [or K(T)] defined above are the objects of our discussion in what follows. As their argument x is the distance between the two points in Euclidean space-time, we
call them point-to-point correlation functions, or correlators.
We use this specific name because in various applications people have used other representations of correlators related to the above ones by some integral transformation. We compare here their definitions and briefly
comment on their advantages and disadvantages. 5
If one makes a Fourier transform of K(x), the resulting function Kmom(q2)
depends on the momentum
transfer q flowing from one operator to another. For

5
As we actually do not use any of them in what follows, the
reader may well skip this section.

Edward V. Shuryak: Correlation functions in the QCD vacuum

clarity we use the following notations, introducing

momentum squared with a negative sign Q2=q2.
We
are interested in spacelike momentum transfers, as in
scattering experiments, for which q2 <0 and Q2 >0.
Due to causality, the Fourier transform of the pointto-point correlation function satisfies the usual dispersion
relation,
Kmom(q2)

= (l/Tr)fds

ImK
{s)
m mom
mz
.
is q )

(1.5)

The numerator on the right-hand side, ImKmom(s),

is the
physical spectral density. It describes the squared matrix
elements of the operator in question between the vacuum
and all hadronic states with the invariant mass sl / 2 , and
is nonzero only for positive s. Because we are considering only negative q2, we never come across a vanishing
denominator and therefore may ignore ie, which is usually put in the denominator. This simplification is possible
because our discussion is restricted to virtual processes,
although in the right-hand side we shall use information
coming from the real processes of particle creation and
annihilation. 6
Equation (1.5) is the basis of the so-called Q C D sum
rules. Their general idea is as follows. Suppose one
knows Kmom(q2) in some region. This implies that some
integral of the spectral density is known, which can be
used to fix a set of physical parameters. Unfortunately,
such finite-energy sum rules are not very productive, because the dispersion integrals are usually divergent, leading to useful sum rules only after some subtractions.
This introduces extra parameters and significantly undetermines their predictive power.
Let us be more specific, taking the mesonic correlation
functions as an example. As mentioned earlier, the
mesonic correlators are given by a simple loop diagram
for small x, corresponding in the coordinate representation to the free-quark propagator squared. It is not
difficult to see that the imaginary part of this diagram,
corresponding to the production of a qq pair, is
ImKmom(s)~s.
This is also obvious on dimensional
grounds. Then from the dispersion integral we see that
at large s the Fourier-transformed correlator depends on
s as Kmom(s)~s
ln( s). However, the dispersion integral is also divergent, which signals that something is
missing in the last argument. One simple way to get
around this difficulty is to consider the second derivative
over Q2: then one deals with the function
Kf^om(s)f
which is defined by a convergent dispersion relation.
However, while going back to the original function
Kmom(s), one has to fix two integration constants corresponding to the missing terms in Kmom(s) of the type

6
In principle, virtual processes contain all the information;
but, of course, in practice, it is much more difficult to go in the
opposite direction and reproduce the physical spectral density
from the point-to-point correlators.

Rev. Mod. Phys., Vol. 65, No. 1, January 1993

cls-\-c2y which have no imaginary part. We can safely

ignore them below in our discussion of K(x), provided x
is never zero, because these correspond in the coordinate
space to contact terms, 8 functions, and their derivatives.
However, in finite-energy sum rules, these two undefined
constants need to be determined also from the data.
Several other ideas have been suggested to improve
these sum rules. First, after taking a sufficient number of
derivatives, one may take Q=0 and arrive at the socalled moments of the spectral density,
Mn=(l/7r)fdsImKmom(s)/sn

+l .

(1.6)

Following ideas presented in the original paper of Shifman, Vainshtein, and Zakharov (1979a), this method is
commonly used in the discussion of "charmonium sum
rules," which are related to correlators of Zc currents.
Another idea, suggested in the same paper (Shifman
et al.y 1979a), is to introduce the Borel transform of the
function Kmom(Q), defined as follows,
Kbor(m)=

lim
n> oo

Q2

\n

(-d/dQ2)"Kmom(Q2)

1 ;!

s* oo

m2 = Q2/n2

(1.7)
Applying this to the dispersion relation (1.5), we obtain
the sum rules in the Borel-transformed representation:
KboT(m) = (l/7r)fdsImKmom(s)exp(-s/m2)

(1.8)

Now the integral is cut off at large s by the exponential

function. This formula also has another useful feature:
usually we know the contribution of the lowest states (the
first resonance) better than the contribution of the multibody high-energy part; so the exponential cutoff hides
our ignorance and is therefore welcomed. Such forms of
the sum rules have been used in many papers based on
the O P E (see, e.g., references in Shuryak, 1984, 1988a
and Shifman, 1992).
However, most of the results obtained by this technique can also be presented in a much simpler way. Instead of the Borel transformation, one can Fourier transform back to coordinate space; then the dispersion relation has the transparent form (Shuryak, 1984, 1988a)
K(x) = (\/7r)fdsImKmom(s)D(sW2yx)

(1.9)

Here function ImKmom(s) describes the amplitude of production of all intermediate states of mass sl/2, while the
function
D(m,x)

= (m/47r2x)Kl(mx)

(1.10)

is nothing more than the propagator of these states to a

point x. In practice there is not much difference between
this equation and Borel sum rules. At large x the propagator goes as exp( mx); therefore one has an exponential cutoff, but with the factor exp( sl/2x)
instead of
exp( s/m2).
However, the space-time dispersion rela-

Edward V. Shuryak: Correlation functions in the QCD vacuum

tion has a much clearer physical interpretation, and we
shall keep to it in what follows.
For completeness, let us also mention one more type of
correlation function, the one traditionally used in lattice
gauge theory. This is the so-called plane-to-plane correlation function obtained from K(x) by an integration
over a three-dimensional plane:
^planetoplane(^)=</^^^(^^^(0,0)> .

(1.11)

has both types of contributions. Lattice calculations deal

mainly with one-loop diagrams, and therefore with the
1 = 1 channels, for technical reasons. Some general statements can be made about the one-loop diagrams, which
we would like to outline here, following Weingarten
(1983). 7
To derive the relations, we first note the following formula for the propagator in the backward direction, 8
S(x9y)=-r5S+(y,x)y5

In other words, a spatial integration selects intermediate

states of momentum zero; so dispersion relations are
done in energy only. The above function can be related
to a physical spectral density by
^pianetopiane( r ) = ( 1 / ^ ) / r f ^ Im/sT m o m (m )exp(-~rm ) .
(1.12)
The mass of the lowest hadron can be obtained directly
from the logarithmic derivative of this function at large
r. However, its essential disadvantage is that it mixes
contributions of small and large distances. This makes it
difficult to match with the OPE-derived functions at
small distances, and also obscures the physics going on at
intermediate distances.
However, in the important case of heavy-light mesons
(see Sec. IV.A), a lattice evaluation of point-to-point
correlation functions has been made. In this case there is
no difference between point-to-point and plane-to-plane
ones, because the super-heavy quark does not propagate
in space; so the integral in (1.11) has only a 8 function
contribution.
To summarize this section, we have noted five different
correlation functions in use: (1) the original point-topoint function K(x) in coordinate space; (2) the Fourier
transform Kmom(q2);
(3) the moments of the spectra density Mn; (4) the Borel-transform function Kmom(m);
and
(5) the plane-to-plane correlation function function
** plane to plane v ^ ' *

Although each correlation function has its advantages,

we suggest that for the understanding of the underlying
physics it is better to use the original point-to-point function K(x)9 and we shall do so in what follows.

(1.13)

One next decomposes it into Dirac matrices S = 2 a / r / ,

where r f = 1 , 7 5 , 7 ^ / 7 5 7 ^ 1 7 ^ 7 , , (f*=v). Finally, one
considers all diagonal one-loop correlators of the type
n = T r [ 5 ( x , ^ ) r / S ( j ; , x )Ft] and evaluates the traces.
The most interesting result appears for the pseudoscalar (pion) correlator: in this case one has a sum of all
coefficients squared,
n P s / n i e = (|a1|2+|a5|2+|aM|2+|aM5l2+|aMV|2)/|a0|2,
(1.14)
while, for example, the scalar one is
n s /n| ee
=

(~|a1|2~|fl5|2+|flM|2+|a^|2-|iiMV|2)/|aol2.
(1.15)

Here we have normalized the correlator to its asymptotically free version, containing free propagators of massless
quarks. Assuming that the propagation takes place in
the time direction, the propagator is S f r e e = 70/(27r2X())
and the only nonzero coefficient is a0 = 1 /(2W2XQ).
Comparing the above two equations, we obtain the Weingarten inequality. This states that the pseudoscalar correlator exceeds the scalar one at all distances, | FI PS > | II S .
The nontrivial thing is that the physical pion is very
light, while scalars are heavy; therefore for x > 0.5 fm the
scalar correlator is practically zero, while the pseudoscalar ratio is very large. This requires a very delicate cancellation between the different a{ in the propagator.
Additional information is provided by similar relations
for vector (p) and axial (Ax) channels,
n F / n f r = (2|a1|2-2|a5|2+|a/i|2-|aiti5|2)/|a0|2,

D. General relations and inequalities

One can classify correlation functions according to
quark paths, recognizing two different types of diagrams:
(1) the one-loop diagrams, in which quarks produced by
one operator go to another one, and (2) the two-loop diagrams, where quark lines are closed on the same operator.
For example, in the isospin 7 = 1 channels like t r + , one
has operators like u Td, where T is any Dirac matrix. In
this case, obviously, only the one-loop diagrams contribute. On the other hand, considering the nondiagonal
correlators in flavor, say, (u(x)Tu(x)d(0)d(0)),
one is
restricted to the second type. For most 7 = 0 cases, one
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

(1.16)
f

n / , / n r = ("2|a1| + 2|a5| 4-|a/x| -|a/i5| )/|a0|2,

(1.17)

7
In preparation of this section J. Verbaarschot has helped a lot
toward the understanding of the meaning of these relations. He
also found a few new ones.
8
Readers who wonder why 75 is needed should take as an example a free massive propagator and notice that the terms proportional to {xy)^Y^ and to m behave differently under the
transformation x<->y.

Edward V. Shuryak: Correlation functions in the QCD vacuum

TABLE I. Set of the operators and correlation functions discussed in this paper.
Channel

Section

Current
{uytlu~-dy^d)/2ul
(u7(1u^dy^d)/2l/2

P
CO

IYVLS

K*
Ax

uy5u-dy5d)(i/2l/2)
uiy5s
(uy5u~dy1d-2Jy5s)/(i/6W2)
(uYsYvU -^-dysy^d +
sy5Ylts)/(1/31/2)
GG
qq
bymub
QTq
T
q CTnQ
{uTCd)u-(uTCy5d)y5u
(u TCytlu )u

77

K
V
V'

v'
scalars
T
J5-type mesons
heavy baryons
N
A

and the following inequalities may be proven

e

nPS /n| > i( n v mf +nA

free_

ee

/nA )

(1.18)

rfree\

(1.19)

n P S /n t P r s e e >i(n F /nr-n^/n t r)

Witten (1983) has found another interesting inequality

between vector and axial correlators, but it applies only
to the momentum representation, and we do not discuss
it here.
As these inequalities are identities, they are satisfied
for any configuration of the gauge field, and they therefore are not very restrictive from a theoretical point of
view. However, they can be used to check consistency of
experimental data, as discussed below.
On the other hand, the diagonal correlators themselves
are positive monotonously decreasing functions, as is
clear from the spectral decomposition discussed in the
previous section. This condition is trivial to satisfy from
the empirical determination of correlators; but from the
theoretical point of view, it produces nontrivial limitations for the ensemble of vacuum fields. Some
configurations do produce negative correlators, especially
in the scalar channel. If their weight in the ensemble of
vacuum fields is too large, the positivity and monotonicity may be violated. These conditions may provide interesting new conditions on the models of the vacuum.
II. PHENOMENOLOGY
OF MESONIC CORRELATION FUNCTIONS
A. Vector currents and correlators
We start the discussion of the correlation functions
with vector currents for an obvious reason: these

9
These were shown to me by J. Verbaarschot (private communication).

Rev. Mod. Phys., Vol. 65, No. 1, January 1993

Info
e + e -^Nir,N
even
e+e~-+N7r,N
odd
e+e--+KK
+ N<rr
decay r>vT-\-K*
decay r>vT+iV17., N odd
pion decay
K decay

II.B
II.C
II.C
II.D
II.E
II.F
II.F
II.F
II.G
II.G
II.H
III.A
IV.A
IV.C
IV.D
IV.D

J/xfj-+y-\-r] etc.
J/ip-^y-\-7]
etc.
masses, generalities
e+e~>BB+
pions
masses of heavy flavored mesons
masses
O P E predictions
O P E predictions

currents really exist in nature, evidenced by their coupling to weak and electromagnetic fields, in contrast to
many other operators to be discussed. In several cases
the complete spectral density of the corresponding correlation functions is experimentally known, subject, of
course, to some experimental uncertainty, from
e^e~
annihilation into hadrons.
The vector currents and their correlation functions to
be discussed below will be denoted by the name of the
lightest meson in the corresponding channel; in particular, we define the p, the co, and the (f> currents as the following quark currents,
y=(l/21/2)[i7yM-Jy^]
x/2

j^(\/2 )[uY^u^dy^d}

or

uy^d

(2.1)
(2.2)
(2.3)

Further definitions may be found in Table I. The electromagnetic current is the following combination of the
quark currents:

j^=\uyfXu-\dy^d
= (l/2

1/2

) y - ( l / 2 1 / 2 3 ) y + .

(2.4)

The vector correlation functions are defined as

nifAiVU)=i<o|r7I>U)7I>(0)|o> ,

(2.5)

and the Fourier transform (in Minkowski space-time) is

traditionally written as
ifd4x

e^n/>v(x) = n/(^2)(^^v-^2g/xv) .

(2.6)

The right-hand side is explicitly transverse, i.e., it vanishes when multiplied by momentum q. This is necessary
for conservation of the vector current.
The dispersion relations for the scalar functions II; (# 2 )
are

Edward V. Shuryak: Correlation functions in the QCD vacuum

Ili(Q2^-q2)

= (l/7r)fds

'

(2.7)

where the physical spectral density Imll/te) is directly related to the cross section of e+e~ annihilation into hadrons. As this quantity is dimensionless, it is proportional
to the normalized cross section
*/<*> = < V . - ^ > / < V e - ^ + / 1 - < * > >

(2.8)

where the cross section of muon pair production (neglecting the muon mass) is just a + + - = (47ra2/3s) and
J

>/i ft

a is the fine-structure constant. If both quarks in the

current considered have the same flavor, as, for example,
the si in the (f> current, one obtains
lmlli(s)=Ri(s)/(l27re2)

(2.9)

As we shall see shortly, these relations are well satisfied

experimentally. In fact, this was historically one of the
first and simplest justifications for Q C D .
Coming back to coordinate representation of the
dispersion relation, one obtains
9

X C^dsRAsWis1'2^)
,
(2.13)
J
o
where, we recall, D{m,x) from Eq. (1.10) is just the propagator of a scalar mass-m particle to point x. Contracting indices and using the equation d2D(m,x)
= m2D(m>x)-\contact term, which we disregard, the
dispersion relation finally becomes
= (l/4ir2)fCOdssRi(s)D(sl/2,x)

ni^(x)

(2.14)

This is our experimental definition of the vector correlawhere eq is quark electric charge. Generalization to p><x>
tion functions.
channels is straightforward: instead of the charge there
A final comment related to our notation: as correlastands a corresponding coefficient in the equation for the
tors are very strongly decreasing functions of x, it is more
electromagnetic current, e.g.,
convenient to plot them normalized to the free propagators, namely, as Tl (x)/Hf^(x)
where nj^ e (x) correJmnp{s) = -^RAs)
.
(2.10)
sponds to the simple loop diagram describing free-quark
propagation. In such ratios all uninteresting normalizaThe reader may wonder how the different vector corretion factors, such as the quark electromagnetic charges,
lators are distinguished experimentally. It is clear
drop out. At small distances these ratios are all close to l
enough for the charge and beauty heavy flavors: if the
due to asymptotic freedom.
final state has a pair of such quarks, it is much more likely that they were directly produced in the electromagnetB. Vector / = 1 (or p) channel
ic current than that they were produced by final-state interactions. We shall also use this argument later for the
Figure 1 shows a sample of experimental data on Rp(s)
strange quark, although it is less justified in that case. To
at low energies. One can see that this function consists of
separate the light quark p,co channels, we make use of
two quite different parts: (1) the prominent p-meson restheir isospin and G parity. The two channels have a
onance, seen in the 2-rr channel, and (2) a mixture of muldifferent isospin 7 = 1 , 0 which is conserved by any strong
tipion states, which starts with (at least) two "primed"
final-state interaction. As it is well known, C parity plus
resonances, p' (1450) and p' (1700), seen mainly in the
isotopic invariance leads to the so-called G-parity conserfour-pion channel. However, taken together with the
vation, and pions have negative G parity. Therefore
six-pion channel, they add up to a rather smooth nonresstrong interactions do not mix states with even and odd
onance "continuum," and already at energies of about
numbers of pions. The currents jp,jm
have fixed G parity
1.5 GeV this spectral density follows the prediction
as well, and therefore pionic states created by them can
Rp = -| made above.
have only even or odd numbers of pions, respectively.
Let us start with a simple example to show how these
We have parametrized the data in Fig. 1 by the followrelations lead to definite predictions. The ratios Rt(s)
ing function, shown as the solid line:
have a very simple limit at high energies s, because in this
limit quarks and antiquarks are produced as free parti9
l+4(E~mp)2/T2p
cles. For currents containing only one quark flavor q, the
only difference with the muon is a different
charge and a color factor:
limRqis)

= e2Nc ,

electric

(2.11)

which for the <f> case gives lims_+O0R(p(s) = j . For the p

and co cases, we expand the electromagnetic current
equation (2.4) in terms of (2.1) and (2.2) and obtain from
that representation
lim Rp(s) = ,
S> 00

limi^(s) = j .
S>

00

Rev. Mod. Phys., Vol. 65, No. 1, January 1993

+ (l+as(E)/ir)

(2.12)

-^
r ,
l + e x p [ ( i s 0 E)/o]

(2.15)

where E0 = l.3 GeV, 8 = 0 . 2 GeV, and as{E) = 0JAn

E/0.2 GeV). This parametrization includes all essential
ingredients of the data: the resonance peak and
(smoothened) transition to the asymptotic behavior, corresponding to the famous cross section of free-quark production. For the high-energy contribution, we used a
smoothed function instead of the 6 function, as it is traditionally done in Q C D sum rules. The physical meaning

Edward V. Shuryak: Correlation functions in the QCD vacuum

2.0

FIG. 1. Ratio of a(e+e~-+mr)/o-(e+e~-+p,+fj,~)

with n
even, as a function of the total invariant mass of the hadronic
system. The data points correspond to the following states: the
error bars without points to 2TT (Barkov et al., 1985); stars to 4-rr
(Cosme et al, 1979; Cordier et al, 1982a; Barkov et al, 1988;
Kurdadze et al, 1988); and triangles to 677" states (Cosme et al,
1979; Dolinsky et al, 1989). The solid line is our fit to the sum
of all contributions with n even for the total cross section in the
I\ channel. We have not shown all data points available near
the top of the p peak, nor the region 2E>2 GeV, where the
agreement between data points and our fitted curve is very
good.

of the parameter E0 is the same: it is the energy above

which the asymptotic freedom is restored and simple
quark model estimates for cross section become valid.
We now calculate the correlation function using this
parametrization and the dispersion relation (2.14), taking
the integral over all energies. The resulting curve is
shown in Fig. 2, where the contributions of two components of the spectral density mentioned above are also

1.0

*\

X,

U'

0.5

.-I"'".

^.
1

continuum
1

0.5

. ~Y--t--J--.---.--.i_

L.-

1.5
x(fm)

FIG. 2. Ratio of the 7 = 1 vector correlation function to that

corresponding to free-quark propagation vs the distance x. The
dot-dashed curve is the p-meson contribution calculated as the
contribution to the integral of the region below a total energy of
1 GeV. The dashed curve labeled "continuum" is the complementary contribution of all hadronic states above 1 GeV, and
the solid curve is their sum.
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

shown separately. The first striking observation is that,

starting with the rather complication function I m l l ^ U ) ,
we arrived at a very smooth function of the separation x.
Clearly, the way back from the coordinate representation
to physical spectral density would be much more
difficult.
The second striking observation (Shuryak, 1989a) is
that the contributions of the lowest meson and continuum complement each other in such a way that the ratio
U(x)/IlfTee(x)
remains close to 1 up to distances as large
as 1.5 fm. We call this fine tuning of all parameters superduality. As we shall show, it persists in all vector
channels. For small distances it is nothing more than
asymptotic freedom. At ; c ~ 0 . 3 fm, it is a consequence
of the so-called duality between hadronic and quark
description. However, from 0.3 to 1.5 fm, where the
correlator drops by more than four orders of magnitude,
it is an unexpected and remarkable phenomenon!
Completing this section, let us examine the errors in
the determination of the correlators. Of course, the experimental uncertainties are there, and their magnitudes
are seen in Fig. 1. In the p region (due essentially to
VEPP-2M data from Novosibirsk), the resulting error at
large x > 0 . 6 fm in the correlator is about 5 % . The
high-energy domain is covered by SPEAR data from
SLAC, which fix the normalization of small-* region also
to within a few percent. However, in the most interesting medium distances, we have contributions from the p'
energy region, and here the situation is actually even
more uncertain than our Fig. 1 indicates: the Frascatti
and Orsay data do not agree, and the problem is not statistical. It is quite probable that in this region our parametrization of the cross section is off by as much as
3 0 % , which may lead to error bars for the ratio
n U ) / I I f r e e ( ; c ) plotted in Fig. 2 of about 15% at x - 0 . 6
fm. In view of the apparent systematic deviation of the
two sets of data, it would not be useful to display statistical error bars on the plots of the correlators.
C. o) and tf> channels
The next channel we discuss is the isoscalar channel
having the quantum numbers of the co meson. The corresponding data for the cross section of e+e~ annihilation
into an odd number of pions, now summed over all channels, are shown in Fig. 3. The top of the co peak is not
shown because here the Breit-Wigner curve (with the
width value taken from Review of Particle Properties,
Hernandez et al., 1990) is very accurate: the peak value
of R^ is about 12. One can also see a trace of the (f> peak
due to the co-(f> mixing, which will be disregarded in our
parametrization. Note the change of scale and the essentially larger error bars compared to the 1 = 1 channel.
Within uncertainties the continuum magnitude approaches its asymptotic value Rco = j at about the same
energies as in the 1=1 channel.
For narrow resonances, we include the resonance contribution in the simple form

Edward V. Shuryak: Correlation functions in the GCD vacuum

1.5

~iiI

r-

0.5

^.continuum
i

0.5

F""i

>---

1.5
x(fm)

0.8

FIG. 4. Same as in Fig. 2, but for the 1 = 0 vector correlator,

the co channel.

I
2E(GeV)

FIG. 3. As in Fig. 1, the isospin 1=0 final states, defined as

those having an odd number of pions. Data points are summed
over all channels, compiled in Dolinsky et al. (1989), while the
curve is our fit discussed in the text.

"/u^^'lres"

(2.16)

res> ' '

where the coupling constants of the currents to mesons

are defined as follows 10 :
<Oljjf (resonance> =fresmresefl

(2.17)

Here the eM is the polarization vector of the vector

meson. These couplings and the partial widths to the
e e channel are related as follows:

and a much smaller cross section, the correlator in the co

channel is similar to the p correlator. Figure 2 for the p
channel and Fig. 4 for co agree to within uncertainties,
and the only difference between them appears at distances as large as about 2 fm!
To understand what this phenomenon means, let us
look at the difference between the p and co correlators.
As the former current has the uudd flavor structure,
and the latter uu -\-dd, this difference is the vector
flavor-changing correlator
K^(x)=(urflu(x)dyfld(0))-

'^^COylJLfJL

A l p ^ ^ ;

(2.20)

Thus the data presented above tell us that this amplitude is for some reason extremely small. Unfortunately,
we do not really know how small it is at intermediate dis3mresr(res->e+e )
2 :
(2.18)
tances, up to 1 fm or so, because it is within the experi/ . res
4wa2
mental uncertainties. Only at distances as large as 2 fm
does the difference between the co and p correlators beFor reference, the accepted values of the coupling constants of the p , <f>, and co mesons are / fi> = 46 MeV, come clearly observable. It means that the flavorchanging correlation function (2.20) becomes comparable
/ ^ 7 9 M e V , / p 152 MeV.
to the flavor-diagonal ones 11 only when the latter drops
Next Fig. 4 shows the correlation
function
by many orders of magnitude.
TLa)(x)/Ilfree(x),
again with contributions from the co resThere are two more striking experimental observations
onance and the continuum state shown separately and in
that
suggest that the famous Zweig rule, forbidding the
sum. The curve corresponds to the following parametrizflavor-changing
transitions, is indeed surprisingly strict
ation of the cross section:
in the vector channels: (1) the p-co mass difference is only
12
12 MeV; (2) the co-(f> mixing angle is only l-3.
RJE) =
2
N o general reasons for such strong suppression of
\+4(E-mJ /ri
flavor-changing
transitions in vector channels are known,
1
although some interesting hints have been suggested. In
+
Ul+aAE)/ir) l+exp[(E -E)/8]
'
0
particular, a perturbative analysis leads to the idea that
in the vector case one needs at least three gluons in the
(2.19)
intermediate state, not two as in the pseudoscalar case.
where now E0 = l.l GeV and S = 0 . 2 GeV. The experiHowever, this argument should not be applicable to dismental error on the <f> contribution is about 3 % , but tances of the order of 1 fm and beyond. In this respect,
about 2 0 % for the continuum.
In spite of completely different final hadronic states
11

10

There will, of course, be some ambiguity in these definitions,

if the resonance is broad.
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

We show below that for pseudoscalar correlators such deviation happens at much smaller distances, where the correlation
function is about four orders of magnitude larger.

Edward V. Shuryak: Correlation functions in the QCD vacuum

10
0.5

0.4 h-

0.5
<z 0.2

FIG. 6. Same as in Fig. 2, but for the <f> correlator.

FIG. 5. Same as in Fig. 1, but for the channels containing a
pair of K mesons. The points marked by crosses, closed dots,
open dots, and triangles correspond to the following final states:
K+K~
(Ivanov et al, 1981);
KsKL,K$K-<n-++K$K+TT~
(Mane et al, 1982); and K^K-^TT'
(Cordier et a/., 1982b),
respectively. The solid curve is our fit to their sum. We do not
show the fit near the top of the (/> peak (which in this case is very
high, R being about 50), because it is perfect there.

an important observation can be made from nonperturbative considerations to be discussed later (Sec. III.C). It
is that vector and axial channels do not have a direct instanton contribution in first order in 't Hooft interaction,
in contrast to pseudoscalar and scalar ones. However,
this argument also cannot account for the smallness of
this transition up to very large distances, where multiinstanton effects become important.
Now we show one more figure related to e+e~ annihilation experiments, Fig. 5, which presents the cross section of the production of channels with KK plus pions.
We assume in this case that the Ty^s current dominates
in strangeness production, which may not be well
justified. As above, we do not display the fit near the top
of the (f> peak, because it is nearly perfect; the maximal
value is R^ ~ 50. Instead we show how our fit reproduces
the sum of all other contributions, shown by the solid
line. The parametrization used here was

B[T-VT+K*)

B(r-+vT+p)

22

tan (<9c)

A
fP

52.4

l+4(-m^) 2 /r 2
+ }{l+as(E)/ir)
3

-~
=^ ,
l+exp[(2?02?)/8]

(2.21)

where E0 ~- 1.5 GeV and 5 = 0 . 4 GeV.

Finally, we present in Fig. 6 the correlator
I I ^ ( x ) / n f r e e ( x ) , which is also surprisingly similar 12 to the
pyco correlators shown above.
D. Strange vector (or K*) channel
For completeness, let us also consider the strange vector channel. Here the current is j ^ uy^s and the
lowest meson is the K*(%92). Phenomenological analysis
in this case is not based on electromagnetic processes, but
rather on the vector part of weak currents. The data
come in this case from the weak decay process
r>v r +hadrons. Since the hadrons are produced from a
virtual W instead of virtual photon, we obtain an admixture of the strange current from the Cabibbo mixing of
the weak current.
Comparing the Cabibbo suppressed production of K *
to the Cabibbo allowed production of p in this decay, one
can obtain the following relation,

(l-m*/m2)2

(l+2m*/m2)

(l-m2/m2)2

(l+2m2/m2)

where 0C is the Cabibbo angle. Inserting on the left-hand

side the experimental ratio 0.0143+0.0031 (from Review
of Particle Properties, Hernandez et al., 1990), one obtains the following ratio of the coupling constants:
12

R^E)

= 1.1+0.1

(2.22)

(2.23)

We do not have sufficient information about the cross

Here our presentation is somewhat illogical, because we still measure the correlator in units of n ^ e , corresponding to the free
propagation of massless quarks. The decrease of R^ix) with distance is partly kinematical, due to nonzero strange quark mass. We
have not included this correction, in order to make comparison with nonstrange correlators in the same figure.
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

Edward V. Shuryak: Correlation functions in the QCD vacuum

I.O

. . . I . . i I . i

I -~X

X.

K*V

X.

0.5

/
n "-<'*.

I
0.5

X
i

continuum

L. i i 1 i "*T~'r--i---J-ni--i.,..
1.5
x(fm)

s(GeV^

FIG. 7. Same as in Fig. 2, but for the K * correlator.

section of weak production of nonresonance states with

such quantum numbers, K and pions. In addition, the r
lepton mass sets a rather restrictive limit on the available
energy. However, one can still make some arguments
based on spectroscopic data. Indeed, in the strange vector channel there are two primed resonances, at 1415 and
1715 MeV, very much similar to two p' resonances near
1600 MeV. We shall thus assume that the nonresonance
part of both K * and p cross sections are similar. Therefore the same nonresonant contribution as that for the p
channel will be taken in the parametrization, scaled, of
course, to a different limit at infinite energies.
The corresponding curves for n^*/II f r e e (jc) are given
in Fig. 7. The resulting curve fits perfectly between the p
and <j> curves discussed above, suggesting that all these
completely different sets of data are, in fact, deeply connected to each other.

FIG. 8. Contribution of the 3rr channels to the spectral density

of the axial current, measured in the r lepton decay by the
ARGUS Collaboration (Albrecht et a/., 1986; the five-pion one
is small and rather uncertain). The curve is just a parametrization used in the theoretical paper by Peccei and Sola (1987),
from which we took this figure, and it is not used here.

The experimentally measured distribution into three

pions as a function of their invariant mass is shown in
Fig. 8. The asymmetric peak around 1.2 GeV is the contribution of the Ax meson. 14 Its dominance is also
confirmed by the observation that two channels with
three pions, ir~Tr~Tr+ and T T W - , have branching ratios
(of all T>vT-f hadrons decays) equal to (6.80.6)% and
(7.5+0.9)%, respectively. They are equal within uncertainties, and this is precisely what should be the case if
they are dominated by A x decays. For those reasons, we
treat the peak seen in the r decays as an "effective" A x
meson.
Let us introduce coupling constants fA
similar to
those of vector resonances:
<0\dy^r5u\Al)=fAimAifl

E. Axial / = 1 (or A^) channel

Now we turn from vector to axial-vector channels,
concentrating on the 1 = 1 channel. This has the quantum numbers of the A i meson and is related to the following current:
(2.24)
Data corresponding to this channel are also obtainable
from the r lepton decay into the corresponding neutrino
and hadrons, because the weak current has both vector
and axial components. Since we deal with the charge
current associated with the W exchange, we do not have
an 1 = 0 component; so production of odd numbers of
pions is now entirely due to the axial part of the
current. 1 3

13
Decays into neutrino and an even number of pions are, as in
the e*e~~ annihilation, related to the vector p-type current.
The corresponding data are consistent with the e+e~ annihilation data, although they are much less accurate.

Rev. Mod. Phys., Vol. 65, No. 1, January 1993

11

(2.25)

From the experimental branching ratios and the theoretical equation

B{T->VT-AX)

B(r->vT+p)
/A

fp

(l-m2A{/m2T)2
(l-m2p/m2T)2

(l +

2m2A{/m2T)
(l+2m2p/m2T)

(2.26)

one deduces a value for the coupling constant,

/ ^ / / p ^ 1.0+0.07 .

(2.27)

Having fixed the resonance contribution, we next

14
The A i shape observed in the r decay and hadronic reactions is somewhat different. This point is discussed in Isgur
(1989), which also contains further references. The data shown
in Fig. 8 seem to suggest an admixture of some nonresonance
background at the largest energies, but the errors are still too
large to allow any definite conclusions.

Edward V. Shuryak: Correlation functions in the QCD vacuum

12

proceed to the continuum states at larger invariant

masses. Unfortunately, the r lepton is not heavy enough
to produce final states in the asymptotic region; direct
observation of the axial spectral density is limited by its
mass, 1784 MeV. Moreover, as one can see from Fig. 8,
the statistics of the existing experiments are only good up
t o s 1 / 2 1 . 4 - 1 . 5 GeV. Therefore we do not see the most
interesting region in which spectral density approaches
its asymptotic limit.
However, we have some general arguments that allow
one to fix the continuum contribution with reasonably
small uncertainties. First, in the chiral limit, the following inequality has been proven 1 5 (Witten, 1983):

n^(g 2 )-n^;(^ 2 )>o (for aiu 2 <o).

(2.28)

This condition should become an equality at large \q2\

because the Od/q2)
terms, corresponding to the
O (1 /x2) terms in the coordinate representation of the
correlators, should be the same for vector and axial
correlators in the chiral limit. This statement is known
as the second Weinberg sum rule 1 6 (Weinberg, 1967):
1
/ ife(Imnw4 -Imnj^) = 0

(2.29)

As the resonance contribution is proportional to

m2
A f2A ~~m2pf1p>^y
the contribution of the nonresonance continuum should be negative. Assuming our previous parametrization of the continuum, we may therefore conclude that asymptotic freedom in the axial channel should be recovered at larger energies, E0 l >E$9
which is indeed the case. The sum rule (2.29) is satisfied

transverse part, namely,

VL^v(a)-{q^lqv-gyLVq2)

(2.30)

This is not the case. The existence of a Goldstone mode,

the massless pion, coupled to the axial current, produces,
in addition, a longitudinal contribution:

n^v{q)=iit(q )^q^v-g^q

)+flq^v/r

(2.31)

In the coordinate representation, the second term just

gives a singularity at x = 0 , which does not spoil current
conservation.
Now we proceed further, discussing the real world in
which quark masses are nonzero. We still have a longitudinal part due to a pion contribution, which now depends
on x as 3 /i 3 v Z>(m 7r ,x). Taking the divergence 3^
(or contracting indices fiv), we obtain
d2D(m1T>x)
2
= m r D ( m 7 r , x ) + c o n t a c t term. Now we have a longitudinal contribution, nontrivially depending on distance,
but it is proportional to m 2 . This result is not unexpected: although in the real world the axial current is not
conserved, its divergence is O ( m q ) = O (m %).
The conclusion from these theoretical considerations is
that one can partially get rid of the pion signal in the A x
correlation function by simply contracting the indices on
the correlator. This will also make better contact to the
correlators for the vector channels. The contraction
leads to the following approximate relation for the axial
correlator:

at E0 l 1 . 5 GeV, if we use the same shape as that used

for the p case, with 8 = 0.2 GeV. This is quite a firm prediction, provided the shape of the continuum spectrum is
the same. However, to show the sensitivity of the correlator to this uncertainty, we shall display two curves for
the axial correlator, with E0 l = 1 . 5 and 1.7 GeV. For a
more detailed discussion of the axial spectral density, including, in particular, its relation t o r n +m 0 , see Peccei and Sola (1987).
Before plotting the correlation function, let us also
clarify a theoretical point related to a general form of the
axial correlators. If chiral symmetry were exact, with all
quark masses zero, the nonsinglet axial currents would be
conserved. Because of that, one might think that the
Fourier transform of their correlators would have only a

15

The author is indebted to S. Nussinov for bringing this

theorem to his attention.
16
This statement can be derived from the fact that, in the
chiral limit, the only dimension-4 scalar operator is a gluonic
field strength squared, which contributes the same amount to
vector and axial correlators (see Sec. III.B). However, for
nonzero quark masses, there appear contributions of the type
mqqq, different for vector and axial correlators.
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

3 r
+ -f4lT J

,_ . _ .
fdEE3D(E,r)1+

l+ax(E)/ir
exp[(E0-E)/8]
(2.32)

The first two terms are the contributions of the A { and

the 7r, and the third term is the nonresonant continuum.
The latter is expressed in our usual way, with a perturbative contribution starting at some E0, taken to be 1.5 or
1.7 GeV. As before, we took 8 = 0 . 2 GeV.
Now everything is fixed, and the resulting correlator is
shown in Fig. 9. Comparing it with the p correlator in
Fig. 2, one observes that it has a completely different
shape. The A { contribution can be slightly larger than
>
the p one at small x (again, because fAmA
/ p m p )> ^ u t
at larger distances it drops due to larger A x mass. Eventually, at large JC, the axial correlator grows again, due to
the long-range pion contribution.
Finally let us emphasize that the difference between
the vector and axial correlators is entirely due to the
chiral asymmetry of the Q C D vacuum. By studying how
this difference develops as a function of distance, one can
hope to learn something about the mechanisms creating
this asymmetry.

Edward V. Shuryak: Correlation functions in the QCD vacuum

terms, its mass is small due to the near vanishing of the

light quark masses. This subject is reviewed in detail by
Gasser and Leutwyler (1987), which also has original
references.
However, taking a closer look at this problem, one arrives at the opposite puzzling conclusion: the pion is
surprisingly heavy, given the light quark masses. Indeed,
the pion mass can be written as

1 . 5 1 i i i | i i i i | i i i i i i i i r

ml = (mu+md)K
0.5

I
x(fm)

I.5

FIG. 9. Same as in Fig. 2, but for the axial current. The dotdashed line is the contribution of the Al meson, while the
short-dashed one is that of the pion. Two long-dashed lines
show the contribution of the nonresonance continuum, if its
threshold is E0 1.5 or 1.7 GeV. Two solid lines show the
sums of all contributions in these two cases; the true correlator
is somewhere between them.

F. Pseudoscalar correlation functions for

the SU(3) octet (the ir,K,r} channels)
Here we consider correlations of the octet pseudoscalar quark-antiquark operators

j K = iuy s

(2.33)

j v = (i/6W2)(uy5u+dy5d-2sy5s)

(2.34)
.

(2.35)

These correlators are very important for the understanding of Q C D vacuum structure. One might naively
think that because the pseudoscalars are the lowest excitations of the Q C D vacuum, they tell us primarily about
its long-range structure. However, as we shall see shortly, they also provide much puzzling information about its
short-range structure as well.
Generally speaking, the pseudoscalar and scalar
mesons are rather exceptional members of the family of
hadrons. There are some surprisingly large numbers attached to them; in particular, the coupling constants to
the corresponding currents are very large. Therefore the
contributions of these particles to the correlators are also
important at small x.
Before we come to correlation functions, some general
comments about pseudoscalars are in order. Throughout
the history of hadronic physics, from naive nonrelativistic quark models to modern lattice calculations, some
puzzles related to these particles have presented
difficulties, and they are in many cases still unexplained.
New, surprising facts are revealed if one considers the
correlation functions.
The well-known observation that the pion is extraordinarily light was, in fact, explained in classical works of
the '60s, even before Q C D was discovered: it is a Goldstone mode associated with chiral symmetry. In Q C D
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

(2.36)

where the constant K is nonzero in the chiral limit. This

constant is related to the quark condensate and the piondecay constant / 7 r = 2 l / 2 F 7 r = I3l MeV by the famous relation (Gell-Mann, Oakes, and Renner, 1968)
K=2\{uu)\/fl

(2.37)

We do not present its derivation here and only note that

the standard values of the quark masses 17 are (Gasser and
Leutwyler, 1987)
md7MeV,

j = (i/2U2)(uy5u-dy5d)

13

m*4MeV.

(2.38)

One then finds a very large value of this constant associated with the quark condensate: K1700
MeV. 1 8
Masses are external to Q C D , but the value of K is an
internal problem, which should be explained by Q C D .
We formulate this question in a slightly more general
way as the first puzzle: (1) Why are the masses of the
pseudoscalar octet mesons so sensitive to small quark
masses?
The second well-known puzzle related to the pseudoscalar channels is the famous Weinberg (1975)
"UA{\)
problem," which is related to the SU(3) singlet channel
and the 77' meson. Ignoring the u,d quark masses and
considering only the effect of ra5, one can easily see that
chiral perturbation theory predicts 77' to be lighter than
the 7] meson: the former has ~ of the "strange" component, while the latter has f of it. 19 Experimentally,
m ^ 9 5 8 MeV, which is much larger than these naive
estimates. Let us now formulate this problem somewhat
more generally: (2) Why is the singlet channel so much

17

Quark masses are not physical, but are instead a kind of

theoretical parameter; so their values depend on their exact
definition. In particular, they have perturbative anomalous dimensions; so the numbers depend on "resolution" (normalization point /x0) used. For example, speaking about bare quark
masses in the lattice Lagrangian, one has resolution on the scale
of lattice spacing fiQ a~x. The numbers mentioned correspond
to the scale fi0= 1 GeV.
18
Accuracy of these "standard" numbers depends on whether
extrapolation of chiral perturbation theory is good for the
strange quark; see details in Gasser and Leutwyler (1987).
19
We simplify discussion of this point for pedagogical reasons.
The reader may consult the original paper (Weinberg, 1975) for
his estimates of the upper limit of the 77', with and without
Oims) effects.

Edward V. Shuryak: Correlation functions in the QCD vacuum

14

different from the octet ones? What is the mechanism responsible for this splitting?
The third problem we address is also an old one, related to the fact that in pseudoscalar channel we do not see
even a trace of the Zweig rule. Namely, flavor changing
is not suppressed in this channel, but rather enhanced:
(3) Why isn't the strange sector in the pseudoscalar multiplet separated from the nonstrange one, as in other multiplets? What is the mechanism of these mixings?
We now proceed to discussion of the pseudoscalar
correlation functions. The main point is that the coupling constants of the mesons to the pseudoscalar
currents also can be expressed in terms of known parameters. For example, starting with the definition of the
pion-decay constant
<0\urfly5d\7rypfl)=if^p^

(2.39)

one takes the divergence of the axial current and obtains

{mu+md)(Q\uir5d\ir,p)=f7rm27r

(2.40)

from which one obtains the needed pseudoscalar coupling constants

k7r==(0\uiy5d\7rypfl)=fnK~(480MeV)2

(2.41)

This kw is a large quantity, due to the large value of the

factor K. This in turn can be traced to a large value of
the quark condensate.
The axial current matrix element to the K meson is
known from its weak decays, and the decay constant is
(2.42)

/^1.24/V.

We shall extrapolate from the TT and K cases to the TJdecay constant with
fr)~jfK

jfir~l'32>frr

(2.43)

We obtained this formula assuming that the deviations

from the SU(3) symmetry were due to strangeness. The 17
is j strange, while the K is only \ strange.
So far we only have information about 77% K, rj contributions to the axial correlators. However, in order to obtain their contributions to the pseudoscalar ones, we have
to make additional assumptions. Here we assume that
they scale in a way similar to the decay constants 2 0 :

^K/K-fK/f^/K-fv/f*

(2.44)

In Fig. 10 we show the resulting w, K, and 77 pseudoscalar

correlators in the form K(x )/KfreQ{x).
Apart from resonance contributions, we assumed a nonresonance continuum and selected a value EQ^1.6
GeV to smoothly
bring the ratio to unity at small distances. In fact, the

20
In any case, a 10-20 % level of accuracy is good enough for
most of our conclusions here, and at this level all couplings can
just be considered as equal.

Rev. Mod. Phys., Vol. 65, No. 1, January 1993

ambiguity in the E0 value is important only in a very

small window at about x = 0 . 2 fm.
Note the marked difference compared to the vector
correlators considered above: instead of changes within
1 0 - 2 0 % in the region x~l
fm, the ratio K/K{ree
has
changed by two orders of magnitude.
The general reason for this behavior is the well-known
feature of pseudoscalar mesons, that they are exceptionally light. In terms of qq interaction, this behavior implies that there is a strong attraction between quark and
the antiquark in this channel, forcing them to move in a
correlated manner. As a result, the correlation function
is larger than the perturbative one.
Note also that up to distances of the order of 0.5 fm,
there is no marked difference between the three curves,
which implies that all effects proportional to the strange
quark mass are irrelevant in this region. In fact, the
heavier mesons have slightly larger couplings, making
the curves for different channels even more similar.
Surprisingly, due to contributions from these lowest
mesons alone, asymptotic freedom is violated at very
small distances, about \ fm. This fact, noticed in Novikov et ah (1981), deserves to be considered as another
general puzzle: (4) Why do deviations from the perturbative behavior start at such small distances in the pseudoscalar channels?
G. The SU(3) singlet correlation functions:
axial, pseudoscalar, and gluonic ones
The SU(3) singlet channel, called 17' for brevity, is traditionally discussed in relation to the axial current
s
s
5
1/!

y=(*r v +rfy y/+sr v)/3

(2.45)

Its matrix element is connected to fv- in the usual way:

<0\j%\vf)=ifv>K

(2.46)

This axial current is subject to the famous Adler-BellJackiw anomaly (Adler, 1969; Bell and Jackiw, 1969),
which means that its divergence is not just proportional
to the quark masses, but it also contains a gluonic operator 21.

V= 3

1/2

2im,Jy
s
s 5r 5

3# 2
~
+ ^ r2 G G

(2.47)

16TT

In the above equation GG = ~a^vGaf3G^v

is the contraction of the gluonic field strength with its dual, analogous
to E B in electromagnetism.
Therefore, sandwiching this relation between vacuum
and 7)' states, one does not find a direct relation between
the couplings to pseudoscalar and axial currents Ay and

Contributions proportional to the light quark masses are ignored here.

Edward V. Shuryak: Correlation functions in the QCD vacuum

Several estimates of f ^ have been put forward by Novikov et aL (1980), all suggesting it to be smaller than
(2.48)

/V =(0.5-0.7)/,

The simplest estimate is related to the J /tfj radiative

decay, which is also important because it provides some
direct information about the matrix elements of the
gluonic operator entering the anomaly (Novikov et aL,
1980). Indeed, if the charmed quarks are sufficiently
heavy, one can describe the cc annihilation in terms of local operators. We do not need to go into detail here, but
only comment that, for the decays
J/xp-^y-\-pseudoscalar meson, one has to deal with the lowest-dimension 22
pseudoscalar gluonic operator GG. The exact coefficient
of this operator in the effective Lagrangian is irrelevant,
because we shall only consider ratios of the decay probabilities:
<0|GG|i7')
(0\GG\r])

(2.49)

The last factor is the phase-space ratio for P-wave decays.

Experimentally the left-hand-side ratio is
4 . 9 + 0 . 5 , 2 3 from which one finds the ratio of the matrix
elements to be
(0\GG\y')

= 2.46+0.1

(2.50)

<0|GG|T7>

Since that work was published, another large contribution in radiative decay of t/> has been found, that of the
decay into photon and T / ( 1 4 3 0 ) (originally called t). Repeating the same argument, one obtains an even slightly
larger 24 matrix element for this particle:
<0lGgl^(1430))=11202
<0|GG|i7'>

This fixes the absolute scale of these matrix elements.

Now, ignoring the 0(ms) term, we obtain the value for
the coupling constant, fv> ^0.74-f^.
Armed with this information, let us return to the
correlation functions. Unfortunately, the coupling constant of the pseudoscalar SU(3) singlet current remains
unknown. Nevertheless, just for the sake of comparison
with other pseudoscalar correlators, we have also plotted
in Fig. 10 the 77' contribution, making an "educated
guess" based on the ratio of f^/f^
just derived,
^0.74Xir
Whatever are the uncertainties in this coupling, a qualitative difference between the SU(3) octet and the singlet
correlators is obvious. Even if the singlet I I ( x ) / n f r e e ( x )
is flat up to x ~ 0.5 fm, the splitting between them seems
to begin at *spiitting ~ 0 . 2 fm. In the whole interval of intermediate distances x = 0 . 3 - 1 . 5 fm, the singlet correlation function is about one order of magnitude smaller
than the 77 correlator.
Now we switch to another interesting subject: the
correlation functions of the pseudoscalar gluonic operators. Generally, we know very little about them; we do
not even have reliable experimental information about
glueball masses. Heated discussions on whether particular hadronic resonances are glueballs take place at specialized conferences on hadronic spectroscopy, and we
cannot go into this question here.
Let us make only a general comment that all glueball
candidates are rather heavy, with masses in the region
1.5-2 GeV. This is qualitatively consistent with L G T

100 z

(2.51)

In order to evaluate the absolute magnitude of all these

matrix elements, Novikov et aL (1980) proposed to
sandwich the anomaly relation (2.47) between the vacuum and the 77 state. If the latter is an ideal member of
the SU(3) octet, 25 it should vanish because the current is
an SU(3) singlet. Thus one should have an exact cancellation between the 0(ms) and anomalous parts, implying

15

'

\y

_
-

71

'/

10 -

V
:

I It r- - ^
7)'

16IT:

(0\GG\y)

=
=

-2ims(0\sy5s\r])

(i) 17 V> 2 .
r 3 Nl/2

22

(2.52)

Others are suppressed by powers of mc.

These numbers are from Hernandez et aL (1990).
24
Actually, this is an even lower limit of the matrix element,
since the decay ratio branching into yrji 1440) that we use actually contains the branching ratio of rj( 1440)>KKIT.
25
In fact, the 77 77' mixing angle is approximately
-20 and corrections are very small, Oi^mixing)23

Rev. Mod. Phys., Vol. 65, No. 1, January 1993

"^

0.1 b-

0.01

1
0.5

1
1.5

x(fm)

FIG. 10. Normalized pseudoscalar correlation functions vs distance x (in fm). The three solid lines show the 7r,K,rj channels,
while the dashed line corresponds to the contribution of the 77'
meson into the SU(3) singlet correlator.

Edward V. Shuryak: Correlation functions in the QCD vacuum

16

(lattice gauge theory) calculations [see reviews in Lattice

88 (1989), Lattice 89 (1990), Lattice 90 (1991), and
Teraflop (1992)]. Quenched calculations also suggest that
the lightest glueball is the scalar, with a mass of about
1.3-1.5 GeV, while pseudoscalars are about twice as
heavy. Of course, results may be modified in calculations
going beyond the quenched approximation.
The general question of why all glueballs have completely different mass scale, distinct from those typical of
hadrons
made
of
quarks,
remains
essentially
unanswered. 2 6
The really relevant question is the contribution of various hadronic states to gluonic correlation functions, independent of whether or not we call them glueballs.
From our discussion above we obtained several matrix
elements of the pseudoscalar gluonic operator. Our estimates discussed above lead to
< 0 | G G | - 7 > 0 . 9 GeV 3 , iO\GG\^)^2.2
< 0 | G G | T ? ( 1 4 4 0 ) > 2 . 9 GeV 3 ,

GeV 3 ,
(2.53)

where we have also taken a to be "frozen" at as 0 . 3 . It

is tempting to examine the contribution of these three
states to the pseudoscalar gluonic correlation function.
As before, in order to get an idea of whether the matrix
elements obtained are large or small, it is instructive to
normalize this contribution to the asymptotically free
gluonic contribution, which is equal to
K0(x)=(0\GG(x)GG(0)\0)

9(N2 1)
%-j .

(2.54)

31,,,,.,,,,J,,,,.,j,

x(fm)

FIG. 11. Normalized pseudoscalar gluonic correlation function

to that corresponding to the propagation of two free gluons.
Three curves correspond to the contributions of 77,77', 7/(1440)
mesons, respectively.
states" contribute roughly an amount that causes the
K(x)/KfreQ(x)
ratio to level off at 1 for x < \ fm. If so,
the threshold E0 is expected to be rather high, of the order 2 GeV or so. In principle, one can tell whether it is
true or not from studies of the radiative decay
Yy-hhadrons(s), in which hadronic systems with corresponding invariant mass are produced. Moreover, the
local annihilation hypothesis is even better fulfilled here
than for charmed quarks.

IT X

This equation is derived by propagating two gluons from

point 0 to x. The x dependence is obvious, since the
gluon operator GG has mass dimension 4. Apart from
the color factor, the formula is the same as in quantum
electrodynamics.
The estimated contributions of the if s to the gluon
correlator are shown in Fig. 11. We see that the 7]r and
7/(1440) matrix elements found above are indeed comparable to the perturbative ones already at distances as
small as \ fm. Moreover, they become about an order of
magnitude larger at only slightly larger distances.
It is amusing to note that the 77' and 7?(1440) together
contribute to the gluonic pseudoscalar correlator in a
way very similar to the ^r,K,rj contributions to quark
pseudoscalar current. The general tendency of a rapid
rise suggests a strong attraction in this channel, starting
at about the same distances.
One may further speculate that the "true gluonic

26
In fact, in the interacting instanton approximation the
difference in mass scales is quite natural. In the IIA the quark
and the gluon fields have completely different roles and different
distribution in space-time. The former are distributed more or
less homogeneously, while glue is concentrated in small spots of
the strong field, the instantons.

Rev. Mod. Phys., Vol. 65, No. 1, January 1993

H. General properties of the scalar correlators

We conclude our survey of the phenomenology of the
Q C D correlation functions with some remarks about the
scalar correlation functions.
From the phenomenological side the situation is far
from clear. Historically, the first candidate for scalar
mesons was the famous enhancement seen in the isoscalar
TTTT scattering near 500 MeV, known in literature as the
"sigma meson." This name was also used in the "sigma
el" (Gell-Mann and Levi, 1960), in which the scalar particle is essentially the "radial" (1=0) oscillation of the
quark condensate. It is not recognized as resonance, but
still can be strongly coupled to the scalar 7 = 0 current.
The next scalar mesons are isovector and isoscalar
pairs of particles, / 0 ( 9 7 5 ) and a 0 (980). Their close
masses and particular decay modes have led to the suspicion that they are not regular qq mesons, but rather
four-body qqqq mesons containing "intrinsic strangeness." 2 7 This latter observation makes it very improbable
that they play any role in the spectral density of nonstrange quark currents uu, dd.

27
The interested reader can consult the proceedings of any
conference on hadronic spectroscopy, where this topic is repeatedly discussed.

Edward V. Shuryak: Correlation functions in the QCD vacuum

The 1=0 scalar channel has two more resonances listed in the Review of Particle Properties, the / 0 ( 1400) and
/ 0 ( 1 5 9 0 ) . The former decays predominantly into two
pions and can therefore be plausibly assigned as a nonstrange qq meson. 28
The f0( 1590) was produced diffractively in one experiment only, and it has a very interesting dominant mode,
rj'rj, in spite of the fact that it is very much suppressed by
small phase space. Taking into account its production
mechanism and specific decay pattern, with a "gluonie
touch," we see that it is a good candidate for scalar
gluonium. Its mass value also fits well with what
quenched lattice data tell us (see, e.g., Teraflop, 1992).
Unfortunately, there are problems with this interpretation o f / 0 ( 1590): in particular, there are strong experimental limitations on J/ip>y -i-rfr}. If correct, they imply that this particle cannot have a sufficiently strong
coupling to the gluonic operator ( 0 | G 2 | / 0 ( 1590)).
N o isovector scalar resonances (other than aQ mentioned) are in the Review of Particle Properties; so one has
to conclude that such mesons probably do not exist, or
they are too heavy and wide.
Let us now discuss the qualitative behavior of
scalar correlation functions. In the 7 = 1 channel,
K(x)/Kfree(x)
should strongly fall off with xy because the
lowest intermediate state has a mass of at least 1 GeV or
more. In the 1 = 0 channel the situation is different, because the correlation function
Ksc&iaTfJ=oM=(qq(x)qq(0))

(2.55)
1

possesses the factorizable contribution, (qq) ,

which
does not fall off at large distances. Large mesonic masses
in this case mean that transition to this region should be
rather sharp.
Our point now is that it is possible to guess at what
distances this transition takes place just by comparison of
perturbative contributions. In terms of the ratio we usually use, K(x)/Kfree,
it means a rapid increase starting
from the point where
^scalar,/=oU)/^free^)==^4<#>^6/3-l

(2.56)

This estimate tells us that this curve probably turns up

starting from rather small distances, about j fm (which is
again related to and from a rather large magnitude of the
quark condensate).
Summarizing, we have two important observations:
one expects the curve for K(x)/Kfree
(a) to curve up in
the 1 = 0 [or the SU(3) singlet] scalar case, but (b) to
curve down in the 1 = 1 [or the SU(3) octet] case. In other terms, one expects the existence of some attraction in
the singlet and a repulsion in the octet channel. Let us
now compare these conclusions with the behavior of the
pseudoscalar channels. Note that their behavior is exact-

28

However, the sigma meson of the sigma model (Gell-Mann

and Levi, 1960) should be much wider at this mass.
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

17

ly the opposite: a similar ratio (c) goes up for the octet

(7r9K9r]), but (d) goes down for the singlet (77') case.
Let us reformulate these four statements in terms of
somewhat different correlation functions. Considering
for simplicity u, d quarks only, we define instead of the
four previous correlation functions, scalar and pseudoscalar with 1 = 0 and 1, the following linear combinations:
K + +=uLuRuRuL~\-dLdRdRdL

(2.57)

K + ._=uLuRuLuR+dRdLdRdL

(2.58)

K-+=uLuRdRdL+dLdRuRuL

(2.59)

(2.60)

uLuRdLdR tuRuLdRdL

Here L,R stand for left and right chirality. The notations are as follows: the first index here corresponds to
flavor, the second to chirality, ( + ) means this quark
property remains unchanged, ( ) means it is changed.
At small distances the dominant contribution comes
from free-quark propagation, which corresponds to dominant J + + .
Based on the discussion above of both scalars and
pseudoscalar correlators, with 7 = 0 , 1 , one may reach
two important conclusions: (1) The qualitative behavior
of those correlation functions is consistent with the assumptions that the dominant term producing splitting in
parity and isospin is K
; and (2) deviations from
asymptotic freedom are much more radical than those in
vector and axial channels, and they show up at much
smaller distances, x \ - \ fm.
Consequences of these observations will be discussed in
the next section, and we note here only that the K
amplitude corresponds exactly to the quantum numbers
of the instanton-induced 't Hooft interaction.

III. THEORY OF MESONIC

CORRELATION FUNCTIONS
A. Potential models and heavy quarkonia
This section is somewhat separate from the others, because it applies to the physics of heavy quarks only. We
have included it mainly for pedagogical reasons: here
one can use simple nonrelativistic language based on the
interaction potential between quarks, which, we hope,
will make the discussion clear.
Our main goal is to show how studies of the correlation functions may help to reveal information that is
nearly impossible to get from an analysis of stationary
states. This discussion is based on a paper (Shuryak and
Zhirov, 1987) that attempted to find experimental evidence for a strong Coulomb law.
Very heavy quarks and antiquarks form nonrelativistic
bound states similar to positronium, with the interaction
described by a Coulomb-type potential (Appelquist and
Politzer, 1975). The force is as fundamental as a

Edward V. Shuryak: Correlation functions in the QCD vacuum

18

Coulomb law of electrodynamics or the Newton law of

gravity; so it is certainly worth trying to measure it more
precisely. In fact, Q C D does not predict exactly a
Coulomb law, because of the running coupling constant,
which effectively depends on the distance between
quarks. The equation derived in Appelquist and Politzer
(1975) for the potential is

4 <*,(*>

^QCD

F=6.87tf

+const ,

~*

(3.2)

(3.3)

where in the last two formulas all units are GeV or inverse GeV. Both potentials give about equally good
descriptions of all states in t h e J/if; and T families.
However, the Martin potential has no Coulomb term at
all! From this experts in quarkonium spectroscopy have
concluded that there is not yet any direct evidence for a
strong Coulomb law.
Now comes the main idea: if the stationary states,
J /if) and T mesons, are not small enough t o be a
Coulomb system, why not consider a virtual system, a
wave packet of any desirable size? In particular, one can
discuss a correlation function in which quarks propagate
any distance (or Euclidean time) we want.
As for the light quarks already considered, these correlation functions can be recalculated from experimental
data on e+e~ annihilation into heavy quarks. What is
important is that these data contain not only resonances
(the upsilons), but also a continuum of excited states
above t h e heavy quark-antiquark threshold. Therefore
one can obtain information not only about lowest bound
states, but also about the unbound (or scattering) states.

29

Some recent work done in the direction of fixing the scale

from charmonium physics can be found in Mackenzie (1991).
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

U2

1-As)D{s

e e~^b(s)^^(l27T2rr/Mr)d(s

and the Cornell potential (Eichten et al., 1980)

V= - 0 . 5 2 / 1 * + 0 . 1 8 * ,

^j[dss2ap+-

G +

R '

where Nc and Nf are the number of colors and flavors.

Note that this potential contains a parameter A C o u l o m b .
Its measurement is crucial for setting the absolute scale
in Q C D , which is also needed for lattice calculations (discussed below). 29 Heavy quarkonium is in principle an
ideal place to measure the Q C D scale, because the potential is perturbative in nature but still produces large observable effects.
Unfortunately, neither c nor even b quarks are heavy
enough for this simple idea to be applicable. However,
these mesons are very well described by an effective potential Vefr(r), a combination of confining and Coulomb
forces. To be specific, let us consider two potentials used:
the Martin potential (Martin, 1981)
01

K(r) =

yr)

(3.4)

We evaluate this using the following equation for the

cross section,

(llNc~2Nf)ln[l/(RACoulomb)]

Realizing all this, let us consider a correlator of two

vector currents made of b quarks placed at the same spatial point and separated by the Euclidean time r . It is
connected to the experimentally measurable cross section
mentioned above by the following formula 30 :

~M\)

Ts

+ (47ra2/3s)Rb6(s-s0)

(3.5)

The partial widths and masses of the four upsilon states

are taken from Particle Data Table; s0 is the threshold of
the open beauty production, (2mB )2; and the constant Rb
is taken from the averaged data to be Rb= 0.31 0 . 0 6 ,
consistent with the free-quark value Rb =j just above the
threshold. Putting all this into the equation above, we
obtain the "experimental" correlation function plotted in
Fig. 12. This shows the logarithmic derivative

F(r)=--fln-^-,

(3.6)

where ^f r ee( r ) corresponds to free propagation of the bb

pair. 31 The function K(r) decays very strongly, due to
factors like exp( 2mbr); so it is more informative to
plot the logarithm. By definition, if the interaction potential is absent, F(r) is just zero. Its physical meaning
is, roughly speaking, the mean energy of the wave packet
existing during the Euclidean time period r .
The function F(r) obtained in this way is shown in
Fig. 12 by the dashed region (representing experimental
error bars). Although the experimental accuracy is not
very good, one can see that this function is decreasing toward the small r , which, of course, means that some attraction is present at small distances. Thus one does observe a manifestation of the strong Coulomb law.
At this point we are finished with our phenomenological input and come to theoretical predictions. In the
nonrelativistic case, with a potential-type interaction
V(R), this can be done just by solving t h e ordinary
Schrodinger equation for the Green's function with this
potential. Another practical way to do it (Shuryak and
Zhirov, 1987) is based on the equation
A

*{r)

free^T>

=(exp[-JrfrK[lrg(r)-rg(r)l]|)
\

**

J / free paths

(3.7)

30
In fact, in the nonrelativistic domain under consideration,
rm 1, the propagator can be taken in the nonrelativistic limit, D(M,r)~Af 1/2 T- 3/2 exp( -Mr).
31
The b quark mass was taken to be 4.9 GeV.

Edward V. Shuryak: Correlation functions in the QCD vacuum

! , , , !

,,,

,, ,, ^

ii

ii

r
V

^Martin T
>\
#

\- j'j
"

_
'

Cornell

that essentially the same potentials as those used above

for heavy quarks produce a reasonable description of the
overall spectroscopy of the strange and even of the light
mesons and baryons (Capstick, Godfrey, Isgur, and Paton, 1986). It is only necessary to introduce a phenomenological "constituent" quark mass for the light quarks.
It would be interesting to know whether such an approach could reproduce the correlation functions associated with the light quarks, especially at smaller distances.

19

r (1/GeV)

FIG. 12. F{r)=-d\og(K/K{ree)/dr,

where K(r) is the vector
bb (or T) correlation function, ^ free is its version corresponding
to free propagation of b quarks with a mass mb =4.9 GeV, and
r is Euclidean time (in GeV" 1 ). The error bars show F(r) as
derived from experimental data (see text). The more negative
values of this quantity at small r correspond to stronger
Coulomb forces at smaller distances between quarks. The solid
line corresponds to the Cornell potential, and the dashed line to
the Martin one.

Here one averages the exponential "interaction factor"

over an ensemble of quantum paths, corresponding to the
motion of free quarks.
We have calculated this correlator using the two phenomenological potentials given above. The resulting
curves are also shown in Fig. 12. Although the experimental accuracy is not really good enough to make a
conclusion, it appears that these data show some preference for the Cornell potential over the Martin one. Improvement in the quality of the data, especially in the
measurement of the actual shape of the nonresonance
continuum containing a pair of 6-flavored hadrons, can
clarify this important issue.
Perspectives of Tt spectroscopy is an interesting subject, which we now address briefly. Because the t quark
mass is at least 100-150 GeV or more, its weak decay is
too rapid to allow the Coulomb bound states to show up
as a set of narrow resonances. However, even if the
separate toponium states cannot be seen as well-separated
peaks, by integrating the corresponding cross section as
above and calculating the correlation function, one may
still observe its derivation from that expected for freemoving t quarks and detect a trace of Coulomb-induced
effects. This may provide a method to measure A Q C D .
We conclude this section with a question. It is known
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

B. Operator product expansion

and QCD sum rules
In this section we turn from the simple potential models to a much more complicated approach, that of applying the operator product expansion to correlation functions at small distances. We present only a few important examples of its applications, but actually about a
hundred papers have been written on this topic, and it
has been reviewed by Novikov et ah (1982, 1984),
Shuryak (1984, 1988a), Reinders et ah (1985), and Shifman(1992).
The general idea (Wilson, 1969) is an expansion of the
bilocal operator
j(x)j(0)

= JtCn(x)On(0)
n

(3.8)

in terms of local operators. Here the Cn(x)

are
coefficients, depending on the distance between the
points, and the On(0) are operators. If one deals with ordinary functions, one might think of a Taylor-series expansion of J(x) in powers of x, but the dependence on x
is quite singular in quantum field theory. However, in
massless Q C D the powers of x are just determined by dimensional arguments, except for some nontrivial powers
ofrn(x).
A formal definition of the O P E is based on a separation between the high-momentum and soft momentum
modes of the quantum fields involved. The operators On
contain a cutoff \x so that they only couple to soft modes,
with Euclidean momentum /? 2 </x 2 . The coefficients Cn
absorb all hard modes of the fields, i.e., those with momenta p2 > /x2. If the value of/x is changed, the whole expansion is redefined, however, the sum remains the same,
because fi is just an artificial parameter without any
physical significance.
There are no averaging symbols in this equation; the
expansion is assumed to be valid for any matrix element
of this equation. In particular, we may average it over
the vacuum state, of course, but the equality should hold
for any configuration of the fields separately. We cannot
go deeper into the theoretical discussion here, but refer
those interested to a recent compilation of main papers
(Shifman, 1992), which also has further references.
Two technical points should be noted here. First, since
unrenormalized Q C D with massless quarks has no di-

Edward V. Shuryak: Correlation functions in the QCD vacuum

20

mensional parameters, the coefficients Cn(x) are just

powers of x, depending on the dimension of the corresponding operator On. For example, if one deals with the
product of two currents, one has an object of total mass
dimension 6. Suppose one is interested in the coefficient
of gluonic operator 0 = (G V ) 2 , which has dimension 4.
Without calculations, one sees that the corresponding
coefficient must have an x dependence
CG2(X)~\/X2,
Our second remark is that, in practice, people have
used a somewhat different form of the OPE, namely, the
expansion of the Fourier transform of the correlators
Kmom(Q2) in powers of l/Q2, where Q is the momentum
transfer. Roughly speaking, large Q corresponds to small
xy but not exactly. Suppose, for example, one has a term
proportional to (l/Q2)n
and takes its Fourier transform.
Then, for n > 2, one gets the power of x2 dictated by
naive dimensional counting times the ln(x 2 ). In fact, this
logarithm is present because all contributions that are
regular at x=0, i.e., proportional to (x2)n without logarithms, are missing in this approach. 3 3
The existence of terms regular at x = 0 is one of the
reasons why people in the past avoided the use of coordinate representation. We return to this point at the end of
this section.
Let us first show some examples of how the Shifman,
Vainshtein, and Zakharov (SVZ) approach works in the
space-time representation. Derivation of the formulas
can be found in the original papers (Shifman, Vainshtein,
and Zakharov, 1979b) or in reviews (e.g., Reinders et al.,
1985; Shuryak, 1988a; Shifman, 1992). For clarity, we
omit some terms that are, in practice, unimportant, like
the mqqq operators. We also did not include the lengthy
expressions for higher-dimension operators, because they
are not actually used in applications.
In Euclidean time r, which is the same as the spatial
distance x used before, the normalized correlation functions for p and A x channels are given by (Shifman et al.,
1979b):
P

, A t1

Ka

W*

<(gG*)2)r4

as(r)

(r)/n22c(r)=l +
A*/*

^ ~

j;.

3X2

AJ>

16

rfi

p,Al

(3.9)
The complicated four-fermionic

operators OpA

are

by
P = ^Y-(ur^stau

-dy^5tad

+ ^^r^au+d7^tad)

)2

l^qr^q)

o.io)

0^i=Op + 27ras(i7Lr^X-^LV^L)
^^RY^auR-dRr^tadR)

(3.11)

where R,L denote right- and left-hand polarization on

the quarks.
Estimates of vacuum expectation values of the above
operators were made using the so-called vacuum dominance hypothesis (Shifman et al., 1979b). 34 The recipe is
as follows: one should try to transform this operator into
a product of two scalars, and just include the vacuum
state in the sum over intermediate states between the two
scalar operators. The vacuum expectation of the scalar
will be nonzero if there is a quark condensate, and it is
evaluated as such.
For the operators mentioned above the answer is (Shifman et al, 1979b)
<0/9)(7X2V/34)a5<^)2 ,

(3.12)

<0^>-(2V/34)as<^>2 ,

(3.13)

where (ifsxp) is the quark condensate. Note that these

two expressions have opposite signs.
We can see how well this works in Fig. 13. The phenomenological correlation functions of the previous section are shown by the solid curves, and the O P E predictions by dashed curves. The curves marked p,SVZ and
Al9SVZ
correspond to the above expression, and, in
particular, the short-dashed line shows the O P E prediction, including perturbative and gluon condensate corrections. One can see that the general behavior of these
correlation functions is reproduced surprisingly well up
to distances of about \ fm. In particular, (1) the splitting
between the vector and axial 1=1 channels happens exactly in the right place, and (2) the magnitude of the
splitting is also correct. Both observations show that the
estimates of the vacuum expectation values of these two
four-fermion operators are probably reliable. 35 As a
third point, note that in the vector case the quark and
gluon corrections nearly compensate each other, so that

different for the vector and axial channels and are given

34

32

Radiative corrections produce terms containing a dimensional parameter AQ C D , but only in the form of some powers of
ln(xA QCD ), the so-called anomalous dimensions. Since these
complications are not very important for our discussion, we
shall not introduce them, for the sake of simplicity of presentation.
33
Their Fourier transform is KmQm{Q2)~exp( QXconst),
and such terms are more difficult to trace.
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

The accuracy of this estimate is an interesting question that

can be addressed with lattice data, the IIA, and other vacuum
models. The IIA (Shuryak, 1989b) strongly contradicts the vacuum dominance hypothesis for gluonic operators, but more or
less agrees with it for the operators in Eqs. (3.10) and (3.11).
35
The general discrepancy between OPE curves and experimental ones, about 10% in absolute normalization, may well be
due to higher-order radiative corrections. They are also comparable to the experimental uncertainties in the axial channel.

Edward V. Shuryak: Correlation functions in the QCD vacuum

i

n
i

/!
/
P,VD / p.svz
/ /
x-^ ^
/"
^--^

l.3h-

~~

I.2

"

^ - ^\ ^ \ _ \ \

y/

1.3

\ANltSV2T
^ s .>v

?7,K,7T
1.2

FIG. 13. Ratio of the correlation function to that corresponding to a free-quark propagation vs distance x (fm). One solid
line shows the p contribution, and two solid lines for the A {
correspond to experimental data as in Fig. 9. Other curves are
different versions of the OPE. The short-dashed line shows the
perturbative correction and that due to the gluon condensate:
those are the same for both channels. The long-dashed lines
marked p,SVZ and AXySVZ correspond to \/Q expansion,
while the dot-dashed ones marked p, VD and A x, VD include
regular terms as well.

at least the beginning of superduality is reproduced.

Encouraged by this success, let us look at the pseudoscalar channels. The O P E expression for the pseudoscalar correlators was also given by Shifman et ah (1979b).
In coordinate representation the correlator can be written as
n7r(r)/nfree(r)=l+a5(r)/7r+<(gG^v)2)r4/384
(3.14)

TK

where the four-fermion operators are defined as follows:

(3.15)

+ (iras/3)[(urfltau)(2gqrfltaq)]

(daflvtad)2]
.

(3.16)

The expectation values of these operator products are

evaluated according to the vacuum dominance hypothesis, and the main one is 36
<02>(567r/27)a5<^)2

36

(3.17)

The operator Ou estimated in the same way, has a smaller

matrix element. We have separated Ox because it is the operator that obtains contribution from instanton zero modes. In the
instanton liquid model, its vacuum expectation value is actually
several times larger than that for 02 (Shuryak, 1983), but it still
does not contribute enough to make a good description of the
data.
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

///
// /

///

'

SVZH

i f

#
///

/ /
/ /
/ /
/

/ /
' '

,'''

=+G2

J''' --""

-~-^"~^sZ^^*

///

^^ ^"

0.75

-Tras{ua^tau)(dcr^d)

VD

I.I

0.5

+ (7r 2 /48)r 6 ln-4 r (0 1 +0 2 ) ,

'

x(fm)

02 = (7ras/2)[(uaflvtau)2

//

''

///

\ ^ \
A,
AltVD\
^ \
0.25

//

1.4

5--'

\ .^ s .

^V*^
', N .

0.9 h-

Ox^

"~~~-^ J X ^ o c s + G 2
-^^"^
^*

"^
~"-^-^

I.L)

-----"''''

0.8,

21

1
0.5

0.25
x(fm)

FIG. 14. Same as Fig. 13, but for the pseudoscalar channel.
The three solid lines close together correspond to the TT, K, and
77 phenomenological correlators, respectively.

The resulting curves are compared to phenomenological

ones for TT,K,7} mesons in Fig. 14. One sees that the SVZ
results have "good intentions" in the sense that the predicted behavior looks similar to the experimental trend,
but numerically these effects are too small to reproduce
the experimental data. (Remember, coming from x ~ \
fm to \ fm means correlators being reduced by
~ 2 6 = 64.)
This failure was discussed in the important paper by
Novikov et al. (1981) entitled "Are all hadrons alike?"
The suggested answer was that all spin-zero correlation
functions are special: new, important effects show up in
these cases, which are not seen in the O P E framework.
Attempts to understand how this could come about have
led to the instanton theory discussed in the next section.
However, before we go into it, we can take one more
step forward. It was noticed at the beginning of this section that the SVZ expressions ignore the part of the
correlation functions that is nonsingular at x=0.
The
natural question is what corrections can arise due to
them. In fact, the theory of regular terms is even simpler
than that of singular ones. In particular, it is not difficult
to understand that the constant term at x>0 is nothing
but the current squared: Kreguiar(x> 0) = ( / 2( 0 ) ) . One
can also easily evaluate their vacuum expectation values
in the same vacuum dominance approximation, gaining
some insight into the effect of those regular terms. Moreover, as all currents considered in this section have the
simple flavor structure ud, two quark lines do not mix
and one can obtain all 0( { ^ ) 2 ) corrections by using the
quark propagator in the form 37 5 = 5 0 + -^ (if>t[>). The re-

37
This simple way of implementing the vacuum dominance
was first used by Ioffe in treating the baryonic sum rules. We
come to them in the next section.

Edward V. Shuryak: Correlation functions in the QCD vacuum

22

suits look as follows:

n^(T)/n^e(r)=l + ^ < ^ >
n/t

(r)/n^e(r)=l-^<^)

,
2

(3.18)
,

(3.19)

nw(r)/nfree(r)=l + ^^<tAV')2

0.20)

nscalar(r)/nfree(r)=l-^-<^)2 .

(3.21)

These corrections are added to the SVZ terms discussed above and shown as the dot-dashed curves in Figs.
13 and 14 marked V D (vacuum dominance). We see that
in all channels these corrections have signs coinciding
with experimental trends. However, the quantitative
comparison is not at all satisfactory: although inclusion
of the regular corrections makes disagreement in the
pseudoscalar case somewhat smaller, it also worsens the
agreement in the vector and axial channels.
In summary, the O P E can be used in two forms: as
1/Q expansion in momentum space and x expansion in
space-time, the latter possessing extra regular terms.
Supplemented by the vacuum dominance hypothesis, it
predicts correct qualitative behavior of correlators, but it
is not able to reproduce them quantitatively.
C. Interacting instanton approximation
A detailed discussion of the theory of instantons and
related phenomena cannot be made here. We simply outline the main steps of the development of this theory
(presenting the references), then briefly consider some
qualitative features of the instanton-induced effects to
first order in the instanton-induced 't Hooft effective Lagrangian, denned in Eq. (3.22). After that, jumping over
a decade of work, we proceed directly to the particular
predictions, including all orders, in this interaction. We
end with a brief discussion of the connections between
the IIA and lattice data.
Since the discovery of the instanton solution in nonAbelian gauge theories (Belavin, Polyakov, Schwartz,
and Tyupkin, 1975), they have been believed to be an important ingredient of strong-interaction physics. Early
applications to hadronic physics are discussed in Callan,
Dashen, and Gross (1978). The theory is very elegant,
using semiclassical theory related to the physics of tunneling.
However, instanton-induced effects appeared to be
"too strong" in the following sense: only a few effects,
such as the short-distance deviation from asymptotic
freedom discussed below, could be understood from a
first-order treatment of instantons. Other properties of
Q C D , such as hadronic masses and coupling constants,
could be described only by including many instantons interacting with each other. That is why only recently has
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

it become possible to create a consistent approximation

treating the interacting instantons (IIA, for short), which
both includes 't Hooft effective interaction to all orders
and is simple enough to make calculations practical.
The main steps in the development of this theory were
as follows. Phenomenological discussions pioneered by
Geshkenbein and Ioffe (1980) and Novikov et aL (1981)
eventually led to the "instanton liquid model" (Shuryak,
1982a, 1983, through which the qualitative picture of the
ensemble of interacting pseudoparticles was significantly
clarified. Surprisingly enough, essentially the same picture emerged from the variational approach to the "instanton liquid" (Diakonov and Petrov, 1984). Encouraged by this agreement, much effort was devoted to
the study of the approximations involved in this analysis
and to the development of a more quantitative theory. A
straightforward numerical solution was obtained by the
present author (Shuryak, 1988b), who then applied the
approximation to the calculation of various mesonic
correlation functions (Shuryak, 1989a, 1989c, 1989d.
These results are discussed below.
Let us now briefly introduce the reader to some qualitative features of the instanton-induced effects. The main
physical phenomenon, a tunneling through a barrier
separating gauge fields of different topology, was shown
Ct Hooft, 1976) to be related to specific rearrangements
of the light quark states: some of them "dive into the
Dirac sea" during this process, and some others emerge
from it. Without going into details, let us only mention
that the tunneling between gauge fields is described by
the 't Hooft effective Lagrangian, which has the structure

/
Here qj is a quark field of flavor / (f = u,d,s), while t/>0 is
the so-called fermionic zero mode, a solution of the Dirac
equation Dif>0(x) = 0 in the field of the instanton. These
zero modes play the role of wave functions of quark
states, in which they are produced or absorbed during the
tunneling; they depend in a known way on collective
variables of the instanton. 3 9 Important for us is the following fact: these zero modes have chirality, directly related to the topological charge of the gauge field: there is
only a left-handed solution for the instantons and a
right-handed one for the anti-instanton. Thus the quark

38

The status of this activity is somewhat intermediate between

what is usually called a "model" and a "theory." Unlike the
"instanton liquid model" of the early '80s, now it does not contain any model-dependent elements; its statistical sum is, in
principle, directly derived from QCD. But it is also not a complete theory: it is restricted to the sum over gauge-field
configurations of a certain type, a superposition of some number of instantons and anti-instantons.
39
There are 12 collective degrees of freedom in QCD: the size
p, the four-dimensional position of the instanton center, and
seven (out of eight) color rotation angles.

Edward V. Shuryak: Correlation functions in the QCD vacuum

Lagrangian generated by a single instanton or antiinstanton has an ^effective interaction like q^qf
(or
R++L), but never ipftf (or tpftfrf). The Dirac structure
of the helicity-flip interaction contributes only to scalar
and pseudoscalar correlators. Noting also the specific
flavor-changing structure of this Lagrangian, we find that
the interaction has the following important properties:
(1) The first-order corrections in the 't Hooft effective
interaction are present in the scalar and pseudoscalar
correlators, but absent in the vector and axial ones.
(2) These corrections have the opposite sign for the
scalar and pseudoscalar channels.
(3) These corrections have the opposite sign for the isospin 1 and 0 channels.
All three points are welcomed to reconcile the
disagreements in the previous section. The first point accounts for the nature of the disagreement in the pseudoscalar case, while preserving the good agreement for the
axial and vector cases. The last two points show how
this interaction has exactly the structure of the amplitude
AT__, which was demanded at the end of Sec. II to provide a qualitative explanation of the behavior of all four
spin-zero correlators. Unfortunately, one can use the
first-order results only at small distances, where the
instanton-induced effects are small corrections to the perturbative correlation functions.
To go beyond the first-order effects, one can numerically model an ensemble of interacting instantons, using a
partition function of the form

X[det(iB+imf)]

(3.23)

Here we have denoted by dit the measure in space

of collective coordinates of the zth instanton.
Si = Sw2/g2(pi)
is the action corresponding to individual
instantons. 5 i n t describes the classical (gluonic) interaction between instantons and anti-instantons. At large
distances it is known to be a dipole interaction (Callan
et al., 1978), but at finite distances this quantity requires
a specific definition. Starting with Diakonov and Petrov
(1984), a set of trial functions of growing complexity was
used (Shuryak, 1988c). However, the most natural collective coordinates for the instanton-anti-instanton
problem can be obtained by "descending down the valley" (Shuryak, 1988d), solving the so-called streamline
equation (Yung, 1988). This was done for gauge fields in
Verbaarschot (1991), where one can find a detailed discussion of this interaction. 4 0
The last (and the most complicated) factor in (3.23) is
the so-called fermionic determinant, which describes

40

Recently this interaction has attracted much attention in

connection with possible baryon number violation by weak interaction (see Khoze and Ringwald, 1991; Shuryak and Verbaarschot, 1992).
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

23

quark-induced interactions. It is evaluated by division

into two terms, that due to zero and nonzero modes. The
former can be written as the NXN matrix in the socalled zero-mode subspace, where N is the number of instantons (and anti-instantons) considered. Its determinant is evaluated directly for each configuration,
which is equivalent to inclusion of all diagrams in the 't
Hooft interaction to Nth order.
This system is somehow more complicated than the
traditional systems considered in statistical mechanics:
the fermion determinant induces a very nonlocal interaction. Therefore in simulations one has to deal with about
2 0 - 6 0 instantons. The question of whether or not chiral
symmetry is broken becomes a matter of calculation,
similar to the question of whether a collection of atoms
behaves as a conductor or an insulator. The situation is
also complicated by the fact that the ensemble of instantons is not "frozen" into a periodic structure, but
remains liquid-like. However, this problem still is enormously simpler than lattice gauge theory. In fact, one
needs only about 10 parameters to describe field in the
volume 1 fm4, instead of 105 or more used for the same
purpose in current lattice calculations. We also have
some evidence that these variables are in fact the main
ones (see below).
After this brief introduction, a sample of results is in
order. In Fig. 15 the 1=1 mesonic correlation functions
presented in Shuryak (1989a, 1989b, 1989c, 1989d) are
shown, in the form of TltM/Yl^ix).
Qualitative behavior of all correlators agrees well with our discussion. In
particular, there is strong attraction in the octet pseudoscalar case, causing the curve to go up starting from rath-

1.4
1.2

0.8

\,

FIG. 15. Ratio of the various correlation functions to that corresponding to the free-quark propagation vs distance xAPV
(normalized to the Pauli-Villars APV parameter, roughly in fm).
The points correspond to the calculation in the IIA framework
(Shuryak, 1989a) for scalar (S), pseudoscalar (PS), vector (V),
and axial (A) channels with the flavor structure udyus (closed
and open points, respectively). The dashed curves are the
three-parameter fit described in the text.

24

Edward V. Shuryak: Correlation functions in the QCD vacuum

er small distances. There is very good superduality in the

vector channels. N o resonance, it is amusing t o note, is
needed to fit the scalar channel: the correlator found is
consistent with just a continuum threshold at sufficiently
high masses. This prediction is intriguing in its correlation with phenomenology.
Moreover, t h e results are in quantitative agreement
with the data. The dashed curves are the three parameter fits for the data points, with resonance mass m, its
coupling, / , and continuum threshold E0 obtained for
each channel. 41 Further studies of a wide set of correlation functions in the IIA are now in progress.
Let us now briefly discuss connections between IIA
and L G T . In order to see the instantons in L G T , one
must "cool" the lattice configurations of gauge fields by
damping out t h e quantum fluctuations. This leaves a
configuration of only classical gauge fields, which can be
examined. Kremer et al. (1988) have shown that t h e
classical fields remaining after cooling are essentially the
instantons, and their density is roughly comparable to
what is used in the IIA. Second, these instantons are related to fermionic zero modes, which strongly contribute
to the quark condensate (Hands and Teper, 1990). Third,
it was recently found (Chu and Huang, 1991) that cooling
affects the correlation functions surprisingly weakly; so
hadronic masses and coupling constants determined from
the full Q C D and from the cooled, instanton-dominated
configurations are practically the same. However, these
studies are in their initial stages and the accuracy of the
numerical results should be improved. Further references on studying the IIA with lattice methods may be
found in the recent lattice conference proceedings [Lattice 90 (1991); Teraflop 1992)].
We conclude this section by listing the main phenomena that can be explained using the IIA:
(1) We found that the IIA accounts for the chiral symmetry breaking of the Q C D vacuum, including the value
of the quark condensate (Shuryak, 1989b) at T = 0, as
well as for the chiral symmetry restoration at higher temperatures (Ilgenfritz and Shuryak, 1989).
(2) As will be discussed in Sees. IV.A and IV.B, t h e
IIA produces an effective mass for the quark of about
300-400 MeV. This is the constituent quark, the main
building block of all hadronic masses.
(3) The IIA gives the absolute magnitude and quantum
numbers of qq effective interaction at small distances,
especially in the scalar and pseudoscalar channels.
(4) The IIA reproduces some delicate effects, such as
the superduality in the vector channels or the absence of
isovector scalar resonances.
It perhaps can account for some other phenomena discussed in literature, such as (5) spin splittings of baryons,
especially explaining why the nucleon is so light (see Sec.
IV.D); (6) part of the NN potential and the experimental

41 A

table of the numbers is given in Shuryak (1989a).

Rev. Mod. Phys., Vol. 65, No. 1, January 1993

absence of a well-bound dihyperon H (Takeuchi and Oka,

1991); and (7) the suppression of the singlet axial charge
of the proton (known as the "spin crisis") or, in other
words, the strong polarization of the gluonic field in the
polarized protons (Forte and Shuryak, 1992).
However, the IIA certainly does not explain the important Q C D phenomenon of quark confinement. This
may be seen from the fact that in the "instanton liquid"
the static potential between a pair of heavy quarks tends
at large distances toward a constant (Shuryak, 1989d). 42
IV. OTHER CORRELATION FUNCTIONS
A. Light-and-heavy mesons
The central idea of our discussion has been that the
correlation functions have quite different properties depending on the quantum numbers of the channel considered, which means that a realistic qq interaction is
rather complicated and that by no means does it reduce
to a simple universal confining potential. In this section
we shall continue these studies, considering now the qq
interactions.
However, before addressing the issue of interquark interaction in two- (or three-) body systems, it would be
logical to study first the propagation in the Q C D vacuum
of a single quark. However, as the quark propagator is
not a gauge-invariant quantity, it cannot by itself have
any physical meaning; one has either to fix the gauge, or,
following Schwinger, to switch from the ordinary propagator S to its gauge-invariant version S:
S(x)=($(x)Pexp \ug/2)fXA^adr}tf;(0))

(4.1)

where ta are Gell-Mann matrices of the color group, and

the path-ordered exponent contains an integral to be taken over the straight line, going from 0 to x. In other
words, one can simply supplement our light dynamical
quark with a static, very heavy antiquark, considering a
heavy-light meson instead of a single quark. Of course,
this now changes the long-distance physics: t h e light
quark becomes bound by the confining potential to the
heavy one. However, at the small distances we are going
to be studying, these confining forces are not crucial, and
the static charge of the heavy quark produces only small

42

The reader may be puzzled about how an approximation can

produce correct hadronic masses without this important piece
of physics. Instanton-induced forces are strong enough to make
effective masses of quarks and to bind them together into
mesons and, perhaps, even baryons. However, even if those
have reasonable parameters, there also should be some false
states related to free propagation of "dressed" quarks. So far
the behavior of the correlation functions in IIA has not revealed this unphysical component, which probably means that
the false states do not contribute very much.

Edward V. Shuryak: Correlation functions in the QCD vacuum

(and calculable) corrections. This is the line of reasoning
that has led us to study such mesons. In a sense, it is the
simplest system with light quarks, so to say, a hydrogen
atom of hadronic physics.
Certainly, there is some empirical information about
charmed and bottom mesons. For simplicity we concentrate on the simplest case, in which the light quark mass
is put to zero and that of the heavy one is considered
large on the hadronic scale: mq A , m g A . Discussion of the corresponding correlation functions in the
O P E framework was started in Shuryak (1982b). In a
more general context, discussion of some effective theory
excluding the heavy quarks was initiated by Isgur and
Weise (1991) and applied to various decay form factors,
etc., which can be related to multipoint correlation functions.
Returning to our main object, the two-point correlators, we shall discuss the correlation functions of the following type,
^(r)=(G(r)r^(r)?(0)r/+e(0)) ,

= m(J7r)-mg

Using the values of s, c, and b quark masses from the

TABLE II. Masses of some heavy-light mesons (in MeV), according to Particle Data Table.
Jp

0~

1"

1+

0+

Ju
cu
bu

497
1865
5278

892
2007
5327

1270,1400
2422

1430

Rev. Mod. Phys., Vol. 65, No. 1, January 1993

Q C D sum-rule analysis m s 1 5 0 MeV, m c 1 2 5 0 MeV,

and m ^ 4 8 0 0 MeV (Shifman, 1986) and assuming that
the spin splitting is caused by the S ^ - t y p e interaction,
we have the following equation for the spin-averaged excess energies of the negative- and the positive-parity
states
E = }E(l)

+ }E(0)

(4.3)

where 1,0 label spin. Using the experimental numbers,

one concludes that the spin-averaged P = 1 state is
separated from the quark mass by i s ~ 6 4 0 , 520, and
475 MeV for s, c, and b quarks, respectively. Splitting of
the two parity states 5E=E + E~ is roughly 600 and
450 MeV for s and c cases. From this we conclude that
for the superheavy mesons the excess energies are probably
" " - 4 5 0 MeV , + ~ 9 0 0 MeV .

(4.4)

(4.2)

where Tt is some gamma matrix, r is the Euclidean time

difference between the two points, and Q,q stand for the
field of heavy (light) quarks. We assume that the heavy
antiquark just stays at x = 0 all the time. All energies are
naturally measured relative to the heavy quark mass MQ .
Additional simplification due to the large MQ limit is the
absence of a spin splitting: the spin direction of the superheavy quark cannot be of any importance. Thus the
pseudoscalar and the vector mesons are degenerate, as
well as the scalar and the axial ones. Therefore, we consider only the splitting in parity P, negative () for the
first two cases and positive ( + ) for the two latter ones,
denoting these correlators by K_(r) and K+(T).
These
functions were evaluated numerically both on the lattice
(Boucaud et al., 1988; Bernard, Heard, Labrenz, and
Soni, 1992) and in H A (Shuryak, 1989d; see below).
Experimental data on heavy-flavored mesons are still
very incomplete. Some masses of flavored mesons have
been compiled in Table II. We included strange mesons
as well as charmed and bottom mesons, although the s
quarks and probably even the c quarks are not sufficiently
heavy to have the above approximation applied to them.
However, the data are useful for showing the trends. We
express these masses in terms of the excess energy above
the heavy quark mass,
E(Jn

25

For the small-distance behavior of the corresponding

correlation functions, we see that the trivial limit is given
by the product of free propagators
^free(r)

= Tr[5free(r)r5free(r)r]

( 4 5 )

where
S f e e ( T ) = - r 0 / ( 2 7 r V ) , S e e ( r ) ~ ( l + r o ) (4.6)
Here we have dropped all unimportant factors in the
heavy quark propagator.
As before, we consider the ratio of the true correlation
function to the free one, R(r)=K(r)/Kfree(r).
First, for
the O P E coefficient of the leading operator, the light
quark propagator is modified due to the presence of the
nonzero quark condensate as follows (Shuryak, 1982b),
Sg(T)=-Y0/(2ir2T3)+(qq)/l2+

(4.7)

This yields
R(r)=l-()^~\<qq)\+0(r5) .
(4.8)
o
In this equation, the stands for the parity of the
state considered, and we use the modulus of the quark
condensate to avoid any sign confusion. Thus one can
see that the nonzero quark condensate naturally produces the splitting of the correlation functions with the
opposite parity. The O P E suggests a simple symmetry
splitting of the two correlation functions considered: the
odd-parity correlator curves up, which means these
mesons are lighter, and the even-parity one curves down,
which means these mesons are heavier. These predictions certainly agree with phenomenology.
Following the usual O P E descriptions, some further
terms were evaluated and the resulting correlator was fit
(Shuryak, 1982b) with the usual parametrization of the

Edward V. Shuryak: Correlation functions in the QCD vacuum

26

evaluate both P = + l correlators and consider their ratio:

then the mass difference between the two lowest mesons
can be calculated without this problem.
Finally, let us turn to the calculations of these correlation functions in the IIA (Shuryak, 1989d). Omitting all
details, we come directly to the results shown in Fig. 17.
Although the splitting between two parity channels still
is significant, the whole picture looks quite different from
the symmetric 1 + const X r 3 behavior suggested by the
OPE. The solid lines are again the three-parameter fit by
the standard equation (4.12), but now with the following
values of the parameters:
- = ( 2 + 0 . 4 ) A , EQ = 3 . 2 A ,
FIG. 16. Correlation functions for the i?-type meson (arbitrary
units) vs Euclidean time t (in lattice units <z), evaluated in
(Maiani, Martinelli, and Sachrajda, 1992).

physical spectral density 43

ImK(E)=f2TesMQ6(E-ET{

,) +

{lE2/Tr)0{E-E0)

= (4.5+l)A,

(4.12)

E=5.5A,

(4.13)

where A is the Q C D scale parameter. With the experimental values A = 150-200 MeV it gives E~ = 330-440
MeV and E + E ~ = 350-460 MeV, in reasonable agreement with the empirical energy excesses.
B. Does the constituent quark model make sense?

(4.9)
The fit Eres corresponds to the meson mass; the corresponding excess masses found were
E~ = 4 0 0 + 1 0 0 MeV ,

(4.10)

E + -E ~ = 800+250 MeV

(4.11)

Comparing this with the data discussed earlier, one can

see that the E~ was evaluated correctly, while the splitting E* E~ was overestimated, roughly by about a factor of 2.
Let us now present an example of point-to-point correlation functions measured on the lattice. In Fig. 16 one
can see a set of early data by Boucaud et al. (1988) corresponding to the 5-type (which means pseudoscalar lightand-heavy quark) meson. 44
At small distances (r\a 3a) they are in a reasonable
agreement with a free propagator; but at large distances
they fall with distance very rapidly, and one finds the excess energy of the lowest state E~~l
GeV, about twice
larger than expected phenomenologically. The reason is
not difficult to trace: the pointlike superheavy quark has
a divergent self-energy, SMQ ~l/a,
and one should subtract this unphysical quantity. Another way out is to

The nonrelativistic quark model of the '60s suggest a

simple picture of hadrons as relatively loose bound states
of constituent quarks with masses of about 350 MeV.
During its long history, this idea obtained impressive
phenomenological support (DeRujula, Georgi, and
Glashow, 1975; Isgur and Karl, 1978), reproducing many
properties of the baryons, such as masses and magnetic
moments.
Another simple argument in favor of the nonrelativistic quark model presents itself when one considers the
correlation functions: although the masses and the cou-

n/n

free
T>>1

T T^

Q>H-

0.5
PS,V
o S,A
i

0.2

0.4

^ e ^

0.6

x APV
43

One may be surprised by the appearance of a large quantity,

the heavy quark mass, in this equation. In the limit MQ > oo,
the combination f^esMQ tends to constant, equal to 12TT, where
n is the density of a light quark at the origin.
^Later data, corresponding to much larger statistics and
larger lattices, can be found, for example, in Allton et al. (1991),
Lubicz et al (1992), and Bernard et al. (1992), together with
further discussion of how the large-mass limit on the lattice can
be reached, Maiani et al. (1992).
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

FIG. 17. Correlation functions for IMype mesons, normalized

to the free propagator, vs distance xAPV (normalized to the
Pauli-Villars APV parameter, roughly in fm). The points correspond to the calculation in the IIA framework (Shuryak,
1989d). The closed and open points correspond to negativeparity channels [pseudoscalar (PS) and vector (V) ones] and
positive-parity ones [the scalar (S) and axial (A) channels], respectively. The solid curves are the three-parameter fit described in the text. The dotted line shows the common contribution of the nonzero fermionic modes.

Edward V. Shuryak: Correlation functions in the QCD vacuum

27

plings of the lowest states strongly depend on the quantum numbers of the channel, the continuum thresholds
are always about 2 ? 0 = 1 . 3 - 1 . 5 GeV.
Now, if
confinement demands production of an extra pair of constituent quarks, these thresholds should be roughly 4M eff
in order to produce an extra pair, which indeed corresponds well with observation. Unfortunately, it is very
difficult to understand how the constituent quark model
may be derived from Q C D .
Considering correlators for light-and-heavy mesons,
one may wonder how a " b a r e " quark becomes a
"dressed" one. For example, what are the distance scales
involved, and what is the spin structure of the light quark
propagator? To be specific, let us consider the following
linear combinations of the two correlation functions discussed above:
+

Saip = Tr[S(T)]~[K

-K-],

(4.14)
x Ac

>

nonflip "

}Tr[roS(r)]~[K

+K~],

(4.15)

where "flip" and "nonflip" refer to the light quark chirality, S is the gauge-invariant quark propagator (4.1), and
K + ,K~ are the correlators for light-heavy mesons with
different parity (discussed above). These quantities are
shown in Figs. 18 and 19, as calculated using the IIA
(Shuryak, 1989d). In the constituent quark model, one
expects that these two amplitudes will be given by the
following simple formulas
SmjT)~Z2mD(m,T)
'nonflip

(4.16)

{^-Z^^Dim^)

(4.17)

where m is the constituent quark mass, D(m,r)

is the

_0.8
"0.6
0.4
0.2
0

0.2

0.4

0.6

0.8

FIG. 19. Same as in Fig. 18, but for the spin-nonflip component
of the propagator S (in units Apv) vs distance x (in units
\/ApV), Dots correspond to IIA calculations (Shuryak, 1989d),
and the dashed and solid lines correspond to the constituent
quark model with m eff =0.1.5A PK , respectively.

propagator of a scalar particle at distance r, and Z is the

coupling constant of the constituent quark to the bare or
"current" quark. Note that m in the propagator and in
the numerator of 5 f l i p should be the same, if this model is
to make sense.
The results of the calculations are indeed reproduced
by these simple formulas well enough, with Z = 1 and a
quark effective mass of about (1.5-2)A. This fact is quite
surprising. Particularly unexpected is the result that
Z = 1 within the error limits; it means that for unknown
reasons the current quarks become constituent quarks
without any additional admixture of more complicated
states!
The fit is not perfect, of course: in fact, looking at 5 f l i p ,
we observe that this model works only for r > 0.5 fm; so
one gets an idea of the distances at which the constituent
quark mass is formed. This agrees with the OPE-based
estimates (Shuryak, 1982b) of the effective mass at small
distances:
mtff(x)

(ir2/3)\(W)\x:

(4.18)

xA PV
FIG. 18. Spin-flip propagator of the light quark normalized to
the value of the quark condensate, j^TTS(x)/{qq ) vs distance
xAPVi normalized to the Pauli-Villars APV parameter. Closed
points are the result of the calculation in the IIA framework
(Shuryak, 1989d; open ones show the contribution of the
nonzero modes). The dashed lines are simply a fit to the points,
while the two solid lines correspond to the constituent quark
model with a fixed value of the quark effective mass meff = 1,1.5
Apy.

Rev. Mod. Phys., Vol. 65, No. 1, January 1993

if one assumes that this equation is valid up to m e f f ~ 3 0 0

MeV is formed.
Recently, lattice measurements of the light quark
propagator, taken in a variety of gauges, were reported
by Bernard, Murphy, Soni, and Yee (1990). Remarkably,
the gauge dependence turns out to be relatively weak,
and the data can be more or less described by a simple
constituent quark model with a similar mass value. We
think this interesting point deserves further study.

Edward V. Shuryak: Correlation functions in the QCD vacuum

28

C. Diquarks and baryons containing a heavy quark

In contrast to the qq channels, we know very little
about the qq interaction. Empirical information must
come from baryons, but it is nontrivial to deal with a system containing three light quarks. The logical extension
of what we did in the preceding section is to work with a
qq pair, adding a heavy static quark Q to produce a physical gauge-invariant object.
Masses of the corresponding charmed baryons are
known:
M(A C ) = 2284.91.5 MeV ,
M ( 2 C + + ) = 2452.21.7 MeV ,

(4.19)

M ( E C ) = 246019 MeV .
If one ignores the kinetic energy and any interactions
with the charmed quark, the following conclusions
emerge.
(1) Subtracting m c , one finds that the excess energy as-

KA(r)=([u^r)Cr5dm(r)]'felmnPexp

\{ig/2) f'A^t'dr

where T means transposition and C charge conjugation.

D. Ordinary baryons. Why is the nucleon so light?
As mentioned earlier, we want to learn about the interaction between quarks by studying the baryonic correlation functions. Unfortunately, we do not have as much
experimental information about baryons as we have
about mesons. What we actually know is essentially only
the baryonic masses 45 ; but as we have those for all
members of octet and decuplet, it is still possible to get
some information from them.
We start this section with a presentation of lattice data
on hadronic masses, using the so-called Edinburgh plot,
the ratio mN/mp
versus m^/mp.
In Fig. 20 are data
taken from Teraflop (1992). 46 None of these points so far
correspond to quarks being as light as they are in Nature.
Nevertheless, one can see that these data are all consistent with the naive quark model,
mN/mp

=\ .

(4.20)

1 [M,r(0)Cr^y(0)]e^ ) ,

(4.21)

(fe/2)/*A^t'dr]
|_

Kx(r)=([uir(r)Cyfldm{T)]fe!mnPcxp

sociated with two light quarks is rather large: it is about

1 GeV. It is also about twice the minimal energy of the
light quark 1_, estimated above, for the light-heavy
mesons; so these numbers are consistent.
(2) The 2-A splitting yields the difference between energies of the spin-0 and spin-l diquarks to be about 170
MeV.
(3) The S 2 difference is 175+20 MeV, more or less
consistent with the standard value of the strange quark
mass.
Unfortunately, we are not aware of any studies of the
corresponding correlation functions (besides some preliminary discussion in Shuryak, 1982b). Therefore we
simply indicate which correlation functions we suggest in
this respect.
For simplicity, let us take two quarks of different
flavors, say, u,d ones. We consider a heavy quark mass
to be infinitely large; so it only propagates in the (Euclidean) time direction. Thus one may consider the following correlation functions:

(4.22)

[u?(0)Cr5dj(0)]eiJk)
J w&

Unfortunately, this is 2 5 % higher than the empirical ratio.

Recently, large-scale lattice simulations with dynamical fermions were reported by the Columbia group
(Brown et al.y 1991). They were able to come down in
quark masses to the values ma = 0 . 0 1 and even to 0.004
(for the valence quarks only) 47 ; and even in this case the
N/p mass ratio is found to be 1.527(11), or still much too
large.
Although the history of lattice simulations has shown
that such problems should disappear if larger and finer
lattices are used, it is always instructive to ask what particular effect requires such a fine-grained lattice. For the
N/p mass ratio, we do not know; but we can speculate.
Note that the mass of the A ( { , | ) is phenomenologically
in better shape. Thus it is more probable that something
is wrong with the nucleon itself: for some reason it is not
as light as it should be on the lattice. The nucleon differs
by having s=0 qq pairs; so our speculation is that there
is a short-ranged attraction between quarks in this channel.
Let us examine the general origin of the spin splitting

45

In principle, exclusive reactions with baryons at highmomentum transfer (e.g., elastic form factors) provide some information about the probability of finding three quarks at the
same point, in definite spin and color state.
46
We discuss recent large-lattice data with dynamical fermions
(Brown et al.y 1991) below. A review of the current situation
can be found in the talk given by Toussaint (1992).
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

47
Those numbers roughly correspond to bare masses at lattice
scales of about 15 and 7 MeV in absolute units. In other words,
in the ma = 0.004 case, the pion mass is already nearly as small
as it is in the real world.

Edward V. Shuryak: Correlation functions in the QCD vacuum

1

H.4

) . . .

'

\^>\
\

C
.#~*^
h Staggered:
L
i

"

Wilson:

DOOXO

1.2

I2 4 m
I2 4 rn
I6 4 m
I6 4 *r
!6 4 *r
1

= O.O!
= 0.025
= 0.01
=0.1585
=0.1600
i

0.5

i
i
1.0

M-rr/Mp

FIG. 20. Compilation of lattice data obtained by the

HEMCGC Collaboration [this figure is taken from the review
Teraflop (1992)] on the plot mN/mp vs mv/mp. The dagger in
the left corner is the experimental point, that at the right is the
one at (1,|) expected for very heavy quarks. The line at large
masses corresponds to the nonrelativistic potential model, while
the line at small ones corresponds to chiral perturbation theory.
Data points correspond to different lattices and fermion masses,
explained in the figure.

in an effort to understand why lattice simulations may

have problems reproducing them. It is instructive to
start with a somewhat simplisitic discussion of lightheavy hadrons. In Sec. IV.A we concluded that the
minimal energy of a light quark is about E~ ~ 450 MeV,
ignoring here the energy of the interaction with the static
quark. In Sec. IV.C we found that an average diquark
has an energy about twice that. Does that imply that the
typical three-quark baryon should have a mass about
3E " 1 . 3 - 1 . 4 GeV, even larger than \Mp1
No, if the mutual qq interaction is significant: in the
three-quark system it may be much more important than
in the two-quark one. Checking whether the simple idea
of binary interactions can explain this difference, we follow a simple analysis by Shuryak and Rosner (1989).
They used a nonrelativisitic quark model, in which light
and strange quarks had some energies Eq,Es together
with additional negative interaction for spin-zero diquarks 8E,8ES. All masses of the baryon octet and decuplet can be very well described by this primitive model,
with the following values of the parameters:
g = 4 1 2 . 9 MeV , Es = 5 5 7 . 5 MeV ,
(4.23)
8E = - 2 0 0 . 5 MeV , 8ES = -132.7

MeV .

We make the following observations on these quantities:

(1) Eq is indeed consistent with the magnitude of the light
quark energy in light-heavy hadrons; and (2) &E is also
close to the binary interaction found from the splitting in
the light-heavy baryons. The strange sector also leads to
no surprises: (3) the difference is given by
Es~Eq^145
MeV, close to the standard estimates of ms; and (4) the
ratio 8ES/8E^0.66
is close to the ratio of the magnetic
moments and to the ratio of the energies themselves.
To summarize, the baryon mass systematics suggests
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

29

that the nucleon is light because it contains spin-zero qq

pairs, which have attractive interactions.
Two mechanisms have been suggested as the cause of
this attraction: (a) a perturbative spin-spin interaction
due to gluon-magnetic moments (DeRujula et ah, 1975);
and (b) the instanton-induced interaction (Betman and
Laperashvili, 1985; Kochelev, 1985; Shuryak and Rosner,
1989; Takeuchi and Oka, 1991).
How can one distinguish these two mechanisms? One
possible way is a comparison of the ud and us pairs, related to point (4) in the discussion above. The conclusion
drawn above is often taken as a proof that t h e spin splitting mechanism is due to a gluon-magnetic spin-spin interaction; but, as shown in Shuryak and Rosner (1989),
the instanton-induced forces lead to precisely the same
prediction.
Another possible way is to look at the corresponding
correlation functions. Even if both mechanisms are
tuned to fit the observed mass splittings, they have, in
general, a different spatial dependence. In fact, both are
rather short-range forces; but still they should have
different systematic errors on the lattice. If, as it follows
from IIA, the typical instanton radius is as small as y fm
(Shuryak, 1982a; Diakonov and Petrov, 1984; Shuryak,
1988b), one still may have problems, even with the lattice
spacing being about a ~ 0 . 1 fm. And, certainly, correlation with instantons can be studied on the lattice on a
configuration-by-configuration basis.
The last topic in this section is the O P E analysis of the
baryonic correlators, as initiated by B. L. IofFe and collaborators (Ioffe, 1981; Belyaev and Ioffe, 1982; see also
Farrar, Zhoang, Ogloblin, and Zhitnitsky, 1981 and
Reinders, Rubinstein, and Yazaki, 1983). The results obtained in these papers show some remarkable features
(which were partly unnoticed in the original papers), well
correlated with our discussion above.
We shall use IofFe's nucleon and A currents, which are
defined as (Ioffe, 1981)
rjN = (uTCd)u(uTCy5d)y5u

(4.24)
(4.25)

where C is the charge-conjugation matrix, and T stands

for transposition. We gave suppressed color indices,
which are, as usual, convoluted with el,j,k. Convolution
of the spinor indices is prescribed by brackets; the
current is a spinor itself, and its index is the same as that
of the quark field in the last position in this formula.
One can consider the nucleon (or the A) correlation
functions with or without a spin flip; so, in fact, one still
has a variety of correlation functions to be discussed.
Deviating slightly from Ioffe in this respect, I consider
the simplest traces 4 8

48

One can consider other traces, say, without y 0 , which actually was used by Ioffe as well. Certainly, our discussion of
baryonic correlators is not complete; we have picked up the
sum rule which is assumed to be the best.

Edward V. Shuryak: Correlation functions in the QCD vacuum

30

KN(T)=Tr[y0(7iN(T)VN(0))]

(4.26)

(4.27)

KA(T)=TT[Y0(7JAJT)VAJ0))}

The two points are separated by the Euclidean time r.

Both correlation functions are nonzero in the freequark approximation; so, as above, we normalize the
correlators to these values:
^

ree

(r)=-

(4.28)

6r_

5h7

'

'

, 1

'

'

''

/ X"7"

Nucleoli

A
BI

47
3 3h

/ /

-\

l_

IT T

18

-**^^'"

TT T

(4.29)

i i

KfF*(T) =

i i

Let us now present the O P E predictions by Ioffe (1981)

and Belyaev and Ioffe (1982) 49 :
RN(T)=1

+i

J<(gO%)2)

^<^)2(l-r2mi/32)r4.4c
52
3X21

UA(r)=l+

3X2

12

*G

(4.30)

)2>

< ^/X/A2
>2(l-7r2mg/96) +

(4.31)

Here ml = ig(tpaflvtaG^vxff)/(xf;tf;)f
according to
Ioffe, it is rather large, of the order of 0 . 5 - 1 GeV 2 . Note
that these (xpip)2 corrections correspond to straightforward application of the vacuum dominance hypothesis
(see Sec. III.B). Radiative corrections to them,
0(as)ln(l/rfi),
which played an important role for vector mesons, are not included.
The corresponding curves are displayed in Figs. 21 and
22. The short-dashed, long-dashed, and dot-dashed lines
are O P E corrections including all terms up to gluonic
condensate, quark condensate, and quark-gluon condensates. One can probably trust the last line up to, say,
x = 0 . 7 fm before it starts to bend too much. 5 0
As we do not have experimental data for this correlator, we can only compare it to predictions for these
correlators, which are actually based on similar equations. If we parametrize the imaginary part in the usual
way, we get

KNU)=0\D'(mN,T)+-j-j

2V4

dss2D'(su\r)

These formulas have been recalculated in the space-time representation from the original expressions given in Boreltransformed form. The vacuum expectation values of operators
were evaluated using the vacuum dominance hypothesis.
50
In general, behavior like this is typical for any power expansion of a rapidly falling function. The subsequent terms are sign
changing and tend to compensate each other.
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

r ~ H - i I - _ . J _ . _

~7*'

0.5
x(fm)

FIG. 21. Ratio of the nucleon correlation function (see text) to

that corresponding to the free propagation of three quarks vs
distance x (in fm). The short-dashed, long-dashed, and shortdot-dashed lines are the OPE predictions including the gluon
condensate, the gluon and quark condensates, and the gluon,
quark, and mixed condensates, respectively. The solid line is
the resulting prediction of Belyaev and Ioffe, (1982), to be compared with the resulting OPE line at x < 1 fm. (The long-dotdashed line shows the contribution of the continuum states,
apart from that of the nucleon itself.)

KA(x) = 2X2D'(mA,T)

+ -^-j 4 J fdss2D'(sW2,T)
2V
^ A

,
(4.33)

where D (m,r)=
d/drD{m,T).
The parameter values
suggested by Belyaev and Ioffe (1982) are
WN**1.5 GeV , /3?0.45 GeV 6 (27r) 4 ,

(4.34)

J F A 2 . 1 GeV , X 2 2 . 3 GeV 6 (2irr) 4 .

(4.35)

i>

<

'

-1

/"
//

//
////
/1//
// / // f
/
/ /

z
z

\_

FZ0Z

i _..+-'' i

/
/ '"1

0.5

Z.

'.

.K

. ^ ^ ^ N

~^
-

BI
1

-i-

\ ,

'

\~
(4.32)

49

0.5

1.5

~^~~~~~~~~.
i

x(fm)

FIG. 22. Same as in Fig. 21, but only for the A correlation
function. In this case we have shown by two solid lines predictions of two groups: BI (Belyaev and Ioffe, 1982) and FZOZ
(Farraref aL, 1981).

Edward V. Shuryak: Correlation functions in the QCD vacuum

The corresponding prediction correlators are also shown
in Figs. 21 and 22 as the solid lines. Here BI stands for
the predictions of Belyaev and Ioffe (1982), while F Z O Z
corresponds to Farrar et ah (1981), where the baryonic
parameters were obtained from a somewhat different set
of sum rules (Chernyak and Zhitnitsky, unpublished).
The latter work obtained similar parameters for the nucleon, but significantly smaller coupling for the A,
A, 2 0.5 GeV 6 (27r) 4 . The agreement between the O P E
and the predicted curves in some regions of r is not
surprising: the parameters of these curves were obtained
from essentially the same sum rules.
What is surprising is the large predicted magnitude of
the nucleon correlator at distances of from 1-2 fm,
where they are essentially larger than the perturbative
ones. Comparing this curve with the mesonic ones discussed in Sec. II, one sees that they are reminiscent of the
pseudoscalars, for which strong attractive forces certainly exist.
It seems the curves for the A have less pronounced
peak than those for the nucleon, if the F Z O Z numbers
are used. Is it because the A contains only diquarks of
spin one, similar to the p, while the nucleon also has scalar diquarks, similar to scalar and pseudoscalar mesons
with strong instanton-induced interactions?
Although these curves seem to support our discussion,
it remains to be proven whether these O P E predictions
are really true, and it is not actually applicable at such
large distances.
Recently the first attempt to calculate the instantoninduced corrections to baryonic correlation functions in
first order in *t Hooft interaction was made. Kochelev
(1990) reported a significant improvement of the sum
rules compared to the Ioffe OPE-based results. Similar
work in the H A framework is now in progress, including
the evaluation of all orders of the 't Hooft interaction.
We have seen that a number of approaches, including
the naive quark model and the OPE, suggest the existence of an attractive qq interaction inside the nucleon.
The attraction is less pronounced inside the A and therefore is attributed to the spin-0 isospin-0 diquarks. If the
attraction is there, it can explain why the nucleon is so
light and its coupling to the Ioffe current is so large. The
nature of these forces remains an open question. We also
do not understand why lattice simulations have problems
in reproducing them.

31

(Shuryak, 1980). This transition can be studied experimentally in heavy-ion collisions at high energies. Such
experiments are under way at C E R N and Brookhaven
[see Quark Matter-89 (1990) and Quark Matter '90
(1991)]. Construction of a large, dedicated facility in
Brookhaven, the Relativistic Heavy Ion Collider (RHIC),
was begun in 1991, and there are plans to use, in the
study of this transition, the future Large Hadronic Collider at C E R N as well.
This problem can be studied theoretically in the framework of various models and, from the first principles of
Q C D , by numerical simulations on the lattice. This work
is very active at the moment [see Lattice 88 (1989); Lattice 89 (1990); Lattice 90 (1991)]. The proposed
T E R A F L O P project 51 (Teraflop, 1992) in the United
States promises to increase the computer power involved
by 2 - 3 orders of magnitude.
Readers interested in finite temperature Q C D in general are directed to Shuryak (1980), Gross, Pisarski, and
Yaffe (1981), Shuryak (1988a), and Hwa (1990). In this
section we shall concentrate on the correlation functions
at finite temperature.
Two main qualitative features of the Q C D vacuum will
disappear at a high enough temperature: one expects
deconfinement and chiral symmetry restoration. Figure
23, taken from a recent paper by the Columbia group
(Brown et ah, 1990), makes a brief summary of our
present understanding of the corresponding phase diagram. Two observations are important: (1) The two
phase transitions mentioned seem to be well separated,
which supports the idea that they are based on completely different physics. (2) The real world seems to be outside of the first-order transition regions, but still very
close to the line of the chiral one. 5 2
Speaking about the Q C D vacuum more quantitatively,
one may ask why typical energy density 53 is needed to
melt it? For deconfinement one may take as a guess
something similar to the M I T bag constant,
MiT~50MeV/fm3 .

(5.1)

For the energy density related to chiral symmetry restoration, one may take the simple estimates (Asakawa
and Yazaki, 1989; Li, Bhalrao, and Bhadury, 1991) in the
Nambu-Jona-Lasinio model. This is essentially an estimate of how much energy will be gained if the quarks in
the negative-energy Dirac sea are correlated, making the
quark condensate and the massive constituent quarks.

V. CORRELATION FUNCTIONS AT NONZERO

TEMPERATURES AND/OR DENSITIES
A. Melting the QCD vacuum
Because the Q C D vacuum is a complicated medium
made out of nonperturbative fluctuations of quark and
gluon fields, one naturally arrives at the idea of trying to
"melt" it. The present understanding is that at some
critical temperature or density it undergoes a transition
to another phase called the quark-gluon plasma
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

Similar projects are developing in Europe and Japan as well,

but none have been approved yet.
52
We shall use the notion of critical temperature Tc, meaning
the one at which specific heat is maximal and chiral symmetry
is "practically restored." Numerically it is expected to be about
150-180 MeV.
53
We do not speak here about the temperature T, because the
energy density is much more relevant for experiments.

Edward V. Shuryak: Correlation functions in the QCD vacuum

32
two
flavors \

pure
gauge \ ^ ^

We have discussed in the preceding sections what kind

of information is currently available about the correlation functions at T 0. Roughly speaking, phenomenology tells us that the physical spectral density can usually
be represented by two distinct components,

00

0.1-

second
order

V ~ -first "
\ _ order _

\L : =

one
flavor

- _first -"x
- order - _. \
0.25- - _ - # " _ - \
0 . 0 1 -J
0 i
1
0.01 0.025

GO
"u,d '

FIG. 23. Diagram of the QCD phase transitions, given in the

mu,ms plane (the light and strange quark masses, respectively)
according to Brown et al. (1990). In the lower left corner, one
observes a first-order chiral restoration transition, and in the
upper right a first-order deconfinement. The experimental
point is shown by the dashed circle.

Without going into details, we conclude only that this

leads to a magnitude of chiral energy density comparable
to the BM1T value.
However, it is important to realize that the total nonperturbative energy density of the Q C D vacuum is much
larger: using the trace anomaly relation and the value of
the gluon condensate (Shifman et ah, 1979b), one obtains
(Shuryak, 1978a)
11^/3-2^/3

~500MeV/fm3 ,
(5.2)
where Nc, Nf are the number of quark colors and flavors.
Therefore one may conclude that some important nonperturbative effects should be present even at T larger
than that needed for deconfinement and chiral restoration; and in order to obtain the quasi-ideal quark-gluon
plasma, one actually needs to surpass this energy density.
That is why the planned energy of the heavy-ion collider
is chosen to be so high.
The hadrons are expected to melt along with the vacuum as one approaches the phase transition. Presumably,
hadrons gradually become unstable in hot matter and
finally fail to represent the main degrees of freedom of
the system. To discuss hadrons at finite temperature is
rather tricky, because we have to define precisely the objects of discussion at nonzero 7\
The correlation functions, on the contrary, have the
same definition below or above Tc and retain essentially
the same physical meaning.
The only obvious
modification in their definition is that the averaging
changes its meaning, the angular brackets here representing the statistical average over the Gibbs ensemble,
characterized by the temperature T and possibly by
chemical potentials, for finite charge or flavor densities.
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

ImK(s)=f2S(s-m2)

+ 6(s-El)ImKpext(s)

(5.3)

describing propagation of bound quarks at large distances and of free quarks at small distances.
Is this general structure preserved at T=01 Let us
take it as a rough working hypothesis. If so, the general
question formulated above is reduced to a question about
the temperature and density dependence of these three
parameters, / , m, and EQ.
We have already mentioned two phenomena that are
expected to occur at high temperaturedeconfinement
and restoration of chiral symmetry. As one signal for
deconfinement, one can consider "hadron swelling,"
presumably detectable in the correlation functions as a
strong decrease of the coupling constants f to local
currents. 5 4 Another possible signal might be a drop of
the observed threshold E0:
the threshold under
confinement may be interpreted as 4m eff (see discussion
in Sec. IV), and we expect it to drop to something like
2m eff at the point where deconfinement takes place.
So far there has not been much discussion in the literature of the effect of phase transitions on the correlators.
Rather, the main focus has been on "dropping masses"
m(T).
There were suggestions that parity doublet hadronic modes could be formed (TT a, p Ax, N N*,
etc.) above Tc (DeTar and Kogut, 1987). Another interesting suggestion is that masses of many hadronic
modes should vanish at the critical point, because they
are related to the vanishing quark condensate (Brown,
1991), and this phenomenon may cause much more
smooth behavior at the phase transitions (Brown, Bethe,
and Pizzochero, 1991) compared to naive estimates with
free pions and quark-gluon plasma.
Let us add to this list of general questions a few more
that are related to specific channels. As we have mentioned, the isovector scalar channel most probably has no
"normal" qq mesons; so these correlators are mainly related to the quark continuum. As T approaches Tc, does
the 1=1 scalar resonance go down in mass to meet its
parity partner, the pion, and become a visible resonance,
or does one have simply a cut with a decreasing threshold, going down as the temperature approaches Tcl Can
the scalar-pseudoscalar mode, representing large fluctuations in qq, persist even above Tcl
The next very interesting channel is the 17', the isoscalar pseudoscalar channel. It is well known that
UA(\)

54

The usual order parameters used for the deconfinement on

the lattice, such as the expectations of Wilson of Polyakov
loops, are not really useful in theories with light quarks, because forces between static quarks are screened.

Edward V. Shuryak: Correlation functions in the QCD vacuum

chiral symmetry is explicitly broken by the instantoninduced interaction, and therefore it is never restored in a
strict sense. However, at some temperature it is restored
for practical purposes, in the sense that differences between singlet (17') and octet (K9n,7]) correlation functions become small.
Thinking about the mixing of hadrons made of quarks
with glueballs, one may naturally ask whether the
quark-related and the gluon-related correlators approach
their high- T limit at the same or at different temperatures. In particular, it is known that quark-made mesons
are much lighter than glueballs: does this imply that one
will need higher temperatures to "melt" the glueballs?
In the remainder of this section we shall review what is
known on this topic. Investigations are only recently
started, so only a few of these questions have answers.

Here A, V are axial and vector currents, and their upper

indices are, for example, those of the SU(2) isospin group.
Then, using commutator relations
[^8,r*]=r*^,

KT(x)=((j(x)j(0)))=l,n(n\j(x)j(0)e ~

\n)

= (\-e)Klv(T

= 0) + K*v(T)

(5.6)

K^(T)

= (l-)K^(T

= 0) + eK^v(T)

(5.7)

The admixture coefficient is simply

(5.4)
If the temperature and baryon density are small enough,
the matter will be normal hadronic, made of wellseparated particles: pions at low T or nucleons at low
densities nb. Therefore, to first order in matter density,
one need consider only the one-body states in the statistical sum. This makes finite density corrections calculable,
provided the corresponding matrix element over the hadronic state can be estimated.
Let us restrict our discussion to the case of zero density of all charges and divide the low T range into two
separate regions: (a) parametrically small temperatures,
meaning that power series in T may be terminated at the
lowest nonvanishing terms, and (b) any T below Tc, excluding the vicinity of the transition region.
In the former case, one can use such general methods
as the partially conserved axial current hypothesis
(PCAC) and the Weinberg effective Lagrangian. In the
latter case, one should use some more involved parametrization of the empirical interaction between particles
involved. Note that in the latter case the conclusions are
essentially model independent, but their accuracy is limited by the accuracy of the corresponding data.
As an example of the general statements valid at
parametrically small temperatures, we consider vector
and axial correlators with p, A x quantum numbers, following Dey, Eletsky, and Ioffe (1990). At low temperature, the vacuum has added a dilute gas of pions; so finite
temperature expectation values can be expressed in terms
of the thermal density of pions nn(T) as
{M)T=(M)+n^(T)(7r\MW)
a

To evaluate <TT| V (x)V (0)\ir),

one can use P C A C and
replace it by vacuum averages of the type ( A VVA ).
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

(5.5)

Klv(T)

{a En)/T

[A^A^f^V^,

it can be further reduced to a combination of T = 0 vector and axial correlators. The results can be written in
the following elegant form, expressing the small- T correlators in terms of vacuum correlators:

B. The tow-temperature and low-density limits

The Gibbs averaging of the correlation functions can
be written, in general, as the sum over stationary states

33

4d3k
(2w)32k

1
_ T2
exp(*/r)-l
6Fl '

As both vector and axial correlators at T = Q are known

(see Sec. II), we therefore know both correlators at finite
but parametrically small T, e 1. In a later paper (Eletsky, 1990) a similar statement was also proven for
positive- and negative-parity baryonic currents. From
this general approach, one has learned that vector and
axial correlators start to merge at small T. By the time
T>Tcy they should become identical.
Unfortunately, this general approach is very restricted
in its applicability, as the following argument shows.
What has been taken into account is essentially the
forward-scattering amplitude of a pion on virtually a p
meson or nucleon. This is indeed known at low pion momenta from quite general considerations. However, this
region ends as soon as the thermal pion energy becomes
large enough to hit the first resonance, say, A x in the irp
channel or A in the TTN case. This condition actually restricts T t o a level below 100 MeV or so.
However, using the experimentally measured forwardscattering amplitudes, one can get realistic estimates of
the modification of hadrons at low T. Such calculations
for ir9 K, p, co, and other mesons, modified in the pion
gas, were made in Shuryak (1991) and Shuryak and
Thorsson (1992). Without going into details here, let us
only mention some main conclusions. In all the cases hadronic dispersion curves co(k) were found to shift down
with increasing T. However, the magnitude of the effect
was found to be rather modest, even at T= 150-200
MeV, where the phase transition is expected. Roughly
speaking, the corresponding collective potentials are of
the order of the nuclear potential well, say, 50 MeV, an
order of magnitude smaller than mesonic masses themselves.
These calculations show that the hadronic masses can
significantly drop only very close to Tc> where the calculations are inapplicable due to possible multiple particle
interactions. Whether the masses really drop or not
remains an open question.

Edward V. Shuryak: Correlation functions in the QCD vacuum

34

C. Quark propagation in the quark-gluon plasma

at high temperatures

ST(x) =

At very high temperatures (and/or densities) matter is

believed to be a nearly ideal gas of fundamental constituents, the so-called quark-gluon plasma. Discussion of
this statement in the framework of the perturbation
theory and beyond it can be found in Shuryak (1980,
1988a) and Gross et ah (1981). Without going into detail
here, many phenomena were found to be analogous to
those in ordinary electromagnetic plasmas. For example,
electric fields are known to be screened, there are plasma
oscillations, etc. Only the long-range gluo-magnetic field
may have a specific nonperturbative structure at the scale
82T.
If hot plasma is made of nearly noninteracting quarks,
one may expect that the correlators correspond to independent propagation of quarks, described by the ordinary loop diagram with thermal quark propagators,
below denoted by ST(x). Such zero-order diagrams were
first evaluated for the vector correlator at 7V=0 by Bochkarev and Shaposhnikov (1986), and for the nucleon
current by Adami and Zahed (1990). These authors used
a Borel-transformed representation of the correlator,
which made the derivation rather complicated.
Calculations are significantly simplified in the coordinate representation. In the zeroth order, one has a completely independent propagation of quarks; so mesonic or
baryonic correlators are reduced essentially to the square
or cube of the thermal quark propagator. For simplicity,
we ignore the quark masses and consider correlators in
the spatial direction only. 55
There are three ways of getting the thermal quark
propagator: (1) use the standard Feynman rules in the
Matsubara formalism and then make a Fourier transform; (2) use a real-time formalism and look at all scattering processes; or (3) solve the Dirac equation in spacetime with Matsubara antiperiodic boundary conditions.
The first way leads to the sum

STM = (yxdx)T^f

d3k
(2TT)

expjikx)
k2+[irT(2n-l)]2

(5.9)

where n ranges over all integers, positive and negative.

The second method leads to
ST(x) = (yxdx)f

d3k
(2v)3k

exp(ikx)

1
l+exp(A:/D
(5.10)

where the last term is the Fermi occupation factor, and

the third to

55

In general, there will be different functions for spatial and

time separations, because the thermal system is not Lorentz invariant.
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

(Yxdx)T2n(-l)n

x2 +

(r-n/T)2

(5.11)

where an antiperiodic scalar propagator is written as a

sum over paths with different "winding number." These
different methods give the same results, which can be expressed in terms of a universal temperature modification
factor / ,
ST(x)=S0(x)f(irTx)

(5.12)

where, at T = 0y one h a s / ( 0 ) = l, S0(x)=l/(2rr2x3)

at 7 ^ 0

and

r. *
,
, z + l + (z l)exp( 2z)
/(z)-zexp(-z)
-*-r
.
[1exp(2z)] z
At small z, / may be expanded as

, lav
(5.13)

/ = l - ( 7 / 3 6 0 ) z 4 + 0(z6)

(5.14)

The meaning of the fourth-order coefficient and the

reason why lower powers vanish will be discussed in Sec.
V.E. Consider now another limit: why does this function decrease exponentially, f(z)=z2
exp( z), at large z?
This behavior implies that the thermal propagator decoupled from low-energy quark states, although they certainly exist in the noninteracting gas of massless quarks.
Some readers are probably satisfied by the formal
answer to this question based on Matsubara formalism
(Eletsky and Ioffe, 1988): the lowest Matsubara frequency for fermions is (irT), It may also be seen in perhaps
more physical terms by looking at the second equation
(5.11) containing the combination [\ rif(k)] with the
Fermi occupation factor. In the case of Bose statistics,
one has instead a much more familiar combination
[j + nB(k)]9 due to zero-point motion and thermal excitations, respectively. In our case these two terms have
the same interpretation as well. At A: = 0 one has exactly
nf = \; there is cancellation due to the destructive interference of zero-point fluctuations and excitations.
One may also rewrite [\ nF(k)} as \[(\ nF) nF],
so that the particle-hole symmetry nF^>{\nF)
becomes
obvious. If the situation near / = 0 and / = l is physically similar, up to a sign of the propagator, it is natural
that the point where occupation is exactly nF = \ is special; here the particle and hole terms compensate each
other.
We now use these formulas to compare the expected
high- T behavior of the correlation functions with what is
known in the vacuum (T=0).
In Figs. 24 and 25 comparison is made for the vector-axial and nucleon-A correlators. This comparison indicates that the axial correlator should show the weakest T dependence in the interval
of distances considered, the p and A ones should show
somewhat stronger T dependence, and the nucleon correlator should depend most strongly on T. We have not
discussed scalar and pseudoscalar cases, but from our discussion in Sec. III.A and Fig. 24 it may be inferred that
in these cases the T dependence should be even more
dramatic.

Edward V. Shuryak: Correlation functions in the QCD vacuum

1.5

niiiIiiiiIiiir

- ^ T=0 vector

looked very mysterious. However, as it was shown in the

previous section, even for independent quark propagation, a destructive interference of two terms in the quark
propagator [nF(E) \] leads to such exponential decay.
A set of lattice data compiled by Gocksch (1991) is
plotted in Fig. 26 in the form of the so-called screening
masses as a function of temperature. The screening mass
is defined as

0.5

M = - lim

x(fm)

FIG. 24. Ratio of correlation function to that corresponding to

free-quark propagation, vs distance x (in fm). The dashed
curves correspond to p and A j channels at zero temperature, as
derived from experimental data in Sec. II. Two solid lines correspond to the factor f2(7rTx), describing modification of quark
propagators in the quark-gluon plasma; they are shown for
r=200and400MeV.

D. Lattice data
1. Screening masses
Lattice studies of correlation functions at nonzero T
were pioneered by D e G r a n d and DeTar (1986), who observed the exponential decay with distance and interpreted it as the existence of hadronic modes, even at high
T>TC. As it was in apparent contradiction with popular
ideas about deconfinement and even with perturbative
Debye screening of color charges in the plasma phase, it

_j

d\nK(x)
dx

(5.15)

Its value at low T should coincide with the masses of the

lowest hadrons in the corresponding channels. A t
T>TC, data for vector mesons and baryons show that
M/T is essentially constant. Moreover, although it is
not exactly 2-TT and 377% they agree reasonably well with
the corresponding values after finite-size corrections
(Bornef a/., 1991).
The following observations can also be made on the
basis of these results.
(a) The lattice results clearly show one of the anticipated phenomena, parity doubling. At T>TC vector-axial
and scalar-pseudoscalar correlators, etc., become identical, which is a direct consequence of the restored chiral
symmetry.
(b) F o r the vector and baryonic correlation functions,
the asymptotically high-7 1 form is smoothly reached already near Tc, without any noticeable discontinuities.
(c) For the scalar-pseudoscalar case, one observes
much smaller screening lengths than those corresponding
to Matsubara frequencies.

15

j _

BI nucleon

35

10

h-

, . . , | . . .
Screening Masses

J
H

'*

~J

-
\-

<X<X-

^ .. .

<>

o.

- -^
-

L i

-2

1 , , , , 1 , , ,

T/T c
x(fm)

FIG. 25. Same as Fig. 24, but for baryonic currents. Two solid
curves for r = 2 0 0 and 400 MeV correspond to the factor
f3(irTx); the dashed ones are Belyaev-Ioffe (BI) predictions (Belyaev and Ioffe, 1982) for the nucleon and Farrar et ah (FZOZ)
ones for the A (Farrar et aL, 1981) discussed in Sec. V.
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

FIG. 26. Compilation of screening masses (Gocksch, 1991) obtained on the lattice for different correlators vs temperature T
(scaled in units of the critical one, Tc). The short-dashed curves
show values 2rrTy 3TTT corresponding to the lowest Matsubara
frequencies in the continuum, while the long-dashed ones show
their values for the finite lattices used.

Edward V. Shuryak: Correlation functions in the QCD vacuum

36

2. Binding of Qq pairs propagating

in spatial direction
The question of whether qq pairs are correlated or
move independently was addressed in the recent lattice
study by Bernard, DeGrand, DeTar, Gottlieb, Krasnitz,
Ogilvie, Sugar, and Toussaint (1991). It was demonstrated that if the qq pair is forced to propagate large distance
in, say, the z direction, it moves in the transverse x,y
plane in a correlated way. The corresponding wave function at finite T happens to be similar to that at T = 0, and
it even has a smaller radius. This result has created some
excitement and we believe, some misinterpretation.
The key technical point, important for the understanding of the terminology used, is an interchange of the time
t and z axis. After transformation to this new space, one
deals with a system at zero temperature in a periodic box
in the z direction, with periodicity /3=\/T.
Using this
language it is obvious that in the high- T limit the socalled dimensional reduction takes place, and a 1 + 3 dimensional gauge theory becomes a 1 + 2 dimensional one
(Appelquist and Carazzone, 1975).
The main physical point is that in the high-T limit the
motion of a quark in the new space is dominated by its
momentum in the z direction, which, as a result of the
antiperiodic boundary conditions, is given by TTT. For
the motion in the transverse directions xyy9 this momentum acts just like a mass:
MeW = V/m2 + 7r2T2 .

(5.16)

As at a high temperature, M eff becomes very large, any

attractive potential can bind the quarks in the transverse
direction.
Qualitative features of the related quantum electrodynamic problem of d = 2 positronium were recently discussed in Hanson and Zahed (1991): apart from the
effective mass just mentioned, another important ingredient is the logarithmic Coulomb potential in two dimensions. More quantitative analysis of this problem
was made by Koch et al. (1991).
Turning to the Q C D case and using the old dimensional reduction argument (Appelquist and Carazzone, 1975),
one realizes that one has to deal in this case with
d = 2 + 1 Yang-Mills theory, which is far from simple. In
particular, as argued in D'Hoker (1982), it has a linear
confining potential, as do d = 1 + 1 and d = 1 + 3 (the real
QCD!) theories.
In the high- T limit quarks have large effective mass,
similar to superheavy quarkonia (Appelquist and Politzer, 1975); so the size of the bound state is small
R ~ 1 /gT and the Coulomb potential is justified. However, one cannot compare this picture with the available
lattice data because the temperature is too low. The lesson learned from quarkonium physics is that the nonrelativistic approach based on an effective potential works
very well. Such effective potential, in fact, was numerically studied some time ago by Manousakis and Polonyi
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

FIG. 27. Wave functions for p,7r channels at I 7 =250 MeV (a)
and T = 350 MeV (b), as a function of r/a (a0.22 fm). The
lines are solutions of the d = 2 Schrodinger equation with the
potential discussed in the text from Koch et al. (1991), while
the points are from lattice measurements by Bernard et al.
(1991). The solid line and the short-dashed line correspond to
p with spin projections *SZ = 1 and 0, respectively, while the
long-dashed line corresponds to the pion wave function. In (a)
normalization is arbitrary, while in (b) all cases are normalized
in the same way.

(1987), who found for the potential 5 6

V=-a/r + ar

(5.17)

with a = 0 . 1 8 4 + 0 . 0 2 , ^ = 0 . 2 2 + 0 . 0 3 . This should be

compared
to
the
T 0
parameters,
a =0.25,
Va = 0 . 2 2 + 0 . 0 2 . Thus the potential measured in the
space direction appears to be essentially T independent.
In Koch et al. (1992) spin-dependent forces due to a
Fermi-Breit interaction are discussed in detail. In particular, this paper points out that the vector states p,co are
split into separate states at high temperature, depending
on the spin projection Sz on the propagation axis. All
these splittings are proportional to one common parameter, the effective color charge, and more detailed lattice
study can check the Fermi-Breit splitting mechanism
well enough. Presumably, it is the only mechanism to
survive at high T, while at low T nonperturbative phenomena discussed in previous sections should also be involved.
With a potential fixed, one can proceed further and
solve numerically the d = 2 Schrodinger equation and
compare it with lattice data, as shown in Fig. 27. Comparing data for the p,rr channels, one can observe a
significant difference between the two cases. The average
size of the pion is significantly smaller and, in particular,
at small distances r < 2a, the pion wave function is much

56

One may compare these data with the potential corresponding to quarks propagating in time direction. Not only are the
confinement forces absent above the deconfinement temperature, but even the Coulomb part is much smaller, due to screening effects.

Edward V. Shuryak: Correlation functions in the QCD vacuum

more prominent, 5 7 with | ^ ( 0 ) | 2 about 4 times greater
than in the p-meson case. Let us also note that spindependent forces explain not only the difference in wave
functions, but also large splitting of the screening masses
discussed in the previous section.
We have discussed in this section correlation functions
in the spatial direction at high temperature. Unlike for
T=0, these spectra are by no means related to the energy
spectra of hadronic modes. However, the corresponding
momentum or "screening masses' spectra, as well as the
corresponding wave functions, happen to be very similar
to those at T = 0. This may be explained by the d2
Schrodinger equation with an effective mass (5.16) and
potential (5.17).

- ~83x4

(a) :

- --continuum

37
_' ' ' ' 1 ' '
- ~83x4

"i

l.5xTc.

Tc
,

1 ,

i 1 ,

. i

t i i i . "

- continuum

':
S

(b):

g -

a
o

':

l.5xTcI
C
". . . i 1 . , . , 1 . , , i, . . . ."
=

FIG. 28. Baryonic charge susceptibility for singlet (a) and nonsinglet (b) definitions (5.17) vs the bare coupling constant g (arrows show positions of the corresponding chiral restoration
temperature Tc and 1.57^, from Gottlieb et al.f 1987).

correlation functions integrated over space-time. 58 T h e

integrated correlators vanish at 7^=0, which can be explained as follows (McLerran, 1987). The general strucAs a final topic, we mention an interesting set of lattice
ture
of t h e vector
correlator
at T=0 is
data (Gottlieb et aL, 1987) related to the so-called baryon
2
II
=0
8
d
g
)H(x),
and
the
density-density
correiUV
AZ
v
flv
number susceptibility (McLerran 1987):
lator is t h e IIOQ component. The two terms generally
compensate time derivatives, and the spatial integral over
Xb=(WT)fd3x((nb(x)nb(0)))
,
(5.18)
spatial derivatives vanishes.
At nonzero temperature this simple form for the vecwhere nb is the baryon density. Such an integral measure
tor
correlator is no longer valid. There exist other transof the SU(3) singlet vector correlator (a combination of
because the thermal rest frame is special.
co,<f> ones) was found to j u m p significantly at Tc> as shown verse tensors,
can
e
Thus
Xb
b
nonzero. However, Fig. 28 suggests that
in Fig. 28, soon reaching about -f- of its asymptotic value
no
significant
modification
in vector correlators takes
corresponding to the free thermal motion of light quarks.
place
until
T
is
very
close
to
T
c.
Shown also in the same figure, the nonsinglet susceptibiliCertainly,
much
more
work
is needed to reach a real
ty with the quantum numbers of p has the same behavior,
understanding of these data. It would be helpful to meaand within the error limits (Gottlieb et al.y 1987). T h e
sure the entire point-to-point correlation functions FT(R)
definitions for singlet and nonsinglet cases can also be
on t h e lattice, not just its derivative, (the screening
written as derivatives with respect to the chemical potenlength) or an integral (the susceptibility).
tials,
3. Baryonic number susceptibility

XsfNS = ^/^u^/^d)(nund)

(5.19)

One may be puzzled by this j u m p after our previous

experience of smooth behavior in the screening lengths.
As the magnitude of xb *s comparable to that expected
for free quarks at high temperature, the real question is,
why is xb so small below Tcl
One can answer this question in two ways. First, the
nucleons are the lowest baryons and they are rather
heavy. Therefore below Tc we expect their density to be
low, and the effects of changes in chemical potential will
be small. Thus the data shown in Fig. 28 suggest t h e
baryonic masses do not drop significantly, until maybe
very close to Tc.
Second, one can relate this susceptibility to the vector

E. Sum rules based on the operator product

expansion at finite temperatures and densities
Generalization of the Q C D sum rules for nonzero temperatures was pioneered by Bochkarev and Shaposhnikov
(1986), and for nonzero baryonic density by Drukarev
and Levin (1990). These two important papers have
created quite a substantial literature, which can only be
discussed in full in a specialized review. Our discussion
below concentrates on a few key points, some of them being firmly established and others remaining open questions. We conclude with a sample of the results. Even
more than for the vacuum case, the status of these predictions is not completely clear, but they are certainly
very interesting. In any case, they have raised many new
physical questions and have given a new perspective in
the area of finite temperature and density Q C D .

57

On the topic of pion wave functions, the T=0 pion wave

function by Bernard et al. (1991) is sensibly the same as the
zero-temperature Bethe-Salpeter amplitude calculated by Chu,
Lissia, and Negele (1991).
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

58

A warning: we have up to now ignored contact terms, but

they do contribute to the integrals.

Edward V. Shuryak: Correlation functions in the QCD vacuum

38

one for deep-inelastic scattering. As at T=0, the Fourier

transform of the correlator

1. Modifications of the operator product expansion

In Sec. III.B we saw that many questions related to the
domain of applicability of the O P E remain open even for
the vacuum. In the case of nonzero temperature or density we certainly have to face extra complications. All
finite temperature complications can be put into two
categories: (a) modification of the O P E itself, because in
matter one has to include a wider set of operators on the
left-hand side; and (b) the physical spectral density on the
right-hand side is more complicated, containing not only
production processes but also scattering on matter constituents. In this section we mainly concentrate on the
first set of questions, holding to our general idea that predictions for the correlation functions are interesting even
without empirical information, for comparison with other (e.g., lattice) studies.
Let us first recall that the O P E has a generic scale ji
used for the separation of the low and high frequencies,
with which the matrix elements and coefficients of the
operators are constructed. At finite temperature or
nonzero chemical potential, one should first decide on the
relationship between T and /x. If T fi, coefficients Ct
are not modified and only the operator matrix elements
{{Ot)) are changed, including new contributions from
the heat bath.
The major difficulty with the finite temperature O P E is
that the set of contributing operators is no longer limited
to Lorentz scalars but includes all symmetric tensors.
This is, of course, due to the existence of a preferred
frame, the thermal rest frame. Let us define the unit
four-vector of the thermal frame, n^. If, for example,
one considers a correlation function of scalar or pseudoscalar operators, the O P E equation would have the form
K(x) = 2Cu)xa....,xa
otj

{{O
u.n

"lu

(5.20)

and the thermal average of the operator would give factors na , . . . ,na . In simpler terms, one simply has to
expand separately in powers of time and the space interval.
The kinematics are slightly more involved for vector
and axial currents, but it is essentially the same as the

K(x)/KfTQe(x)=l+x4[Ci(E2-B2)

+x6[CA$02

T^v(q) = ifd4x

^((;^);v(0)

(5.21)

can be considered as a physical scattering amplitude, say,

of forward photon scattering on the heat bath, and its
imaginary part has all the usual analytic properties, although it depends on the energy transfer co = (qn) and
momentum transfer separately. Standard notations for
vector current describe the corresponding physical spectral density in terms of the two functions Wx and W2'

(5.22)
Here the transverse part of the vector n is
njl=nll {nq )qVL /q2.
This parametrization satisfies
gauge invariance of electrodynamics, 59 which demands
that (190 = 0.
The main new operators are the so-called leading twist
quark operators (Politzer, 1974) of the kind

ou

=#a,...,a, ^.

(5.23)

The origin of these operators is easily explained. These

operators are associated with a process in which a quark
is created at point 0, travels from 0 to x, and then is returned to the heat bath. Its amplitude can be written as
^ ( 0 ) r ( r M x M ) / ( 2 i r 2 x 4 ) r ^ ( x ) , where T^y^
or y5 for vector and pseudoscalar currents, respectively. All that
remains to be done is to simplify y matrices and expand
i^(^;)ina Taylor series in x.
The lowest of these operators, with only two indices,
has dimension 4 and is nothing other than the quark
stress tensor. That is why our small-distance expansion
of the quark propagator (5.14) started with the x4 term.
Furthermore, the 0(T4) coefficient is nothing more than
a pressure of the plasma made of free quarks.
Now let us display few O P E terms of the lowest dimension, measuring the correlation function in a spatial
direction x9

C2(E2+B2)+C3(t/jyldli>)]

+ C5(($yltP)($yltf;))+C6($yld]0

This equation includes some operators that have not

yet been studied in the literature, in particular, the gluonic stress tensor (C 2 ) and the square of the vector current
(C 5 ).

] .

(5.24)

I
ing, one cannot answer these questions without an understanding of the underlying nonperturbative dynamics.
However, some statements can be made about the matrix

2. Operator expectation values

59

The next set of questions concerns the temperature

dependence of the operator averages. Generally speakRev. Mod. Phys., Vol. 65, No. 1, January 1993

For the nonconserved axial current there are four functions:

two additional ones can be chosen to be proportional to q^qv
2indqlxnv + nyLqv.

Edward V. Shuryak: Correlation functions in the QCD vacuum

elements, and we briefly review some ideas suggested in
the literature.
Let us start with the leading twist operators just discussed. In Bochkarev and Shaposhnikov (1986) and in
many later papers on the subject, the diagrams were evaluated using the thermal quark propagator ST(x) instead
of SQ(x). It seems like a logical thing to do, but actually
this prescription has a limited region of validity. It involves the operators mentioned above, but perturbatively, which means that their average values are automatically taken for an ideal quark gas.
Such estimates of these matrix elements may be used at
high T>TC, where matter is indeed the quark plasma,
but they are not valid otherwise. Therefore the simplified
procedure based on thermal propagators overestimates
the temperature dependence of the correlation functions
at T < Tc. In the opposite limit of low Ty one can evaluate the quark part of the thermal energy density as
q(T)=xqw(T),
where ew(T) is the energy density of a
dilute pion gas and xq *s the quark share of the pion momenta. This share is about xq ~ y > inferred from empirical pion structure functions.
In OPE-based sum rules, one of the essential ingredients is the gluon condensate, { G^ v ) . In matter one
has to include two separate operators: essentially the
electric- and magnetic-field strength squared. An argument was proposed (Bochkarev and Shaposhnikov, 1986)
that the correction is connected with the average at small
T o v e r the pion state ( T T I G ^ J T T ) are small because they
are proportional to the pion mass. 6 0 In Adami, Hatsuda,
and Zahed (1991) lattice data were used to conclude that
although these quantities decrease with T, the decrease
occurs near Tc and is roughly a factor of 2. The IIA also
predicts a qualitatively similar behavior of the instantoninduced gluonic fields. It can be parametrized, for example, as follows:
((E2))

= ((H2))~exp(~T2/T2)

(5.25)

where T0 is of the order of Tc, but actually has nothing

to do with it. For example, there is a systematic suppression of the instantons of size pT> 1, but they certainly do
not disappear at the chiral symmetry restoration point Tc:
the so-called instanton-anti-instanton molecules do exist
even in the chirally symmetric phase.
The next important ingredient of the OPE-based sum
rules is specific four-fermionic operators. Their analogs at
finite T are, generally speaking, a set of all symmetric
tensors, and the corresponding O P E formulas are very
complicated. 6 1 In Bochkarev and Shaposhnikov (1986),
estimates of their expectation values were obtained as fol-

60

We recall that due to an anomaly relation this operator can

be considered as a trace of the energy-momentum tensor.
61
The relevant formulas were derived for higher twist corrections to deep-inelastic scattering (Jaffe and Soldate, 1982;
Shuryak and Vainshtein, 1982a, 1982b).
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

39

lows: they were rewritten as a sum of squares of some

currents, with quantum numbers of different mesons. At
small F, one therefore may start with a simple pion loop
diagram, making a unique estimate of the T dependence
of the corresponding matrix element. For example, for
the operator 0 entering the p-meson sum rules, this
pion loop leads to a negative correction; so its expectation value diminishes with temperature.
Another simple way of reasoning (Furnstahl, Hatsuda,
and Lee, 1990; Adami et al., 1991) leading essentially to
the same conclusion is as follows. According to the vacuum dominance hypothesis (Shifman et ah, 1979b), VEV
of four-fermionic operators can be expressed via quark
condensate squared. If so, it is natural to expect all of
them to vanish at the chiral symmetry restoration point
T=TC. The specific parametrization usually used is then
fermion

04

fermion(T

<r>
=0)

^m 2
^(0)2
i-r

/r

(5.26)

The last parametrization is used because it is simple and

reproduces lattice data.
3. Some results
We now apply the above approximations to the vector
and axial correlators. Specifically, we start with the O P E
expression at T = 0, modify the quark loop with the factor f2(irTR ), and take the condensates to have a temperature dependence given by Eq. (5.26). The results are
shown in Fig. 29. We see that both vector and axial
correlators converge rapidly and join smoothly to the expected high-temperature curves. Next, one can make fits
to such curves 6 2 to a parametrized spectral density, deriving the temperature dependence of the parameters involved.
As discussed earlier, the spectral density is usually expressed with three parameters, the resonance mass m, its
coupling to the current / , and the threshold energy E0.
Bochkarev and Shaposhnikov (1986) first analyzed the
temperature dependence of the p channel in this way, and
they concluded that all three parameters decrease with T,
with E0 dropping much faster than the others. However,
several later papers addressing the same issue (Dosch and
Narison, 1988; Furnstahl et al., 1990; Adami et a/.,1991)
found somewhat weaker temperature dependence of all
three parameters in the vector channel, still with the
main T dependence seen in E0. Figure 30 shows the typical results taken from Adami et al. (1991).

62

Actually, it is traditionally done in Borel representation,

with the so-called Borel parameter playing the role of distance
in our approach. Let us repeat that there are no real advantages behind this trick, which makes the whole presentation
much less transparent.

Edward V. Shuryak: Correlation functions in the QCD vacuum

40
1.0

0.6

j*

0-4

0.8

1 T

(a)
Model II

0.2
I

0.00.0I

0.4

J L.

0.6

1.0

T/Tc

2.5
2.0

P
EF

,,..., i

I ' 'J A A 1 ' ' i

~~-~l

>^

o \

|-

CO

0.5
0.0

b
t.. J

o \ j

~E,

Model II
i

0.0

1 t
0.2

(b)

i
0.4

i i i i i
0.8

0.6

1.0

T/Tc
0.08
0.06
.> 0.04

F7
h

t j

1_|.

-5
-^ (c)
H

p
"

0.02 ZL
0.00

Model II

t i
0.0

-"

u a

'_

1 1 1 1 1 ! 1
0.4
0.2

1 i
0.6

1 i
0.8

i H

1.0

T/Tc

FIG. 29. Schematic OPE predictions for vector (solid lines) and
axial (dashed ones), given by the ratio of the correlation function to that for free quarks vs distance x (in fm). Four solid
lines correspond to TQ, 140, 200, 400 MeV (from upper curve
down), and the dashed lines correspond to T = 0 and 140 MeV
(in the last two cases, above Tc, they coincide with solid ones).

FIG. 30. Solid lines show predictions of the temperature dependence of the p meson mass and of the threshold parameter of
the continuum S0 and the coupling constant Fpi derived in
Adami et ah (1991) from the OPE-based sum rules. (The points
are a slightly different calculation and can be disregarded.) The
"Model I I " used corresponds roughly to a condensate
modification discussed in the text and to the correlation function shown in Fig. 29.

compared to the pion gas of comparable density, and

makes
the nucleon lighter. Whether the next-order
Effects in the p channel may be weaker than the other
corrections
produce positive shifts in the nucleon mass,
channels, because the p correlator is close to free-quark
as would be required to saturate nuclear matter (see
propagation, as we saw in Sec. II. It would be interesting
Drukarev and Levin, 1990), is still unclear.
and relatively simple to generalize t h e analysis done for
The current state of the art of Q C D sum rules at finite
vector current to the axial case, where we have IT, A x
T
does not allow firm conclusions. However, there is
contributions and nonperturbative corrections acting in
some understanding of how the correlation functions are
the same direction.
Another suggestion is t o focus on ca,<f> mesons, for modified with temperature, and also what physical meaning these modifications may have. A lot of work, both
reasons related to experiment (Shuryak, 1991). Their
analytic and numerical, is needed and is currently under
widths are small enough so that even a relatively small
way.
shift of the mass could be observable. On the other hand,
their widths are large enough t o give them a chance to
decay inside the "fireball" created in heavy-ion collisions.
VI. SUMMARY AND DISCUSSION
Even in the environment of heavy-ion collisions, one can
+
observe two decay modes, e e~ and K*K~, and thus
A. Summary of phenomenological observations
measure the coupling to the current. Therefore in these
cases the resonance modification with temperature can be
In Sec. I I we discussed the phenomenology of mesonic
subjected to direct experimental testing.
correlation functions and showed that existing experiFinally, consideration has also begun of correlation
mental data not only fixed the logarithmic derivative of
functions in dense nuclear matter: Drukarev and Levin
the correlation functions at large distances, the hadronic
(1990), Cohen, Furnstahl, and Griegel (1991), and Hatsumasses, but, in several important cases, they also providda and Lee (1991). There are corrections of first order in
ed a good description of the whole function.
density that also contain certain nucleon matrix eleUsing these data one can conclude that a realistic qq
ments. Fortunately, these can be determined empirically
interaction is much more complicated than just a univerfrom deep-inelastic scattering. T h e results suggest that
sal confining potential. Various channels show very
nuclear matter produces significantly larger corrections,
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

Edward V. Shuryak: Correlation functions in the QCD vacuum

different trends, and their deviations from a perturbative
picture of free-quark propagation in some cases occurs at
such small distances, where confinement effects are yet
unimportant. Several phenomena have been observed, in
particular:
(1) We have seen the superduality in the p and other
vector channels, giving unexpectedly small deviations
from the free-quark behavior. These deviations are
within 1 0 - 2 0 % up to very large distances 1-1.5 fm,
while the correlation function changes in this interval by
several orders of magnitude. In other words, in all vector
cases all nonperturbative corrections in this range of distance are remarkably canceled. N o explanation for this
phenomenon is known.
(2) It was shown that p and co correlators are identical
within errors in a wide range of distances. Not only the
masses of these particles, but also the coupling constants
and even the production amplitudes of the nonresonance
states are very similar. This means that the famous
Zweig rule forbidding flavor mixing is unexpectedly strict
in the vector channels: it holds up to distances x ~ 2 fm,
where the correlations are extremely small.
(3) The last interesting observation related to vector
channels is that even for the K*^
channels involving
strange quarks a similarity to all other vector correlators
persists up to distances of about 1 fm. The effect of
larger masses is partially compensated by larger coupling
constants, and all curves go together as shown in Fig. 31.
In other words, all corrections proportional to the
strange quark mass are canceled separately.
Taken together, observations (l)-(3) show that all
point-to-point vector correlators in coordinate representations are more similar than the cross sections from

41

which they were calculated. This clearly demonstrates

our main point: these correlators are more fundamental
objects than particular hadronic states and their excitation amplitudes involved.
(4) Comparing the axial and vector correlators, one observes quite different behavior.
This axial-vector
difference is due to chiral asymmetry of the Q C D vacuum and should therefore gradually decrease with temperature, disappearing at the chiral restoration point T=TC.
(5) The octet pseudoscalar correlators, including the TT,
K, and 17 channels, are very similar at distances up to
x = 0 . 5 fm. In all cases the effect of a very strong qq attraction is seen already at x ~ y fm.
(6) The SU(3) singlet (77') pseudoscalar correlator shows
completely different behavior. N o trace of attractive
quark interaction is observed, at least at distances x > 0.5
fm, where it is presumably dominated by the evaluated 17'
contribution.
Let us also point out that in none of the pseudoscalar
channels is any trace of the Zweig rule seen: on the contrary, the flavor-changing amplitudes seem to be the
dominant effect.
(7) Radiative J /xf) decays provide valuable information
on the 77,77',^(1440) couplings to gluonic
pseudoscalar
operator GG. Deviations from asymptotic freedom in this
case take place at very small distances x ~ \ fm, as in
quark-related pseudoscalar channels. Similarly, also the
sign of these deviations indicates the presence of a strong
attraction.
(8) Qualitative comparison of all four spin-zero channels (scalars and pseudoscalars, octets and singlets) made
in Sec. II.H leads to the following conclusion: all deviations from the asymptotic freedom can presumably be explained by the dominance of one particular amplitude
(denoted K~~) changing both quark flavor and chirality.
B. What new experiments are needed?

x(fm)

FIG. 31. Phenomenological information on the various mesonic correlation functions discussed (same as in Fig. 2, etc.).
Different vector correlators were derived from completely
different sets of data, but they are very consistent with one
another and demonstrate a systematic trend.
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

Electromagnetic and weak currents are the only local

probes available in nature, and their relations to fundamental quark fields are by now well established. The
cross sections of e + e ~ h a d r o n s and r>v r +hadrons
are giving us fundamentally important information, in
particular, information about the correlation functions in
the Q C D vacuum discussed in this review.
In view of significant efforts devoted to understanding
Q C D (heavy-ion collider projects like R H I C , or new
large-scale lattice simulations like the T E R A F L O P project), one may also think about new generations of e + e ~
and r-lepton-related experiments providing more accurate data than those available today. Discussion of various c and b "factories" is under way in many places
around the world, which may be the basis for some optimism in this respect. Here are some possibilities related
to our discussion.
(1) In principle it seems feasible to achieve an experimental accuracy of a few percent in the e+e~ cross section resolved into 7 = 0 , 1 channels. If so, one will have

42

Edward V. Shuryak: Correlation functions in the QCD vacuum

vector (p,co) correlation functions with similar accuracy.

At the moment, the uncertainty in the correlators is
dominated by the poor accuracy ( ~ 3 0 % ) of the data in
the energy region E = 1.5-3 GeV. This corresponds to
the very interesting intermediate-distance
region,
x = 0 . 2 - 0 . 6 fm. If the gap is filled, we shall be able to
tell whether these correlators follow the O P E fomulas
and, if so, to measure directly the vacuum expectation
values of the operators involved.
(2) A new generation of r lepton hadronic decay experiments is very interesting as the best source of information on axial and strange vector correlators. The particular region of hadronic states with invariant mass above
the A! mass is of main interest, since the available data
are very poor in this region.
(3) Our discussion in the 17' correlator, the axial anomaly, and the gluonic matrix elements was based on a rather old set of data on J/ip radiative decays. These studies,
done mainly at SPEAR a decade ago, have found many
interesting phenomena, but many related questions
remain unanswered. As it is the best way we know to approach the mysterious world of gluonic states and corresponding matrix elements, further experiments are
justified.
(4) Much better data on bb production above the BB
threshold can help to measure the magnitude of the
strong Coulomb potential. This is a new and potentially
fruitful method of measuring A Q C D . Observable effects
are large, but they are of a perturbative nature and allow
for reliable evaluation by standard methods.
(5) Although the t quark seems to be too heavy to form
narrow states, the shape of the ft production threshold
can still be used to measure Coulomb-type strong forces.
(6) Last, but not least, hadron modification at finite
temperatures and densities can be studied in high-energy
collisions of heavy ions. Instead of lepton annihilation
into hadrons, one observes the inverse process, a hadronic production of lepton pairs, to detect the melting of the
p , o)9 <f>, and $ mesons.
C. Further lattice studies
The main suggestion for immediate study is as follows:
It would be very interesting to supplement current efforts
oriented to measurements of hadronic masses by a set of
data on point-to-point correlation functions. The most
interesting region is at intermediate distances x = 0 . 2 - 1
fm, which corresponds to a few lattice spacings. For obvious reasons, such data can be statistically more accurate and also less affected by the finite-size effects,
presumably making their comparison with phenomenology more conclusive. Some details of this suggestion are
discussed below.
(1) Generally speaking, any integration of the correlation functions leads to some loss of information, and
measurements with nonlocal sources certainly cannot tell
us much about the short-range interaction of quarks and
gluons. Separation of the contributions of short- and
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

long-range correlations in the vacuum is the natural

thing to do. In particular, it is important to understand
at which distances one should trust OPE-based expressions in various channels.
(2) It has been shown that several correlation functions
are experimentally known at all distances; so their comparison with lattice data can provide a much better test.
In particular, one may wonder if lattice simulation can
reproduce such delicate phenomena as superduality in
the vector channels, their splittings from the axial one, or
short-range attraction in pseudoscalar channels.
(3) In this paper the ratio of the correlation functions
to those corresponding to free-quark motion is systematically used. If one normalizes the lattice data in a similar
way, some systematic errors, such as finite-size corrections, can be canceled or reduced. Moreover, using such
normalized correlators, one can avoid the problem of the
absolute scale of lattice units.
(4) It would be very interesting to study flavorchanging correlation functions, the so-called two-loop
diagrams have not yet been studied because of the problems with statistics, but presumably correlation functions
at intermediate distances are still measurable. Hadronic
phenomenology tells us that these correlators are strongly suppressed in all vector channels, but enhanced in the
pseudoscalar ones. It would be interesting to see whether
L G T reproduces these observations, even qualitatively.
(5) Studies of light-heavy systems can easily be extended to various angular momenta and parities, but so far
most of the work has been concentrated on the pseudoscalar channel. It was argued above that the pattern of
splittings of such mesonic correlators in parity could
shed some light on the old question of the applicability of
the constituent quark model. The point is whether the
model can describe both the spin-flip and the spin-nonflip
part of the propagator. Data on the light-heavy baryons
can similarly clarify the properties of the qq interaction.
In particular, studies defining at which distances and how
the X and A-type correlators become different can tell a
lot about the mechanism of spin splitting in baryons.
There are two candidates: gluon exchanges and
instanton-induced forces, and on the lattice one can invent a number of ways to tell the difference between
them. For example, one may study spin splittings by applying the so-called lattice cooling, thereby killing the
perturbative component but preserving instantons.
(6) A technical point: the three-parameter fit for the
spectral density, with a 8 plus a 6 function, has proved to
be very accurate in most cases studied. It is even more
accurate than a four-parameter fit with two exponents
used in some lattice works, because it has both long- and
short-distance limits right. We recommend using it as
the standard parametrization of the correlators.
D. Theoretical problems
The list of questions formulated below is certainly related to phenomena discussed above, but they are listed

Edward V. Shuryak: Correlation functions in the QCD vacuum

in order of their theoretical importance, from the more
general to the more specific.
(1) Why are the quark-made hadronic states much
lighter than the glueballs? Or, speaking in terms of the
correlation functions, why are gluonic fields uncorrelated
at much smaller distances than the quark ones? Certainly, there should be a big difference between the spacetime distribution of gluon and quark fields in the Q C D
vacuum. One example is provided by IIA, which suggests a picture of the vacuum as a very inhomogeneous
instanton liquid. According to it, gluonic fields are concentrated in small fluctuations, the instantons, while
quarks have another role: they j u m p from one instanton
to another, filling the whole space-time more or less
homogeneously. L G T also reproduces these qualitative
features of correlators, but it is at the moment very
difficult to say whether it is consistent with this explanation.
(2) It was argued earlier that correlation functions tell
us that, in fact, qq and qq interaction is much more complicated than just simple universal confining forces.
After many years of studies of individual hadrons, only
now is an attempt being made to understand first the
much simpler objects, the small-size wave packets that
are the intermediate states of the correlators. The longrange effects (confinement) are in this case much less important, but some others (like spin-spin interactions) are
enhanced. We have found the strongest deviation from
perturbative behavior at small distances in the case of octet pseudoscalars, where attraction dominates at distances as small as \ fm. The O P E formulas do not reproduce the effect.63 What is the physical nature of the shortrange qq attractive interaction?
(3) Considering these data (which are also combined
with some limiting information about other spin-zero
channels, the r( one and the scalars), we have concluded
that the quantum numbers of them in corrections point
toward the amplitude K
, which changes both chirality
and flavor of participating quarks. The quantum numbers of this interaction fit into the
instanton-induced
't Hooft interaction; so it is the best candidate we know.
However, to make our arguments convincing, it would be
interesting to measure all these correlators on the lattice.
Is there a window in which one may use the 9t Hooft interaction in first order, avoiding complicated IIA calculations?
(4) Now we come to a set of questions related to the qq
interaction and baryons. Why is the nucleon so light,
compared to current lattice calculations? Is the OPEbased conclusion really true, and is there a maximum in
KN(x)/K$ee(x)
in which the nucleon contribution
is
several times larger than the perturbative one? Is it also

63

Unless one increases VEV of some four-fermion operators by

more than one order of magnitude compared to the "standard"
estimates.
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

43

true for the I^(L) correlator? More generally, how do

the spin-splitting interactions depend on interquark distances?
(5) Let us ask some questions about the propagation of
a single quark in the Q C D vacuum, assuming such questions can be given physical meaning. Does the constituent
quark model make sense? Is it indeed true that quarks
are "dressed smoothly," with Z factors close to 1, and
more or less independently from one another? Is it true, as
is sometimes conjectured, that the size of the constituent
quark is significantly smaller than 1 fm, a hadronic radius?
(6) Modification of all correlators with temperature
and density is a vast region for investigation. In particular, the correlators were decomposed above out of several
components: resonance and nonresonance hadronic
states. Do these components depend on temperature similarly, or quite differently? What is "melted" first, the resonance contribution or the continuum threshold?
Which
correlation functions
are discontinuous following
the
chiral or confinement phase transitions? What is the nature of strong deviation of the scalar-pseudoscalar screening length from 2irT, which is not observed in other channels?
In conclusion of this review, let us return to a general
point considered in the Preface. The QCD-related studies is a vast field. Many particular problems were analyzed in detail, and we have a vast phenomenology and
great potential for better experiments and lattice simulations. Even so, these new efforts will be more successful
if they are also supplemented by some work aimed on
their synthesis, consolidating all studies to the common
set of observables and problems. Whatever goals this paper has reached, it is an attempt in this direction.
Notes added in proof
Since this paper was submitted, some extremely important new developments have taken place in the field, and
in this note we cite the main ones.
One of the main suggestions of this paper is to perform
detailed lattice studies of point-to-point correlation functions. Now the first such measurements have been made
by (Chu, Grandy, Huang, and Negele, 1992). Another
idea emphasized above was to get much more accurate
results from the instanton-based models, and those are
also now available, both for the simplest case or "random
instanton liquid" (Shuryak and Verbaarschot, 1992a,
1992b, 1992c), and interacting instantons (Shuryak and
Verbaarschot, I992d).
A selection of the most important data for lattice and
random instantons is shown in Fig. 32: they are surprisingly consistent with each other in all cases. For mesons,
they are also very close to phenomenological expectations. For baryons, both show a remarkably different behavior for nucleons and A's, strongly supporting
Chernyak-Zhitnitsky
predictions
(Farrar,
Zhoang,
Ogloblin, and Zhitnitsky, 1981) over Belyaev-Ioffe ones

Edward V. Shuryak: Correlation functions in the QCD vacuum

44

OUVJ

100
30
10
3
1

1 '
/

*J>*

1.5

.5

1.5

very critical. A recent paper (Hatsuda, Koike, and Lee,

1992) has addressed this criticism and deals with the
problem in a much more consistent way. However, it
still is incomplete (tensor operators are not really included) and, by construction, restricted to the very-low- T region. Therefore, we still think some main conclusions of
this paper (e.g., qualitatively different modification of p
and co mesons, predicted there) still may be revised by
further works.

1 ' 1*"-

'

="^T

ACKNOWLEDGMENTS

1
5

r 1I

.5
8 -'N
6

1 ' 1

ri
1.5

P l

.1 _

4
2
n

T T.5,

1i

1
x

1 ~

1.5

.5

I "I

1
x

1-5

FIG. 32. Correlation functions for the isovector pseudoscalar

(77), vector (p), scalar (8), axial-vector {Ax), nucleon (N), and
delta (A) channels, as a function of distance x in fm, and normalized by corresponding correlators for free massless quarks.
Squares and triangles show lattice results by Chu et ah (1992)
and random instanton vacuum (Shuryak and Verbaarschot,
1992b, 1992c), respectively. Solid lines correspond to experimental data; long-dashed (Farrar et ah, 1981) and short-dashed
(Belyaev and Ioffe, 1982) ones show different predictions of the
QCD sum rules.

(Belyaev and IofFe, 1982). As argued above, this provides

a completely new perspective in our understanding of
baryon structure.
At the same time more detailed studies of interacting
instantons (Shuryak and Verbaarschot, 1992d) have revealed a problem: the "streamline"-generated interaction
leads to too strongly bound instanton-anti-instanton
pairs, which spoils a good description of correlators
reached in a random ( = noninteracting) instanton vacuum. However, an additional repulsive core-type interaction does solve a problem, and leads to even better agreement with data, especially for scalar and 77' channels.
New important results concerning the quark-gluon
plasma phase have been reported by Boyd, Gupta, and
Karsch (1992). In agreement with our arguments, they
have clearly observed the appearance of effective quark
mass TTT if the quark propagates in spatial directions, but
only small perturbative mass for quarks propagating in
the time direction.
Our last comment deals with Q C D sum rules at low
temperatures: our discussion of all these works above is
Rev. Mod. Phys., Vol. 65, No. t , January 1993

Most of what I have learned about correlation functions has been gained through discussions with my old
friends A. I. Vainshtein, M. A. Shifman, V. I. Zakharov,
and V. L. Chernyak. The idea for writing this paper
presented itself naturally, because its substance was part
of a course on nonperturbative Q C D in Stony Brook in
1990, a project which would never have materialized
without the practical help and everlasting curiosity of G.
E. Brown. I am also much indebted to S. I. Eidelman,
who supplied relevant experimental data, and to G.
Bertsch and J. Verbaarschot, who took on the painful
task of reading this voluminous manuscript and making
numerous suggestions. I should also mention that this
paper was finished at the Aspen Summer Institute, which
I thank for its hospitality. This work is partly supported
by the U.S. Department of Energy under Grant No. D E FG02-88ER40388.

REFERENCES
Adami, C , T. Hatsuda, and I. Zahed, 1991, Phys. Rev. D 43,
921.
Adami, C , I. Zahed, 1990, "Finite temperature QCD sum rules
for the nucleon," SUNY preprint NTG-90-37.
Adler, S. L., 1969, Phys. Rev. 177, 2426.
Albrecht, H., et ah, 1986, Z. Phys. C 33, 7.
Allton, C. R., C. T. Sachrajda, V. Lubicz, L. Maiani, and G.
Martinelli, 1991, Nucl. Phys. B 349, 598.
Appelquist, T., and J. Carazzone, 1975, Phys. Rev. D 11, 2856.
Appelquist, T., and H. D. Politzer, 1975, Phys. Rev. Lett. 34,
43.
Asakawa, M., and K. Yazaki, 1989, Nucl. Phys. A 509, 608.
Barkov, L. M., et ah, 1985, Nucl. Phys. B 256, 365.
Barkov, L. M., et ah, 1988, Sov.. J. Nucl. Phys. 47, 248.
Belavin, A. A., A. M. Polyakov, A. A. Schwartz, and Y. S. Tyupkin, 1975, Phys. Lett. B 59, 85.
Bell, J. S., and R. Jackiw, 1969, Nuovo Cimento A 60, 47.
Belyaev, V. M., and B. L. Ioffe, 1982, Sov. Phys. JETP 83, 976.
Bernard, C , D. Murphy, A. Soni, and K. Yee, 1990, Nucl.
Phys. B Proc. Suppl. 17, 593.
Bernard, C , T. DeGrand, C. DeTar, S. Gottlieb, A. Krasnitz,
M. Ogilvie, R. Sugar, and D. Toussaint, 1991, "The spatial
structure of screening propagators in hot QCD," preprint
AZPH-TH/91-60.
Bernard, C , C. Heard, J. Labrenz, and A. Soni, 1992, "Decay
constants and wave functions of heavy-light pseudoscalars,"
Brookhaven National Laboratory preprint BNL-45097. Also
in Proceedings of Lattice 91, Tsukuba, Japan.

Edward V. Shuryak: Correlation functions in the QCD vacuum

Betman, R. G., and L. V. Laperashvili, 1985, Sov. J. Nucl. Phys.
41, 295.
Bochkarev, A. I., and M. E. Shaposhnikov, 1986, Nucl. Phys. B
268, 220.
Born, K. D., S. Gupta, A. Irback, F. Karsch, E. Laermann, B.
Petersson, and H. Satz, 1991, Phys. Rev. Lett. 67, 302.
Boueaud, P., et aL, 1988, "JS-meson mass and decay constant
from lattice QCD," CERN-TH.5269/88.
Boyd, G., S. Gupta, and F. Karsch, 1992, "The quark propagator at finite temperature/' preprint CERN-TH.6482/92.
Brown, G. E., 1991, Nucl. Phys. A 522, 397.
Brown, G. E., H. A. Bethe, and P. M. Pizzochero, 1991, Phys.
Lett. B263, 337.
Brown, F. R., F. P. Butler, H. Chen, N. H. Christ, Z. Dong, W.
Schaffer, L. I. Unger, and A. Vaccarino, 1990, Phys. Rev. Lett.
65,2491.
Brown, F. R., F. P. Butler, H. Chen, N. H. Christ, Z. Dong, W.
Schaffer, L. I. Unger, and A. Vaccarino, 1991, Phys. Rev. Lett.
67, 1062.
Callan, C. G., R. Dashen, and D. J. Gross, 1978, Phys. Rev. D
17,2717.
Capstick, S., S. Godfrey, N. Isgur, and J. Paton, 1986, Phys.
Lett. 175B, 457.
Capstick, S., and N. Isgur, 1986, Phys. Rev. D 34, 2809.
Chu, M.-C, J. Grandy, S. Huang, and J. Negele, 1992, "Lattice
calculation of QCD vacuum correlation functions," CTP 2113,
Phys. Rev. Lett., in press.
Chu, M.-C, and S. Huang, 1991, "The relevance of dilute instanton ensemble to light hadrons," preprint MAP-138.
Chu, M., M. Lissia, and J. Negele, 1991, Nucl. Phys. B 360, 31.
Cohen, T., R. Furnstahl, and D. K. Griegel, 1991, Phys. Rev.
Lett. 67, 961.
Cordier, A., D. Bisello, J.-C. Bizot, J. Buon, B. Delcourt, L. Fayard, and F. Mane, 1982a, Phys. Lett. B109, 129.
Cordier, A., D. Bisello, J.-C. Bizot, J. Buon, B. Delcourt, L. Fayard, and F. Mane, 1982b, Phys. Lett. Bl 10, 335.
Cosme, G., et aL, 1979, Nucl. Phys. B 152, 215.
DeGrand, T. A., and C. E. DeTar, 1986, Phys. Rev. D 34, 2469.
DeRujula, A., H. Georgi, and S. L. Glashow, 1975, Phys. Rev.
D 12, 147.
DeTar, C , and J. B. Kogut, Phys. Rev. D 36, 2828.
Dey, M., V. L. Eletsky, and B. L. Ioffe, 1990, Phys. Lett. B 252,
620.
D'Hoker, E., 1982, Nucl. Phys. [FS4] B 200, 517.
Diakonov, D. I., and V. Y. Petrov, 1984, Nucl. Phys. B 245,
259.
Dolinsky, S. I., et aL, 1989, "Review of e+e~~ experiments with
neutral detector
," INP preprints 89-68 and 89-104 (in
Russian), Novosibirsk.
Dosch, H. G., and S. Narison, 1988, Phys. Lett. B 203, 155.
Drukarev, E. G., and E. M. Levin, 1990, Nucl. Phys. A 511,
679.
Eichten, E., K. Gottfried, T. Kinoshita, K. D. Lane, and T. M.
Yan, 1980, Phys. Rev. D 21, 203.
Eletsky, V. L., 1990, Phys. Lett. B 245, 229.
Eletsky, V. L., and B. L. Ioffe, 1988, Sov. J. Nucl. Phys. 48, 384.
Farrar, G., H. Zhoang, A. A. Ogloblin, and I. R. Zhitnitsky,
1981, Nucl. Phys. B 311, 585.
Forte, S., and E. Shuryak, 1991, Nucl. Phys. B357, 153.
Furnstahl, R. J., T. Hatsuda, and S. H. Lee, 1990, Phys. Rev. D
42, 1744.
Gasser, J., and H. Leutwyler, 1987, Phys. Lett. B 184, 83.
Gell-Mann, M., and M. Levi, 1960, Nuovo Cimento 16, 705.
Gell-Mann, M., R. Oakes, and B. Renner, 1968, Phys. Rev. 175,
Rev. Mod. Phys., Vol. 65, No. 1, January 1993

45

2195.
Geshkenbein, B. V., and B. L. Ioffe, 1980, Nucl. Phys. B 166,
340.
Gocksch, A., 1991, "Chiral symmetry in hot QCD,''
Brookhaven National Laboratory preprint BNL-46286.
Godfrey, S., and N. Isgur, 1985, Phys. Rev. D 32, 189.
Gottlieb, S., W. Liu, D. Toussaint, R. L. Renken, and R. L. Sugar, 1987, Phys. Rev. Lett. 59, 2247.
Gross, D. J., R. D. Pisarski, and L. G. Yaffe, 1981, Rev. Mod.
Phys. 53, 43.
Hands, S., and M. Teper, 1990, Nucl. Phys. B 347, 819.
Hanson, T., and I. Zahed, 1991, Stony Brook preprint SUNYNTG-91-44.
Hatsuda, T., and S. H. Lee, 1991, preprint YSTP-91-10.
Hatsuda, T., Y. Koike, and S. H. Lee, 1992, "Finite temperature QCD sum rules reexamined," University of Maryland preprint 92-203.
Hernandez, J. J., et aL, 1990, Phys. Lett. B 239, 1.
Hwa, R., 1990, Ed., Quark-gluon plasma, Advanced Series on
Directions in High Energy Physics, Vol. 6 (WSPC, Singapore).
Ilgenfritz, E. M., and E. Shuryak, 1989, Nucl. Phys. B 319, 511.
Ioffe, B. L., 1981, Nucl. Phys. B 188, 317.
Isgur, N., 1989, Phys. Rev. D 39, 1357.
Isgur, N., and G. Karl, 1978, Phys. Rev. D 18, 4187.
Isgur, N., and M. B. Weise, 1991, Phys. Rev. Lett. 66, 1130.
Ivanov, P., L. M. Kurdadze, M. Yu, Lelchuk, V. A. Sidorov, A.
N. Skrinsky, A. G. Chilingarov, Yu. M. Shatunov, B. A.
Shwartz, and S. I. Eidelman, Phys. Lett. B 107, 297.
Jaffe, R. L., and M. Soldate, 1982, Phys. Rev. D 26, 106.
Khoze, V., and A. Ringwald, 1991, Phys. Lett. B 259, 106.
Koch, V., E. V. Shuryak, G. E. Brown, A. D. Jackson, 1992,
Phys. Rev. D 46, 3169.
Kochelev, N. L, 1985, Sov, J. Nucl. Phys. 41, 291.
Kochelev, N. I., 1990, Z. Phys. C 46, 281.
Kremer, M., A. Kronfeld, M. Laursen, G. Schierhiltz, C.
Schleiermacher and U. J. Wiese, 1988, Nucl. Phys. B 305, 109.
Kurdadze, L. M., M. Yu. LePchuk, E. V. Pakhtusova, V. A.
Sidorov, A. N. Skrinskii, A. G. Chilingarov, Yu. M. Shatunov,
B. A. Shvarts, and S. L. Eidel'man, 1988, JETP Lett. 47, 512.
Lattice 88, 1989, Nucl. Phys. B Proc. Suppl. 9, 1. Fermilab,
Chicago, 1988.
Lattice, 89, 1990, Nucl. Phys. B Proc. Suppl. 17, 1. Capri, Italy,
1989.
Lattice 90, 1991, Nucl. Phys. B Proc. Suppl. 20, 1. Tallahassee,
Florida, 1990.
Lattice 91, 1992, Nucl. Phys. B Proc. Suppl. 26, 1. Tsukuba,
Japan, 1991.
Li, S., R. S. Bhalrao, R. S. and R. K. Bhadury, 1991, Int. J.
Mod. Phys. A 6, 501.
Lubicz, V., G. Martinelli, M. McCarthy, and C. T. Sachrajda,
1992, Phys. Lett. B 274, 415.
Mackenzie, P., 1992, in "Lattice 91," Tsukuba, Japan.
Maiani, L., G. Martinelli, and C. T. Sachrajda, 1992, Nucl.
Phys. B 368, 281.
Mane, F., D. Bisello, J.-C. Bizot, J. Buon, A. Cordier, and B.
Delcourt, 1982, Phys. Lett. B 112, 178.
Manousakis, E., and J. Polonyi, 1987, Phys. Rev. Lett. 58, 847.
Martin, A., 1981, Phys. Lett. B 100, 511.
McLerran, L., 1987, Phys. Rev. D 36, 3291.
Novikov, V. A., M. A. Shifman, A. I. Vainshtein, and V. I. Zakharov, 1980, Nucl. Phys. B 165, 55.
Novikov, V, A., M. A. Shifman, A. I. Vainshtein, and V. I. Za~
kharov, 1981, Nucl. Phys. B 191, 301.
Novikov, V. A., M. A. Shifman, A. I. Vainshtein, and V. I. Za-

46

Edward V. Shuryak: Correlation functions in the QCD vacuum

kharov, 1982, Sov. Phys. Usp. 136, 553.

Novikov, V. A., M. A. Shifman, A. I. Vainshtein, and V. I. Zakharov, 1984, Fortschr. Phys. 32, 585.
Peccei, R., and J. Sola, 1987, Nucl. Phys. B 281, 1.
Politzer, H. D., 1974, Phys. Rep. 14, 130.
Quark Matter '88, 1989, Nucl. Phys. A 498, 1.
Quark Matter '90, 1991, Nucl. Phys. A 525, 1.
Reinders, L. J., H. Rubinstein, and S. Yazaki, 1983, Phys. Lett.
B 120, 209.
Reinders, L. J., H. Rubinstein, and S. Yazaki, 1985, Phys. Rep.
127, 1.
Shifman, M., 1992, Ed., Vacuum Structure and QCD Sum Rules
(North-Holland, Amsterdam), in press.
Shifman, M. A., 1986, Charmed and Beautiful Hadrons, International Symposium on Production and Decay of Heavy Hadrons, Heidelberg, 1986, Sov. Phys. Usp. 30, 91 (1987).
Shifman, M. A., A. I. Vainshtein, and V. I. Zakharov, 1979a,
Nucl. Phys. B 147, 385, 448, 519.
Shifman, M. A., A. I. Vainshtein, and V. I. Zakharov, 1979b,
Nucl. Phys. B 147, 385, 448, 519.
Shuryak, E., 1978a, Phys. Lett. B 79, 135.
Shuryak, E., 1978b, Zh. Eksp. Teor. Fiz. 74, 408.
Shuryak, E., 1980, Phys. Rep. 61, 72.
Shuryak, E., 1982a, Nucl. Phys. B 203, 116, 140, 237.
Shuryak, E., 1982b, Nucl. Phys. B 198, 83.
Shuryak, E., 1983, Nucl. Phys. B 214, 237.
Shuryak, E., 1984, Phys. Rep. 115, 151.
Shuryak, E., 1988a, The QCD Vacuum, Hadrons and the Superdense Matter (World Scientific, Singapore).
Shuryak, E., 1988b, Nucl. Phys. B 302, 559, 574, 599, 621.
Shuryak, E., 1988c, Nucl. Phys. B 302, 574.
Shuryak, E., 1988d, Nucl. Phys. B 302, 621.
Shuryak, E., 1989a, Nucl. Phys. B 328, 102.
Shuryak, E., 1989b, Nucl. Phys. B 319, 541.
Shuryak, E., 1989c, Nucl. Phys. B 319, 102.

Rev. Mod. Phys., Vol. 65, No. 1, January 1993

Shuryak, E., 1989d, Nucl. Phys. B 328, 85.

Shuryak, E., 1991, Nucl. Phys. A 533, 761.
Shuryak, E., and J. L. Rosner, 1989, Phys. Lett. B 218, 72.
Shuryak, E., and V. Thorsson, 1992, Nucl. Phys. A 536, 739.
Shuryak, E., and A. I. Vainshtein, 1982a, Nucl. Phys. B 199,
451.
Shuryak, E., and A. I. Vainshtein, 1982b, Nucl. Phys. B 201,
141.
Shuryak, E., and J. J. M. Verbaarschot, 1992, Phys. Rev. Lett.
68, 2576.
Shuryak, E., and J. J. M. Verbaarschot, 1992a, "Quark propagation in random instanton vacuum," Stony Brook preprint
SUNY-NTG-92-39.
Shuryak, E., and J. J. M. Verbaarschot, 1992b, "Mesonic correlation functions in random instanton vacuum," Stony Brook
preprint SUNY-NTG-92-40.
Shuryak, E., and J. J. M. Verbaarschot, 1992c, "Baryonic correlation functions in random instanton vacuum," Stony Brook
preprint SUNY-NTG-92-41.
Shuryak, E., and J. J. M. Verbaarschot, 1992d, "Interacting instantons in the QCD vacuum," Stony Brook preprint SUNYNTG-92-42.
Shuryak, E., and O. V. Zhirov, 1987, Nucl. Phys. B 292, 714.
Takeuchi, S., and M. Oka, 1991, Phys. Rev. Lett. 66, 1271.
Teraflop, 1992," Physics Goals of the QCD Teraflop Project,"
Int. J. Mod. Phys., in press.
't Hooft, G., 1976, Phys. Rev. D 14, 3432.
Toussaint, D., 1992, Talk given at Lattice 91, Tsukuba, Japan.
Verbaarschot, J., 1991, Nucl. Phys. B 362, 33.
Weinberg, S., 1967, Phys. Rev. Lett. 18, 507.
Weinberg, S., 1975, Phys. Rev. D 11, 3583.
Weingarten, D., 1983, Phys. Rev. Lett. 51, 1830.
Wilson, K. G., 1969, Phys. Rev. 179, 1499.
Witten, E., 1983, Phys. Rev. Lett. 51, 2351.
Yung, A., 1988, Nucl. Phys. B 297, 47.