Theoretical Physics Institute

University of Minnesota

TPI-MINN-98/01-T
UMN-TH-1622-98
January 1998
arXiv:hep-ph/9802214v1 2 Feb 1998

Snapshots of Hadrons
or the Story of How the Vacuum Medium Determines the
Properties of the Classical Mesons Which Are Produced, Live
and Die in the QCD Vacuum

M. Shifman
Theoretical Physics Institute, Univ. of Minnesota, Minneapolis, MN 55455

Lecture given at the 1997 Yukawa International Seminar Non-Perturbative QCD –

Structure of the QCD Vacuum, Kyoto, December 2 – 12, 1997.
Contents
1 QCD Sum Rules: Twenty Years After 2

2 QCD Vacuum and Basics of the SVZ Method 3

2.1 General ideas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Getting started/Playing with toy models . . . . . . . . . . . . . . . . 7

3 Vacuum Condensates 22

4 ρ Meson in QCD 32

5 Basic Theoretical Instrument – Wilson’s OPE 40

6 Practical Version of OPE 45

7 Low Energy Theorems 48

8 Are All Hadrons Alike? 51

9 Ecological Niche 53

10 New Developments 54
10.1 Light-cone sum rules . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
10.2 Heavy flavor sum rules . . . . . . . . . . . . . . . . . . . . . . . . . . 59

11 Sum Rules and Lattices 63

12 Vacuum Fluctuations Are Subtle Creatures 65

13 Instead of Conclusions 68

14 Appendix. Some Useful Definitions and Formulae. 70

1
1 QCD Sum Rules: Twenty Years After
I will discuss a method of treating the nonperturbative dynamics of QCD which
was created almost twenty years ago [1] in an attempt to understand a variety of
properties and behavior patterns in the hadronic family in terms of several basic
parameters of the vacuum state. The method goes under the name QCD sum rules
– rather awkward, for many reasons. First and foremost, it does not emphasize
the essence of the method. Second, in Quantum Chromodynamics there exist many
other sum rules, having nothing to do with those suggested in Ref. [1]. Finally,
some authors add further confusion by using ad hoc names, e.g. the Laplace sum
rules, spectral sum rules, and so on, which are even foggier and are not generally
accepted.
It would be more accurate to say “the method of expansion of the correlation
functions in the vacuum condensates with the subsequent matching via the dis-
persion relations”. This is evidently far too long a string to put into circulation.
Therefore, for clarity I will refer to the Shifman-Vainshtein-Zakharov (SVZ) sum
rules. Sometimes, I will resort to abbreviations such as “the condensate expansion”.
Twenty years ago, next to nothing was known about nonperturbative aspects of
QCD. The condensate expansion was the first quantitative approach which proved
to be successful in dozens of problems. Since then, many things changed. Various
new ideas and models were suggested concerning the peculiar infrared behavior
in Quantum Chromodynamics. Lattice QCD grew into a powerful computational
scheme which promises, with time, to produce the most accurate results, if not for
the whole set of the hadronic parameters, at least, for a significant part.
It seems timely to survey the ideas and technology constituting the core of the
SVZ sum rules from the modern perspective, when the method became just one
among several theoretical components in a modern highly competitive environment.
An exhaustive review of a wealth of “classical”, old elements of the method and
applications was given in Ref. [2]. There is hardly any need in an abbreviated
version of such a report. New applications which were worked out in the last decade
or so definitely do deserve a detailed discussion. As far as I know, no comprehensive
coverage of the topic exists in the literature. Unfortunately, in these lectures I will
not be able to provide such a coverage, which thus remains a task for the future
1
. Instead, I will focus on those qualitative aspects where understanding became
deeper. This is the first goal. Secondly, selected new applications will be considered
to the extent that they illustrate the theoretical ideas of the last decade. And last
but not least, I will try to outline an ecological niche which belongs to the SVZ
method today. As a matter of fact, over the years, slow but steady advances were
taking place in our knowledge of the hadronic world. Some old and largely forgotten
1
Work on systematically reviewing a variety of developments that took place since the mid-
1980’s and numerous new applications is under way (a private communication from B.L. Ioffe).
A survey devoted to the relation between the sum rule and lattice results is being written by A.
Khodjamirian.

2
predictions of the SVZ sum rules were confirmed recently by other investigations
based on totally different principles. These predictions are extremely nontrivial.
For instance, about 15 years ago, it was discovered [3] that not all hadrons are alike;
there are remarkable distinctions between them, especially in the glueball sector.
The fact that not all hadrons are alike is now becoming more and more evident
from the lattice results as well. Other examples of this type are known too. By
confronting them with alternative sources of information, such as lattices, we get
a much fuller picture of the QCD vacuum. This process might be very beneficial
to both sides. Unfortunately, at present the lattice and analytic QCD communities
are largely disconnected, and rarely talk to each other. My task is to show that
borrowing from each other makes everybody richer!

2 QCD Vacuum and Basics of the SVZ Method

2.1 General ideas
The color dynamics described by QCD is very peculiar. If we have two probe color
charges, their interaction approaches the Coulomb law at short distances, with a
weak coupling constant. The Coulomb interaction is due to the one gluon exchange
(Fig. 1). At larger distances, the gluon starts branching (Fig. 2), which leads to
a remarkable phenomenon known as antiscreening, or asymptotic freedom [4]. In
normal theories, like QED, the virtual cloud screens the bare charge making the
charge seen at larger distances smaller than the bare one. This situation is perfectly
transparent intuitively.
In QCD, instead of screening, the branching processes result in a totally counter-
intuitive behavior – the antiscreening. Those of you who would like to know the
physical origin of antiscreening are referred to the very pedagogical review [5], Sect.
1.3. If the distances are not too large one can apply perturbation theory to quanti-
tatively describe the gluon branchings, and derive the famous formula of asymptotic
freedom
r0
αs = Const/ ln (1)
r
where r0 is a dynamically generated scale parameter of QCD, r0 ∼ 1 fm. At small
separations the effective coupling constant dies off logarithmically.
Usually one considers the running coupling constant in the momentum space.
Then in the leading (one-loop) approximation corresponding to Eq. (1),
4π 11 2
αs (µ2 ) = 2 2
, b = Nc − Nf , (2)
b ln(µ /Λ ) 3 3
where Nc is the number of colors and Nf is the number of flavors and b is the first
coefficient in the Gell-Mann-Low function. In the low-energy domain we will be
interested in (below the threshold of the charm production) the number of active
flavors is Nf = 3, and, hence, b = 9.

3
Figure 1: Heavy color charges (solid lines) interacting through the one-gluon ex-
change (dashed line).

In perturbation theory the effective coupling is calculated order by order. The

result is known up to three loops, and for Nc = Nf = 3 can be conveniently written
as [6]
2
 
ln ln µ 2 (0.79)2 µ2 µ2
" #
1 1
a(µ2 ) = µ2 −0.79 2 µΛ2 + 3 µ2 (ln ln 2 )2 − ln ln 2 + 0.415 +O  4 µ2  ,
ln Λ2 ln Λ2 ln Λ2 Λ Λ ln Λ2
(3)
where we define
b αs (µ2 )
a(µ2 ) ≡ , (4)
4 π
and Λ is the scale parameter of QCD introduced in a standard way [6]. In the third
and higher loops the law of running of αs becomes scheme-dependent. The third
term in Eq. (3) refers to the so called modified minimal subtraction (MS) scheme.
More exactly, since Eq. (3) describes running with three flavors, the parameter Λ
(3) (3)
is actually Λ . Below we will deal exclusively with Λ ; therefore, not to make
MS MS
the notation too clumsy, we will suppress the sub(super)scripts.
Let us leave a while the issue of the effective coupling constant, with the intention
of returning to it later. The only lesson one should remember at this stage is that
the effects caused by perturbative gluon exchanges are logarithmic.
Equations (1) or (2) imply that the effective color interaction becomes stronger
as the separation between the probe color charges increases. Being remarkable by
itself, this phenomenon carries the seeds of another, even more remarkable property
of QCD. When the distance becomes larger than some number times Λ−1 , and
exceeds a critical one, the branchings of gluons become so intensive (Fig. 3) that it
makes no sense to speak about individual gluons.
Rather, we should phrase our consideration in terms of the chromoelectric and
chromomagnetic fields. In “normal” theories, like QED, the field induced by the
probe charges separated by a large distance, is dispersed all over space. In QCD
it is conjectured that a specific organization of the QCD vacuum makes such a
dispersed configuration energetically inexpedient [7]. Rather, the chromoelectric

4
Figure 2: Gluon branching responsible for the logarithmic running of the effective
gauge coupling constant in perturbation theory (shown is a sample one-loop graph).

Figure 3: When the separation between the probe color charges exceeds critical the
nonlinearity becomes so strong that it makes no sense to speak about individual
gluons. The color fields between the probe charges form a flux tube with transverse
dimensions ∼ Λ−1 .

5
field between the probe charges squeezes itself into a sausage-like configuration. The
situation is reminiscent of the Meissner effect in superconductivity. As is well-known,
the superconducting media do not tolerate the magnetic field. If one imposes, as
an external boundary condition, a certain flux of the magnetic field through such
a medium, the magnetic field will be squeezed into a thin tube carrying all the
magnetic flux; the superconducting phase is destroyed inside the tube, and there is
no magnetic field outside.
Superconductivity is caused by condensation of the Cooper pairs – pairs of elec-
tric charges. A phenomenologically acceptable picture of the infrared dynamics in
QCD requires chromoelectric flux tubes, not magnetic ones. If something like con-
densation of “chromomagnetic monopoles” took place in the QCD vacuum then this
would naturally explain nonproliferation of the chromoelectric field from the exter-
nal probe color charges in the entire space – a dual Meissner effect would force the
field to form flux tubes in the QCD vacuum, in this way ensuring color confinement
[7].
Nobody ever succeeded in proving that the dual Meissner effect does indeed take
place in QCD. There are reasons to believe, however, that it does. First, some
evidence came from the numerical study in the lattice QCD (see e.g. [8]). Quite
recently, a breakthrough [9] was achieved in understanding of the infrared dynamics
in a theory which might be considered a relative of QCD, the so called N = 2
supersymmetric (SUSY) Yang-Mills theory, in which it was analytically shown that
the monopoles do condense in the strong coupling regime, thus providing a basis for
the dual Meissner mechanism.
Even if, as a result of a breakthrough sometimes in the future, it becomes clear
how a similar mechanism might develop in QCD, we have a long way to go from a
qualitative picture of the phenomenon to a quantitative approach allowing one to
exactly calculate all the variety of the hadronic properties from the first principles.
It may be that such string-based exact approach will never emerge. While the issue
of the exact solution is still under investigation, let us see what analytic QCD can
offer as an approximate solution.
The basic idea lying behind the SVZ method is quite transparent. The quarks
comprising the low-lying hadronic states, e.g. classical mesons or baryons, are not
that far from each other, on average. The distance between them is of order Λ−1 .
Under the circumstances, the string-like chromoelectric flux tubes, connecting well-
separated probe color charges, hardly have a chance to be developed. Moreover,
the valence quark pair injected in the vacuum, in a sense, perturbs it only slightly.
Then we do not need the full machinery of the QCD strings, whatever it might
mean, to approximately describe the properties of the low-lying states. Their basic
parameters depend on how the valence quarks of which they are built interact with
typical vacuum field fluctuations.
It is established that the QCD vacuum is characterized by various condensates
[1]. Half-dozen of them are known: the gluon condensate G2µν , the quark condensate
q̄q, the mixed condensate q̄σGq, and so on. The task is to determine the regularities

6
and parameters of the classical mesons and baryons from a few simple condensates.
Of course, without invoking the entire infinite set of condensates one can hope
to capture only gross features of the vacuum medium, at best. Correspondingly, any
calculation of the hadronic parameters of this type is admittedly approximate. Since
hadronic physics is deprived of small parameters, in the vast majority of problems
high accuracy is by far not the most desired requirement to calculation, however.
Does it make any difference if, say, the proton magnetic moment is predicted theo-
retically to two or to three digits? I do not think so. Rather, it is high reliability of
predictions in a wide range of problems where the answer is not known a priori and
theoretical control over qualitative aspects that are the primary goals of the theory
of hadrons. Below I will try to demonstrate that the SVZ method is both a reliable
and controllable approach which has enjoyed enormous success in dozens and dozens
of instances. There are a few cases where it fails [3], but we do understand why.
And the very fact of the failure of the standard strategy teaches us a lot. This hap-
pens for specific, “nonclassical” hadrons, with a very strong coupling to the vacuum
fluctuations, which are very different from say the ρ meson or nucleon. So, we can
reverse the argument and convert the failure into success by saying that the method
predicts that not all hadrons are alike.

2.2 Getting started/Playing with toy models

The basic microscopic degrees of freedom of Quantum Chromodynamics are the
quarks and gluons. Their interaction is described by the Lagrangian
1
L = − Gaµν Gaµν +
X
q̄i D
6 q (5)
4 f

where f is the flavor index, and the color index of the quark fields q is suppressed.
For simplicity we will assume that the quark mass terms vanish. In other words,
we will work in the chiral limit. In this limit the pion is massless, m2π = 0. This is
known to be quite a good approximation to the real world.
Neither quarks nor gluons are asymptotic states. Experimentally observed are
hadrons – color-singlet bound states. In order to study the properties of the classical
hadrons it is convenient to start from the empty space – the vacuum – inject there
a quark-antiquark pair, and then follow the evolution of the valence quarks injected
in the vacuum medium. The injection is achieved by external currents. The most
popular are the vector and axial currents. Their popularity is due to the fact that
they actually exist in nature: virtual photons and W bosons couple to the vector
and axial quark currents. Therefore, they are experimentally accessible in the e+ e−
annihilation into hadrons or hadronic τ decays.
Thus, the objects we will work with are the correlation functions of the quark
currents. More concretely, let us consider the vector current with the isotopic spin

7
I = 1,
¯ µd
ūγµ u − dγ
Jµ = √ , or Jµ = ūγµ d. (6)
2
The first current shows up in the e+ e− annihilation while the second is relevant to
the τ decays. The two-point function Πµν is defined as
Z
Πµν = i eiqx d4 xh0|T {Jµ(x)Jν† (0)}|0i . (7)

where q is the total momentum of the quark-antiquark pair injected in the vacuum.
Due to the current conservation Πµν is transversal and, hence,

Πµν = (qµ qν − q 2 gµν )Π(q 2 ) . (8)

For historical reasons Π(q 2 ) is often called the polarization operator; we will use
this nomenclature in what follows. It is easy to count that the function Π(q 2 ) is
dimensionless. The imaginary part of Π(q 2 ) at positive values of q 2 (i.e. above the
physical threshold of the hadron production) is called the spectral density,
12π
ρ(s) = Im Π(s) , s ≡ q 2 . (9)
Nc
Up to normalization it coincides with the cross section of e+ e− annihilation into
hadrons (measured in the units σ(e+ e− → µ+ µ− )) or the τ decay distribution func-
tion. The numerical factor in the definition of ρ(s) in Eq. (9) is introduced for
convenience, as will become apparent shortly. In the real world Nc = 3 but we will
keep this factor explicit for a while, to keep track of the Nc dependence.
The spectral density carries full information about the spectrum and widths of
hadrons with given quantum numbers. Every QCD practitioner dreams of the exact
calculation of the spectral density. Later on we will see how the SVZ sum rules
constrain the parameters of the lowest-lying state, the ρ meson in the case at hand.
But first, prior to submerging into technical aspects of the SVZ approach, let us
make an educated guess of how the spectral density might look, in gross features,
to get an idea of what is expected for Π(q 2 ).
To begin with, consider the limit Nc → ∞. As is well-known [10], in this limit
all hadrons are infinitely narrow, since all decay widths are suppressed by powers of
1/Nc . Correspondingly, ρ(s) is a sum of delta functions.
Moreover, there are good reasons to believe that these delta functions must
be approximately equidistant, at least asymptotically, for highly excited states. A
string-like picture of color confinement naturally leads to (approximately) linear
Regge trajectories, and, hence, m2n ∝ n at large n where mn is the mass of the n-th
ρ-meson excitation. In the world of purely linear Regge trajectories there are an
infinite number of daughter trajectories associated with each Regge trajectory. The
daughter trajectories are parallel to the parent trajectory and are shifted by integers.
Thus, in the old Veneziano model (a review is given e.g. in [11]) m2n = m2ρ + n/α′

8
 ρ (s)

15

12.5

10

7.5

2.5

2.5 5 7.5 10 12.5 15

s

Figure 4: The spectral density in the toy model. For clarity I gave a tiny width to
the δ functions in Eq. (10).

(α′ is the slope of the Regge trajectory). In some of the later versions, which are
more “QCD-friendly”, m2n = m2ρ + 2n/α′ , see e.g. [12]. Since for now the focus is on
a qualitative picture, the distinction between these two scenarios is not essential for
our illustrative purposes. For definiteness, let us accept that the distance between
two consecutive excitations contributing to ρ(s) is 2/α′ ≈ 2 GeV2 .
Assembling all these elements together one obtains a sketch of the spectral den-
sity (Fig. 4),
∞
X
ρ(s) = 3 δ(s − 1 − 3n) . (10)
n=0

Here m2ρ is set equal to unity; all dimensional quantities are measured in these units,
for instance, 2/α′ ≈ 3. The couplings of all mesons to the current Jµ are chosen to
be equal. This choice is represented by the overall numerical factor 3 in the sum
(10). The explaination for this will become clear momentarily.
Now, given the imaginary part (10), it is not difficult to reconstruct the polar-
ization operator Π itself,
Nc
Π(Q2 ) = − ψ(z) + Const. , (11)
12π 2
Q2 + 1
Q2 ≡ −q 2 , z = ,
3

9
where ψ is the logarithmic derivative of the Γ function,
∞ 
1 1
X 
ψ(z) = −C + − .
k=0 k+1 z+k

The constant on the right-hand side of Eq. (11) is irrelevant, since it can be always
eliminated by an appropriate subtraction. It will be omitted hereafter.
Although our desired target is the spectral density on the physical cut, i.e. at
positive values of q 2 , all theoretical calculations in Quantum Chromodynamics are
carried out off the physical cut, say, at negative values of q 2 (positive Q2 ). The
reason is obvious: the QCD Lagrangian (5) is formulated in terms of quarks and
gluons, not hadrons. In the absence of the final solution of QCD we can deal only
with the quark and gluon fields. Working in the Euclidean domain, off the physical
cuts, we can calculate in terms of quarks and gluons. This is a common feature of
the SVZ sum rules and lattice strategies.
At positive values of Q2 an asymptotic representation exists for the ψ function,
∞
1 X B2n −2n
ψ(z) = ln z − − z , (12)
2z n=1 2n

where B2n stand for the Bernoulli numbers,

2(2n)!
B2n = (−1)n−1 ζ(2n) ; (13)
(2π)2n

here ζ is the Riemann function. (In some textbooks (−1)n+1 B2n is called the n-th
Bernoulli number and is denoted by Bn .)
Equations (11) and (12) at large positive Q2 imply that in the leading approxi-
mation
Nc
Π(Q2 ) → − ln Q2 . (14)
12π 2
This logarithmic formula for the polarization operator exactly matches what we
expect from perturbation theory. Indeed, at large Q2 , in the deep Euclidean domain,
the points of injection and annihilation of the quark pair in the vacuum are separated
by a small space-time interval. The quarks have no time to interact with the vacuum
medium. They propagate as free objects (Fig. 5). For free quarks, obviously,
Z
Πµν (q) = i eiqx d4 x Tr {γµ S0 (x, 0)γν S0 (0, x)} , (15)

where S0 (x, y) is the free-quark Green function describing propagation from the
point y to the point x,
1 ∆ 6
S0 (x, y) = , ∆ ≡ x−y. (16)
2π (∆2 )2
2

10
Figure 5: The two-point function of the vector currents in the free-quark approxi-
mation. The external current injecting the quark-antiquark pair in the vacuum (and
then annihilating it) is denoted by wavy lines.

It is trivial to obtain
Nc 2xµ xν − x2 gµν
Tr {γµ S0 (x, 0)γν S0 (0, x)} = − . (17)
π4 x8
Note that the expression on the right-hand side is automatically transversal. Sub-
stituting Eq. (17) in Eq. (15) and doing the Fourier transformation, we arrive at
Eq. (14).
Thus, the leading asymptotic behavior of the polarization operator stemming
from the free-quark graph of Fig. 5 is purely logarithmic. An infinite comb of in-
finitely narrow resonances, taken with one and the same residue, see Eq. (10), yields
the same logarithm in the deep Euclidean domain. This explains why all resonance
coupling constants are set equal in Eq. (10). I hasten to add that the spectral
density in the real world is much more contrived. The toy spectral density (10) is
supposed to be a caricature, only roughly reminding the actual one. Nevertheless,
the exercise with the toy spectral density is instructive in several respects.
It is clear that the leading logarithmic term (14) carries absolutely no information
on the mass of the lowest-lying state or on the spacing between the resonances: it
has no built-in dimensional parameter. This information is encoded in the power
corrections.
Examining the 1/Q2 expansion in Eq. (11), 2 we conclude that the high-order
terms have factorially divergent coefficients. The factorial divergence at large n is
due to the factorial growth of the Bernoulli numbers B2n in Eq. (12). In QCD
the expansion in powers of 1/Q2 is nothing but the condensate expansion. The
example considered teaches us that the power expansion of Π(Q2 ) has zero radius
of convergence. By itself, this is no disaster, one can work with asymptotic series
as long as the expansion parameter is small. In our case the expansion parameter
is τ ≡ 1/Q2 , and we are interested in the expansion in the vicinity of Q2 ∼ 1.
(Remember that we put m2ρ = 1. If Q2 is substantially larger than 1, one gets very
little information on the ρ meson from Π(Q2 ).)
In the ideal world we would calculate Π(Q2 ), and, hence, the spectral density
proportional to ImΠ(s), exactly. In reality we must settle for Π(Q2 ), calculated in
2
In the case at hand we deal with pure powers of 1/Q2 . Our toy model is too rude to reproduce
the logarithmic corrections and logarithmic anomalous dimensions of the condensates typical of
QCD.

11
the Euclidean domain (positive Q2 ), in the form of a truncated series. The dispersion
relation provides a bridge between the Euclidean calculation and the spectral density,

1 Im Π(s)
Z
Π(Q2 ) = ds + Const. (18)
π s + Q2
This explains why the method is referred to as the sum rules.
The first five terms in the power expansion of the polarization operator (11) are

τ 2 τ 3 13 4
" #
2 Nc 2 τ
Π(Q ) = − ln Q − + + − τ − τ 5 + ... , τ ≡ 1/Q2 . (19)
12π 2 2 4 3 40

From the structure of the series it is seen that at τ ∼ 1 the polarization operator
Π(Q2 ) is determined by its expansion with very poor accuracy, up to a factor of 2,
at best. This accuracy is obviously insufficient to allow one to estimate, say, m2ρ
with a reasonable precision.
One can drastically improve the accuracy by considering the Borel-transform
of Π(Q2 ) [1]. The Borel transformation can be defined in various ways. For the
functions obeying dispersion relations, the most convenient definition is through the
following limiting procedure:
!n
1 d Q2
B̂ = lim (Q2 )n − 2 , Q2 → ∞, n → ∞, ≡ M 2 fixed . (20)
(n − 1)! dQ n

M 2 is called the Borel parameter. It is not difficult to show that applying B̂ to

(1/Q2 )n we get !n
1 1 1 n
 
B̂ = , (21)
Q2 (n − 1)! M 2
which entails, in turn !
1 1 −s/M 2
B̂ = e . (22)
s + Q2 M2
Another trivial but useful relation we will need is

B̂(ln Q2 ) = −1 . (23)

The Borel-transformed dispersion relation (18) takes the form

1
Z
2 2 2
Π̃(M ) ≡ {B̂Π(Q )} = dsIm Π(s) e−s/M . (24)
πM 2
The advantages one gains in dealing with the Borel-transformed sum rules are
obvious. First, we improve, factorially, the convergence of the power series. Thus,
in the toy model we are playing with – an infinite comb of infinitely narrow peaks

12
presented in Eq. (10) – the 1/Q2 expansion for Π(Q2 ) is asymptotic, while the 1/M 2
expansion of Π̃(M 2 ) is convergent! Indeed,
2
2 Nc 1 e−1/M
Π̃(M ) = 2 2 , (25)
4π M 1 − e−3/M 2
and the radius of convergence of the 1/M 2 expansion is
3
|M∗2 | = . (26)
2π
The factor 2π in the denominator ensures that we can go down to a remarkably low
value of M 2 ; the point M 2 ∼ 1 is well inside the convergence radius. Let us examine
the first five terms of the truncated series,

τ 2 τ 3 13τ 4 τ 5
" #
2 Nc τ
Π̃(M ) = 2
1+ − − + + + ... , τ ≡ 1/M 2 . (27)
12π 2 4 6 240 24

It is seen that in the vicinity of τ ∼ 1 our “theoretical prediction” is expected to

carry an error of order ∼ 2%. Such accuracy is already good enough to get the mass
and residue of the lowest-lying state with a reasonable “resolution”.
On the phenomenological side, proceeding to the Borel-transformed dispersion
relations we automatically kill possible subtraction constants. What is even more
important, the exponential weight function in Eq. (24) makes the integral over the
imaginary part well-convergent. Thanks to the improved convergence the relative
role of the ρ-meson contribution is strongly enhanced, as we will prove shortly by
quantitative estimates.
To see how it all works let us take a closer look at Π̃(M 2 ); to avoid cumbersome
numerical factors we will change the overall normalization and will deal with
12π 2 1
Z
2 2
I(M ) = Π̃(M 2 ) = 2 dse−s/M ρ(s) . (28)
Nc M
In the toy model under consideration everything is known explicitly: the exact
expression for I(M 2 ), its power expansion, and the ρ-meson contribution to I(M 2 ).
The corresponding plots are displayed in Fig. 6. The exact curve has a typical shape:
at small M 2 it is exponentially suppressed, approaching zero as exp(−m2ρ /M 2 ); at
large M 2 it is flattens off and slowly approaches its asymptotic value, which is equal
to unity thanks to a smart choice of the normalization factor in Eq. (28). This unity
is in one-to-one correspondence with the free quark result (14). It is convenient to
normalize the sum rules to the free-quark result at asymptotically large M 2 , and
we will always do that. The exact curve has a steep (left) shoulder, and a shallow
(right) one, with a maximum at M 2 slightly above m2ρ . Remember this pattern since
we will encounter with a very similar picture more than once in real QCD.
The ρ meson mass and residue follow immediately from consideration of I(M 2 )
in the small M 2 domain (the left shoulder). Indeed, in this domain all higher states

13
are exponentially suppressed (for instance, the contribution of the first excitation
relative to that of ρ is ∼ exp(−3/M 2 ). Therefore, I(M 2 ) is fully saturated by ρ in
the limit M 2 → 0. Fitting the left shoulder by C1 exp(−C2 /M 2 ) we would find C1,2 ,
the ρ-meson residue and mass, respectively.
Alas... In the real world I(M 2 ) is not known at small M 2 . All that is known
is the large-M 2 asymptotics, in the form of the expansion in the coupling constant
αs (M 2 ), and a (truncated) condensate expansion. Therefore, let us pretend that in
our toy model we have the same: a truncated power series (27). The corresponding
curve is shown in Fig. 6 (where I have actually included also the O(τ 6 ) term). The
expansion is convergent only to the right from M 2 ≈ 0.5. If we want the truncated
series, with a few first terms included, to accurately represent the theoretical curve,
we have to stop at M 2 ≈ 0.7 (arrow A in Fig. 6). This is the left boundary of what
the sum-rule practitioners usually call the window or working window.
The right boundary of the window (arrow B) is provided by the requirement that
the first and higher excitations show up at the level not exceeding, say, 20 to 30 %.
In this way we ensure sufficient sensitivity to the ρ-meson parameters.
The larger the M 2 , the weaker the ρ-meson dominance will be. The requirement
of the ρ-meson dominance forces us to move to smaller values of M 2 , while keeping
control over the power expansion suggests that we move to larger values of M 2 .
Thus, these two requirements are self-contradictory. A priori, in any given problem,
it is not evident that the AB window exists at all. If it does, we can obviously use
this fact to approximately calculate the mass and the residue of the lowest-lying
state. The reasons why the window exists in the problems of the classical mesons
and baryons will be discussed later.
To get an idea of the accuracy one can expect from the sum rule calculations of
the ρ-meson parameters we must investigate how sensitive our results are to various
assumptions regarding the spectral density. The spectral density presented in Fig.
4 is not fully realistic, for many reasons. If minor details affected the results too
strongly, the SVZ method would be useless. The right edge of the window (arrow
B) is chosen in such a way as to minimize the impact of the (unknown) details
of the spectral density above the ρ-meson peak. Yet, it is important to obtain a
quantitative measure of the corresponding uncertainty.
The most obvious unrealistic feature of the spectral density (10) is the infinitely
narrow width of all resonances from the comb. It is true that in the multicolor QCD
the width-to-mass ratio of all hadrons built from quarks is proportional to 1/Nc [10]
and, thus, vanishes at Nc → ∞. If Nc is fixed, however, and we are interested in the
widths of the excited states, there is another large parameter, the excitation number,
which compensates for the effect of large Nc . Moving to the right along the s axis in
Fig. 4, very soon we find ourselves in the situation with the overlapping resonances
– the δ functions in the comb representing the spectral density are smeared and the
spectral density becomes perfectly smooth. Let us discuss the consequences of such
dynamical smearing.
First of all, we must establish the dependence of the width Γn on the excitation

14
 1.2 exact I(M 2 )

ρ meson
0.8 expansion

0.6

A B
0.4

window
0.2

0 0.5 1 1.5 2 2.5 3

M2

Figure 6: The sum rule for the ρ meson in the toy model.

number n. If one believes that a string-like picture of color confinement develops in

QCD, an asymptotic estimate of Γn at large n becomes a simple exercise. Unlike the
ρ meson and other low-lying classical states, for highly excited states the string-like
picture of the chromoelectric tubes is expected to be fully relevant.
When a highly excited meson state is created by a local source, it can be consid-
ered, quasiclassically, as a pair of (almost free) ultrarelativistic quarks; each of them
with energy mn /2. These quarks are created at the origin, and then fly back-to-
back, creating behind them a flux tube of the chromoelectric field. The length of the
tube L ∼ mn /Λ2 where Λ2 represents the string tension. The decay probability is
determined, to order 1/Nc , by the probability of producing an extra quark-antiquark
pair. Since the pair creation can happen anywhere inside the flux tube, it is natural
to expect that
1 B
Γn ∼ LΛ2 = mn (29)
Nc Nc
where B is a dimensionless coefficient of order one. I used here the fact that within
the quasiclassical picture the meson mass mn ∼ LΛ2 . Thus, the width of the √ n-th
excited state is proportional to its mass which, in turn, is proportional to n for
the linear Regge trajectories 3 .
This simple estimate was obtained in Ref. [13] long ago. Since the argument
bears a very general nature, the square root dependence of Γn should take place in
3
Let us note in passing that the 1/Nc2 corrections
√ due to creation of two quark
√ pairs are of
order L2 /Nc2 within this picture. Since L ∼ mn ∼ n, the expansion parameter is n/Nc .

15
all models with linear confinement. And it does, indeed! Recently, the square root
formula for Γn was obtained numerically [14] in the ’t Hooft model [15].
If Eq. (29) does indeed take place, it is not difficult to find the impact of the
resonance widths on the spectral density. The infinitely narrow pole is substituted
with
1 1 1
→ → →
Q2 + m2n Q2 + m2n − imn Γn Q2 + m2n − im2n B/Nc
1 1
2 2 2 2
→ 2 1−γ
, (30)
Q + mn − γQ ln Q (Q ) + m2n
where
B
γ= , (31)
πNc
and I used the fact that 1/Nc is a small parameter. This explains the last transition
in Eq. (30). For the same reason, in the third transition I replaced m2n /Nc with
Nc−1 π −1 Q2 ln Q2 . Near the pole (i.e. at Q2 = −m2n ) both expressions give the same
in the leading 1/Nc approximation, and are equally legitimate 4 .
As a result, all poles, which lied previously at positive real q 2 , are now shifted
onto unphysical sheets, away from the physical cut (Fig. 7). The δ functions in the
spectral density are replaced by peaks with finite widths. Correspondingly, the toy
model number two, which replaces Eq. (11) is
Nc 1
Π(Q2 ) = − 2
ψ(z) + Const. , (32)
12π 1 − γ
where now z is given by the following formula
(Q2 )1−γ + 1
z= . (33)
3
The additional factor (1 − γ)−1 in the overall normalization is chosen to reproduce
the free quark result (14) at large Euclidean Q2 . Note that the polarization operator
defined by Eqs. (32) and (33) has the correct analytical structure: it is non-singular
in the whole complex Q2 plane, apart from the cut at real positive q 2 . The singu-
larities associated with the poles of the ψ function lie on the unphysical sheet (Fig.
7).
The spectral density (i.e. the imaginary part of Π at q 2 = s + iǫ) takes the form
of the sum of (modified) Breit-Wigner peaks
∞
3 X s1−γ sin πγ
ρ(s) = . (34)
π(1 − γ) k=0 [s1−γ cos πγ − (1 + 3k)]2 + [s1−γ sin πγ]2
It is depicted in Fig. 8 for B = 0.6 and Nc = 3. Although the choice of the
parameter B is rather arbitrary, it is not unreasonable; it corresponds to Γ/m ≈ 0.2.
Experimentally the ρ meson has a close width-to-mass ratio.
4
The logarithm introduces not only the imaginary part, i.e. the width, but a shift in the real
part, equivalent to a shift in the resonance mass. Both effects are of order 1/Nc .

16
 2
2 Q
Lower side of the cut Im z z
Im Q

Cut 2
Re Q Re z
0
- m 2n

Upper side of the cut

(a) (b)

Figure 7: Analytical structure of the polarization operator. (a) The polarization

operator must be analytic everywhere in the complex Q2 plane, except the cut
running on the negative real semi-axis of Q2 (positive real semi-axis of q 2 ). The
imaginary part of the polarization operator must be positive at the upper side of
the q 2 cut; (b) The mapping of the Q2 plane onto the z plane, Eq. (33). The physical
sheet on the Q2 plane corresponds to the z plane with the shaded sector removed.
The boundaries of the sector correspond to the lower and upper sides of the q 2 cut.
The angle of the removed sector is proportional to γ.

17
 ρ(s)
5
toy model # 2
(modified) Breit-Wigner
4

0 2 4 6 8 10
s

Figure 8: The spectral density corresponding to a sum of (modified) Break-Wigner

peaks with one and the same width-to-mass ratio 0.2. All residues are set equal,
and γ ≈ 0.06369, see Eq. (34). The s axis is in the units m2ρ .

We can see on the picture a pronounced ρ-meson peak, accompanied by a deep

dip, and then an almost smooth curve which approaches the asymptotic value (unity)
in an oscillating mode. The smearing of the δ functions into a smooth curve occurs
because the width of the excited states is proportional to mn while the distance
between two neighboring states ∆mn ∼ 1/mn . It is assumed, of course, that Nc
is fixed. The larger the value of Nc , the further we must go, to higher excitation
numbers, for the resonances to overlap. Figure 8 shows that at Nc = 3 a smooth
curve starts right after the ground state. The first excitation already belongs to
the smooth curve. Remember this pattern of the spectral density – a conspicuous
first peak accompanied by a smooth (and oscillating) curve approaching its constant
asymptotic value. The gross features of this picture follow from very general theo-
retical arguments [16]. We will see shortly that experimental data, being different
in fine details, exhibits the very same type of behavior.
The spectral densities shown in Figs. 4 and 8 at first sight do not produce
an impression of close relatives. Let us integrate them over, with the exponential
weight function, and compare I(M 2 ) obtained in this way. The comparison is shown
in Fig. 9. The dotted curve corresponds to the comb of δ functions. It was already
presented in Fig. 6. The solid curve corresponds to Eq. (34) and Fig. 8. Finally,

18
 2
I(M )
1.2

Breit-Wigner peaks
0.8

comb of δ functions
0.6
SVZ

0.4

0.2 A B
2
0 0.5 1 1.5 2 2.5 3
M

Figure 9: I(M 2 ) for three models of the spectral density.

the dashed curve is obtained with an extremely crude model of the spectral density
depicted in Fig. 10,
ρ(s) = 3δ(s − 1) + θ(s − s0 ) , s0 = 2.7 . (35)
It presents an infinitely narrow ρ, plus a gap, plus all higher excitations fused into
“continuum” which starts abruptly at s = s0 and coincides (at s > s0 ) with the
asymptotic free-quark expression for the spectral density. This is the original SVZ
model [1]. It, obviously, captures only gross features, and misses all details: the fact
that the gap at s > m2ρ is not quite empty, the onset of continuum is not so abrupt,
and the limit ρ(s) → 1 at s → ∞ is achieved through oscillations. Nevertheless,
it is seen that I(M 2 ) in all three models come out very close to each other. The
Breit-Wigner model of Eq. (34) gives a little bit less steep fall off to the left of the
window, at M 2 → 0. This is quite understandable, since in this model there is a
non-vanishing spectral density at s < m2ρ , due to the ρ meson width.
Comparison of three curves in Fig. 9 teaches us a lesson: by examining I(M 2 ),
calculated approximately inside the window, one cannot expect to predict, with any
reasonable accuracy, such fine structure of the spectral density as the width of the
ρ meson, and the shape of the curve in the continuum (i.e. higher excitations).
As Fig. 9 shows, it is hardly possible to distinguish between three models under
discussion, which, in a sense, represent extreme situations. The best one can hope
for is estimating m2ρ , its residue, and, to a lesser extent, an effective value of s where
the dip following the ρ-meson peak ends (i.e. s0 ).

19
 ρ(s)
17.5

"delta+continuum" model
15

12.5

10
δ function
7.5

2.5

s
1 2 3 4 5 6 7

Figure 10: The original SVZ model of the spectral density.

In a broader context, this message is instructive for the lattice practitioners too.
Doing calculations in the Euclidean domain – a common feature of the sum rules
and lattice technology – derives virtually no information about the spectral density
beyond the ground state in the given channel. Indeed, the original SVZ model and
the comb of the δ functions, representing drastically different ρ(s) point-by-point,
lead to I(M 2 ) coinciding with each other to a high accuracy. To distinguish between
these two scenarios one needs exponential precision going far beyond the level so far
achieved in the lattice calculations.
Now we are ready to explore the sensitivity of I(M 2 ) to the ρ-meson parameters.
To this end we will take the SVZ model in the form

ρ(s) = C1 δ(s − C2 ) + θ(s − s0 ) (36)

and let the parameters float by, say, 10% around their “reference” values, C1 =
3, C2 = 1 and s0 = 2.7. We then compare I(M 2 ) obtained in this way with I(M 2 )
emerging in our (most realistic) Breit-Wigner toy model, see Eq. (34).
Figure 11 illustrates the sensitivity of the sum rule to the ρ meson residue (cou-
pling constant). The upper and lower dashed curves correspond to C1 = 3.3 and 2.7,
respectively. Figure 12 illustrates the sensitivity to the ρ meson mass. The upper
and lower dashed curves correspond to m2ρ = 0.9 and 1.1 of the experimental value,
respectively. It is seen that it is quite realistic to expect to get gρ2 and Mρ2 with
the accuracy ∼ 10% by inspecting I(M 2 ) in the window and fitting the theoretical

20
 1.2 C 1 =3.3
1
C 1 =2.7
0.8

0.6

0.4

2
0.2
I versus M
0 0.5 1 1.5 2 2.5 3

Figure 11: Changing the ρ meson residue. The solid curve corresponds to a “refer-
ence” spectral density presented in Eq. (34). The dashed curves correspond to Eq.
(36) with C1 = 3.3 and 2.7, respectively. The reference value is C1 = 3.

prediction by the model (36). The estimates of s0 are less accurate. One should not
be surprised, of course, since the sum rules were designed to be most sensitive to
the ground-state parameters and relatively insensitive to the details of the spectral
density in the continuum. The exponential weight in the definition of I(M 2 ) takes
care of this feature. Deviations of s0 from its reference value at the level of 10%
lead to rather insignificant changes in I(M 2 ). Deviations of s0 at the level of 20%
are quite noticeable.
I hasten to add, however, that the 10% accuracy is not a typical outcome of the
sum rule analysis. The ρ-meson channel is most favorable from the point of view of
applications of the SVZ sum rules. In a sense, this is a dream case: the role of con-
tinuum with respect to ρ is as tempered as it can possibly be, and higher (unknown)
condensates in the truncated condensate expansion show up at remarkably low val-
ues of M 2 so that the working window is comfortably wide. In many other channels,
mesonic and baryonic, we have to deal with a narrower window. The general rule
is: the narrower the window the worse the accuracy. In some channels, as we will
see later, the window shrinks to zero. Then the SVZ method fails. The reasons why
it is successful in some cases and fails in others, when properly understood, give us
a unique hint as to the structure of the QCD vacuum (see Sect. 8).

21
 1.2
C 2 =0.9
1
C 2 =1.1
0.8

0.6

0.4

2
0.2
I versus M
0 0.5 1 1.5 2 2.5 3

Figure 12: Changing the ρ meson mass. The dashed curves correspond to the 10%
variation of m2ρ , around its reference value (m2ρ = 1).

3 Vacuum Condensates
It is high time to abandon our baby version of QCD and proceed to the real thing.
What can be said about I(M 2 ) now?
Needless to say that no analytic calculation of I(M 2 ) at M 2 ∼ < Λ2 , that would
be based entirely on the first principles, exists. QCD at large distances is not
yet solved analytically. Under the circumstances it seems reasonable to start at
short distances and advance to larger distances “step by step”, staying on the solid
ground – working with the microscopic variables, quarks and gluons, where it is fully
legitimate. At short distances the quark-gluon interactions are adequately described
by perturbation theory. Certainly, as we have just learned by inspecting the toy
models, the (truncated) perturbative series is not going to yield us any estimates
relevant to the ρ meson parameters. To get them we should add at least some
information regarding the large distance dynamics.
In the ideal world this information would be obtained from the theory per se.
In the real world we may try to parametrize the effects caused by the vacuum
fields. If the quark-antiquark pair injected in the vacuum by the current Jµ does
not propagate too far, its impact on the vacuum fields is, hopefully, not drastic. This
means that the polarization operator can be well approximated by the interaction
of the valence quarks with a few vacuum condensates. For the validity of this
assumption it is necessary that the characteristic frequencies of the valence quarks
inside the ρ meson ω be larger than the characteristic scale parameter of the vacuum

22
medium µ. Certainly, we can count only on an interplay of numbers: since the only
dimensional parameter of QCD is Λ, all quantities of dimension of mass are of order
Λ. It may well happen, however, that, say ω ∼ 3Λ while µ ∼ Λ/2. In a sense, the
success of the SVZ sum rules confirms this assumption a posteriori.
On general grounds it is expected that all gauge invariant Lorentz singlet local
operators built of the quark and/or gluon fields develop non-vanishing vacuum ex-
pectation values (VEV’s). It is convenient to order all operators (and their VEV’s)
according to their normal dimensions. The higher the dimension, the higher the
power of 1/M of the corresponding coefficient. This fact allows one to control the
condensate expansion inside the working window.
The lowest dimension (zero) belongs to the unit operator. The unit operator is
trivial. When one calculates the perturbative contribution to I(M 2 ), one actually
calculates the coefficient of the unit operator.
The operator of the next lowest dimension (three) is the quark density operator,
Oq = q̄q . (37)
No other gauge and Lorentz invariant operators of dimension three exist.
There is one operator of dimension four, the gluon operator
αs
OG = Gaµν Gaµν . (38)
π
The factor αs /π appears in the definition naturally. Since both operators, (37) and
(38), play a very special role we will interrupt our excursion before passing to higher
dimensions, to make several important remarks.
In the chiral limit (i.e. when the quark mass term in the Lagrangian is put to
zero) Oq is the order parameter – its VEV signals the spontaneous breaking of the
axial SU(Nf ) symmetry and the occurrence of the corresponding massless pions. In
perturbation theory the chiral symmetry remains unbroken, of course, hOq i vanishes
identically. It is tempting to say then that hOq i = 6 0 measures deviations from
perturbation theory, and the same refers to all other condensates.
In the early days of non-perturbative QCD such an understanding was widely
spread. I hasten to say that this is a wrong understanding. It is impossible to define
the condensates as “truly non-perturbative” residues obtained after subtracting from
them a “perturbative part” [17]. A heated debate took place in the literature in the
eighties regarding the possibility of defining the condensates, in a rigorous way, by
subtracting a “perturbative part”. Glimpses of the debate can be seen in Ref. [18]
where it was clearly emphasized, for the first time, that the procedure of isolating
“purely non-perturbative” quantities does not (and must not) work 5 . I will outline
5
Surprisingly, the debate was revived recently, over a decade after the issue had been seemingly
settled, in connection with the heavy quark theory. This theory, as it exists today, is also based
on the operator product expansion and is a close relative [19] of the SVZ sum rules. One can ask
there the very same questions that are usually raised in connection with the condensate expansion
in the SVZ method. Later on, in Sect. 10.2, we will return to the issue and discuss some elements
of the heavy quark theory.

23
the proper procedure below using the gluon condensate as the most instructive
example.
Unlike Oq , the gluon operator OG is not an order parameter. There is no known
symmetry whose spontaneous breaking would generate a vacuum expectation value
hOG i. Correspondingly, hOG i is generated in perturbation theory. Physically, hOG i
measures the vacuum energy density εvac . To see that this is indeed the case let us
use the fact that the trace of the energy-momentum tensor
b αs a a
θµµ = − G G . (39)
8 π µν µν
This expression is nothing but the scale anomaly of QCD [20]. Strictly speaking, the
coefficient on the right-hand side contains higher orders in αs ; they are not essential
for our purposes, however, and will be ignored. The vacuum expectation value of
the left-hand side is obviously 4εvac . Hence
b
εvac = − hOG i . (40)
32
Everybody knows that the vacuum energy density badly diverges in field theory,
as the fourth power of the cut-off 6 . If so, how can one make sense of the gluon
condensate?
The vacuum fields fluctuate; these fluctuations contribute to the vacuum energy
density. High frequency modes of the fluctuating fields belong to the weak coupling
regime; their contribution in the correlation functions is described by perturbation
theory, giving rise to the standard perturbative expansions. What we are interested
in are the low frequency modes, soft vacuum fields that are responsible for the
peculiar properties of the vacuum medium. By saturating the condensate by the
soft modes only we get a consistent definition of the condensates that automatically
solves the problem of divergencies.
Specifically, we must introduce a somewhat artificial boundary, µ: all fluctuations
with frequencies higher than µ are supposed to be hard, those with frequencies lower
than µ soft. The parameter µ is usually referred to as normalization point. Only
soft modes are to be retained in the condensates. Thus, the separation principle
is “soft versus hard” rather than “perturbative versus non-perturbative”. This is
the basic principle of Wilson’s operator product expansion, which, in turn, is the
foundation of the SVZ sum rules. Being defined in this way the condensates are
explicitly µ dependent. All physical quantities are certainly µ independent; the
normalization point dependence of the condensates is compensated by that of the
coefficient functions.
Generalities of Wilson’s approach and its particular implementation in QCD will
be further discussed in Sects. 5 and 6. Here we complete our consideration of the
gluon condensate.
6
Let me parenthetically note that in supersymmetric glyodynamics the vacuum energy density
vanishes, and so does the gluon condensate.

24
 Figure 13: The one-gluon correction to the D function.

It is instructive to trace how the gluon condensate emerges from a slightly dif-
ferent perspective. A natural starting point is the diagram of Fig. 13 presenting the
one-gluon correction to Π(Q2 ) (two other similar graphs with the one-gluon exchange
are not shown). For convenience we will deal with the so called D function,

D(Q2 ) = −(4π 2 )Q2 (dΠ/dQ2 ) , (41)

rather than with the polarization operator Π itself. In terms of D(Q2 ) the bare
quark loop of Fig. 5 becomes just unity.
If the gluon momentum is denoted by k the one-gluon term in D(Q2 ) takes the
form Z
1
∆D(Q ) = αs k 2 dk 2 2 F (k 2 , Q2 )
2
(42)
k
where it is implied that the angular integration over the orientations of the vector
k with respect to q is already carried out. The factor 1/k 2 comes from the gluon
propagator, while F (k 2 , Q2 ) represents the rest of the graph. The function F (k 2 , Q2 )
was calculated in Ref. [21] (it can also be extracted from Appendix B of Ref. [22]),
8 1 7
  
F (k 2 , Q2 ) = − ln τ τ + (1 + τ ) [L2 (−τ ) + ln τ ln(1 + τ )] , (43)
3π Q2 4
where
k2
τ= ,
Q2
and Z x
L2 (x) = − dyy −1 ln(1 − y)
0

is the dilogarithm function. The analytic continuation of the function (43) to τ > 1
can be more conveniently rewritten as
8 1 3 1 1
  
2 2
F (k , Q ) = 1 + ln τ + + ln τ +
3π Q2 4 2 τ
h io
(1 + τ ) L2 (−1/τ ) − ln τ ln(1 + τ −1 ) . (44)
Substituting Eqs. (43) and (44) in Eq. (42) and doing the integral over k 2 in
the most straightforward manner one gets ∆D = αs /π, the standard well-known
one-gluon correction to the D function.

25
 2
0.12
Q F

0.1

0.08

0.06

0.04

0.02

0 0.5 1 1.5 2 2.5 3

τ
Figure 14: The distribution function Q2 F (k 2 , Q2 ) defined in Eq. (42) versus τ .

The plot of the function F (k 2 , Q2 ) is shown in Fig. 14. Although it has a clear-
cut peak at τ ∼ 0.5, it presents a distribution spanning the entire range from k 2 = 0
to k 2 = ∞. The gauge coupling constant αs in Eq. (42) is literally a constant
provided one limits oneself to the graph of Fig. 13. In higher orders, however, αs
starts running, and it is evident that αs → αs (k 2 ). Then αs (k 2 ) should be placed
inside the integral in Eq. (42) rather than outside. By doing so we effectively resum
a subset of the multi-gluon graphs.
Certainly, putting
#−1
bαs (Q2 ) k2
"
αs (k 2 ) = αs (Q2 ) 1 + ln 2 (45)
4π Q

inside the integral in Eq. (42) we do not account for all higher-order corrections.
Only those are included that are responsible for the running of αs in the leading
logarithmic approximation. Usually they say that the corresponding multigluon
graphs are of the bubble chain type (Fig. 15). Strictly speaking, in the covariant
gauges the gauge coupling renormalization (45) in QCD (unlike QED) does not
reduce to the bubble insertions in the gluon propagator depicted in Fig. 15. Various
graphs of a more complicated structure are involved too; they combine together to
produce a gauge-invariant expression (45). The bubble chain saturation of Eq. (45)
takes place in the physical gauges, e.g. in the Coulomb gauge. In the covariant
gauges one can artificially reduce the problem to the bubble chains, without explicit
identification of the full set of graphs, by exploiting the so called “large negative
Nf trick”. We will not dwell on this technical issue, since it is irrelevant for our
purposes. The interested reader is referred to Refs. [21, 22].
Even though the bubble chain is a very specific subset of graphs, and the majority
of the multigluon graphs are left aside in this procedure, it still makes sense to
substitute the running αs (k 2 ) inside the integral in Eq. (42), since in this way we take

26
into account an essential part of the underlying dynamics: the fact that the quark-
gluon interaction becomes weak if the gluon is far off-shell, and, on the contrary,
becomes stronger as the gluon virtuality decreases. If we follow this strategy, then
Z ∞
∆D(Q2 ) = αs (k 2 )dk 2 F (k 2 , Q2 ) =
0
#−1
bαs (Q2 ) k2
"
Z ∞
2 2
αs (Q ) dk 1+ ln 2 F (k 2 , Q2 ) . (46)
0 4π Q
We immediately run into a problem here. The running αs has the Landau pole at
small k 2 , the square bracket in the integrand explodes right inside the integration
interval. The explosion occurs at
" #
2 4π
2
k = Q exp − = Λ2 . (47)
bαs (Q2 )

Although the function F (k 2 , Q2 ) is suppressed at small k 2 , i.e. τ → 0, (see Fig.

14), it by no means vanishes at τ = Λ2 /Q2 , rendering the integration impossible.
In order to do the k 2 integral literally, as it is given in Eq. (46), we must say how
the singularity at k 2 = Λ2 is to be treated. Some people say: “let us take the
principal value”, others bypass the singularity by shifting the integration contour
in the complex k 2 plane, still others do something else. All these prescriptions are
arbitrary and physically meaningless for obvious reasons. Indeed, let us introduce
the normalization point µ = several units ×Λ and split the integration over k 2 in
two parts: from zero to µ2 in the first integral and from µ2 to ∞ in the second.
Then at k 2 > µ2 , in the weak coupling domain, we do trust Eq. (46) while below µ2
we do not. At best, below µ2 one can use Eq. (46) for the purpose of orientation.
At small k 2
2 k2
F (k 2 , Q2 ) → , (48)
π Q4
implying that
#−1
2αs (Q2 ) 1 Z µ2 2 2 bαs (Q2 ) k2
"
2
∆{k2 <µ2 } D(Q ) = k dk 1 + ln 2 ∼
π Q4 0 4π Q

8 1 Z µ2 2 2 Λ2 8π Λ4
k dk ∼ . (49)
b Q4 0 k 2 − Λ2 b Q4
This is a very crude estimate; I substituted the square bracket by the corre-
sponding pole, and then evaluated the integral by equating it to its imaginary part,
i.e. substituting (k 2 − Λ2 )−1 → πδ(k 2 − Λ2 ).
From this exercise we learn the following lesson. The domain k 2 < µ2 gives a
contribution to D(Q2 ) which scales as 1/Q4 . The coefficient in front of 1/Q4 cannot
be reliably calculated and must be parametrized by the gluon condensate – that is
the best we can do. It is not accidental that the k 2 expansion of F (k 2 , Q2 ) starts

27
 Figure 15: One of the graphs from the bubble chain set.

from k 2 , while the term of the zero-th order is absent. If it were not the case, the
condensate series would have to start from O(Q−2 ). This is impossible, however,
since no gauge invariant local operator of dimension two exists in QCD. The absence
of such operator and the absence of terms k 0 in the distribution function F (k 2 , Q2 )
are in one-to-one correspondence.
At first sight, one could try to circumvent the problem of the Landau pole inside
the integration domain by expanding the square bracket in Eq. (46) in the series
in αs (Q2 ). This is one of the “remedies” sometimes cited in this context. I put the
word remedy in the quotation marks since in fact it does not work. True, in every
given order of the αs expansion the integral (46) becomes well-defined. It is easy to
see, however, that the expansion is going to be factorially divergent in high orders.
Indeed,
2 n
Q2
Z µ2 !
Q
k 2 dk 2 ln 2 ∼ 2−n n! , n ≫ ln 2 . (50)
0 k µ
It is seen that perturbation theory per se carries the seeds of the gluon con-
densate. The perturbative series cannot be consistently defined unless the gluon
condensate is introduced. Once we cut off the k 2 < µ2 tail of the integral from
the perturbative expression obtained from Eq. (46), the factorial divergence of the
αs series at high n disappears, the series becomes well-defined, and so is the gluon
condensate.
The factorial divergence of the perturbative series of the type we have just dis-
cussed is called the infrared renormalon [23, 24] for the reasons which need not
concern us here. It is associated with the bubble chain graphs of Fig. 15 and is
inevitable if one forces the perturbative integrals to run all the way down to k 2 = 0.
Bounding the integration interval in the perturbative part from below by µ2 we get
rid of the infrared renormalon altogether. The domain below µ2 is not lost; it is
fully represented by the gluon condensates. Higher condensates appear as higher
order terms in the k 2 expansion of F (k 2 , Q2 ). The intimate relationship between
the infrared renormalons and condensates was first revealed by Mueller [25].
I hasten to warn that one should not literally equate the infrared renormalons
to the condensate expansion. Even if one sticks to a certain particular definition
for summation of the bubble chains, that eliminates ambiguities, say the principal
value prescription in the integral (46), one typically gets a numerical value of the
Q−4 term which is grossly off compared to the gluon condensate, let alone ambigu-
ities associated with various possible choices of the regularization prescription. For

28
instance, in the case at hand,
2π 2
D(Q2 ) = 1 + hOG i + ... , (51)
3Q4
and the O(1/Q4 ) term is 0.08 GeV4 /Q4 under the standard choice of the gluon con-
densate (see below). Equation (49) yields, on the other hand, only 0.004 GeV4 /Q4 at
Λ = 0.2 GeV (0.01 GeV4 /Q4 at Λ = 0.25 GeV). Within the SVZ method, based on
Wilson’s expansion, the role of the infrared renormalons is purely illustrative. They
show, in a very straightforward manner, that without the condensates the pertur-
bative series cannot be defined and, on the contrary, introducing the condensates
one eliminates ambiguities associated with the high-order tails of the perturbative
series 7 .
On the other hand, in some (perturbatively infrared stable) processes that have
no operator product expansion, most notably in the jet physics, the infrared renor-
malons, in spite of all their obvious limitations, become poor man’s substitute for
the condensate expansion. Although, unlike OPE, the renormalon analysis does
not provide us with the proper numerical coefficients, we may infer from the renor-
malons what powers of 1/E (or 1/Q) appear in the nonperturbative parts of such
quantities as, say, thrust. Whether the nonperturbative corrections are O(1/E) or
O(1/E 2 ) is clearly a question of paramount importance for phenomenology. Using
the infrared renormalons for counting the powers of 1/E in the processes without
OPE is a totally fresh idea which can be considered as a distant spin off of the SVZ
method, see Ref. [27] for a review. Previously the nonperturbative corrections were
just ignored since nobody knew how to approach the issue scientifically.
Now, when the meaning of the condensates appearing in Wilson’s expansion is
hopefully clear, we can proceed to a brief discussion of their numerical values. In
principle, two alternative approaches are conceivable: (i) calculating the condensates
from first principles (e.g. on the lattices); (ii) extracting the condensates from the
sum rules themselves. Attempts of the lattice calculation of the condensates were
reported in the literature (for a review see [28]). The main difficulty, which, to a
large extent, devaluates this work is the fact that the strategy of “isolating a non-
perturbative residue from the perturbative background” was adopted. As we already
know, this strategy is doomed to failure. Instead, one should have consistently
implemented Wilson’s procedure. This has never been done, however. Some initial
ideas as to how Wilson’s procedure can be implemented on the lattices are presented
in Ref. [29].
7
The infrared renormalon is not the only source of the factorial divergence of the αs series in
QCD. The αs expansion of Eq. (46) contains also the so called ultraviolet renormalon – factorial
divergence of high-order αs terms coming from the large k 2 domain. This factorial divergence is
readily summable, however, as is perfectly clear from the unexpanded expression (46). Indeed, the
integrand has no singularities at large k 2 and is well convergent. Below, the ultraviolet renormalons
will never be mentioned again. Those readers who want to familiarize themselves with this subject
are referred to an excellent review [26]. A factorial divergence of a different nature showing up in
the coefficient functions is briefly discussed in Sect. 5.

29
 Therefore at present, as twenty years ago, one has to exploit the sum rule them-
selves in order to determine the condensates. One picks up certain channels where
sufficient experimental information on the corresponding spectral densities is avail-
able. These channels are sacrificed, i.e. the analysis goes in the direction opposite
to conventional: the values of the condensates are extracted from the Euclidean
correlation functions obtained through dispersion integrals. The first determination
of the gluon condensate was carried out exactly in this way, from the sum rules in
the J/ψ channel [30, 1]. The corresponding value will be referred to as standard,

hOG i = 0.012 GeV4 .

Since then this estimate was subject to multiple tests, sometimes with conflicting
results. A brief discussion can be found in the beginning of Sect. 2 in Ref. [2].
The overall conclusion is that certain deviations from the standard value (say, at
the level of ∼ 30%) are not ruled out 8 . One of the reasons limiting the accuracy
is the fact that the gluon condensate was determined within a simplified (the so
called practical) version of the operator product expansion, which is strictly speaking
somewhat ambiguous. We will dwell on the issue in Sect. 6. At the present level
of understanding it is highly desirable to repeat the analysis within the framework
of consistent Wilson’s procedure. This is not done so far. To take into account a
certain degree of numerical uncertainty it is reasonable to accept that

hOG i = 0.012 xG GeV4 , (52)

where a dimensionless numerical factor xG is allowed to float in the vicinity of unity.

This parametrization will be used below in the ρ-meson sum rule.
Let us pass now to the quark condensate hOq i. As was mentioned, this con-
densate is the order parameter for the spontaneous breaking of the chiral symme-
try. Hence, its value can be independently determined from the corresponding phe-
nomenology. In particular, the celebrated Gell-Mann–Oakes–Renner formula relates
the product of the light quark masses and hOq i to observable quantities,
¯ = −M 2 f 2 ,
(mu + md )hūu + ddi (53)
π π

where fπ is the pion constant (fπ ≈ 133 MeV) and Mπ is the pion mass. The original
estimate of hOq i used in Ref. [1] was obtained by substituting the crude estimates
of the light quark masses that existed at that time [31, 32],

mu + md ≈ 11 MeV .

In this way one arrives at

hOq i = −(250 MeV)3 . (54)
8
I leave aside extremist and to my mind unfounded statements in the literature, that the
standard value underestimates the gluon condensate by a factor of 2 to 5.

30
Since then considerable work has been carried out to narrow down the theoretical
uncertainty. First of all, chiral corrections to the Gell-Mann–Oakes–Renner formula
were derived and analyzed. Moreover, much effort has been invested in perfecting
our knowledge of the quark masses. The progress is summarized in Ref. [33], where
an extensive list of references can be found. No dramatic changes occurred. The
value of mu + md extracted from phenomenology of the chiral symmetry breaking
went up by about 30%, with an error at the level of 15%. Thus, one is tempted to
say that the actual value of the quark condensate is close to its “standard” value
quoted in Eq. (54); perhaps, somewhat suppressed 9 .
Events took a dramatic turn in 1996, after results of the lattice calculations of
the light quark masses (with two dynamical quarks) became known. The lattice
result for mu + md lies significantly lower (by a factor of two to three, see e.g. [34])
than any of the reasonable analytic estimates! Were these results confirmed, this
would mean that the quark condensate is significantly larger than the number given
in Eq. (54).
I would caution against hasty conclusions at this point. The discrepancy between
the analytic and lattice estimates may well be an artifact of the lattice procedure.
Putting dynamical almost massless quarks on the lattice is a notoriously difficult
task. On the other hand, the discrepancy, if persists, may prove to be a signature of
a very interesting physical phenomenon – strong dependence of physical quantities
on the number of light dynamical quarks. The surprisingly low lattice result for
mu + md was obtained under the assumption that the number of light quarks is two,
while in fact it is three. We will return to the issue of possible strong dependence
on Nf in Sect. 12.
Under the circumstances it seems reasonable, analogously to the gluon conden-
sate, to parametrize the quark condensate as
hOq i = −(250 MeV)3 xq , (55)
and let xq vary around unity.
The full catalog of all relevant operators up to dimension six, was worked out
in Ref. [1]. It is not as large as one might think at first sight. There is only one
operator of dimension five, the mixed quark-gluon operator
OqG = ig q̄σµν Gµν q (56)
where Gµν ≡ Gaµν T a (T a are the color generators). This operator plays no role in the
ρ-meson sum rules, to be considered below: because of its “wrong” chirality it can
9
The quark mass is definition dependent. Usually the MS scheme and a “reference” normal-
ization point µ = 1 GeV are implied. Not to confuse the reader we will follow this convention, in
spite of shortcomings of the MS scheme. One cannot avoid explicitly introducing the normaliza-
tion point µ in the case of the quark condensate: Oq has a non-vanishing anomalous (logarithmic)
dimension. The q̄q vertex dressed by gluons is logarithmically suppressed. Including the gluons
with off-shellness from M0 down to µ suppresses the vertex by a factor ∝ [αs (µ)/αs (M0 )]4/b . This
formula explicitly demonstrates why µ cannot be put to zero in the case at hand: the logarithm
explodes.

31
enter only being multiplied by the quark mass mq . It is very important, however,
in a wide range of problems involving baryons [35], and mesons built from one light
and one heavy quarks [36].
At the level of dimension six, there is one operator built from three gluon field
strength tensors (symbolically GGG) and several four-quark operators of the type

O4q = (q̄1 Γ1 q2 )(q̄3 Γ2 q4 ) (57)

where Γ1,2 denote certain combinations of the Lorentz and color matrices and qi
(i = 1, ..., 4) are the light quark fields of different flavors, u, d and s. The overall
flavor must be conserved of course, but q1 need not coincide with q2 and so on.
The three-gluon operator is expected to have a significant impact in heavy
quarkonium [37]; it does not appear in the ρ-meson sum rules, as was first shown in
Ref. [38], by virtue of a very elegant theorem.
As for the four-quark operators, their vacuum matrix elements are not known
independently. It became standard to evaluate them in the factorization approx-
imation, which was first applied in this context in [1]. Note that using the Fierz
identities for the color and Lorentz matrices one can always arrange the operators
O4q to have a natural color flow. By natural I mean that Γ1,2 do not contain color
matrices, and the color indices of quarks are contracted inside each bracket in Eq.
(57). The color of q1 flows to q2 and that of q3 flows to q4 . If the number of colors
Nc were large, then the expectation value of O4q would be totally saturated by the
vacuum intermediate state,

hO4q i = hq̄1 Γ1 q2 ihq̄3 Γ2 q4 i . (58)

Corrections to this equation are formally of order of 1/Nc2 . It is clear that in the
factorization approximation, after the natural color flow is achieved by the Fierz
rearrangement, the only surviving structure is that with Γ1,2 = 1.
In the real world Nc = 3, and we certainly do expect deviations from factorization
at a certain level. As far as we can tell today, factorization works surprisingly
well at least for those four-quark operators that appear in the ρ-meson sum rule.
All attempts to detect deviations, both on the lattices [39] and in the sum rules
themselves [40], gave results consistent with zero, within errors that typically lie
in the 10% ballpark. This is quite nontrivial, considering the fact that quite a
few examples are known where large numerical coefficients neutralize formal 1/Nc
suppression factors. Whatever the reasons might be, we will accept Eq. (58) in
what follows.

4 ρ Meson in QCD
Experience accumulated in Sect. 2 will now be applied for explorations in actual
QCD. As we learned from the toy models, the polarization operator Π(Q2 ) (or its
Borel transform Π̃(M 2 )) in the deep Euclidean domain has an expansion in ln Q2

32
(perturbation theory) plus power terms (1/Q2 )k or (1/M 2 )k (the condensates). If in
the toy models the power terms (truncated in certain order) can be made as large or
as small as we want, in QCD they are determined dynamically; through interactions
of the valence quark pair with the soft vacuum fields. This interaction gives rise
to the condensate expansion. The first power term is due to the gluon condensate,
while the next one is associated with the four-quark condensate. In principle, higher
condensates were analyzed in the literature too, but we will not discuss them in this
Lecture. √
The quark-antiquark pair q q̄ with the total energy s is injected in the QCD
vacuum either by the virtual photon (e+ e− annihilation) or W boson (τ decays).
The q q̄ pair, being injected, starts evolving according to the dynamical laws of QCD.
At first the quarks do not feel the impact of the vacuum “medium”. As separation
between them grows, the effects of the medium become more and more important,
so that eventually it prevents quarks from appearing in the detectors. The injected
quarks get dressed and materialize themselves in the form of hadrons. In the sum
rule approach we control only the beginning of this process.
In the sum rule framework, the condensates become important in the vicinity of
the left edge of the window. To the left of the arrow A they explode, and theoretical
control is lost. On the other hand, the value of I(M 2 ) near the right edge of the win-
dow is determined mostly by ordinary perturbation theory. All perturbation theory
sits in the unit operator. Technically it is convenient to formulate the corresponding
calculation in terms of the perturbative corrections to the spectral density. We take
the graph of Fig. 5, and attach to it various gluon and quark loops (e.g. Figs. 13
and 15). In this way we get the perturbative part of ρ(s); at present in the vector
isovector channel it is known up to third order in αs (s) [41]. In the MS scheme the
result takes the form
ρ(s) = 1 + k1 a + k2 a2 + k3 a3 (59)
where a(s) is given in Eq. (3). In the low-energy domain the number of active
flavors is Nf = 3. For Nf = Nc = 3 the parameter b = 9, and the values of ki are
4
k1 =
, k2 = 0.729k1 , k3 = −2.03k1 . (60)
9
Equation (59) can be immediately translated into the prediction for the per-
turbative part of I(M 2 ). The perturbative expansions for I(M 2 ) and ρ(s) do not
coincide. To get the former one must integrate the latter with the exponential weight
function. The transition is readily carried out with the aid of the formula [42]
  
2 2
ds exp(−s/M ) 1 ν(ν + 1) π /6 1
Z
=  M 2 ν 1 + 2 + O   4  ,
  
 ν
M 2 s 2
 2 2
ln Λ2 ln Λ2 eγ ln ΛM2 eγ ln ΛM2 eγ
(61)
where γ denotes Euler’s constant, γ = 0.577. Expanding this expression near integer
values of ν one also gets necessary expressions for the integrals containing logarithm
of logarithm (ln ln).

33
 In this way we arrive at [42]
!#2 !#3
M2 M2 π2 M2
! " "
2
Ipc (M ) = 1 + k1 a + k2 a + (k3 + k1 ) a . (62)
eγ eγ 6 eγ

Similar expressions were obtained in Ref. [43], were a(M 2 /eγ ) was reexpressed in
terms of a(M 2 ). It is more convenient to work directly with the expansion (62).
The quark mass corrections in ρ(s) and I(M 2 ) are also known, they are propor-
tional to m2u,d /M 2 and are negligibly small due to the smallness of the current quark
masses. They will be discarded since the corresponding uncertainty is invisible in
the background of other uncertainties.
Now we return to the condensates. The modern perturbative calculations of the
coefficients Cn are based on the background field technique. This is an important
aspect of the SVZ approach as it exists today. The method is quite elegant, to say
nothing that it is much more economic than the original “brute force” calculations of
Ref. [1]. We will not discuss the technique per se, however. Conceptually everything
is clear here. If you need to do a calculation, you just go and learn the corresponding
technique using, for instance, the review paper [44].
The coefficients CG and C4q , to the leading order, are known from the ancient
times [1],

π2 8π 3
 2 
¯ α γ5 T a d

Inpc (M 2 ) = 1 + hOG i − αs ūγα γ5 T a u − dγ −
3M 4 M6
16π 3
* +
¯ α T ad
 X
a a
αs ūγα T u + dγ (q̄γα T q) . (63)
9M 6 q

The four-quark structures appearing in the angle brackets, as well as the factor αs
in front of them, are normalized at Q. The anomalous (logarithmic) dimension of
this particular combination nearly vanishes [1]; we will ignore it in evolving the
four-quark operator at hand down to µ. Applying factorization, as was explained in
Sect. 3, we get for the four-quark term in Inpc (M 2 )

448π 3 2 1
C4q hO4q i = − αs (µ)hq̄q(µ)i → − 0.03 x4q GeV6 , (64)
81M 6 M6
where, as previously, we have introduced a factor x4q to allow for possible theoretical
uncertainties. Under the standard choice of parameters [1] this factor is equal to
unity. If we are on the right track it is reasonable to expect that x4q is close to unity.
A huge amount of work was carried out in calculating two-loop corrections in the
coefficients of various condensates. The results can be inferred, say, from Ref. [45]
or from the reviews [46]. Unfortunately, the existing uncertainty in the condensates
themselves precludes us from taking advantage of the precision achieved in pertur-
bative corrections. Therefore, we will limit ourselves to the leading order results for
CG and C4q .

34
 Now, the stage is set and we can finally present the sum rule for I(M 2 ),
!  !#2 
M2 M2 M2
" !# "
4  
I(M 2 ) = 1 + a 1 + 0.729 a − 0.386 a +
9 eγ  eγ eγ 

0.6 2 0.6 3
   
0.1xG − 0.14x 4q , (65)
M2 M2
where M in the second line of Eq. (65) is measured in GeV. The term M −4 is
due to the gluon condensate, while the term M −6 is due to the quark condensate.
With the “standard” numerical values of the gluon and quark condensates in the
the nonperturbative part xG = x4q = 1; the dimensionless constants xG and x4q
allow the gluon and four-quark condensates to “breathe” to a certain degree. For
instance, xG > 1 would imply that the actual value of the gluon condensate is larger
than the “standard” value, and so on.
We are ready to examine the situation in the ρ-meson channel. Figure 16 shows
experimental data on the spectral density in the vector isovector channel, measured
in the τ decays. The data points are from Ref. [47]. Note a remarkably close
resemblance with our toy model number two. If we take the beginning of Fig. 8
and expand it to make the energy scales coincide 10 the curves will essentially repeat
each other. There exist a few extra points in the τ decays, spanning the interval
of s from 2.7 to 3 GeV2 , and some data above 3 GeV2 from the e+ e− annihilation,
but the corresponding error bars are so large that plotting these points would just
obscure the picture.
There is a subtle point which I have to discuss here. Directly measurable in the
hadronic τ decays is the sum of the vector and axial spectral densities. To obtain
the spectral densities separately one has to sort out all decays by assigning specific
quantum numbers to each given final hadronic state. In the majority of cases such
an assignment is unambiguous. For instance, two pions (whose contribution is the
largest) can be produced only by the vector current. Some processes, however, can
occur in both channels, for instance, the KKπ production. Using certain theoretical
arguments it was decided in Ref. [47] that around 3/4 of all KKπ yield should be
ascribed to the vector channel. Other theoretical arguments [42], which seem to
be much more convincing to me, tell that virtually all KKπ production should
take place in the axial channel. Therefore, in dealing with the data, I will subtract
the KKπ yield from the ALEPH data points. I hasten to add that the subtracted
quantity is small (see Fig. 16), and this subtraction is not very essential for a general
picture I draw here by broad touches.
The solid curve in Fig. 16 is a best fit by a smooth curve representing the sum of
two Breit-Wigner peaks (the first one is a modified Breit-Wigner taking into account
threshold effects important for the ρ meson). The dashed curve is what I believe
10
Warning: the energy scale (the horizontal axis) in Fig. 8 presents s in the units of m2ρ while
that in Fig. 13 is in GeV2 .

35
 ρ (s), vector isovector
5

4 ALEPH

0 0.5 1 1.5 2 2.5

2
s, GeV
Figure 16: The spectral density in the vector isovector channel measured in τ decays.
The data points belong to ALEPH. The solid curve is a best fit. The dashed curve
is the best fit minus KKπ (the latter contribution is shown by the dotted curve)
which, I think, should not have been included in the vector spectral density, see
text.

36
 5

ρ (s) versus s, in GeV 2

4

3 perturbative asymptotics

0 1 2 3 4 5 6 7

Figure 17: “Theoretical experimental” spectral density in the ρ meson channel. The
solid curve presents experiment (see text), while the dashed curve is the perturbative
result for ρ(s), including terms up to O(αs3), with Λ = 0.2 GeV.

the actual spectral density is. The data points above 1 GeV2 carry a noticeable
uncertainty. Since we are going to use the spectral density only in the integrals,
where many data points are summed over, these individual errors can be neglected,
since they will be statistically insignificant in the integrals. Systematic uncertainty
might be important, but since nobody knows how to estimate it, I will ignore it for
the time being. For our limited purposes we can consider the dashed curve in Fig.
16 as an exact experimental result for the spectral density.
How does the experimental spectral density might look above 2.7 GeV2 ? This
question is irrelevant for calculating the ρ-meson parameters. It is very instructive,
however, to have a broader perspective. I will show you what I call a “theoretical
experimental” spectral density, and then will make a short digression explaining
where the answer comes from.
My expectations are summarized in Fig. 17. The curve to the left of the arrow
is just the fit to the experimental data, as explained above (it identically coincides
with the dashed curve in Fig. 16.) To the right of the arrow I have attached a
“tail” which approaches a smooth asymptotic prediction for ρ(s) (the dashed line)
given in Eq. (59), in an oscillating manner. The tail and the experimental data are
smoothly matched at 2.65 GeV2 .

37
 Why I think that the actual spectral density, when and if it is measured, will
approach the smooth asymptotic curve with oscillations rather than monotonously?
The reason is of a very general nature [16]. In a nut shell, there are exponential
terms that are not seen in the truncated OPE series for Π(Q2 ) in the Euclidean
domain. When these terms are analytically continued to the Minkowski domain,
where Im Π(s) belong, they show up as oscillating terms. The smooth asymptotic
prediction is obtained from the truncated series and, thus, carries no traces of the
exponential/oscillating terms. We will return to the issue later on (see Sect. 5).
Although the very existence of oscillations in the spectral densities may be consid-
ered firmly established, the question of how rapidly they die off is highly non-trivial
and has no unambiguous answer in today’s theory. In resonance-inspired models the
damping factor is ∼ exp(−const
√ s) (see Sect. 2.2), in the instanton-based models
[48] it is√∼ exp(−const s). The tail added in Fig. 17 is taken to be proportional to
exp(−π s) cos(πs + const). This rather eclectic formula is chosen for sophisticated
reasons which need not concern us here. The task which we address is illustrative,
anyway. I certainly cannot guarantee that around 3.5 GeV2 the value of ρ(s) is 0.97,
as it is shown in Fig. 17, but I could bet that a shortage of the spectral density in
this domain will be observed in precision measurements, so that ρ(s) at 3.5 GeV2
will lie somewhere around unity. I emphasize again that the precise form of the tail
does not affect the sum rule calculation of the ρ meson parameters, and is discussed
only for completeness of the picture.
With the experimental spectral density in hands, we can calculate the “experi-
mental” value of the integral I(M 2 ). The corresponding result is presented in Fig.
18 by the solid curve.
The agreement is excellent; it is even better than one could expect apriori. We
see that the conspicuous ρ peak and all further twiddles characteristic to ρ(s) are
washed out. If we descend from larger to smaller values of M, the behavior of I(M)
is flat down to M ∼ 0.8 GeV, i.e. down to the ρ meson mass. At M ∼ 0.8 GeV
the regime smoothly changes, the curve dives down, and at still lower values of M,
approaches the exponential asymptotics (this domain is not shown in the figure).
The boundary value of M where the regime changes is correlated with the mass
of the lowest state, the ρ meson in the case at hand. This is practically the only
characteristic dimensional parameter in I(M).
Let us pretend now that we do not know the spectral density and want to use
the sum rules to determine the ρ meson mass. It is quite clear that the truncated
condensate series does not allow one to go to the limit M → 0 where this determi-
nation would be exact: the series explodes. We must stay inside the window where
the expansion is still under control. Correspondingly, our determination of Mρ is
going to be approximate. The strategy can be formulated as follows. Let us assume
that somebody shows us a sketch of the spectral density presented in Fig. 10 with
the numbers along the horizontal and vertical axes erased. This sketch is known to
correctly reproduce the basic features of the actual spectral density but leaves us
ignorant as to where the ρ-meson peak lies and what is its height. We insert this

38
 1.2

0.8
experiment

0.6
theory

0.4

ρ meson sum rule

0.2
ρ
0 1 2 3 4 5 6

Figure 18: I versus M 2 (in GeV2 ) in the ρ meson channel: confronting experiment
and theory. The theoretical curve corresponds to xG = 0.8, x4q = 1.3 and Λ = 0.2
GeV.

39
sketch in our sum rule, do the integral
1
Z
2
ds ρ(s)e−s/M ,
πM 2
fit the numbers in such a way as to be as close to the theoretical prediction (65)
inside the window – click, click – the scales are restored, and there come out the
ρ mass and the coupling constant. Since inside the window the ρ meson saturates
the integral at the level ∼ 90% , the fact that our continuum model is a caricature
(minor details are lost) is unimportant. Even if we are off by a factor of two in this
model this will affect our estimates referring to the ρ meson at the level of ∼ 10%.
This example is quite typical. A similar situation takes place for all classical
low-lying hadrons: those built from the light quarks, heavy, and light and heavy.
I do not define here precisely what the “classical meson” means but will return to
this point later, after considering some nonclassical channels. I must admit that the
ρ meson is the example where the SVZ method demonstrates its best facets. The
window is sufficiently broad, everything is clean. As we proceed to higher spins,
the mesons become larger in size. A snapshot of a high-spin meson would show a
string-like picture, with a longitudinal size of the “sausage” much larger than its
transverse size. Under the circumstances one could hardly expect that the SVZ
method would work. And, indeed, it ceases to be informative for spins higher than
two [49].

5 Basic Theoretical Instrument – Wilson’s OPE

The theoretical basis of any calculation within the SVZ method is the operator prod-
uct expansion (OPE). It allows one to consistently define and build the (truncated)
condensate series for any amplitude of interest in the Euclidean domain. The phys-
ical picture lying behind OPE was described above: consistent separation of short
and large distance contributions. The former are then represented by the vacuum
condensates while the latter are accounted for in the coefficient functions. Thus, the
operator product expansion in the form engineered by Wilson [50] is nothing but
a book-keeping procedure. Wilson’s idea was adapted to the QCD environment in
Ref. [18].
Although OPE is used in an innumerable amount of works since the mid-1970’s,
there are no good text-books or reviews devoted to this issue. Moreover, one typi-
cally encounters a lot of confusion in the literature. Below I will try to explain the
Wilsonean procedure using, as an example, the T product of the two vector currents
defined in Eq. (7).
Technically it is more convenient to deal with the D function (41) rather than
Πµν ,
D(q 2 ) =
X
Cn (q; µ)hOn (µ)i (66)
n

40
Figure 19: A typical multiloop graph contributing to the D function. The number
of such graphs grows factorially with the number of loops.

where the normalization point µ is indicated explicitly. The sum in Eq. (66) runs
over all possible Lorentz and gauge invariant local operators built from the gluon
and quark fields. The operator of the lowest (zero) dimension is the unit operator I,
followed by the gluon condensate G2µν , of dimension four. The four-quark condensate
gives an example of dimension-six operators.
At short distances QCD is well approximated by perturbation theory. The co-
efficient functions Cn absorb the short-distance contributions. Therefore, as a first
approximation, it is reasonable to calculate them perturbatively, in the form of ex-
pansion in αs (Q) ∼ (ln Q)−1 . This certainly does not mean that the coefficients Cn
are free from non-perturbative, non-logarithmic terms of the type ∼ Q−γ where γ
is a positive number.
To substantiate the point let us consider the coefficient of the unit operator.
The Feynman graphs for D(q 2 ) of the lowest order are depicted in Figs. 5 and
13. Assume that the momentum flowing through the graph is Euclidean and large,
Q → ∞. If all subgraphs in these and similar graphs are renormalized at Q, the final
result, being expressed in terms of the running coupling constant αs (Q), is finite.
The virtual momenta saturating the loop integrals scale with Q as the first power
of Q. At any finite order the perturbative series is well-defined.
At the same time, if the number of loops n becomes very large, in a random
graph (Fig. 19) each loop carries momentum of the same order of magnitude, and
the total momentum Q is shared between many lines, so that the characteristic
virtual momentum is proportional to Q/n. The contribution of each graph to CI is
of order [αs (Q)]n , but the number of distinct diagrams ν grows factorially, ν ∼ n!,
so that the expansion in αs (Q) is asymptotic (for an exhaustive discussion of this
phenomenon see Ref. [51]).
Note that this factorial divergence has nothing to do with the renormalon di-
vergence considered in Sect. 3. There, we had one specific graph of the n-th order
whose contribution was proportional to n!. This factorial was a spurious artifact
of an improper calculation where the domain k ∼ Λ was included. Introducing a
lower cut-off at µ = several units times Λ and discarding the contribution coming
from k 2 < µ2 , as it should be done in the consistent calculation of CI (µ), we would
eliminate the renormalon divergence. At the same time, the factor n! counting the
number of graphs of the n-th order certainly cannot be eliminated in this way.
If n is not too large, so that Q/n > µ, the graphs of Fig. 19 present a legit-

41
imate contribution to CI (µ). It is quite clear that the tail of the αs series with
n∼> 1/αs (Q) ∼ ln Q generates a non-perturbative contribution of the type
!bγ
κΛ
∆CI ∼ exp (−2πγ/αs (Q)) ∼ . (67)
Q

Here γ and κ are numerical constants. An explicit example of such terms is provided
by the so called direct instantons [3]. Integration over the sizes ρ of the direct
instantons is saturated at ρ ∼ Q−1 . The constant γ = 1 for one instanton, 2 for two
and so on.
The correspondence between the Feynman graphs with 1/αs (Q) loops and the
small-size instantons is not straightforward. I will not go further in this issue,
referring the interested reader to Ref. [51]. At the qualitative level the inevitability
of occurrence of the hard non-perturbative terms (67) is evident. At the moment
the only semi-quantitative framework we have at our disposal for their evaluation is
the instanton mechanism.
Equation (67) gives a non-condensate power term. Generically we will call such
terms hard non-perturbative. They may or may not be important numerically de-
pending on the particular value of Q under consideration. The value of bγ need not
be integer, generally speaking; numerically it is very large. This means that once
we cross the boundary where the argument in the brackets in Eq. (67) becomes less
than one, Q > κΛ, these terms immediately become totally unimportant. On the
other hand, below this boundary they are so large that no expansion is possible.
It is assumed that inside the window the terms (67) are unimportant; they are
completely disregarded in the practical version of OPE (Sect. 6) which, thus, applies
to the values of Q above the critical point. Note that the large value of bγ results
in a large degree of suppression of such terms after borelization.
Although the assumption above works very well in many “classical” channels,
it seems to fail in exceptional cases of “non-classical” mesons. This issue will be
addressed in more detail in Sect. 8.
I pause here to make a remark concerning non-dynamical power terms in CI (µ)
whose origin is related to the introduction of the normalization point µ. Let us
return to the one-loop graph of Fig. 13 and the corresponding expression (42). One
should not forget that in doing the loop integrations in CI (µ) we must discard the
domain of virtual momenta below µ, by definition. Subtracting this domain from
the perturbative loop integrals we introduce power corrections of the type (µ2 /Q2 )n
in CI (µ), by hand. For instance, in the first order in αs

αs (Q) αs µ4 2π 2
" #
2
D(Q ) = 1 + − + hOG (µ)i + ... , (68)
π π Q4 3Q4

where I combined Eqs. (48) and (51). The expression in the square brackets is the αs
part of the coefficient CI (µ). The explicit µ dependence of this expression conspires
with an implicit µ dependence residing in hOG (µ)i to ensure the µ independence of

42
the physical quantity (i.e. the D function). If higher-order terms in the k 2 expansion
of the F function (43) were taken into account, we would get higher powers of µ/Q in
the square brackets, which then should have been combined with higher condensates,
e.g. GD 2 G and so on.
If µ can be chosen sufficiently low, the explicit and implicit µ-dependent terms
in Eq. (68) may be numerically insignificant and can be ignored.
The coefficient functions in front of other operators generally speaking have the
same structure as CI . The general situation is quite similar. For higher dimension
operators it may happen (and, in fact, happens) that some of the loop integrations
are not saturated at Q or µ but are, rather, logarithmic. Logarithms [ln(Q/µ)]γ
occurring in this way are associated with the anomalous dimensions of the operators
at hand.
Let us have a closer look at the graph depicted in Fig. 13. As we already know,
some of the lines can be in a special regime – the corresponding virtual momenta
are soft and their off-shellness does not scale with Q. If the gluon line is soft, this
gives rise to the gluon condensate, if some of the quark lines are soft we deal with
the four-quark condensate. Graphically one distinguishes the soft lines by cutting
them; then the remainder of the graph shrinks to a point while those lines that are
cut form a local operator. The question is: what happens if we pass to the multiloop
graph of Fig. 19 and start cutting more and more lines?
Naturally, in the power series we proceed to the operators of higher and higher
dimension. When the number of the soft lines becomes very large (of order of Q/Λ)
it is conceivable that there are no hard lines at all: the external momentum Q is
transferred from the initial to the final vertex through a very large number of quanta
(growing like a power of Q), and none of the quanta carries momentum scaling with
Q. Of course, in this situation one can not speak of individual quanta, one should
rather use the language of typical field fluctuations transmitting the momentum Q
without having any Fourier components with frequencies of order Q. Contributions
of this type are not seen in OPE truncated at any finite order. They correspond to
the high-order tail of the condensate series which, analogously to the αs series, is
factorially divergent. In this way we obtain the so-called exponential terms [16].
The fact that the condensate series is factorially divergent in high orders is rather
obvious from the analytic structure of the polarization operator Π(Q2 ). In a nut
shell, since the cut in Π(Q2 ) runs all the way to infinity along the positive real
semi-axis of q 2 , the 1/Q2 expansion cannot be convergent. The actual argument
is more subtle than that, but I would not like to go into details here referring the
reader to Ref. [16] where a careful consideration of the issue is carried out. The final
conclusion is perfectly transparent. It is intuitively clear that the high-order tail of
the (divergent) power series gives rise to exponentially small corrections ∼ exp(−Qσ )
where σ is some critical index.
The numerical value of σ is correlated with the rate of divergence of high or-
ders in the power series. At the moment very little is known about this rate from
first principles, if at all. The best we can do is to rely on toy models. The sim-

43
plest example is again provided by instantons. This time one has to fix the size
of the instanton ρ by hand, ρ = ρ0 . Then the fixed-size instanton contribution is
O(exp(−Qρ0 )). The exponential factor is the price we pay for transmitting the large
momentum Q through a soft field configuration whose characteristic frequencies are
or order ρ−10 .
One can look at this example from a broader perspective. Consider the polariza-
tion operator Π in the coordinate rather than the momentum space. If the regular
terms in OPE are in one-to-one correspondence with singularities of Π(x) at x = 0,
the exponential terms, invisible in the truncated OPE, are related to the singulari-
ties of Π(x) located at a finite distance from the origin [52]; the distance from the
origin plays the same role as the instanton radius ρ0 . For our illustrative purposes it
is sufficient to consider the singularities of the simplest possible structure, namely,
1
, ln (x2 + ρ20 ), (x2 + ρ20 ) ln (x2 + ρ20 ), and so on . (69)
x2 + ρ20
The Fourier transforms of these expressions have the generic form

(Qρ0 )−n Kn (Qρ0 ), n = 1, 2, ... (70)

where Kn is the McDonald function. At large Euclidean Q2 the corresponding

contribution dies off exponentially,

∆Π = (Qρ0 )−n Kn (Qρ0 ) ∝ (Qρ0 )−n−1/2 e−Qρ0 , (71)

in full accord with intuition regarding transmitting a large momentum through a

soft field fluctuation.
The terms that are exponentially small for Euclidean values of Q2 become oscil-
lating upon analytic continuation to the Minkowski domain, Q → iQ. Analytically
continuing ∆Π given in Eq. (71) we arrive at
π √ √ π √ √
Im ∆Π = (−1)n+1 ( sρ0 )−n Jn ( sρ0 ) ∝ (−1)n+1 ( sρ0 )−n−1/2 cos( sρ0 − δn ) ,
2 2
(72)
2
where Jn is the Bessel function. The oscillating character at q > 0 is the most
important signature of the exponential terms representing the high-order tail of the
power series.
Having fixed ρ0 is unrealistic. We would do much better than that by allowing
ρ0 to vary around a typical value of order of Λ−1 . In other words, any reasonable
model will introduce a distribution over ρ0 , with a weight function w(ρ0 ) peaked at
Λ−1 , Z √ √ dρ0
hIm∆Πi ∝ ( sρ0 )−n Jn ( sρ0 )w(ρ0 ) (73)
ρ0
where the angle brackets denote ρ0 smearing. The weight function w(ρ0 ) is model-
dependent. Particular details of the result will certainly depend on w(ρ0 ), but

44
the general pattern – the exponential fall-off of Im∆Π modulated by oscillations –
remains the same under any reasonable choice of w(ρ0 ). This is the reason why the
experimental spectral density in Sect. 4 is extrapolated beyond the range of the
direct measurements in an oscillating mode (Fig. 17). The exponential suppression
factor at large E takes the form

exp(−kE σ ) , (74)

where the critical index σ in various models lies in the interval 0 < σ < 2.
It remains to be added that the exponential terms are usually ignored in the sum
rule analyses one encounters with in practical applications.

6 Practical Version of OPE

In practical calculations one can hardly go beyond several lowest-order terms in
the αs expansion of the coefficient functions and in the condensate expansion. For
instance, in the ρ meson channel, which is the most advanced, four terms of the
perturbative series are known in the unit operator.
Thus, the approximation we make at the very first stage of building OPE is
truncating both the perturbative and condensate series. By doing so we expect that
those few terms that are kept represent D(Q2 ) or I(M 2 ) inside the window with
sufficient accuracy.
Besides truncating the series, one usually ignores the hard non-perturbative
terms in the coefficient functions. The reason why these terms are neglected is
quite obvious: no framework allowing one to reliably calculate them was worked
out so far. From the instanton studies we know [3] that such terms must have an
extremely steep M dependence, so that above a critical value of M their impact is
expected to be totally negligible. It is assumed that the critical value of M, where
the hard non-perturbative terms die off, lies to the left of the window.
This is not the end of the story, however. The third key element inherent to
the practical OPE is a simplified treatment of the µ dependence in the coefficient
functions and the condensates. The perturbative parts of the coefficient functions
are often borrowed from earlier calculations of relevant Feynman graphs which,
sometimes, date back to QED. These calculations make no distinction between small
and large virtual momenta. Moreover, in analytic QCD it is technically not always
easy to explicitly carry out the Wilsonean separation of virtual momenta (higher
than µ and lower than µ). At the very least, such a separation requires dedicated
analyses.
This is an obvious task for the future. A simplified strategy of today is as follows.
Assume that we do not plan to go beyond the one-loop correction of Fig. 13 in the
perturbative part of CI and beyond the leading-order result for CG (and higher

45
condensates). The corresponding proper expression for the D function is
∞ π3 F ′
Z
D = 1 + αs dk 2 F (k 2 , Q2 ) + hOG (µ)i (75)
µ2 3
(cf. Eq. (68)). Here F ′ is dF/dk 2 at k 2 = 0. The dimension-six condensate GD 2 G
enters with the coefficient proportional to F ′ ′ at k 2 = 0, and so on.
Now, let us add and subtract to the right-hand side the integral
Z µ2
αs dk 2 F (k 2 , Q2 ) . (76)
0

This integral looks exactly as the one-gluon contribution naively continued below
µ2 . Of course, below µ2 the gluon propagator has nothing to do with 1/k 2 , but since
the transformation is identical, we do nothing wrong. Equation (75) then takes the
form
π3 F ′
" #
Z ∞ 3αs
Z µ2
2 2 2 2 2
D = 1 + αs dk F (k , Q ) + hOG (µ)i − 3 k dk (77)
0 3 π 0

where in the square brackets, I expanded the integrand in k 2 and kept only the
leading term. The next term proportional to k 4 would be important at the level of
dimension-six condensate, etc.
The first integral on the right-hand side is the full perturbative one gluon cor-
rection to CI , with no Wilsonean separation (it is equal to αs /π). The term in the
square brackets can be called an effective one-loop gluon condensate. It is obtained
from the genuine condensate by subtracting from it its perturbative one-loop expres-
sion. By construction, the µ dependence of the effective one-loop gluon condensate
cancels provided we do not go beyond the accuracy specified above.
Adding the integral (76) to the right-hand side we achieve the desired goal. Now
the coefficient CI , to order O(αs ), is given by full perturbative graphs of Fig. 13.
To avoid double counting we subtract the same contribution from the condensates.
More exactly, we subtract almost the same contribution. Since the sum over con-
densates is truncated at some finite order n0 , the best we can do is to subtract from
the condensate part the integral
µ2 n0
Z
2 1 (n) 2
F (k = 0, Q2 ) (k 2 )n .
X
αs dk (78)
0 n=1 n!

In this way we subtract from each condensate its “one-loop perturbative value”.
It is quite clear that this strategy of converting the genuine Wilson OPE into the
practical version is admittedly approximate and works only if we limit ourselves to a
given number of loops and given number of terms in the condensate expansion. The
effective condensates obtained in this way should be supplemented by a superscript
indicating the number of loops in the subtracted part. Their numerical value is µ
independent but does depend on the number of loops.

46
 This approximate procedure (i) leads to a loss of factorization of short and large
distance contributions inherent to Wilson’s OPE; (ii) can not be systematically
generalized to arbitrary number of loops since there is no way to unambiguously
define the integrand to all orders in the small k 2 domain 11 ; (iii) only approximately
avoids double counting due to the necessity of truncating the condensate series at
a finite order (the lower the order we truncate the larger the error). Moreover,
the effective condensates are not universal, strictly speaking; they become process-
dependent. If this non-universality were numerically significant, the SVZ method,
as it exists now, would be undermined.
Therefore, the practical version is useful in applications only provided µ2 can
be made small enough to ensure that the “one-loop perturbative” contributions to
the condensates are much smaller than their genuine values and, at the same time,
αs (µ2 )/π is small enough for the expansion to make sense. The existence of such
“µ2 window” is not granted a priori and is a very fortunate feature of QCD. We
do observe this feature empirically although the roots of the phenomenon remain
unclear.
Although the origin of the µ2 window, ensuring the applicability of practical
OPE 12 is not understood, one can find similar situations in simplified settings. The
best example of which I am aware is the two-dimensional O(N) sigma model in the
limit of large N. In this limit, the model is exactly solvable; therefore any question
of interest can be exhaustively answered. The model bears remarkable parallels
to QCD – it is asymptotically free, and the mass scale is generated dynamically.
A surprising result was established some time ago [53]. In the large N limit all
vacuum condensates in this model are µ independent, and the practical version of
OPE becomes exact; µ dependence in the condensates and in the coefficient functions
comes only at the level of 1/N. If N is large these terms are parametrically small.
(It would be great to find a hidden parameter which would explain the suppression
of the µ2 /Q2 terms in QCD in the same way 1/N explains it in the sigma model.)
It is often asked whether it is legitimate to retain in the truncated expansions
logarithmic and power terms simultaneously 13 . Indeed, formally any 1/(ln Q)n term
is parametrically larger that, say, 1/Q4 (moreover, any power term is parametrically
larger than the exponential ones). If so, the theoretical uncertainty due to truncation
of the logarithmic series is formally larger than even the lowest-order condensate
correction.
The logarithmic and power terms in practical OPE have distinct physical nature.
In some instances power terms describe effects that do not show up in perturbation
theory. This is valid in all cases where the chiral quark condensate is involved. In
11
Within practical OPE we automatically forbid to ourselves questions concerning any aspects
of the high order behavior.
12
A natural abbreviation for practical OPE would be POPE, but I do not risk to put this
abbreviation into circulation.
13
In discussing the issue of duality violations one deals with the logarithmic, power, and expo-
nential terms simultaneously, see Ref. [48].

47
other cases, when the power terms are not different in their structure from those
occurring in perturbation theory, there is a strong numeric enhancement of the
power terms (see Sect. 3). This mysterious fact – the numerical enhancement of the
vacuum condensates – explains why Wilson’s OPE can be substituted in QCD by
the practical version, at least in the classical channels .
In the early days of QCD, in 1970’s, it was common to assume that the coefficient
functions in the operator product expansion are saturated by perturbation theory,
with no separation of virtual momenta. Although the difference between Wilson’s
OPE and the practical version was realized long ago [18], very little has been done
to investigate this difference and perfect the procedure. Only now we are witnessing
attempts in this direction, mostly in connection with the heavy quark theory (e.g.
[19, 48, 54]).

7 Low Energy Theorems

Although the low-energy theorems are independent of the sum rules, at least some
of them were derived in connection with the development of the SVZ method, and
were later combined with the sum rules giving rise to an intriguing observation of
the hadronic non-universality. This observation is the topic of the next section. Here
we will dwell on the low-energy theorems related to the trace anomaly of QCD.
In the limit of massless quarks (and we will never leave the limit mq = 0)
the classical Lagrangian of QCD is free from dimensional parameters. The scale
parameter appears at the quantum level,

8π 2
!
Λ = M0 exp − 2 (79)
bg0

where M0 is the ultraviolet cut off, g0 is the bare coupling constant, and I ignored the
second and higher-order terms in the β function. (The corresponding modifications
are quite trivial and are suggested to the reader as an exercise.)
The occurrence of the scale parameter (79) is in one-to-one correspondence with
the trace anomaly (39). Let us denote the trace operator

θµµ ≡ σ . (80)

Then, it is not difficult to show [3] that for arbitrary local operator O
 Z 
lim i d4 x eiqx hT {O(0), σ(x)ic = −d hOi , (81)
q→0

where d is the (normal) dimension of the operator O, and the subscript c denotes
the connected part (it will be omitted hereafter). If

O = σ ∝ G2 ,

48
then the dimension d = 4, so that
* ( )+ * +
−bαs 2 −bαs 2 −bαs 2
Z
iqx 4
i e dx T G (x), G (0) = (−4) G at q 2 = 0 .
8π 8π 8π
(82)
For the quark operator q̄q the dimension d = 3, and so on.
a
The derivation is straightforward. First, we rescale the gluon field, Gµν = g0 Gaµν ,
so as to factor out the g0 dependence in the Lagrangian. Instead of Eq. (5) we now
have
1 a a
L = − 2 Gµν Gµν + quark term .
4g0
Then we observe that
∂
Z
i d4 x hT {O(0), G 2(x)}i = − hOi .
∂(1/4g02 )

Since hOi = Λd , we use Eq. (79) to perform differentiation on the right-hand side.
In this way we arrive at the formula (81).
In fact, the simple derivation outlined above should be supplemented by some
regularization since the matrix elements hOi, as well as the two-point functions (81),
are divergent unless the cut-off at µ is introduced. In Ref. [3] (see Appendix B)
it is shown that the regularization procedure does not affect Eq. (81) provided the
cut-off is introduced in a concerted way.
Denote the two-point function on the left-hand side of Eq. (82) as Πs (Q2 ); the
subscript s marks the scalar glueball channel. If Πs (0) is known, this knowledge can
be immediately traded for a power term in the Borel-transformed sum rule. Indeed,
in this case, instead of borelizing the dispersion relation for Πs (Q2 ), one can deal
with the dispersion relation for

Πs (Q2 ) − Πs (0)
.
Q2
After borelization one gets a prediction for the integral
1 ds
Z
2
ImΠs (s) e−s/M .
πM 2 s
Note the occurrence of an extra factor s−1 in the weight function in the integrand
(cf. Eq. (24)). At the very end

b
Πs (0) = hOG i
2
is transferred to the right-hand side of the sum rule where it acts as a coefficient in
the corresponding power correction.

49
 In this way one arrives at

32π 3 16π 4
" #
ds 1
Z
2
ImΠs (s) e−s/M = αs2 (M) 1 + 2 hOG i + ... , (83)
b2 M 4 s bαs (M) M 4

where the first term in the square brackets corresponds to the free gluon loop,
the second term is due to Πs (0), while the ellipses stand for all other corrections,
perturbative and non-perturbative. Comparing the gluon condensate correction to
that in the ρ meson sum rule, Eq. (63), we observe, with astonishment, a very
strong numerical enhancement of the relative coefficient, by a factor
−2
48 αs

∼ 102 .
b π
The standard condensate expansion, with a few first terms kept, does not match
this huge low-energy constant. Attempts to smoothly match this gigantic power
term with other non-perturbative corrections and to saturate the sum rule in the
scalar glueball channel led us [3] to the conclusion of extremely violent distortions
of the spectral density up to the scale s0 ∼ 25 GeV2 , to be compared with s0 ∼ 1.5
GeV2 in the ρ-meson sum rule, see Eq. (36) and Fig. 10. The scale up to which
strong violations of asymptotic freedom extend in the scalar glueball channel is
unconventionally large.
Thus, a new (numerically) large scale was discovered, that does not show up in
the classical channels. It was predicted to have a major impact on the properties
of the “non-classical” mesons. The physical picture lying behind this phenomenon
is discussed in Sect. 8. The quantum numbers of the mesons affected by the “su-
perstrong” interaction with the vacuum medium were found to be in one-to-one
correspondence with the presence of the direct instantons. This observation served
as an initial impetus for the development of the instanton liquid model of the QCD
vacuum [55] (for a review see [56]).
The topic of the QCD low-energy theorems is very vast; it reaches out to such
profound issues as the subtleties of the θ dependence, the η ′ problem, effective dilaton
Lagrangians, the ’t Hooft matching condition, etc. It has never been reviewed in
full, although it certainly deserves attention. This endeavor goes far beyond the
scope of the present lecture, and I will limit myself to fragmentary remarks on the
literature.
The low-energy limit of the n-point functions generated by GG̃ was studied in
[57]. A celebrated mass formula for η ′ was obtained in [58, 59]. An elegant low-
energy theorem for the coupling of the Goldstone mesons to the gluonic sources
was derived in Ref. [60]. The couplings of photons to the gluonic sources were
elaborated in Refs. [3, 61]. Low-energy theorems instrumental in the sum-rule
analyses, additional to those considered in [3], were found in [62]. Selected aspects
of the subject are reviewed in [63].

50
8 Are All Hadrons Alike?
The first, superficial, impression of the hadronic family may leave one quite bored.
There are no obvious small or large parameters, all quantities are of order one in the
appropriate units, and nothing special attracts attention. In particular, one might
think that the characteristic hadronic scale is given in all cases by the slope of the
Regge trajectory, i.e. universal α′ ∼ 1GeV−2 . The masses of the “most” typical
hadrons, nucleons and ρ mesons, are of this order of magnitude. Roughly at the
same momentum transfers scaling sets in in deep inelastic scattering heralding the
onset of asymptotic freedom – the observer sees (almost) free quarks with modest
corrections due to the gluon exchanges.
A careful reflection shows, however, that the hadronic family is far from being
that monotonous. Messages of its non-universality have been coming from various
sides, although it was not so easy to decipher them.
Many years ago Witten posed a question [58]:
“There is an obvious troublesome question. If m2η′ ∼ 1/Nc and the 1/Nc expan-
sion is usually a good approximation, why is mη′ not much smaller than its actual
value of almost one GeV?
This is not merely a problem of 1/Nc expansion. It is a more general phenomeno-
logical problem. The η ′ − π mass splitting violates Zweig’s rule, since it involves q q̄
annihilation. Since Zweig’s rule is usually rather good, why is the η ′ − π splitting
so large?”
A number of the glueball candidates have been experimentally observed in the
last decade. Still, none was unambiguously proved to be a glueball. Let us ask
ourselves:
Why, unlike the quark mesons, the masses of the glueballs are unusually heavy
or they have an unusually large mixing with the quark states which makes their
widths large enough to escape clear experimental identification, in defiance with the
1/Nc counting?
The answers to all these questions naturally come with the observation [3] of large
scales in the non-classical channels. Yes, m2η′ and Zweig’s rule violating parameters
are suppressed by 1/Nc and are small, but they are small compared to the intrinsic
scale s0 ∼ 25 GeV2 , rather than 1 GeV2 .
The existence of the diverse scales in the hadronic family is due to the fact that
various currents injecting the valence quarks/gluons in the vacuum medium interact
with this medium non-universally. If in the ρ meson (and similar classical mesons)
the impact of the vacuum fields on the valence quark-antiquark pair is modest and
is reasonably well described by the average densities of the vacuum fields, in the
glueballs the coupling to the vacuum fields is much stronger, so that they actually
feel fine details (a grain structure) of the vacuum medium. The strongest coupling
to the vacuum fields takes place in the channels with the total spin zero, both quark
and gluon (scalar and pseudoscalar quarkonia and glueballs). Phenomenologically
deviations from Zweig’s rule are most drastic here.

51
 The occurrence of the “superstrong” coupling manifests itself in an abnormally
large critical scale which may or may not materialize as the mass of the lowest-lying
state with the given quantum numbers. By critical scale I mean the energy at which
the asymptotic perturbative regime sets in in the spectral density.
A combined analysis of the sum rules and the low-energy theorems has led us [3]
to a number of qualitative (and, in many instances, semi-quantitative) predictions.
First of all, the hierarchy of glueball masses was established. The scalar glueball was
predicted to be the lightest, with mass M(0+ ) ∼ 1.5 GeV (see [64]). Its coupling to
the two-pion state was shown to be very large, with no trace of the Zweig suppression.
The pseudoscalar and tensor 2+ glueballs were predicted to lie in the vicinity of 2
GeV.
In both scalar and pseudoscalar channels the valence gluons injected in the vac-
uum are affected by the vacuum medium in the most drastic way, and the critical
scale is the largest. In the tensor glue channel the critical scale was expected to
be smaller than in the scalar/pseudoscalar channels, although larger than for the
classical mesons.
Why the 0± channels are so special, with the strongest reaction of the medium?
It was natural to assume that the phenomenon had to do with the structure of the
dominant vacuum fluctuations. It was noted [3] that instantons would do the job,
since the direct instantons show up exactly in the 0± channels. This was sufficient
for a qualitative classification. Quantitative instanton-based models of the QCD
vacuum appeared later [55, 56]; they automatically incorporate all relevant features
of this peculiar picture. Hunting for instantons in the vacuum is the vogue of the
day in the lattice community. There are indications that they are abundant and
almost saturate the string tension.
It remains to be added that, many years after these predictions, we are witnessing
how other approaches started recovering the very same pattern of non-universality.
Recent calculations in the instanton liquid model produce – not surprisingly – a
very close hierarchy of masses and mixing parameters [65]. What is interesting, this
model explains where the large non-perturbative scale in the scalar glueball channel
is hidden. The violent vacuum fluctuations squeeze this state to an unusually small
radius because of a very strong attraction. The radius of the 0+ glueball in this
model is only 0.2 fm, to be compared with 0.6 fm in the ρ meson case [65]. The
ratio of the radii squared is ∼ 1/10. It is clear that no constituent model compatible
with quantum mechanics can accommodate a state light and narrow simultaneously
(see the end of this section).
This part of the hadronic family is being intensively explored on the lattices
too. Of course, the light quarks are usually non-dynamical on today’s lattices, and
a typical lattice site of 0.2 fm is comparable with the radius of the 0+ glueball.
With all these reservations in mind it is still instructive to compare what the lattice
community learned about the non-classical hadrons. To this end I borrowed relevant
numbers from the talks [66] and original publications to be cited below:
(i) The lightest glueball is scalar, and its mass is close to 1.6 GeV;

52
 (ii) The tensor glueball is significantly heavier; its mass is slightly above 2 GeV;
the pseudoscalar glueball is found at approximately the same place, with larger
errors;
(iii) The sizes of scalar and tensor glueballs were found to be drastically different.
This can be inferred from the different magnitude of the finite size effects (see
e.g. [67]), or seen directly, in the glueball Bethe-Salpeter amplitudes (or “wave
functions”) [68, 69]. The radius of the scalar glueball was found to be 0.2 fm,
while that of the tensor one is much larger, approximately 0.8 fm. A similar lattice
measurement of the ρ meson size yields approximately 0.5 fm.
Needless to say that the lattice calculations provide us only with final numbers,
with no insight as to the mechanisms of the non-universality. To the best of my
knowledge, no lattice publication on this topic mentions the previous sum rule pre-
dictions. True, the lattice results have a potential to be perfected in the future,
while the accuracy of the sum rule method is admittedly limited.
To make the highly nontrivial nature of our findings more contrast let us conclude
this section by comparing them to the expectations of conventional models, say, the
bag model. The quark and (electric) gluon modes in the spherical cavity have
energies 2.04/R and 2.7/R, respectively, where R is the cavity radius. Thus, the
glueball states are expected to be only marginally heavier than the quark states.
In particular, if the spin-dependent forces are neglected the model predicts that
M(2++ ) ≈ M(0++ ) ≈ 1 GeV and M(0−+ ) ≈ 1.3 GeV. Including the spin-dependent
forces changes these numbers rather insignificantly [70]. The complex picture of the
vacuum medium, which most strongly affects the scalar and pseudoscalar mesons,
especially glueballs, is completely missed.

9 Ecological Niche
In the 1980’s the original strategy of the SVZ method was tested, with a remarkable
success, in analyzing practically every static property of all established low-lying
hadronic states. The method was developed in various directions – three-point
functions, inclusion of external electromagnetic and other auxiliary fields, light-cone
modifications and so on – which allowed one to expand the range of applicability to
such advanced problems as magnetic moments [71], form factors [72] at intermediate
momentum transfers, weak decays [73] and structure functions [74] of deep inelastic
scattering at intermediate x. At the initial stage the method was essentially unchal-
lenged since the lattice calculations were lagging far behind burdened by multiple
internal problem of lattice QCD.
Now the situation has changed. The machinery of lattice QCD was bettered, and
it became a powerful tool in many problems where qualitative insight is unimportant,
and the prime emphasis is on numbers. Let us remember that the SVZ method is
admittedly approximate: typically, theoretical accuracy is at the level of 20%. In
some cases it does not work at all.

53
 In order to survive, the approach based on OPE and the sum rules has to find a
range of applications in the hadronic physics, where it could successfully compete.
I must say that it copes with the challenge. Its strongest advantage is the analytic
character of calculations which can be continued to the Minkowski space, term by
term. Thus, the most promising directions of growth are those where the Minkowski
kinematics is entangled in this or that way and/or the processes under consideration
develop in more complicated media, rather than in the vacuum medium. (Such
problems are very difficult for lattice QCD which operates with vacuum correlation
functions numerically evaluated in the Euclidean space.) Let me give three examples.
1) Particle decays with a large spatial momentum carried away by produced
hadrons, for instance, B → ρℓν. If the invariant mass of the lepton pair is small,
q 2 ∼ 0, the spatial momentum of the ρ meson is of order MB /2 ≫ Λ. Treating
the problem in the Euclidean domain we will have to face the necessity of an ex-
tremely remote analytic continuation. At the same time the light-cone sum rule, to
be discussed in Sect. 10, are perfectly fit to deal with this problem.
2). Heavy flavor sum rules. They allow one to determine basic parameters of the
heavy quark theory – the masses of the heavy quarks, form factors at zero recoil,
and so on – in terms of directly measurable quantities. This is a close relative of the
SVZ sum rules. The main distinction is that instead of the vacuum condensates one
deals with the expectation values of various local operators over the heavy hadron
state, for instance the B meson. This approach will be also considered in brief in
Sect. 10.
3) Pre-asymptotic effects in inclusive processes at high energies. A typical prob-
lem from this class is determination of the lifetime differences in the b quark family.
Although the task is somewhat different from those usually treated in the SVZ sum
rules, the method of its solution is exactly the same. The key technical element is
the operator product expansion. In the given context it is imperative to address,
additionally, an issue going beyond OPE (more exactly, its practical version) – du-
ality violations. The topic of the deviations from duality is of paramount practical
importance; it came under renewed scrutiny recently [16].
In all these and other instances the OPE-based methods remain to be the only
theoretical tool available on the market today. We see that the approach is suffi-
ciently rich and flexible to remain viable in the future. I am sure it will continue to
play a key role in solving many applied problems in the hadronic physics in the next
decade. The potential for further growth is there, the active stage is not yet over.

10 New Developments
To give you a flavor of what is going on in the this field I will pick up two or three
subjects which come to my mind first. My selection by no means presents the full
picture.

54
 The heavy quark mass 1/mQ expansions that flourished in the 1990’s revived
interest in the SVZ sum rules for hadrons containing one heavy quark 14 (in practice,
b quark). It was realized – and this is a new element – that (i) the hybrid logarithms
occur in the theoretical part of the sum rules, and (ii) they may result in a strong
enhancement of the gluon radiative corrections. Hybrid logarithms depend on the
ratio mQ /µ, they were unknown before 1988. A typical example where summation of
the hybrid logarithms significantly shifts the answer is fB . I single out this problem
because this parameter is fundamentally important in the heavy quark physics and
because of an instructive story attached to it. Early sum rule calculations of fB ,
which missed then undiscovered hybrid logarithms, yielded a number close to 130
MeV [76]. The main impact of the hybrid logarithms is a change of the argument
of the running coupling constant in the sum rule: instead of αs (mb ) the radiative
corrections, as it turns out [77], are governed by αs (1 GeV). As a result, the sum
rule prediction jumped up to fB = 160 MeV, with the theoretical uncertainty 20 to
30 MeV [77].
Meanwhile, fB was the object of the multiple lattice studies. The first lattice
measurements produced an unbelievably huge value, somewhere around 250 MeV.
With elimination of various sources of the systematic uncertainties the lattice num-
ber has been systematically decreasing. The modern number, quoted at recent
conferences, is close to 180 MeV, with the error in the same ballpark as the sum rule
uncertainty. Thus, the initial contradiction between the two methods faded away
giving place to perfect agreement.
Along with the technical developments of the type I have just mentioned the
method experienced ideological developments too. Below we will discuss some of
them.

10.1 Light-cone sum rules

This formalism was designed to overcome difficulties in the traditional sum rules
for three-point functions. Assume we want to calculate a cubic coupling constant
corresponding to the amplitude A → B + C, where A, B and C stand for either
hadrons or external currents, say, electromagnetic. Typical examples are D ∗ → Dπ
or γ ∗ + γ ∗ → π (here γ ∗ denotes a virtual photon). In many instances in order
to get the desired constant it is necessary to apply borelization with respect to two
momenta, pA and pB , independently. Then the components of the third momentum,
pC , can not be small compared to p2A,B . Indeed, pC (pA + pB ) = p2A − p2B ∼ p2A,B . The
standard condensate expansion contains a series of operators with derivatives which,
eventually, give rise to the expansion parameter of the type pC (pA + pB )/p2A,B ∼ 1.
In other words, all terms in this subseries must be summed over.
This (partial) summation is carried out automatically if, instead of the three-
point function hjA , jB , jC i one considers the correlation function of the currents jA
14
The 1/mQ expansion of the SVZ sum rules per se dates back to the work of Shuryak [75].

55
and jB sandwiched between the vacuum and the state |Ci. The vacuum expecta-
tion values of local operators (condensates) in the SVZ sum rules are substituted by
the light-cone wave functions (distribution amplitudes in the American literature)
describing the momentum fraction distribution of the components in the Fock wave
function of the hadron C. If the SVZ approach is based on the short-distance expan-
sion of the current T products in terms of local operators, and the expansion runs
in dimensions of the local operators, the light-cone sum rules exploit the expansion
in nonlocal “string” operators on the light-cone; the expansion runs in twists, not
dimensions. From the viewpoint of the standard SVZ sum rules, in the light-cone
approach one performs a partial summation of an infinite chain of operators of ar-
bitrary dimension, but given twist. In effect, the light cone sum rules combine the
SVZ approach with the technique used in the description of hard exclusive processes
(see [78]). The price we have to pay for the partial summation is rather high: the
large distance dynamics is parametrized not by numbers, as in the SVZ condensates,
but by functions – the leading twist, next-to-leading twist, and so on distribution
functions.
A Russian proverb says that it is better to see once than to hear hundred times.
Let us have a closer look at a specific example. Below I will outline the calculation of
the D ∗ Dπ constant. This example, as well as the explanations below, are borrowed
from Ref. [79].
The D ∗ Dπ coupling constant g is defined as follows:

hD ∗ (p)π(q)|D(p + q)i = −g qµ ǫµ ,

where the hadrons’ momenta are indicated in the parentheses, and ǫµ is the D ∗
polarization vector. To determine this constant from the light cone sum rules one
considers the correlation function
Z
Fµ (p, q) = i d4 x eipx hπ(q)|T {d̄(x)γµ c(x), c̄(0)iγ5u(0)}|0i . (84)

The first operator in the T product is the interpolating current for D ∗ and the
second is the interpolating current for D.
The correlation function (84) depends on two variables, p2 and (p + q)2 (note
that q 2 = 0 since the pion is on mass shell). The contribution proportional to the
D ∗ Dπ amplitude has two poles, in p2 and (p + q)2 , corresponding to the ground
state mesons both in the vector and pseudoscalar channels. Namely,

MD2 MD∗ fD fD∗

Fµ (p, q) = g qµ + ... . (85)
mc (p2 − MD2 ∗ )((p + q)2 − MD2 )

The ellipses on the right-hand side denote a term proportional to pµ irrelevant for
our analysis, as well as terms with single pole and non-pole continuum contributions,
that will be exponentially suppressed under double borelization (with respect to p2
and (p + q)2 ). All other notations are self-explanatory.

56
 p p+q
c

d ξq u

111111111111111
000000000000000
000000000000000
111111111111111
000000000000000
111111111111111
000000000000000
111111111111111
000000000000000
111111111111111
000000000000000
111111111111111
000000000000000
111111111111111
q π

Figure 20: Light cone sum rule for D ∗ Dπ in the leading order. The current labeled
by p + q interpolates D while that labeled by p interpolates D ∗ .

On the theoretical side we assume that p2 −MD2 ∗ and (p+q)2 −MD2 are Euclidean
and large enough to allow for the operator product expansion of the correlation
function (84). As usual, the corresponding Borel parameters will lie in the window.
The simplest graph pertinent to OPE we have to build is depicted in Fig. 20. The
large distance part of the process is denoted by the shaded area. The light quark
lines form the pion, rather than the quark condensate, that was discussed in Sect.
4. More exactly, we deal here with the leading twist pion wave function whose
definition will be given shortly.
If ξ is the fraction of the pion momentum carried by the d quark, the graph of
Fig. 20 gives rise to the following expression:
1 φπ (ξ)
Z
2 2
Fµ ≡ qµ F (p , (p + q) ) = mc fπ qµ dξ + ... , (86)
0 m2c − (p + ξq)2

where φπ (ξ) is the distribution function, and the ellipses denote corrections, both
perturbative (the αs expansion) and non-perturbative (due to higher twists). The
c quark Green’s function is substituted, in the leading approximation, by that of a
free quark. Emission of gluons from this line would result either in αs corrections
or in the distribution functions of non-leading twists.
For very small values of q oneRcould have substituted the denominator in Eq. (86)
by m2c − p2 . Then the integral dξφπ (ξ) → 1, and Fµ would reduce to the matrix
element of the operator dγ¯ µ γ5 u sandwiched between the pion state and vacuum.
Expanding the denominator in Eq. (86) in q we readily identify higher dimension
←
operators that are summed over: the term linear in q originates from d¯ Dα γµ γ5 u,

57
 ← ←
quadratic in q from d¯ D α D β γµ γ5 u, etc. The actual expansion parameter is
pq
∼ 1.
p − m2c
2

The (leading twist) light-cone wave function φπ (ξ) is formally defined as

Z x Z 1
¯
hπ(q)|d(x)γ µ γ5 exp{ig Aα (y)dyα}u(0)|0i = −iqµ fπ dξ eiξqx φπ (ξ) . (87)
0 0

¯ pair.
It parametrizes the large-distance dynamics of the du
If we do the double Borel transformation, with respect to p2 and (p + q)2 the
denominator in Eq. (86) becomes
M12 M22 2 2
M12
! !
2 M1 + M2
exp −mc δ u − , (88)
M12 + M22 M12 M22 M12 + M22
where M1,2 are the corresponding Borel parameters, that are expected to lie close to
each other inside the window. The sum rule is obtained by matching the double Borel
transform of Eq. (85) with the double Borel transform of Eq. (86). Thus, we see
that the theoretical part of the sum rule is determined by φπ (0.5). Of course, there
are corrections of higher order in αs and from higher twist distribution functions,
but as usual inside the window they should be kept at a modest level. In addition
we must check the stability of the result inside the window. Provided all this is
properly done, one obtains [79] g in terms of φπ (0.5). The latter quantity must be
extracted from independent sources.
(The situation we face here is not quite generic. A priori one could expect the
coupling g to be expressible in terms of φπ (ξ), rather than one constant, φπ (0.5).)
Information on φπ (ξ) is quite abundant; there is some limited information on the
higher-twist wave functions too. The issue of φπ (ξ) is almost as old as the SVZ sum
rules themselves. The key element of the corresponding theory, which was studied
in great detail in connection with hard exclusive processes, is the (approximate)
conformal symmetry of QCD. Other elements are provided by the sum rule and
lattice calculations and models. The issue is not completely settled, although it
seems that a heated debate of the 1980’s gradually wanes. Out of two competing
scenarios – a double hump distribution function of Chernyak and Zhitnitsky [78]
and a narrow distribution of Radyushkin and collaborators [80] that is close to the
asymptotic form of φπ (ξ) – the second is gaining more recognition now. Experts
lean towards a narrow almost asymptotic distribution. As usual, the best strategy
is sacrificing one of the sum rules in order to extract φπ (ξ) from data. A recent
attempt based on the γ ∗ + γ → π 0 data has been undertaken in Ref. [81]. I will
not go into further details here, referring the reader to a very rich original literature
devoted to this topic.
It remains to be added that (i) the term “light-cone sum rules” first appears in
[82]; (ii) the idea of the approach dates back to Refs. [83]; and (iii) a review of this
topic, with a representative list of references was published recently [84].

58
 If the partial summation of the higher-dimension operators turns out to be so
successful in the light-cone sum rules, it is natural to ask [85] why it is not applied
in the vacuum correlation functions appearing in the SVZ sum rules. In this case,
in order to partially sum, say, the quark operators we must substitute the quark
condensate hq̄qi by a “ string expectation value” in the vacuum,
Z x
hq̄(x) exp{ig Aα (y)dyα}q(0)i , (89)
0

which is a function of x.
In the light cone sum rules one can ascribe certain twist to each given distribu-
tion function. Each distribution function sums up a tower of the local operators,
with all dimensions but given twist. Therefore, the distribution functions can be
systematically ordered in twist. Twist is obviously irrelevant in the vacuum cor-
relation functions. Therefore, the partial summation implemented by the “string
expectation values” (89) is hard to justify theoretically. A principle that would
allow us to order distinct “string expectation values” has to be elaborated.
Another major difference with the light cone sume rules is that the pion distri-
bution function is known at least at a certain level, while next to nothing is known
about the vacuum function (89). Until recently even the large x asymptotic behav-
ior of this function was erroneously assumed to be Gaussian. We are at the initial
stage, when first observations are being made regarding properties of the function
(89) following from the general structure of QCD [16]. I do not rule out that, as our
understanding progresses, the corresponding modification of the sum rule approach
will be elaborated.

10.2 Heavy flavor sum rules

Although the character of problems in which this formalism is instrumental is some-
what different from the classical applications of the SVZ sum rules, the basic tech-
niques – OPE, matching of the theoretical expansions with the phenomenological
expressions represented by dispersion integrals over the observable spectral densi-
ties, etc. – are the same. The vacuum condensates are replaced by the expectation
values of local operators over the heavy flavor states, say, B mesons. All questions
that can be asked in connection with the vacuum condensates exist here too; the
answers are more transparent, however. We will consider below a sample applica-
tion. Our main goal is discussing subtle nuances associated with the condensate
expansion (Wilsonean OPE versus practical version, see Sect. 6).
The heavy flavor sum rules were engineered [19] to treat the inclusive heavy
flavor decays, for instance the semileptonic decays B → Xc ℓν where Xc is arbitrary
hadronic state containing c quark. The lowest-lying state is the D meson, then
comes D ∗ and then higher excitations/continuum. By varying the momentum of
the lepton pair one can measure various spectra in the transition b → c induced by

59
 q1
00
1
b c
0000011111111
00000000
111
000 1b
111111111111111111
11111
000000000000000000
111
000 11111
00000
0
000000000000000000
111111111111111111
B { 000000000000000000
111111111111111111
000000000000000000
111111111111111111
000000000000000000
111111111111111111
} B
1111111111
0000000000
11111111111
00000000000
Figure 21: The operator product expansion for hab . For large negative ǫ the c quark
is far off shell, and the points of the current emission and absorption are close. The
B meson in the initial and final states is assumed to be at rest.

the weak current. The object of the theoretical study is the transition operator
Z
T̂ab (q) = i d4 x eiqx T {ja† (x)jb (0)} , (90)

that will be eventually sandwiched between the states |Bi and hB| (see Fig. 21).
Here ja denotes a current of the type c̄Γa b with an arbitrary Dirac matrix Γa ; q is
the momentum carried away by the lepton pair.
The average of T̂ab over the heavy hadron state B with the momentum pB rep-
resents a forward scattering amplitude (the so-called hadronic tensor),
1
hab (pB , q) = hB|T̂ab |Bi . (91)
2MB
The structure functions wab are obtained from the hadronic tensor by taking its
imaginary part,
wab = (1/i) disc hab .
All observable spectra are represented as certain integrals over wab . The hadronic
tensor hab and the structure functions wab can be expanded in terms of various
kinematic structures. For each set of the currents generically we may have up to five
distinct kinematic structures; these details need not bother us here. For illustrational
purposes it is sufficient to limit ourselves to the axial currents, so that Γa,b → γµ γ5 ,
and to the first function,

hµν = −h1 gµν + ... , w1 = (1/i) disc h1 .

The lowest-lying state produced by the axial current from B is D ∗ . To end up with
the kinematical aspects we must specify the lepton pair momentum q. It will be
assumed that the spatial momentum ~q is small but non-vanishing, Λ ≪ |~q| ≪ MD .
The terms O(~q2 ) will be kept while those of higher order in |~q| will be neglected.
Since ~q is fixed, hAA
1 becomes a function of one (complex) variable, q0 . Instead of q0
it is more convenient to work with ǫ defined as

ǫ = q0max − q0 (92)

60
where
~q2
q0max = MB − ED∗ , ED∗ = MD∗ + . (93)
2MD∗
When ǫ is real and positive we are on the physical cut where the imaginary part of
hAA
1 is measurable. For negative ǫ we are below the cut, in the Euclidean domain,
where the T product (90) can be computed as an expansion in 1/mc,b . The leading
operator in this expansion is b̄b. It has dimension three. No operators of dimension
~ 2 b and
four exist. At the next level, of dimension five, there are two operators, b̄(iD)
µ ν
(ig/2)b̄γ γ Gµν b. The first operator (sometimes called the kinetic energy operator)
is not a Lorentz scalar. In the problem considered the operators to be retained in
the expansion need not necessarily be Lorentz scalars since, unlike the vacuum case,
the very presence of the B meson in the initial/final state singles out a reference
frame.
The whole procedure is perfectly analogous to the 1/Q2 expansion of the polar-
ization operator discussed in Sect. 4. The operator b̄b occupies the same position in
the hierarchy as the unit operator in the expansion of the vacuum correlators. The
operators of dimension five and higher generate 1/mc,b corrections analogous to the
condensate corrections. In the problem at hand all local operators appearing in the
expansion of the T product (90) are averaged over the B meson state. In this way
we get a theoretical expression for hAA1 off the cut.
A bridge between h1 at negative ǫ and w1AA at positive ǫ is provided by the
AA

dispersion relation. Expanding the denominator in the dispersion relation in 1/ǫ we

obtain an infinite set of the sum rules. The third of them is as follows [86]:
1 1
Z
dǫ ǫ2 w1AA (ǫ) = µ2π ~v 2 + ... . (94)
2π 3
Here ~v is the velocity of the produced heavy hadron (in the B rest frame), and µ2π
is defined as the expectation value
1 ~ 2 b|Bi .
µ2π = hB|b̄(iD) (95)
2MB
This parameter, µ2π , has the meaning of the average spatial momentum squared
of the heavy quark inside B meson, and is one of the fundamental parameters of
the heavy quark theory. One encounters µ2π in a large number of applied problems.
It is fair to say that in the heavy quark physics µ2π plays the same role as the gluon
condensate in classical applications of the SVZ method.
The ellipses in Eq. (94) denote corrections suppressed by powers of 1/mc,b.
Since the weight function in Eq. (94) is proportional to ǫ2 , the “elastic” B → D ∗
transition drops out, and the integral on the left-hand side is saturated by “inelastic”
transitions, i.e. decays of the B meson in the excited Xc states. For such transitions
w1AA (ǫ) is proportional to ~v 2 at |~v| ≪ 1. It is convenient then to introduce a spectral
function σ(ǫ) which does not vanish in the limit |~v| → 0,
σ(ǫ) = |~v|−2 w1AA (ǫ) , ǫ > 0 . (96)

61
 ε2 σ(ε) versus ε

Λ µ
Figure 22: A sketch of ǫ2 σ(ǫ) in the b → c transition at small |~v| (solid curve). The
spectral density σ(ǫ) is defined in Eq. (96). The dashed curve represents ǫ2 σ(ǫ) in
perturbation theory at one loop.

Our brief excursion in the heavy flavor sum rules is almost over. In terms of the
new spectral density the sum rule we are interested in takes the form
1 Z 1
dǫǫ2 σ(ǫ) = µ2π . (97)
2π 3
As with any sum rule it can be read in both directions: if µ2π is known this is a
prediction for the spectral density. On the other hand, if the spectral density is
measured we can take the integral and get µ2π . Below we will pursue the latter
strategy assuming that σ(ǫ) is measured. In actuality this is not the case, but we
have a pretty good idea of how σ(ǫ) will look like when it is measured. First of
all, it is positive-definite. At small values of ǫ, of order of a few units time Λ,
there is one, perhaps, two resonances; then at higher values of ǫ the spectral density
must approach its perturbative asymptotics which is known too: σ(ǫ) → Cαs /ǫ,
where C is a constant which was calculated in Ref. [19]. The numerical value of
C is unimportant for our purposes. The asymptotic regime is achieved through
oscillations – the reasons are the same as in Sects. 2.2 and 4. A schematic plot of
ǫ2 σ(ǫ) is depicted in Fig. 22.
If you think about Eq. (97) you will, probably, find it remarkable: it gives a
condensate-like non-perturbative parameter in terms of an integral over the observ-
able spectral density. Further examination raises a question. Indeed, if the integral
extends to infinity, it badly diverges.
One need not be concerned, however. This infinity reflects the fact that the
proper definition of µ2π requires a normalization point, just in the same way as the
for the gluon condensate. In the case at hand this normalization point has a very
transparent physical meaning – µ is the maximal excitation energy of the hadronic

62
state belonging to the class of “soft excitations” (Fig. 22). Higher excitations
are considered to be hard, and are well approximated by perturbation theory. I
remind that the hard contributions do not belong to µ2π ; rather, they determine
the coefficient functions in OPE. In the perturbative calculation of the coefficient
functions one must remove the part with ǫ < µ. In fact, the cut off in the excitation
energy as a normalization point was suggested in the given context long ago in Ref.
[87]. Thus, in the Wilsonean approach µ2π (µ) is determined as
3 µ
Z
µ2π (µ) = dǫǫ2 σ(ǫ) . (98)
2π 0

In practical OPE we deal, instead, with subtracted “surrogates”. Say, if one

limits oneself to one-loop accuracy,
3 µ
Z
“µ2π ” = dǫǫ2 [σ(ǫ) − σ1−loop (ǫ)] .
2π 0

The right-hand side in this expression is µ independent at one loop. Its µ dependence
shows up only at two loops. If we agree to discard all effects O(αs2 ) and higher,
just totally ignore them, then the integration in “µ2π ” can be extended to infinity.
Correspondingly, the coefficient functions must be calculated at one-loop, following
the standard Feynman rules, with no removal of the ǫ < µ part.

11 Sum Rules and Lattices

In this Lecture it has been already noted, more than once, that certain aspects
of the condensate approach, as well as a rich experience in the analysis of various
spectral densities, might be useful for lattice practitioners. Remarks to this effect
are scattered here and there. Of particular importance are qualitative observations,
such as an enhanced sensitivity to the vacuum fermion loops in certain channels (this
effect is neglected in the quenched approximation), the existence of new, abnormally
large scales in the 0± glueball channels, and so on. Surprising though it is, all these
hints, which might have produced a strong impact on the lattice theory, are so far
almost totally ignored 15 . The reaction of rejection of any ideas coming from outside
the lattice theory per se plagues the lattice community. This happens even in those
cases where the overlap is quite obvious.
Let me give an almost anecdotal example. As was mentioned in Sect. 7, in
1981 a wide class of “QCD scale anomaly” low-energy theorems was obtained [3] in
connection with the SVZ sum rules. These theorems relate to each other n-point
functions with arbitrary number of insertions of the operator σ(x) = θµµ (x) ∝ G2
(see Eq. (39)) at vanishing momentum. One of this theorems can be cast in the
15
I am aware of two dedicated works where the sum rule results where analyzed in parallel with
the lattice calculations [88]. This seems to be a rare exception.

63
 x3

R/2

x4
-T/2 T/2

-R/2

Figure 23: Rectangular contour in the Wilson loop operator. One can consider it
as a trajectory of infinitely heavy quarks (in the imaginary time). R is the distance
between the quark and antiquark, T → ∞.

form of a prediction for the expectation value of G2µν in the presence of the Wilson
loop operator W ,
∂V (R)
Z 
hW i−1 d3 xσ(x)W = V (R) + R (99)
E,c ∂R
where V (R) is the static potential between the heavy quarks separated by distance
R, I
W = Tr exp(igAµ (x)dxµ ) ,
C
the contour C in the definition of the Wilson loop is depicted on Fig. 23, and the
subscript E reminds us that all definitions and calculations refer to the Euclidean
space. Moreover, the integration in Eq. (99) is performed in three directions per-
pendicular to the Euclidean time; the subscript c refers to the connected part, i.e.
the disconnected part of the correlation function (99) must be discarded. One can
interprete the operator σ(x) in Eq. (99) as a probe of the energy density of the vac-
uum medium at the point x in the presence of a pair of heavy static quark sources
separated by distance R [3]. In the case of purely linear potential the right-hand
side reduces to 2V (R). In this form the theorem (99) was derived in Ref. [3] (see
Appendix B). The generalization to the case of arbitrary potentials is due to Dosch
et al. [89].
Derivation of the theorem (99) is very similar to other theorems discussed in Sect.
7. The functional-integral representation for hW i is differentiated with respect to
1/g02. We then use the fact that

32π 2
!
∂V (R) ∂V (R)
2
= − V (R) + R . (100)
∂(1/4g0 ) b ∂R
The latter equality, in turn, follows from consideration similar to that presented in
Sect. 7. Thus, the low-energy theorem (99) is a reflection of the scale anomaly of
QCD, and as such, represents a fundamental aspect of the theory.
Now I come to a culmination of the story. In 1987 the theorem (99) was redis-
covered in the lattice community [90], where it goes under the name of the “action

64
Michael sum rule”. The original derivation was erroneous, the second term on the
right-hand side of Eq. (99) was omitted. The error was corrected in Ref. [89]
which was specifically devoted to the issue. Nevertheless, the theorem continues to
circulate in the lattice community as the “Michael sum rule” , see e.g. [91].
This example of a “lattice xenophobia” is by no means unique. Needless to say
that I consider the situation as absolutely unhealthy. The problems we face in the
hadronic world are too complicated to afford neglecting the work done in related
areas. The analytic and numerical approaches should complement each other. I
sincerely hope that the attitude of the lattice community will start changing after
recent breakthrough discoveries in supersymmetric gauge theories. Supersymmetry
is a powerful tool which allows us to reveal features of strong coupling gauge dynam-
ics we never suspected of. One of the lessons which seems undeniable is the proof
that massless (or light) quarks play much more significant a role than is universally
believed among the lattice practitioners. In the quenched approximation, the most
wide spread today, the dynamical quarks are neglected altogether. Even the most
advanced lattice investigations, which do not rely on the quenched approximation,
routinely use extrapolations in the number of the massless quarks. Supersymmetric
gauge theories teach us that increasing the number of massless quarks from two to
three or changing the color representation of the fermion fields from fundamental to
adjoint is not necessarily a smooth process. Dynamics may change drastically!
Penetration of ideas and insights obtained analytically, in the practice of lattice
calculations, where they can be used, at the very least, to test reliability of various
approximations, is unavoidable. First steps in this direction have been already
reported [92].

12 Vacuum Fluctuations Are Subtle Creatures

The rich SVZ phenomenology confirms the wide-spread belief that the structure of
the hadronic family, in all its gross features and intricacies, derives from a compli-
cated organization of the vacuum medium. The vacuum fluctuations are violent, but
fine-tuned in a very specific way. Within the method, as it exists today, we do not
calculate the vacuum condensates. This does not mean, of course, that the origin
of the condensates, and their relative values, are of no interest. I merely wanted to
say that the problem of the vacuum structure in analytic QCD is not solved in full.
There are indirect methods, however, which show that this structure is very subtle
– much more subtle than one might think of a priori – and depends on seemingly
insignificant details of the theory. For instance, changing the number of colors from
three to two dramatically changes the pattern of the spontaneous chiral symmetry
breaking resulting in a drastic rearrangement in the Goldstone sector of the hadron
family. Instead of Nf2 − 1 QCD “pions” (Nf stands for the number of the light quark

65
flavors) we get 2Nf2 − Nf − 1 massless Goldstone hadrons 16 . This conclusion can be
achieved [93] from consideration of the ’t Hooft matching conditions [94] (for earlier
analyses based on less rigorous arguments see Ref. [95]).
Let us dwell on this issue. Assume that we have Nf massless (light) quarks
but the color gauge group is SU(2). The quarks belong to the fundamental rep-
resentation of SU(2). Unlike QCD, in which the color triplets and antitriplets are
independent representations, the SU(2) doublets are essentially the same as anti-
doublets. In other words, all representations of SU(2) are (quasi)real.
This simple observation has far reaching consequences. Indeed, the flavor sym-
metry of the Lagrangian is SU(2Nf ) now, rather than SU(Nf ) × SU(Nf ) we deal
with in QCD. Out of 4Nf2 − 1 currents 2Nf2 − Nf − 1 are axial; all correspond to
the spontaneously broken symmetries since there is no way one can achieve the
matching of the triangular AVV anomalies otherwise [93]. This means that the
pattern of the chiral symmetry breaking is SU(2Nf ) → Sp(2Nf ) rather than the
SU(Nf ) × SU(Nf ) → SU(Nf ) we got used to in QCD. The very gross features of
the vacuum structure do depend on the number of colors!
Not only do they depend on the number of colors, they are also sensitive to the
fermion contents of the theory.
Indeed, let us substitute the standard quarks of QCD with one Majorana fermion
λ in the adjoint representation of the color group. The theory thus obtained is
nothing but supersymmetric gluodynamics. Superficially it looks very similar to
conventional QCD. At first sight, one can hardly expect any conspicuous deviations
from the conventional pattern of behavior.
It has been known for many years that the supersymmetric gauge theories posses
some miraculous properties. In particular, in supersymmetric gluodynamics the
Gell-Mann–Low function is known exactly [96]

αs2 3N
β(αs ) = − , (101)
2π 1 − (Nαs /2π)

where the SU(N) gauge group is assumed.

As QCD, supersymmetric gluodynamics is believed to be confining in the in-
frared domain. Unlike QCD, however, the condensate hλ2 i may or may not develop.
It was argued recently [97] that two distinct phases coexist in this theory – the con-
ventional chirally asymmetric phase with hλ2 i =6 0, that reminds QCD, and a very
unusual chirally symmetric phase where hλ2 i = 0. The arguments leading to this
conclusion are too technical to be reproduced here. They would lead us far astray.
The interested reader is referred to Ref. [97].
Instead, let us dwell on another example which is pretty close to actual QCD. In
fact, we will consider just conventional QCD with more than three massless flavors.
16
It is curious to note that the existence of the extra massless states was discovered in the lattice
SU (2) theory, which is being used as a theoretical laboratory for decades, only recently. The result
is known in analytic QCD since the mid-eighties. This is another manifestation of the “lattice
xenophobia”, see Sect. 11.

66
 The Gell-Mann-Low function in QCD has the form [6]

αs2 α3
β(αs ) = −b − b1 s2 − ... ,
2π 4π
2 19
b = 11 − Nf , b1 = 51 − Nf . (102)
3 3
At small αs it is negative (asymptotic freedom!) since the first term always domi-
nates. With the scale µ decreasing the running gauge coupling constant grows, and
the second term becomes important. Generically the second term takes over the first
one at αs /π ∼ 1, when all terms in the αs expansion are equally important, i.e. in
the strong coupling regime. We do not know what exactly happens at αs /π ∼ 1. As-
sume, however, that for some reasons the first coefficient b is abnormally small, and
this smallness does not propagate to higher orders. Then the second term catches up
with the first one when αs /π ≪ 1, we are in the weak coupling regime, and higher
order terms are inessential. Inspection of Eq. (102) shows that this happens when
Nf is close to 33/2, say 16 or 15 (Nf has to be less than 33/2 to ensure asymptotic
freedom). For these values of Nf the second coefficient b1 turns out to be negative.
This means that the β function develops a zero in the weak coupling regime, at
αs∗ b
= ≪ 1. (103)
2π −b1
(Say, if Nf = 15 the critical value is at 1/44 and is indeed small.) This zero is
nothing but the infrared fixed point of the theory. At large distances αs → αs∗ , and
β(αs∗ ) = 0, implying that the trace of the energy-momentum vanishes. Then the
theory is in the conformal regime. There are no localized particle-like states in the
spectrum of this theory; rather we deal with massless unconfined interacting quarks
and gluons; all correlation functions at large distances exhibit a power-like behavior.
In particular, the potential between two heavy static quarks at large distances R
will behave as ∼ αs∗ /R, i.e. we have a pure Coulomb behavior. The situation is not
drastically different from conventional QED. As long as αs∗ is small, the interaction of
the massless quarks and gluons in the theory is weak at all distances, short and large,
and is amenable to the standard perturbative treatment (renormalization group,
etc.). The chiral symmetry is not broken spontaneously, the quark condensates do
not develop. QCD becomes a fully calculable theory.
The fact that at Nf close to 16 QCD becomes conformal and weakly coupled in
the infrared limit is known for about 20 years [98, 99]. If the conformal regime takes
place at Nf = 16 and 15, let us ask ourselves how far can we descend in Nf without
ruining the conformal nature of the theory in the infrared? Certainly, when Nf is
not close to 16 the gauge coupling αs∗ is not small. The theory is strongly coupled,
as conventional QCD, and, simultaneously, conformal in the infrared. The chiral
symmetry presumably is unzbroken, hq̄qi = 0. The vacuum structure is totally
different, and so are all properties of the theory.

67
 The range of Nf where this phenomenon occurs is called the conformal window.
The right edge of the conformal window is at Nf = 16. There are indications that
the left edge of the conformal window Nf∗ lies not too far from actual QCD (i.e. not
too far from Nf = 3).
First, in the instanton liquid model it was found [100] that at Nf = 5 the chiral
condensate disappears, hq̄qi = 0. Although nothing is said about the onset of the
conformal regime in this work, the vanishing of hq̄qi may be a signal. More definite
is the conclusion of the lattice investigation [101]. It is claimed that at Nf = 7
not only the quark condensate disappears, hq̄qi = 0, but the decay law of the
correlation functions at large distances changes from the exponential to power-like,
i.e. the theory switches from the confining regime to the conformal one. One could
speculate that at Nf = 5 and 6 we are in the intermediate phase, when confinement
is still operative at large distances but the chiral symmetry is unbroken. Previously
it was believed that that confinement of color in QCD implies the chiral symmetry
breaking [102]. Today’s wisdom tells us that this need not be necessarily the case.
At least in some supersymmetric gauge theories we have both: confinement and
unbroken chiral symmetry (see e.g. [97]). In any case, we see that ascending from
Nf = 3 to Nf = 5 or 7 – quite a modest variation– we drastically change the vacuum
structure of the theory, with the corresponding abrupt change of the picture of the
hadron world.
An interesting question is what happens slightly below the left edge of the con-
formal window Nf∗ . If Nf were a continuous parameter, and the phase transition
in Nf were of the second order, slightly below Nf∗ the string tension σ would be
parametrically small in its natural scale given by Λ2 . In actuality Nf changes dis-
cretely. Still it may well happen that at Nf = Nf∗ − 1 the ratio σ/Λ2 is numerically
small. This would mean that the string and its excitations, whose scale is set by
σ, are abnormally light. Under the circumstances one might hope to build effective
low-energy approaches analogous to the chiral Lagrangians of actual QCD. In the
limit σ/Λ2 → 0 the string becomes, in a sense, classic; the quantum corrections are
unimportant. I do not rule out that the string representation of QCD – the holy
grail of two generations of theorists – is easier to construct in this limit. One could
even dream of an expansion in Nf∗ − 3. At this moment, this is a pure speculation,
however, and I have to wind up.

13 Instead of Conclusions
As time passes the hopes that the full analytic solution of QCD will be found fade
away. After all, the problem is with us for over a quarter of the century, and none
of numerous theoretical attacks reached the goal. Many theorists whose philosophy
is “all or nothing” abandoned the field 17 . Does it mean that the theory of hadrons
17
The “all or nothing” philosophy is wide-spread and, unfortunately, not only in theoretical
physics. This is a favorite child of the so called revolutionaries in all times and in all countries.

68
is an aging science at the verge of submerging into a permanently dormant state?
I do not think so. Those theorists who stayed in the field can claim many partial
successes. The field is messy and down-to-earth but – what can we do? – this is the
only world God gave to us. The theory of hadrons is alive. Our understanding of
the vacuum structure of quantum chromodynamics and various dynamical regimes
it can support continues to grow. Especially fruitful were the last years, with many
breakthrough discoveries in supersymmetric gauge theories. These discoveries are
to be used as hints and insights in actual QCD.
OPE-based methods – and the SVZ sum rules is one of them – continue to grow
too. They are admittedly approximate, but their analytic nature and a graphic phys-
ical interpretation make them indispensable in many practically important problems.
In some issues they share their role with other methods, which were developed later,
for instance, lattice QCD. Since the theory of hadrons is so difficult it is in our best
interests to combine all sources of information. We are in no position to neglect one
of them in favor of the other. Many ideas and theoretical devices surfaced in con-
nection with the SVZ method (e.g. the low-energy theorems, Sect. 7); the method
also produced a number of variations and spin-offs. This is a healthy process which
will hopefully continue.

Acknowledgments
It is my pleasure to thank Prof. T. Suzuki and Prof. T. Kunihiro who were my
hosts during my 1997 visit to Japan, for their kind hospitality. Prof. T. Suzuki
kindly arranged this lecture in the framework of the International Workshop Non-
Perturbative QCD – Structure of the QCD Vacuum, Yukawa Institute for Theoretical
Physics, Kyoto, December 2–12, 1997.
Useful communications with V. Braun, H.G. Dosch, B.L. Ioffe, A. Kaidalov, A.
Khodjamirian, A. Radyushkin and E. Shuryak are acknowledged. I would like to
thank A. Vainshtein for numerous illuminating discussions.
This work was supported in part by DOE under the grant number DE-FG02-
94ER40823.

The misfortunes it brought to our world are innumerable. Needless to say, I personally think this
is a rotten philosophy.

69
14 Appendix. Some Useful Definitions and For-
mulae.
The Gell-Mann–Low function is defined as
∂αs α2 α3
≡ β(αs ) = −b s − b1 s2 + ... (A.1)
∂ ln µ 2π 4π
This definition is standard except that in some sources the first coefficient is called
β0 , the second β1 , and so on (cf. Ref. [6]). At two loops the β function is scheme
independent.
The coefficients are
2 19
b = 11 − Nf , b1 = 51 − Nf . (A.2)
3 3
The running coupling αs (µ) is parametrized as follows
 2

4π  2b1 ln ln Λµ 2
αs (µ) = µ2
1 − 2 µ2
+ ... . (A.3)
b ln Λ2 b ln Λ2

A more convenient form of the very same expression is

2π µ b1 µ2
= b ln + ln ln 2 + ... . (A.4)
αs (µ) Λ b Λ

Being expanded, at two loops, the latter expression is equivalent to the former; Eq.
(A.4) effectively sums up some higher order terms.
The scale parameter Λ following from Eq. (A.3) is
 b1 /b ( " #)
b 2 b 2π b1 2π
Λ = µ exp − − ln . (A.5)
b αs (µ) b αs (µ)

The first factor on the right-hand side is an (inconvenient) artifact of the definition.
It could have been easily avoided. Since the definition is standard, we will keep it
not to cause confusion.
Perturbative calculations in QCD, as a rule, are carried out in the so called
modified minimal subtraction (MS) scheme [103], using dimensional regularization.
For massless quarks and gluons it works nicely; the treatment of the heavy quark
mass thresholds in this approach is ugly. The most common is the step-function ap-
proximation [104]: immediately above the threshold the quark is declared massless,
while below the threshold it is treated as infinitely heavy, so that it is frozen out and
does not participate in loops. This is equivalent to treating Nf as a function of µ
constructed from several step functions, Nf = 3 below mc , then it jumps to Nf = 4
between mc and mb , etc. Matching conditions at thresholds require equivalence of

70
one effective theory with Nf massless quarks to another effective theory with Nf − 1
massless quarks.
If the step-function approximation is applied, at one loop the coupling constant
is obviously continuous, while the first µ derivative experiences a jump. At two and
three loops the coupling itself becomes discontinuous if the matching is done at the
quark masses [105]. Several suggestions as to how one can smooth out the running
coupling constant using various “physical” definitions were presented in the litera-
ture [106]. I propose a somewhat different approach inspired by supersymmetry. In
supersymmetric theories the quark mass thresholds can be accounted for exactly, to
all orders [107]. For instance, consider supersymmetric QCD with one flavor, with
the mass term m. Assume that we start our evolution at a high normalization point
M0 (the corresponding coupling constant is αs0 ), pass the threshold and descend
down to µ ≪ m. A continuously running αs can be obtained from the formula
2π 2π M0 M0
= − 3N ln 1/3
+ ln , (A.6)
αs (µ) αs0 µ(αs0/αs (µ)) µZ(µ)
for SU(N) gauge group. Here Z(µ) is the Z factor of the matter fields and µ is
arbitrary: larger or smaller than m. At µ = m and below the second logarithm
freezes at ln M0 /m0 where m0 = mZ is the mass parameter normalized at M0 . The
one loop Z factor implies αs (µ) at two loops. Needless to say that Z(µ) varies
continuously at one loop. At the two-loop level the high- and low-energy scale
parameters are related as follows:
N  2C2
N − 3N−1 " 2C2 /(3N −1) #
2 3N − 1 2π
 
Λ3N 3N −1
LOW = ΛHIGH m0 , (A.7)
3N 2 αs0

where
N2 − 1
C2 = .
2N
At this level of accuracy m0 and αs0 in the square brackets can be treated in the
leading logarithmic approximation. In this approximation the expression in the
square brackets is renormalization-group invariant. In principle, the procedure can
be extended to all loops, but we will not pursue this goal here.
In non-supersymmetric QCD the exact treatment of the mass thresholds seems
impossible. However, numerically the situation is quite close to what we have in
supersymmetric QCD. One can work out compact formulae applicable at two loops.
Say, for the charm threshold
2π 2π M0 2 M0
= − 9 ln 32/81
+ ln , (A.8)
αs (µ) αs0 µ(αs0 /αs (µ)) 3 µY (µ)
where " #3y/25
αs0
Y =
αs (µ)

71
and
107
y= .
18
Correspondingly,
 32/9  77/25 " 12/25 #2/3 " #7/45
2 25 2π 2π

25/3
Λ9LOW = ΛHIGH m0 . (A.9)
9 6 αs0 αs (m)

If it were not for the last factor (which is quite close to unity) the relation between
the low- and high-energy scales would be similar to that in supersymmetric QCD.
Matching at two thresholds, charm and beauty, yields
 32/9  58/23 " #21/115 " #7/45
23/3 2 23 2π 2π
Λ9LOW = ΛHIGH [m∗1 m∗2 ]2/3 ,
9 6 αs (m2 ) αs (m1 )
(A.10)
where 12/23
2π

m∗i = mi0 .
αs0

References
[1] M. Shifman, A. Vainshtein and V. Zakharov, Nucl. Phys. B147 (1979) 385;
448.

[2] Vacuum Structure and QCD Sum Rules, Ed. M. Shifman (North-Holland, Am-
sterdam, 1992).

[3] V. Novikov, M. Shifman, A. Vainshtein and V. Zakharov, Nucl. Phys. B191

(1981) 301.

[4] D. Gross and F. Wilczek, Phys. Rev. Lett. 30 (1973) 1343;

H. D. Politzer, Phys. Rev. Lett. 30 (1973) 1346.

[5] V. Novikov et al., Phys. Rep. 41 (1978) 1.

[6] See e.g. the review paper I. Hinchliffe, Phys. Rev. D54 (1996) 77 and references
therein.

[7] G. ’t Hooft, The Confinement Phenomenon in Quantum Field Theory, 1981

Cargese Summer School Lecture Notes on Fundamental Interactions, Eds. M.
Levy and J.-L. Basdevant, NATO Adv. Study Inst. Series B: Phys. vol. 85, p.
639 [reprinted in G. ’t Hooft, Under the Spell of the Gauge Principle (World
Scientific, Singapore, 1994), p. 514];
S. Mandelstam, Phys. Reports 23 (1976) 245.

72
 [8] For a review see e.g. A. Di Giacomo, Nucl. Phys. Proc. Suppl. 47 (1996) 136
[hep-lat/9509036]; Monopole Condensation in Gauge Theory Vacuum, in Con-
finement 95, Proc. International Workshop on Color Confinement and Hadrons
Japan, March 1995, Eds. H. Toki, Y. Mizuno, H. Suganuma, T. Suzuki, O.
Miyamura (World Scientific, Singapore, 1995) [hep-lat/9505006]; Mechanisms
for Color Confinement, in Selected Topics in Nonperturbative QCD, Proc. Int.
School of Physics “Enrico Fermi”, Eds. A. Di Giacomo and D. Diakonov (IOS
Press, 1996), page 147.
[9] N. Seiberg and E. Witten, Nucl. Phys. B426 (1994) 19; (E) B430 (1994) 485;
Nucl. Phys. B431 (1994) 484.
[10] G. ’t Hooft, Nucl. Phys. B72 (1974) 461;
E. Witten, Nucl. Phys. B160 (1979) 57;
for a brief review see Chapter 8 in Ref. [2].
[11] A.B. Kaidalov, Usp. Fiz. Nauk 105 (1971) 97 [Sov. Phys. Uspekhi 14 (1972)
600];
P.D.B. Collins, An Introduction to Regge Theory and High Energy Physics
(Cambridge Univ. Press, 1977).
[12] A.Yu. Dubin, A.B. Kaidalov, and Yu.A. Simonov, Phys. Lett. B323 (1994) 41.
[13] A. Casher, N. Neurberger and S. Nussinov, Phys. Rev. D20 (1979) 179.
[14] B. Blok, M. Shifman and Da-Xin Zhang, hep-ph/9709333 (Phys. Rev. D, to
appear).
[15] G. ’t Hooft, Nucl. Phys. B75 (1974) 461 [Reprinted in G. ’t Hooft, Under the
Spell of the Gauge Principle (World Scientific, Singapore 1994), page 443]; see
also F. Lenz, M. Thies, S. Levit and K. Yazaki, Ann. Phys. (N.Y.) 208 (1991)
1; C. Callan, N. Coote, and D. Gross, Phys. Rev. D13 (1976) 1649; M. Einhorn,
Phys. Rev. D14 (1976) 3451; M. Einhorn, S. Nussinov, and E. Rabinovici, Phys.
Rev. D15 (1977) 2282; I. Bars and M. Green, Phys. Rev. D17 (1978) 537.
[16] M. Shifman, Theory of Preasymptotic Effects in Weak Inclusive Decays, in
Proc. Workshop on Continuous Advances in QCD, ed. A. Smilga (World Sci-
entific, Singapore, 1994), page 249 [hep-ph/9405246]; Recent Progress in the
Heavy Quark Theory, in Particles, Strings and Cosmology, Proc. of the XIX
Johns Hopkins Workshop on Current Problems in Particle Theory and the V
PASCOS Interdisiplinary Symposium, Baltimore, March 1995, Ed. J. Bagger
(World Scientific, Singapore, 1996), page 69 [hep-ph/9505289].
[17] F. David, Nucl. Phys. B209 (1982) 433; B234 (1984) 237.
[18] V. Novikov, M. Shifman, A. Vainshtein and V. Zakharov, Nucl. Phys. B249
(1985) 445.

73
[19] I. Bigi, M. Shifman, N. Uraltsev and A. Vainshtein, Phys. Rev. D52 (1995)
196.

[20] R. Crewther, Phys. Rev. Lett. 28 (1972) 1421;

M. Chanowitz and J. Ellis, Phys. Lett. B40 (1972) 397; Phys. Rev. D7 (1973)
2490;
J. Collins, L. Duncan and S. Joglekar, Phys. Rev. D16 (1977) 438.

[21] M. Neubert, Phys. Rev. D51 (1995) 5924.

[22] P. Ball, M. Beneke, and V. Braun, Nucl. Phys. B452 (1995) 563.

[23] G. ’t Hooft, Can We Make Sense out of “Quantum Chromodynamics”? in The

Whys Of Subnuclear Physics, Erice 1977, ed. A. Zichichi (Plenum, New York,
1977), p. 943 [reprinted in G. ’t Hooft, Under the Spell of the Gauge Principle
(World Scientific, Singapore, 1994), p. 547].

[24] B. Lautrup, Phys. Lett. 69B (1977) 109;

G. Parisi, Phys. Lett. 76B (1978) 65; Nucl. Phys. B150 (1979) 163;
A. Mueller, Nucl. Phys. B250 (1985) 327.

[25] A. H. Mueller, in Proc. Int. Conf. QCD – 20 Years Later, Aachen 1992, eds. P.
Zerwas and H. Kastrup, (World Scientific, Singapore, 1993), vol. 1, page 162.

[26] R. Akhoury and V. I. Zakharov, The Physics of the Ultraviolet Renormalon,

hep-ph/9710257.

[27] R.Akhoury and V. Zakharov, Nucl. Phys. Proc. Suppl. 54A (1997) 217;
M. Beneke, hep-ph/9706457; V. Braun, hep-ph/9708386; B.R. Webber, hep-
ph/9712236.

[28] A. Di Giacomo, Lattice Gauge Theories and SVZ Expansion, in Nonperturba-

tive Methods, Proc. 1985 Montpellier Int. Workshop, Ed. S. Narison (World
Scientific, Singapore, 1986), page 135.

[29] X. Ji, hep-ph/9506216 (unpublished).

[30] M. Shifman, A. Vainshtein, M. Voloshin, and V. Zakharov, Phys. Lett. B77

(1978) 80.

[31] J. Gasser and H. Leutwyler, Nucl. Phys. B94 (1975) 269.

[32] S. Weinberg, in A Festschrift for I.I. Rabi, ed. L. Motz, Trans. New York Acad.
Sci. Ser. II 38 (1977) 185.

[33] H. Leutwyler, hep-ph/9609467 and a forthcoming book.

74
[34] T. Bhattacharya, and R. Gupta, Talk at Lattice 97, XV Int. Symp. on Lattice
Field Theory, Edinburgh, Scotland, July 1997 [hep-lat/9710095].
[35] For a review see B.L. Ioffe, Acta Phys. Polon. B16 (1985) 543.
[36] V. Novikov et al., in Neutrinos 78, Proc. VIII Int. Conference on Physics and
Neutrino Astrophysics, West Lafayette, April 1978, Ed. E. C. Fowler (Purdue
University, 1978), page C278.
[37] M. Voloshin, Nucl. Phys. B154 (1979) 365;
L.J. Reinders, H. Rubinstein, and S. Yazaki, Phys. Reports 127 (1985 ) 1;
S.N. Nikolaev and A.V. Radyushkin, Nucl. Phys. B213 (1983) 285; Phys. Lett.
B124 (1983) 243.
[38] M. Dubovikov and A. Smilga, Nucl. Phys. B185 (1981) 109.
[39] M. Crisafulli et al, Phys. Lett. B369 (1996) 325.
[40] K.G. Chetyrkin et al, Phys. Lett. B174 (1986) 104; L. Reinders and S. Yazaki,
Nucl. Phys. B288 (1987) 789; K. Chetyrkin and A. Pivovarov, Nuov. Cim.
A100 (1988) 899. Literally speaking, these authors test factorization for the
four-quark operators averaged over the Goldstone mesons, π, K, etc. rather
than over the vacuum state. Within the soft pion technique the former matrix
elements are reducible to the latter.
[41] The two-loop correction was found by K. Chetyrkin, A. Kataev and F. Tkachev,
Phys. Lett. B85 (1979) 277; M. Dine and J. Sapirstein, Phys. Rev. Lett. 43
(1979) 668; W. Celmaster and R. Gonsalves, Phys. Rev. Lett. 44 (1980) 560.
Three-loop calculations are due to S. Gorishny, A. Kataev, and S. Larin, Phys.
Lett. B259 (1991) 144; L. Surguladze and M. Samuel, Phys. Rev. Lett. 66
(1991) 560; (E) 66 (1991) 2416; K. Chetyrkin, Phys. Lett. B391 (1997).
[42] S. Eidelman, L. Kurdadze, M. Shifman, and A. Vainshtein, unpublished.
[43] S. Gorishny, A. Kataev, and S. Larin, Pisma ZhETF 53 (1991) 121 [JETP Lett.
53 (1991) 127].
[44] V. Novikov, M. Shifman, A. Vainshtein and V. Zakharov, Fortsch. Phys. 32
(1984) 585.
[45] L. Surguladze and F. Tkachev, Nucl. Phys. B 331 (1990) 35.
[46] D. Broadhurst, A.G. Grozin, in Proc. Int. Workshop on Software Engineering
and Artificial Intelligence and Expert Systems for High Energy and Nuclear
Physics New Computing Techniques in Physics Research IV, Pisa, Italy, April
1995, Eds. B. Denby and D. Perret-Gallix (World Scientific, Singapore, 1995),
page 217 [hep-ph/9504400];
L. Surguladze and M. Samuel, Rev. Mod. Phys. 68 (1996) 259.

75
[47] R. Barate et al. (ALEPH Collab.), Z. Phys. C76 (1997) 15.

[48] B. Chibisov, R.D. Dikeman, M. Shifman, and N. Uraltsev, Int. J. Mod. Phys.
A12 (1997) 2075.

[49] M. Shifman, Yad. Fiz. 36 (1982) 1290 [Sov. J. Nucl. Phys. 36 (1982) 749].

[50] K. Wilson, Phys. Rev. 179 (1969) 1499;

K. Wilson and J. Kogut, Phys. Reports 12 (1974) 75.

[51] Large-Order Behaviour of Perturbation Theory, Eds. J.C. Le Guillou and J.

Zinn-Justin, (North-Holland, Amsterdam, 1990).

[52] M.S. Dubovikov and A.V. Smilga, Yad. Fiz. 37 (1983) 984 [Sov. J. Nucl. Phys.
37 (1983) 585].

[53] V. Novikov, M. Shifman, A. Vainshtein and V. Zakharov, Phys. Reports 116

(1984) 103.

[54] X. Ji, hep-ph/9506216, hep-ph/9507322 (unpublished).

[55] E. Shuryak, Nucl. Phys. B198 (1982) 83;

D. Diakonov and V. Petrov, Nucl. Phys. B245 (1984) 259.

[56] E. Shuryak, The QCD Vacuum, Hadrons and the Superdense Matter (World
Scientific, Singapore, 1988) [World Scientific Lecture Notes in Physics, v. 8];
T. Schäfer and E. Shuryak, hep-ph/9610451 (Rev. Mod. Phys., to appear) and
references therein.

[57] R. Crewther, Phys. Lett. B70 (1977) 349;

for a review see G.A. Christos, Phys. Reports 116 (1984) 251.

[58] E. Witten, Nucl. Phys. B156 (1979) 269.

[59] G. Veneziano, Nucl. Phys. B159 (1979) 213.

[60] M. Voloshin, Yad. Fiz. 44 (1986) 738 [Sov. J. Nucl. Phys. 44 (1986) 478]; Yad.
Fiz. 45 (1987) 190 [Sov. J. Nucl. Phys. 45 (1987) 122].

[61] H. Leutwyler and M. Shifman, Phys. Lett. B221 (1989) 384.

[62] I. Balitsky, D. Diakonov, and A. Yung, Phys. Lett. B112 (1982 ) 71; Z. Phys.
C33 (1986) 265;
for a review see D. Diakonov, in Proc. XXI Winter School of Physics (LINP,
Leningrad, 1986), page 3.

[63] M. Shifman, Phys. Reports 209 (1991) 341.

[64] M. Shifman, Z. Phys. C9 (1981) 347.

76
[65] T. Schäfer and E. Shuryak, Phys. Rev. Lett. 75 (1995) 1707.

[66] C. Michael, in Proc. X Les Rencontres de Physique de la Vallee d’Aoste: Results

and Perspectives in Particle Physics, La Thuile, Italy, March 1996 Results and
Perspectives in Particle Physics, Ed. M. Greco (Istituto Naz. Fis. Nucl., 1996)
[hep-ph/9605243];
T. DeGrand, hep-th/9610132.

[67] P. van Baal and A. Kronfeld, Nucl. Phys. B (Proc. Suppl.) 9 (1989) 227.

[68] K. Ishikawa, G. Schierholz, H. Schneider, and M. Teper, Nucl. Phys. B227

(1983) 221;
T. DeGrand, Phys. Rev. D36 (1987) 176;
R. Gupta et al., Phys. Rev. D43 (1991) 2301.

[69] P. de Forcrand and K.-F. Liu, Phys. Rev. Lett. 69 (1992) 245.

[70] C.E. Carlson, T.H. Hansson and C. Peterson, Phys. Rev. D27 (1983) 1556; (E)
D28 (1983) 2895;
M. Chanowitz and S. Sharpe, Nucl. Phys. B222 (1983) 211.

[71] B.L Ioffe and A.V. Smilga, JETP Lett. 37 (1983) 298; Nucl. Phys. B232 (1984)
109;
I. Balitsky and A. Yung, Phys. Lett. B129 (1983) 328;
see also Chapter 7 in Ref. [2].

[72] V. Nesterenko and A. Radyushkin, Phys. Lett. B115 (1982) 410;

B.L Ioffe and A.V. Smilga, Nucl. Phys. B216 (1983) 373;
see also Chapter 6 in Ref. [2].

[73] B. Blok and M. Shifman, Sov. J. Nucl. Phys. 45 (1987) 135; 301; 522; Sov. J.
Nucl. Phys. 46 (1987) 1310;
for a review and a representative list of references see A. Khodjamirian and R.
Rückl, hep-ph/9801443.

[74] V. Belyaev and B.L. Ioffe, Int. J. Mod. Phys. A6 (1991) 1533;
B.L. Ioffe and A. Khodjamirian, Phys. Rev. D51 (1995) 3373.

[75] E. Shuryak, Nucl. Phys. B198 (1982) 83.

[76] T. Aliev and V. Eletsky, Sov. J. Nucl. Phys. 38 (1983) 936.

[77] E. Bagan, P. Ball, V. Braun, and H.G. Dosch, Phys. Lett. B278 (1992) 457.

[78] V.L. Chernyak and A.R. Zhitnitsky JETP Lett. 25 (1977) 510;
A.V. Efremov and A.V. Radyushkin, Phys. Lett. B94 (1980) 245;
G.P. Lepage and S.J. Brodsky, Phys. Lett. B87 (1979) 359; Phys. Rev. D22

77
 (1980) 2157
for a review see V.L. Chernyak and A.R. Zhitnitsky, Phys. Reports 112 (1984)
173.

[79] V.M. Belyaev, V.M. Braun, A. Khodjamiryan and R. Rückl Phys. Rev. D51
(1995) 6177.

[80] A.V. Radyushkin, Nucl. Phys. A532 (1991) 141; hep-ph/9707335.

[81] A. Khodjamirian, hep-ph/9712451.

[82] P. Ball, V.M. Braun and H.G. Dosch, Phys. Rev. D44 (1991) 3567.

[83] N.S. Craigie and J. Stern, Nucl. Phys. B216 (1983) 209;
I. Balitsky, V. Braun, and A. Kolesnichenko, Nucl. Phys. B312 (1989) 509 and
earlier works of the same authors cited therein.

[84] V. Braun, hep-ph/9801222.

[85] S.V. Mikhailov and A.V. Radyushkin, Phys. Rev. D45 (1992) 1754;
A.V. Radyushkin, Phys. Lett. B271 (1991) 218;
A.P. Bakulev and A.V. Radyushkin, Phys. Lett. B271 (1991) 223.

[86] I Bigi et al., Phys. Lett. B339 (1994) 160.

[87] N. Isgur and M. Wise, Phys. Rev. D43 (1991) 819.

[88] M.C. Chu, et al., Nucl. Phys. Proc. Suppl. 30 (1993) 495;
J. W. Negele, M. Burkardt, and J. M. Grandy, Ann. Phys. (N.Y.) 238 (1995)
441.

[89] H.G. Dosch, O. Nachtmann and M. Reuter, hep-ph/9503386 (unpublished).

[90] C. Michael, Nucl. Phys. B280 (1987) 13.

[91] H. Rothe, Phys. Lett. B364 (1995) 227.

[92] N. Evans, S. Hsu and M. Schwetz, hep-th/9707260.

[93] M. Vysotsky, I. Kogan and M. Shifman, Yad. Fiz. 42 (1985) 504 [Sov. J. Nucl.
Phys. 42 (1985) 318].

[94] G. ’t Hooft, in Recent Developments in Gauge Theories, Eds. G. ’t Hooft et al.

(Plenum Press, New York, 1980).

[95] S. Dimopoulos, Nucl. Phys. B168 (1980) 69;

M. Peskin, Nucl. Phys. B175 (1980) 197.

78
[96] V. Novikov, M. Shifman, A. Vainshtein and V. Zakharov, Nucl. Phys. B229
(1983) 381; Phys. Lett. B166 (1986) 329;
M. Shifman and A. Vainshtein, Nucl. Phys. B296 (1988) 445.

[97] A. Kovner and M. Shifman, Phys. Rev. D56 (1997 ) 2396.

[98] A. Belavin and A. Migdal, Pis’ma ZhETF 19 (1974) 317 [JETP Lett. 19 (1974)
181]; Scale Invariance and Bootstrap in the Non-Abelian Gauge Theories, Lan-
dau Institute Preprint-74-0894, 1974 (unpublished).

[99] T. Banks and Zaks, Nucl. Phys. B196 (1982) 189.

[100] T. Schäfer and E. Shuryak, Phys. Rev. D53 (1996) 6522.

[101] Y. Iwasaki et al., Z. Phys. C71 (1996) 343.

[102] A. Casher, Phys. Lett. B83 (1979) 395.

[103] W.A. Bardeen et al., Phys. Rev. D18 (1978) 3998.

[104] W. Marciano, Phys. Rev. D29 91984) 580;

R.M. Barnett, H.E. Haber and D.E. Soper, Nucl. Phys. B306 (1988) 697;
G. Rodrigo and A. Santamaria, Phys. Lett. B313 (1993) 441.

[105] W. Wetzel, Nucl. Phys. B196 (1982) 259; W. Bernreuther and W. Wetzel,
Nucl. Phys. B197 (1982) 228; W. Bernreuther, Ann. Phys. (N.Y.) 151 (1983)
127; Z. Phys. C20 (1983) 331; S. Larin, T. van Ritbergen, and J.A.M. Ver-
maseren, Nucl. Phys. B438 91995) 278; K. Chetyrkin, B.A. Kniehl, and M.
Steinhauser, Phys. Rev. Lett. 79 (1997) 2184.

[106] S. Brodsky, M. Gill, and J. Rathman, hep-ph/9801330, and references therein.

[107] M. Shifman, Int. J. Mod. Phys. A11 (1996) 5761.

79
 Recommended Literature

Two review papers

B.L. Ioffe, Acta Phys. Polon. B16 (1985) 543;
L.J. Reinders, H. Rubinstein, and S. Yazaki, Phys. Reports 127 (1985) 1
were published in the mid-1980’s. They may serve for an initial exposure but they
do not provide an overview of modern developments. Neither do they reflect mod-
ern understanding of conceptual issues. A dedicated comprehensive review on this
subject is long overdue.
A brief survey of the light-cone sum rules is given in
V. Braun, hep-ph/9801222.
SVZ sum rules in the heavy quark theory are discussed in
M. Neubert, Phys. Reports 245 (1994) 259.
For a review of the sum rule applications in weak decays see
A. Khodjamirian and R. Rückl, hep-ph/9801443.
Some topics related to the SVZ sum rules are covered in
M. Shifman, Phys. Reports 209 (1991) 341;
T. Schäfer and E. Shuryak, hep-ph/9610451 (Rev. Mod. Phys., to appear).
A collection of the original papers with an extended commentary reflecting the state
of the art in the late 1980’s can be found in
Vacuum Structure and QCD Sum Rules, Ed. M. Shifman (North-Holland,
Amsterdam, 1992).
Some useful formulae are compiled in
S. Narison, QCD Spectral Sum Rules, (World Scientific, Singapore, 1989).
I do not agree with the treatment of many conceptual issues and technical details
in this book.

80