Lectures On The Functional Renormalization Group Method: Janos Polonyi

CEJP 1 (2003) 1{71
Lectures on the functional renormalization

group method
¤
Janos Polonyi1;2
1
Laboratory of Theoretical Physics,
Louis Pasteur University, Strasbourg, France
2
Department of Atomic Physics,
L. EÄotvÄos University, Budapest, Hungary
Received 20 June 2002; revised 7 December 2002
Abstract: These introductory notes are about functional renormalization group

equations and some of their applications. It is emphasised that the applicability of this
method extends well beyond critical systems, it actually provides us a general purpose
algorithm to solve strongly coupled quantum eld theories. The renormalization group
equation of F. Wegner and A. Houghton is shown to resum the loop-expansion. Another
version, due to J. Polchinski, is obtained by the method of collective coordinates and can
be used for the resummation of the perturbation series. The genuinely non-perturbative
evolution equation is obtained by a manner reminiscent of the Schwinger-Dyson equations.
Two variants of this scheme are presented where the scale which determines the order of
the successive elimination of the modes is extracted from external and internal spaces.
The renormalization of composite operators is discussed brie®y as an alternative way
to arrive at the renormalization group equation. The scaling laws and xed points are
considered from local and global points of view. Instability induced renormalization and
new scaling laws are shown to occur in the symmetry broken phase of the scalar theory.
The ®attening of the e¬ective potential of a compact variable is demonstrated in case of
the sine-Gordon model. Finally, a manifestly gauge invariant evolution equation is given
for QED.
®c Central European Science Journals. All rights reserved.
Keywords: strongly coupled theories, non-perturbative approximation, universality

PACS (2000): 25.75.-q
¤
E-mail: polonyi@fresnel.u-strasbg.fr
2 J. Polonyi / Central European Journal of Physics 1 (2003) 1{71
Contents
1 Introduction 3
2 Functional RG equations 4
2.1 Resumming the loop expansion 5
2.1.1 Blocking in continuous space-time 5
2.1.2 Local potential approximation 7
2.1.3 Gradient expansion 10
2.2 Resumming the perturbation expansion 14
2.2.1 Polchinski equation 15
2.2.2 Gradient expansion 18
2.3 Composite operator renormalization 19
2.3.1 Toy model 20
2.3.2 Quantum Field Theory 21
2.3.3 Parallel transport 23
2.3.4 Asymptotical scaling 24
2.3.5 Perturbative treatment 25
2.4 Continuous evolution 26
2.5 Blocking in the internal space 29
2.5.1 Wilsonian action 30
2.5.2 E¬ective action 31
3 Applications 33
3.1 Fixed points 34
3.1.1 Rescaling 34
3.1.2 Reparametrizing 35
3.1.3 Local potential approximation 36
3.1.4 Anomalous dimension 37
3.2 Global RG 38
3.2.1 RG in Statistical and High Energy Physics 38
3.2.2 Scalar model 42
3.2.3 RG microscope 46
3.2.4 The Theory of Everything 50
3.3 Instability induced renormalization 50
3.3.1 Unstable e¬ective potential 52
3.3.2 Tree-level WH equation 53
3.3.3 Plane wave saddle points 53
3.3.4 Correlation functions 56
3.3.5 Condensation as crossover 56
3.4 Sine-Gordon model 58
3.4.1 Zoology of the sine-Gordon model 58
3.4.2 E¬ective potential 60
3.4.3 Breakdown of the fundamental group symmetry 61
3.4.4 Lower critical dimension 62
3.5 Gauge models 63
3.5.1 Evolution equation 63
3.5.2 Gauge invariance 64
4 What has been achieved 65
References 66
J. Polonyi / Central European Journal of Physics 1 (2003) 1{71 3
1 Introduction
The origin of renormalization goes back to hydrodynamics, the de¯nition of the mass
and other dynamical characteristics of bodies immersed into °uids. The more systematic
elaboration of this concept is based on the (semi) group property of changing the observa-
tional scale in a theory. The renormalization group (RG) method has already been used
with a number of di®erent goals. Some of the more important directions are (i) to remove
U.V. divergences [1], (ii) to describe the scale dependence of physical parameters and to
classify the parameters of a theory around a critical point according to their impact on
the dynamics [2], (iii) to express the highly singular product of local ¯eld variables [3],
(iv) to resum the perturbation expansion in Quantum Field Theory [4] and in the case of
di®erential equations [5] and ¯nally (v) to solve strongly coupled theories. Furthermore,
the RG method o®ers a systematic way of organizing our understanding of a complicated
dynamics by successively eliminating degrees of freedom. The distinctive feature of the
method is that it retains the in°uence of the eliminated subsystem on the rest.
The main subject of these lectures is application (v) in the framework of the functional
formalism, the search of a general purpose algorithm to handle non-perturbative and not
necessarily critical theories.
The strategy of the RG scheme for the resummation of the perturbation expansion is to
introduce and to evolve e®ective vertices instead of dealing with higher order corrections.
The result is an iterative procedure where the contributions to the e®ective vertices
computed at a given step in the leading order of the perturbation expansion and inserted
in a graph at the next step reproduce the higher order corrections. In the traditional
implementation of the RG procedure one follows the evolution of few coupling constants
only. Our concern will be an improvement on this method in order to follow the evolution
of a large number of coupling constants. This is realized by means of the functional
formalism where the evolution is constructed for the generator function for all coupling
constants.
Before starting let us review the conventional multiplicative RG schemes as used in
Quantum Field Theory. This scheme is based on the relation
Ã ! Ã 2!
(n) (n) · 2 pj
G R (p1 ; ¢ ¢ ¢ ; pn ; gR ; · )= Z ¡n=2 GB (p1 ; ¢ ¢ ¢ ; pn ; g B ; ¤) +O +O ; (1)
¤2 ¤2
for the Green functions, where ¤ and · denote the U.V. cut-o® and the substraction
(observational) scales, respectively. Z depends on either gR or gB and the ratio ¤=· (the
mass is treated as one of the parameters g). The ignored terms stand for non-universal,
cut-o® dependent interactions.
Bare RG equation: The renormalized theory is independent of the choice of the
cut-o®,
d (n) d h n=2 (n)
i
¤ G R (p1 ; ¢ ¢ ¢ ; pn ; gB ; ¤) = ¤ Z (g B ; ¤)GB (p1 ; ¢ ¢ ¢ ; pn ; g B ; ¤) = 0; (2)
d¤ d¤
where the renormalized coupling constants and the substraction scale are kept ¯xed in
computing the derivatives. The equation involving the renormalized quantities is not
too useful. It is the multiplicative renormalization scheme which leads to the second
equation, expressing the possibility of compensating the change of the substraction scale
by the modi¯cation of the coupling constants gB (¤) arising from the solution of the
di®erential equation by the method of characteristics. Notice that the compensation
is possible only if the list of the coupling constants fg B g includes all relevant coupling
constants of the appropriate scaling regime, around the U.V. ¯xed point.
Renormalized RG equation: The bare theory is independent of the choice of the sub-
straction scale,
d (n) d h ¡n=2 (n)
i
· G B (p1 ; ¢ ¢ ¢ ; pn ; g R ; · ) = · Z (g R ; · )GR (p1 ; ¢ ¢ ¢ ; pn ; gR ; · ) = 0; (3)
d· d·
where the bare coupling constants and the cut-o® are kept ¯xed.
Callan-Symanzik equation: The change of the mass in the propagator is governed by
the expression
d 1 1 1
= 2 ¢ 2 (4)
dm p ¡ m
2 2 2 p ¡ m p ¡ m2
2
which can be used to ¯nd out the dependence on the renormalized mass when the cut-o®
and all bare parameters except the bare mass are kept ¯xed. The resulting evolution
equation [6] is similar to the two previous RG equations except that the derivative is
with respect to the bare mass instead of the cut-o® or the substraction scale and the
right hand side is non-vanishing. This latter feature indicates that contrary to the ¯rst
two schemes the evolution gR (m2 ) in the Callan-Symanzik equation is not a renormalized
trajectory, it connects theories with di®erent mass.
The serious limitation of these equations is that they are asymptotic, i.e., are appli-
cable in the regime · 2 ; p2 << ¤2 only. In fact, the omission of the non-universal terms in
the multiplicative renormalization scheme (1) requires that we stay away from the cut-o®.
In models with IR instability (spontaneous symmetry breaking, dynamical mass gener-
ation,...) another limitation arises, m 2dyn << · 2 ; p2 , to ensure that we stay in the U.V.
scaling regime. This is because the IR scaling regime may have relevant operators which
are irrelevant on the U.V. side [7]. In more realistic models with several scaling regimes
which have di®erent relevant operator sets we need non-asymptotic methods which can
interpolate between the di®erent scaling regimes. This is achieved by the functional
extensions of the RG equation based on the in¯nitesimal blocking steps.
2 Functional RG equations
There are di®erent avenues to arrive at a functional RG equation. The simplest is to follow
Wilson-Kadano® blocking in continuous space-time. It leads to a functional di®erential
equation describing the cut-o® dependence of the bare action. Better approximation
schemes can be worked out when the modes are suppressed smoothly as the cut-o®
changes. The parameters of the bare action which are followed by these RG equations
are related to observables qualitatively only. It is more advantageous to transform the RG
equation for the e®ective action whose parameters have direct relation with observables.
For the sake of simplicity we consider an Euclidean theory in d dimensions for a scalar
¯eld ¿ x governed by the action SB [¿ ]. An O(d) invariant U.V. cut-o® ¤ is introduced in
the momentum space by requiring ¿ p = 0 for jpj > ¤ to render the generator functional
Z
1 1 1
e h
·
W [j]
= D[¿ ]e¡ h· SB[Á]+ h· j¢Á (5)
¯nite. We shall use the following notation conventions, stated below: In order to render
the functional manipulations well de¯ned we always assume discrete spectrum, i.e., the
presence of U.V. and IR regulators, say a lattice spacing a and system volume V = ad N d .
R P R R
The space-time integrals are d d x = ad x = x , x fx g x = f ¢ g,
R d P R R R
d p=(2º )d = V ¡1 p = p, fp = x e¡ipx fx , and fx = p eipx fp . The Dirac-deltas
¯ x;y = a¡d ¯ x;y
K K
and ¯ p;q = V ¯ p;q K
are expressed in terms of the Kronecker-deltas ¯ x;y and
K
¯ p;q .
2.1 Resumming the loop expansion

We start with the simplest form of in¯nitesimal blocking-step RG equations which is the
functional extension of the bare RG scheme mentioned above [8].
2.1.1 Blocking in continuous space-time
We shall denote the moving U.V. cut-o® by k. Its lowering k ! k ¡ ¢k leads to the
blocking transformation of the action which preserves the generator functional (5). Due
to the presence of the source this blocking introduces an explicit source dependence in
the action, Z Z
1 1 1 1
Dk [¿ ]e¡ h· Sk[Á;j]+ h· j¢Á = D k¡¢k [¿ ]e¡ h· Sk¡dk[Á;j]+ h· j¢Á; (6)
where Dk [¿ ] stands for the integration measure over the functional space Fk consisting
of functions whose Fourier transform is non-vanishing for jpj µ k. In order to avoid
this complication one usually assumes j 2 Fk¡¢k . In this case the blocking becomes
a mapping Sk [¿ ] ! Sk¡¢k [¿ ] and it is enough impose the invariance of the partition
function Z
1 1 ~
e ¡h
· Sk¡¢ k [Á]
= D[¿ ~]e¡ h· Sk[Á+Á]; (7)
where ¿ 2 Fk¡¢k and ¿ ~ 2 Fk nFk¡¢k .

The evaluation of the path integral by means of the loop-expansion gives immediately
the functional RG equation,
¹h ¯ 2 Sk [¿ + ¿ ~cl] ³ ´
Sk¡¢k [¿ ] = Sk [¿ + ¿ ~cl ] + tr ln + O ¹h2 ; (8)
2 ¯ ¿ ~¯ ¿ ~
where ¿ ³~cl denotes
´ the saddle point. The trace is over the functional space ¿ ~ 2 Fk nFk¡¢k
2
and O h ¹ represents the higher loop contributions. We write this equation as
¹
h ¯ 2 Sk [¿ + ¿ ~cl] ³ ´
Sk [¿ ] ¡ Sk¡¢k [¿ ] = Sk [¿ ] ¡ Sk [¿ + ¿ ~cl] ¡ tr ln + O ¹h2 : (9)
| {z } 2 ¯ ¿ ~¯ ¿ ~
tree level | {z }
loop contributions
(a)
(b)
Fig. 1 Graphs contributing to the blocking, (a): tree-level, (b): one-loop corrections. The
dashed line stands for a particle of momentum jpj = k and the solid lines represent ¿ .
When can one safely assume that the saddle point is trivial, ¿ ~cl = 0? Let us suppose
that ¿ x is weakly and slowly varying, ie ¿ x = © + ² x where ² x º 0 and the characteristic
momentum of ² is small with respect to the cut-o® k. Then the °uctuating compo-
nent ² appears as an external source breaking translation invariance. The saddle point,
¿ ~ 2 Fk nFk¡¢k , being inhomogeneous exists as ² ! 0 only if the external space-time
symmetry is broken dynamically.
But a ¢k-dependent small saddle point, ¿ ~cl = O (¢k n ), n > 0 may occur with-
out breaking external symmetries. In this case we expand in ¿ ~cl and ¯nd the Wegner-
Houghton (WH) equation [8]
Ã !¡1
1 ¯ Sk [¿ ] ¯ 2 Sk [¿ ] ¯ Sk [¿ ]
Sk [¿ ] ¡ Sk¡¢k [¿ ] = ¢ ¢ (10)
2 ¯ ¿~ ¯ ¿ ~¯ ¿ ~ ¯ ¿~
¹h 2
¯ Sk [¿ ] ³ ´ ³ ´
¡ tr ln + O ¹h2 + O ¢k 2 ;
2 ¯ ¿ ~¯ ¿ ~
where all functional derivatives are taken at ¿ ~ = 0. The saddle point can be omitted in
the argument of the logarithmic function because the trace brings a factor of ¢k as we
shall see below. The discussion of the tree-level renormalization when ¿ ~cl is non-vanishing
and ¯nite as ¢k ! 0 is deferred to section 3.3.
One can understand the loop contribution better by splitting the action into the sum
of the quadratic part and the rest, S = 12 ¿ ¢ G¡1
0 ¢ ¿ + S i and expanding in S i ,
1
Ã !n
¯ 2 Sk [¿ ] X (¡ 1)n ¯ 2 Sk [¿ ]
tr ln = tr ln G ¡1
0 ¡ tr ¢ G0 : (11)
~
¯ ¿ ¯ ¿ ~ n=1 n ¯ ¿ ~¯ ¿ ~
We recovered the sum of one-loop graphs. The loop corrections close on all possible
pair of ¿ ~ legs with the propagator G0 . The tree-level piece describes the feedback of the
change of the cut-o® on the dynamics of the classical background ¯eld ¿ . This is classical
contribution because the cut-o® controls the con¯guration space, the number of degrees
of freedom. Some of the graphs contributing to the right hand side are shown in Fig. 1.
The point of central importance is that ¢k serves here as a new small parameter
to suppress the higher-loop contributions. As an example consider the simplest case,
¿ x = ©, ¿ ~cl = 0 with the ansatz

Z · ¸
1 2
Sk [¿ ] = Z k (¿ x )(@¹ ¿ x ) + U k (¿ x) ; (12)
x 2
for the action with Z(¿ ) = 1,
Z ³ ´
¹
h
U k¡¢k (©) = U k (©) + ln[p2 + U k00 (©)] + O (¹h¢k)2 : (13)
2 k¡¢k<jpj<k
Each loop-integral is over the shell k ¡ ¢k < jpj < k in momentum space. So long
the propagator is non-singular in the integration domain the n-loop integrals will be
proportional to the n-th power of the integration volume, giving a dimensionless small
suppression parameter º (¢k=k)n ,the proportion of the modes eliminated with those
left. The question of the singularity will be considered in section 3.2.2. The higher loop
contributions are suppressed in the in¯nitesimal blocking step limit and the one-loop
evolution equation turns out to be an exact functional equation. The limit ¢k ! 0 is
safe for the loop contributions but more care is needed when the saddle point is non-trivial
because the tree-level contribution might be singular in this limit.
Notice that the convergence of the loop expansion was assumed in the argument
above. Another word of caution is in order about Eq. (9). Not only the trace but the
logarithm function itself is considered within the subspace ¿ ~ 2 Fk nFk¡¢k , a rather formal
observation which becomes essential later.
It is sometimes easy to arrive at exact, but unmanageable equations. The problem
is rather to ¯nd an approximation, an acceptable compromise between precision and
simplicity. By assuming the absence of non-local interactions and the homogeneity of the
vacuum the blocked action is usually expanded in the gradient of the ¯eld, leading to the
ansatz (12) for the action where contributions O (@ 4 ) are neglected. This step is the key
to the success of the RG method. The point is that the blocked action can be identi¯ed
by evaluating the integral (7) for di®erent background con¯gurations. The determination
of the blocked action is an over-constrained problem because there are ’more’ background
¯eld con¯gurations than coupling constants in a given ansatz. The tacit assumption, that
the e®ective action is well de¯ned, i.e., independent of the set of con¯gurations used to
read o® its coupling constants, gives the real power of the RG method: it enables us to
make predictions.
2.1.2 Local potential approximation
In order to relate the WH equation to conventional perturbation expansion we shall use

the ansatz (12) with Z = 1, in the local potential approximation. In order to determine
the only non-trivial piece of this action the potential U k (¿ ) it is su±cient to consider a
homogeneous ¯eld ¿ (x) = © in (9). The saddle point is vanishing as mentioned above,
¿ ~cl = 0, so long as the external space-time symmetries are unbroken. The blocking
equation simpli¯ed in this manner yields Eq. (13), the projection of the WH equation
into the restricted space of actions,
¹ dkd
h (2)
U_ k (©) = ¡ ln[k 2 + U k (©)] (14)
2(2º )d
Fig. 2 The rst four one-loop graphs contributing to the WH equation in the local potential
approximation when the potential U (¿ ) is truncated to the terms ¿ 2 and ¿ 4 . The dashed line
corresponds a particle of momentum jpj = k.
where the dot stands for k@k and d = 2º d=2 =¡(d=2) denotes the solid angle.
The expansion of the logarithmic function leads to the series
8 0 1 9
n
(2)
¹ dkd <
h 1
X (¡ 1)n U (©) ¡ m2 A =
U_ k (©) = ¡ ln[k 2 + m2 ] ¡ @ k ; (15)
2(2º )d : n=1 n k 2 + m2 ;
where m2 = U (2) (0), the sum of the one-loop graphs whose external legs have vanishing
momentum and the internal lines carry momentum jpj = k, cf Fig. 2, the leading order
contributions to the renormalization of the coupling constants gn in the loop expansion.
The partial resummation of the perturbation series performed by the WH equation
can easily be seen by comparing the solution of Eq. (14) with those obtained in the
independent mode approximation, where the k-dependence is ignored in the right hand
side of the equation,
Z
h
¹ (2)
U k (©) = U ¸ (©) ¡ ln[p2 + U ¤ (©)]: (16)
2 k<jpj<¤
One recovers the one-loop e®ective potential at the IR ¯xed point, Ve® (©) = U k=0(©).
This is not by accident and remains valid for the complete solution, as well, because
e¡V Ve¬ (©) is the distribution of the homogeneous mode and the blocking (7) at k = 0
retains the homogeneous mode only, Sk=0[©] = V U k=0(©). We shall see in section 3.3
that this simple argument is valid in the absence of large amplitude inhomogeneous
instabilities only.
Let us consider as an example the di®erential equation
x_ k = f (xk ); x¤ = xB ; (17)
where f (x) is weakly varying analytic function given in its expanded form around the
base point x = 0,
X f (n) (0)
x_ k = xnk : (18)
n n!
The RG method allows us to perform the expansion at the current point,
³ ´
xk¡¢k = xk ¡ ¢kf (xk ) + O ¢k 2 ; (19)
cf Eq. (8). The k-dependence of the right hand side represents the accumulation of
the information obtained during the integration. In a similar manner, the virtue of the
functional RG scheme (9) is the expansion of the path integral around around the current
action, instead of the Gaussian ¯xed point.
A more detailed picture of the RG °ow can be obtained by inspecting the beta func-
tions. For this end we parametrize the potential as
X gn (k)
n
U k (¿ ) = (¿ ¡ ¿ 0) ; (20)
n n!
and write the beta functions corresponding to ¿ 0 = h¿ i as
¹ dkd n
h
n = k@k gn (k) = k@k @Án U k (¿ 0 ) = @Án k@k U k (¿ 0) =¡ @ ln[k 2 + @Á2 U k (¿ )] (21)
2(2º )d Á
It is easy to see that n is an n-th order polynomial
¹ dkd
h
n=¡ Pn [G 3 ; ¢ ¢ ¢ ; Gn+2 ]; (22)
2(2º )d
of the expression G n = gn =(k 2 + g2 ), eg
P2 = G 4 ¡ G 23 ;
P3 = G 5 ¡ 3G3 G4 + 2G 33 ;
P4 = G 6 ¡ 4G3 G5 + 12G23 G4 ¡ 3G 24 ¡ 6G43 ; (23)
P5 = G 7 ¡ 5G3 G6 + 20G23 G5 ¡ 10G4 G5 ¡ 60G 33 G4 + 30G 3 G 24 + 24G 53 ;
P6 = G 8 ¡ 6G3 G7 + 30G23 G6 ¡ 5G 5 G6 ¡ 120G 33 G5 + 120G3 G4 G5
¡ 270G 23 G 24 ¡ 10G 25 ¡ 10G 4 G 6 + 360G43 G4 + 30G34 ¡ 120G 63 ;
etc. The correspondence between these expressions and Feynman graphs contributing to
the beta functions in the traditional renormalization scheme is indicated in Fig. 3.
We shall need later the beta functions corresponding to coupling constants made
dimensionless by the running cut-o®. By means of the parametrization ¿ = k (d¡2)=2 ¿ ~,
gn = k [gn] g~n , Ã !
d
[g n ] = d + n 1 ¡ ; (24)
2
U (¿ ) = k d U~ (¿ ~) one ¯nds
¹ d
h
~n (¿ ~) = g~_ n (k) = k ¡[gn] n (¿ ) ¡ gn = ¡
[g n ]~ ~3; ¢ ¢ ¢ ; G
Pn [G ~ n+2] ¡ [g n ]~
gn ; (25)
2(2º )d
where G~ n = g~n =(1 + g~2 ).

An instructive way to read Eqs. (20) and (14) is that the potential U (¿ ) is the
generator function for the coupling constants. The functional RG method is an econom-
ical book-keeping method for the computation of graphs and their symmetry factors by
keeping track of their generator function(al).
(a)
(b)
(c)
Fig. 3 Graphs contributing to n , (a): n = 2, (b): n = 3 and (c): n = 4.
Fig. 4 A two-loop contribution to g2 .
The higher loop contributions to the running coupling constants are generated by the
integration of the di®erential equations k@k gn = n . The us consider the two-loop graph
depicted in Fig. 4 and expand the subgraph in the square in the momenta of its external
legs. The zeroth order contributions is the last graph of Fig. 3b. By going higher order
in the gradient expansion one can, in principle, generate the full momentum dependence
of the subgraph. In general, the n-th loop contributions appear in the integration after
n step k ! k + ¢k and all loops are resummed in the limit ¢k ! 0.
2.1.3 Gradient expansion
The natural expansion for the long distance properties of systems with homogeneous
ground state is the expansion in the inverse characteristic length of the °uctuations, in
the gradient operator acting on the ¯eld variable.
Functional derivatives of local functionals: We consider the following generalization
of the O (@ 2 ) ansatz (12),
Z · ¸
1 ¡1
S[¿ ] = Z(¿ x )@ ¹ ¿ x K (¡ h )@¹ ¿ x + U (¿ x) : (26)
x 2
We shall use the notation Zx = Z(¿ x ), U x = U (¿ x ) for the coe±cient functions and
Zx(n) = @Án Z(¿ x ), U x(n) = @Án U (¿ x ) for their derivatives. The ¯rst two functional derivatives
of the action are

¯ S X ¯ S
V = eipy
¯ ¿ p y ¯ ¿ y
Z "
ipy (1) 1 (1) ¡1
= e ¯ x;y U x + ¯ x;y Z x @ ¹ ¿ x K @¹ ¿ x (27)
x;y 2
#
1 ³ ¡1 ¡1
´
+ Zx @¹ ¯ x;y K @¹ ¿ x + @¹ ¿ xK @¹ ¯ x;y ;
2
and
Z (
2¯ 2S ipy+iqz 1 (2) ¡1
V = e ¯ x;y ¯ x;z Zx @¹ ¿ x K @¹ ¿ x
¯ ¿ p¯ ¿ q x;y;z 2
1 ³ ´
+ Z x @¹ ¯ x;y K ¡1 @¹ ¯ x;z + @¹ ¯ x;z K ¡1 @¹ ¯ x;y
2 ·
1 ³ ´
+ Z x(1) ¯ x;y @¹ ¯ x;z K ¡1 @¹ ¿ x + @¹ ¿ x K ¡1 @¹ ¯ x;z
2 ¸
³ ´
¡1 ¡1
+¯ x;z @¹ ¯ x;y K @¹ ¿ x + @¹ ¿ xK @¹ ¯ x;y
)
(2)
+¯ x;y ¯ x;z U x : (28)
We split the ¯eld into a homogeneous background and °uctuations, ¿ x = ©+² x and ¯nd
up to O (² 2 ),
Z ( µ ¶ "
¯ S 1 1
V = e ipy
¯ x;y U (1)
+U (2)
² x + U (3) ² 2
x + ¯ x;y Z
(1)
@¹ ² x K ¡1 @¹ ² x
¯ ¿ p x;y 2 2
#)
³ ´
(1) ¡1 ¡1
+(Z + Z ² x ) @¹ ¯ x;y K @¹ ² x + @¹ ² x K @¹ ¯ x;y
Z (µ Z Z ¶
1 (3)
= eipy ¯ x;y U (1) + U (2) eir¢x ² r + U ei(r+s)¢x ² r ² s
x;y r 2 r;s
" Z
1 (1)
¡ Z ei(r+s)¢x r ¢ s² r K s¡1 ² s
2 r;s
µ Z ¶Z #)
³ ´
+ Z + Z (1) eir¢x ² r p ¢ seis¢x ² s Ks¡1 + Kp¡1
r s
1 ¡1
=¯ p;0 U
(1)
+ U (2) ² ¡p + p2 Z(K ¡p + K p¡1 )² ¡p
Z 2
1 h ³ í
+ ¯ p+r+s;0 ² r ² s U (3) ¡ Z (1) r ¢ sKs¡1 + p ¢ rKr¡1 + p ¢ sKp¡1 (29)
2 r;s
and
Z ( µ ¶
2 ¯ 2S ipy+iqz 1 (4)
V = e ¯ x;y ¯ x;z U (2) + U (3) ² x + U ² 2
x
¯ ¿ p¯ ¿ q x;y;z 2
"
1
+ ¯ x;y ¯ x;z Z
(2)
@¹ ² x K ¡1 @¹ ² x
2
³ ´
(1) ¡1 ¡1
+¯ x;y (Z + Z (2) ² x ) @¹ ¯ x;z K @¹ ² x + @¹ ² x K @¹ ¯ x;z
³ ´
¡1
+¯ x;z (Z
(1)
+ Z (2) ² x ) @¹ ¯ x;y K @¹ ² x + @¹ ² x K ¡1 @¹ ¯ x;y
µ ¶ #)
1 ³ ´
+ Z + Z ² x + Z (2) ² x2 @¹ ¯ x;z K ¡1 @¹ ¯ x;y + @¹ ¯ x;y K ¡1 @¹ ¯
(1)
x;z
2
Z ( " Z
1 (2)
= e ipy+iqz
¯ x;y ¯ x;z ¡ Z r ¢ seirx ² r K s¡1 eisx² s
x;y;z 2 r;s
µ Z ¶Z ³ ´
+ Z (1)
+Z (2)
e irx
² r eisx ² s q ¢ s(Ks¡1 + K q¡1 ) + p ¢ s(Ks¡1 + K p¡1 )
r s #
µ Z Z ¶
1
+p ¢ q Z + Z e ² r + Z (2) (1)
ei(r+s)x ² r ² s (K p¡1 + Kq¡1 )
irx
r 2 r;s
µ Z Z ¶)
(2) (3) irx 1 (4) i(r+s)x
+ U +U e ² r+ U e ² r² s
r 2 r;s
· ¸
1
= ¯ p+q;0 U (2) ¡ Zp ¢ q(K p¡1 + K q¡1 )
Z
2· µ
1 (1)
+ ¯ p+q+r;0 ² r U ¡ (3)
Z p ¢ q(Kp¡1 + Kq¡1 )
r 2 ¶¸
+q ¢ r(Kr¡1 + K q¡1 ) + p ¢ r(K r¡1 + K p¡1 )
Z · µ
1 1
+ ¯ p+q+r+s;0 ² r ² s U (4)
¡ Z (2)
p ¢ q(K p¡1 + Kq¡1 ) + r ¢ sK s¡1
2 r;s 2 ¶¸
+q ¢ s(K s¡1 + Kq¡1 ) + p ¢ r(Kp¡1 + K r¡1 ) ; (30)
where Z (n) = Z (n) (©), U (n) = U (n) (©) and K s = K (s2 ).

WH equation in O (² 2 ) : We set K = 1 in (26), use sharp cut-o®, i.e. all momentum
integral is for k ¡ ¢k < jpj < k,
Z " #
¯ 2S ³ ´
2 (2) (3) (1)
V =¯ p+q;0 U ¡ Zp ¢ q + ¯ p+q+r;0 ² r U ¡ Z (p ¢ q + q ¢ r + p ¢ r)
¯ ¿ p¯ ¿ q r
Z " #
1 (4) (2)
+ ¯ p+q+r+s;0 ² r ² s U ¡ Z (p ¢ q + r ¢ s + q ¢ s + p ¢ s + q ¢ r + p ¢ r)
2 r;s
= G¡1
p;q + Ap;q + B p;q ; (31)
with
· ¸
1
G¡1
p;q = ¯ p+q;0 U + Z(p2 + q 2 ) ; (2)
Z " 2 #
(3) 1 (1) 2 2 2
A p;q = ¯ p+q+r;0 ² r U + Z (p + q + r ) ;
r 2
Z " #
1 (4) 1 (2) 2 2 2 2
Bp;q = ¯ p+q+r+s;0 ² r ² s U + Z (p + q + r + s ) : (32)
2 r;s 2
As emphasised above, the logarithmic function on the right hand side of the WH
equation is computed within the space of ¯eld con¯gurations whose Fourier transform is
non-vanishing for k ¡ ¢k < jpj < k only. In order to keep this constrain in an explicit
manner we introduce the projection operator P corresponding to this space and write
the rest of the O (² 2 ) evolution equation as
h
¹
@k S[¿ ] = ¡ tr log[G¡1 + A + B]
2¢k
h
¹ h
¹ h
¹
=¡ trP log G ¡1 ¡ tr[PG(A + B)] + tr[PGAPGA]: (33)
2¢k 2¢k 4¢k
Since the propagator does not lead out from the function space, [P; G] = 0, it is su±cient
to make the replacement G ! PG on the right hand side of the evolution equation.
The term O (² ) drops owing to the vanishing of the homogeneous component of the
°uctuation, ² p=0 = 0 and we ¯nd
Z Z (
1 h i ¹kd
h
² ¡r ² r
_ + U_
Zr 2 (2)
+ V U_ = ¡ de V log[Zk 2 + U (2) ]
2 r 2(2º )d
· ¸2
Z U (3) + 12 Z (1) (k 2 + (ek ¡ r)2 + r 2 )
¡ ² ¡r ² r Pr¡ek;ek¡r
r 2(Zk 2 + U (2) )(Z(ek ¡ r)2 + U (2) )
Z )
U (4) + Z (2) (k 2 + r2 )
+ ² ¡r ² r ; (34)
r 2(Zk 2 + U (2) )
including an integration over the direction of k, the unit vector e.

The identi¯cation of the terms O (² 0 r 0 ) results the equations
¹ kd d
h
U_ = ¡ log[Zk 2 + U (2) ]; (35)
2(2º )d
c.f. Eq. (14). The terms O (² 2 r 2 ) give

Z Z Z (
¹ kd
h ² ¡r ² r
Z_ ² ¡r ² r r 2 = ¡ de Z (2) r 2
r 2(2º )d r Zk 2 + U (2)
·
U (3) + Z (1) k 2 2 (U
(3)
+ Z (1) k 2 )2
¡ Pr¡ek;ek¡r 2r 2 Z (1) ¡ Zr
Zk 2 + U (2) (Zk 2 + U (2) )2
Ã ! ¸)
2 2 Z (1)2 (1) U
(3)
+ Z (1) k 2
+k (er) ¡ ZZ : (36)
Zk 2 + U (2) (Zk 2 + U (2) )2
Finally, in O (² 2 r0 ) we ¯nd
Z Z Z ·
¹kd
h ² ¡r ² r
U_ (2) ² ¡r ² r =¡ de Z (2) k 2
r 2(2º )d r Zk 2 + U (2)
¸
(Z (1) k 2 + U (3) )2
+U (4) ¡ Pr¡ek;ek¡r : (37)
Zk 2 + U (2)
Let us make the replacement Pr¡k;k¡r ! 1 + [Pr¡k;k¡r ¡ 1] in Eqs. (35)-(37) and call
the contributions corresponding to 1 and Pk¡r;k¡r ¡ 1 regular and irregular, respectively.
The regular contribution can be obtained from the one-loop graphs. This has already
been seen for Eq. (35) and the regular part of Eq. (37) is just the second derivative of
Eq. (35).
But there are problems with the irregular contributions which represent the cut-o® in
the multi-loop integrals. One obvious problem is the inconsistency between Eqs. (35) and
(37). Another problem is that the irregular contributions include a truncated r-integration
and as a result the left and right hand sides of the equations have di®erent ² -dependence.
This could be avoided by imposing r << ¢k for the momentum of the °uctuations. But
the price is unacceptable high since it would indicate that the radius of convergence for
the gradient expansion is smaller then the in¯nitesimal ¢k.
Another problem the projection operator leads to is non-locality. It is known that
the bare action is non-local at the scale of the U.V. cut-o®, i.e. it contains higher order
derivatives up to a ¯nite order. In fact, the gradient @ n =@x¹n represents couplings up to n
lattice spacing distances when lattice regularization is used. But terms in the action which
are non-polynomial in the gradient induce correlations at ¯nite, i.e. cut-o® independent
distances and are not acceptable. Returning to the gradient expansion we note that the
operator Pr¡k;k¡r restricts the integration domain for r in such a non-isotropical manner
that not only r 2 but jrj appears, too. This latter corresponds to the non-local operator
p
h in the gradient expansion.
This problem, inherent in the sharp cut-o® procedure in the momentum space is
rather general. The one-loop correction for the two-particle scattering amplitude in the
¿ 4 model, represented by the fourth graph in Fig. 3c has a non-local contribution when
the momenta in both internal lines are properly restricted by the U.V. cut-o®. Such non-
local e®ects, disregarded in the usual textbooks, might well be negligible in a renormalized
theory because they are represented by non-renormalizable, irrelevant operators such as
p
¿ 3 h ¿ . Therefore one hopes that their e®ects become weak when the cut-o® is removed.
The more careful analysis is rather involved since the details of the cut-o® procedure
within the momentum regime (1 ¡ ° )k < p < (1 + ° )k in°uences the dynamics on the
distance scale ¹h=(2° k) which is a quantity of the form 0 ¢ 1.
But it is not so easy to dispel the doubts. First, the irrelevance and unimportance of a
coupling constant are quite di®erent concepts. An irrelevant coupling constant approaches
its I.R. ¯xed point value as we move in the I.R. direction. But its value at the ¯xed point
may be strong, indicating that the coupling constant in question is important. It is the
scale dependence only what becomes unimportant for an irrelevant coupling constant and
not its presence. The higher order vertices induced at low energies in a strongly coupled
self interacting scalar ¯eld theory in d < 4 may serve as an example of this di®erence.
Second, the global view of the renormalization group, outlined in Section 3.2 suggest that
a coupling constant which is irrelevant at the U.V. ¯xed point does not necessarily remain
so at around the I.R. ¯xed point. Finally, this problem is obviously present in e®ective
theories where the cut-o® reaches physical scales.
2.2 Resumming the perturbation expansion
The basic idea leading to the WH equation is the successive elimination of the degrees
of freedom. This procedure produces the blocked action, the integrand of the functional
integral of the blocked theory. This integrand must be well de¯ned, ie each mode must
either be left intact or be completely eliminated during the blocking. This is the positive
side of the sharp cut-o® in momentum space. The negative side is that it generates non-
local interactions and spoils the gradient expansion. The simplest cure, to smear out the
regulator and to use smooth cut-o® procedure, is not a valid alternative. This is because
the blocking based on smooth cut-o® suppresses the modes partially and the integrand for
the functional integral of the blocked theory is ill de¯ned. Despite this general sounding
argument one can proceed and generalize the successive elimination process for smooth
cut-o® in a rather surprising manner.
2.2.1 Polchinski equation
Let us start with the partition function

Z
1 ¡1 1 I
Zk = D[©]e¡ 2·h ©Gk ©¡ h
· Sk [©] ; (38)
where the interaction functional SkI [©] corresponds to the U.V. cut-o® k. We split the
propagator and the ¯eld variable into IR and U.V. components,
~ k ; © = ¿ + ¿ ~;
G k = G k¡¢k + G (39)
with the intention that the ¯elds ¿ and ¿ ~ should propagate with G k¡¢k and G ~ k , respec-
tively. In other words, the kinetic energy contribution to the action is supposed to be of
1 ~ ~ ¡1 ~ ~
the form (¿ G¡1k¡¢k ¿ + 2 ¿ G k ¿ )=2 when written in terms of ¿ and ¿ .
The introduction of the ¯elds ¿ and ¿ ~ separated by a smooth cut-o® makes both
¿ p and ¿ ~p non-vanishing and appears as a double counting of the degrees of freedom.
In order to ensure the proper integration measure we introduce a dummy ¯eld, ©, ~ in
such a manner that a suitable, momentum dependent linear combination of ¿ p and ¿ ~p
reproduces ©p and © ~ p,
Ã ! Ã !Ã !
© A 1;1 (h ) A 1;2 (h ) ¿
= (40)
~
© A 2;1 (h ) A 2;2 (h ) ¿~
~
and decouple ©,
1 ¡1 1 ~ ¡1 ~ 1 1 ~ ~ ¡1 ~
¿ G k¡¢k ¿ + ¿ ~G ¡1
k ¿ = ©G k © + ©G D ©; (41)
2 2 2 2
with a freely chosen dummy propagator G ~ D . Owing to the second equation in (39)
A1;1 = A 1;2 = 1. The condition (41) gives
Ã ! Ã !Ã !Ã !
G¡1
k¡¢k 0 1 A 2;1 (h ) G¡1
k 0 1 1
~ = ~ ; (42)
0 G¡1
k 1 A 2;2 (h ) 0 G ¡1
D A2;1 (h ) A 2;2 (h )
q q
~DG
whose solution is A 2;1 (h ) = G ~ k =G k¡¢k G k , A2;2 (h ) = G ~ D Gk¡¢k =G
~ k G k . The trans-
formation (40) is non-singular so long as G ~ k 6= Gk¡¢k , the propagations of ¿ and ¿ ~ are
distinguishable.
We are ¯nally in the position to describe the blocking procedure which consists of the
following steps:
(1) Redouble the number of degrees of freedom, © ! (©; ©). ~
(2) Rotate the new, dummy degrees of freedom into the dynamics in such manner that
the rotated ¯elds split the original ¯eld variable and follow the prescribed propaga-
tion.
(3) Integrate out ¿ ~.

In order to complete step 3. we insert a ¯eld independent constant into the partition
function, Z
~ ¡ 2·1h ©~ G~ ¡
Z k = D[©]D[©]e
1~ 1
D ©¡ 2·
¡1 1 I
h ©Gk ©¡ h
· Sk [©] : (43)
The redistribution of the degrees of freedom by the inverse of the transformation (40)
yields Z
1 ¡1 1 ~ ~¡1~ 1 I ~
Zk = D[¿ ]D[¿ ~]e¡ 2·h ÁGk¡¢ kÁ¡ 2·h ÁGk Á¡ h· Sk[Á+Á] ; (44)
up to a constant. This motivates the introduction of the blocked action Sk¡¢k [¿ ] de¯ned
as Z
1 ~ ~¡1~ 1 I ~
e ¡Sk¡¢ k [Á]
= D[¿ ~]e¡ 2·h ÁGk Á¡ h· Sk[Á+Á] : (45)
The higher loop corrections were suppressed in the WH equation by restricting the
functional space of the blocking, the volume of the loop integration in momentum space.
But there is another way to suppress the radiative corrections, to decrease the propagator.
Such a suppression method is better suited for smooth cut-o® where the U.V. ¯eld can not
be constrained into a restricted, ’small’ functional space. This strategy is implemented
by requiring that in the so far arbitrary split (39) G ~ k (p) = O (¢k). The obvious choice
is
G~ k (p) = ¢k@k G k (p): (46)
By assuming the convergence of the perturbation expansion we expand the action in ¿ ~
and perform the integration over ¿ ~. Since the new small parameter is the propagator we
retain graphs where the propagator appears in a minimal number. By assuming that the
saddle point is at most O (¢k) we can expand the action in ¿ ~,
Z ¯ SI [¿ ] 2 I
~~ ¡1~ 1 ~ 1 ~ ¯ Sk [¿ ] ~
e
1 I
¡h
· Sk¡¢ k [Á] =
1
D[¿ ~]e¡ 2·h ÁGk
1 I
Á¡ h
· Sk [Á]¡ h
· Á¢
k
¯ ¿
¡ 2·
h Á¢ ¯ ¿ ¯ ¿
¢Á+O (Á~3)
µ ¶¡1 µ ¶
¯ SI [¿ ] ¯ 2 SI [¿ ] ¯ SI [¿ ] ¯ 2 SI [¿ ]
1 I
¡h 1 k ¢ ~¡ 1 k ¢ k ¡ 12 tr log ~¡ 1 k +O (¢k2)
· Sk [Á]+ 2·
h ¯ ¿ G k + ¯ ¿ ¯ ¿ ¯ ¿ G k + ¯ ¿ ¯ ¿
=e · µ ¶¸
I ¯ SI [¿ ] h ¯ 2SI [¿ ]
1 ¯ Sk [¿ ]
1 I
¡h S [Á]+ ¢
· k h
·
k
2 ¯ ¿
¢@kGk¢ ¯ k¿ ¡ ·2 tr @kGk k
¯ ¿ ¯ ¿
+O (¢k 2)
= e" #
¹h¢k ¯ ¯ 1 I
³ ´
= 1¡ ¢ @k G k ¢ e¡ h· Sk [Á] + O ¢k 2 : (47)
2 ¯ ¿ ¯ ¿
The third equation was obtained by expanding the logarithmic function and gives the
di®erential equation [9], [10]
" #
1 ¯ SkI [¿ ] ¯ S I [¿ ] ¹
h ¯ 2 SkI [¿ ]
@k SkI [¿ ] = ¢ @k G k ¢ k ¡ tr @k Gk : (48)
2 ¯ ¿ ¯ ¿ 2 ¯ ¿ ¯ ¿
The fourth line leads to a linear equation,
1 I 1 I 1 ¯ ¯
k@k e¡ h· Sk[Á] = h
¹ Be¡ h· Sk [Á]; B = ¢ k@k G k ¢ : (49)
2¯ ¿ ¯ ¿
It is instructive to notice the similarities and di®erences with the WH equation.
Eq. (45) looks as the starting point for the derivation of the WH equation and could
have been obtained in a trivial manner for sharp cut-o®. The crux of the argument
leading to this relation is that this simple-looking equation is actually valid for smooth
cut-o®. But in a surprising manner it is just the sharp cut-o® limit when the rest of
the derivation of Polchinksi equation is invalid. This is because G ~ k as given in (46) is
¡1 0
¢kO (¢k ) = O (¢k ).
Eq. (48) is a leading order equation in the perturbation rather than the loop ex-
pansion. It is formally similar to the WH equation, (10) when only the leading order is
retained in the expansion (11) which corresponds to keeping graphs with a single vertex
only in Fig. 1(b) and using the dashed lines to denote the propagation of ¿ ~. Notice that
even if the higher order contributions of the perturbation expansion are suppressed in
the limit ¢k ! 0, the convergence of the perturbation expansion was assumed.
The di®erence between WH equation and (48) is that the tree-level ¯rst term on the
right hand side is always non-vanishing in Eq. (48), contrary to the case of the WH
equation. It is easier to understand the reason of the non-trivial saddle point when the
blocking (6) is considered in the presence of the source term. The source can be omitted
and the cut-o® independence of a single quantity, the partition function can be considered
as the basic equation when the source has vanishing component in the functional subspace
to integrate over in the blocking. In case of a smooth cut-o® the background ¯eld has
arbitrary momentum component and we can not really omit the source. One ought to
follow the more cumbersome blocking which keeps the generator functional RG invariant.
When classical physics (action) is modi¯ed by changing the cut-o®, the ¯eld induced by
the source changes, as well. This is the origin of the tree-level renormalization. But we
made a short-cut and omitted the source. The price is that the tree-level renormalization
can only be understood by letting the background ¯eld, ¿ vary freely.
The understanding how the ¿ -dependence may give rise the tree-level renormaliza-
tion in (47) is an interesting demonstration of the well know theorem stating that the
expansion in loops and in h ¹ are equivalent. In fact, the ¯rst equation in (47) introduces
~
an e®ective theory for ¿ and the ¯rst two terms in the exponent of the second line give
the e®ective action corresponding to a given background ¯eld ¿ in the classical approxi-
mation. In other words, the action with lowered cut-o® which is supposed to reproduce
the action with the original cut-o® for arbitrary con¯guration ¿ requires tree-level adjust-
ment. The second term in the exponent has two factors of ¹h¡1 coming from the ’source’
¹h¡1 ¯ SkI [¿ ]=¯ ¿ of the ¯eld ¿ ~ and a factor of h¹ from the ¿ ~ propagator. The overall 1=¹ h
disappears in the di®erential equation (48) which displays³ ´ h
¹ times the exponent. This is
0
the way the graphs of Fig. 1(b) appear on the O h ¹ tree-level in Eq. (48). Another
reason is that they are graphs with no freely changing loop-momentum variable. One
can understand in a similar manner that both the tree- and the loop-level graphs appear
with a common factor of h ¹ in Eq. (49).
2.2.2 Gradient expansion
We shall use here the results obtained in section 2.1.3 with the ansatz
Z · ¸
I 1 ¡1 2
S [¿ ] = (Z(¿ x) ¡ 1)@¹ ¿ xK (¡ h =k )@¹ ¿ x + U (¿ x) (50)
x 2
to derive the gradient expansion for Eq. (48). For the left hand side we ¯nd in the given
order of the gradient expansion
Z Z Z
Z_ p2 Z¡ 1 U_ (2)
² ¡p ² p + ² ¡p² p p2 K_ p¡1 + V U_ + ² ¡p ² p
2 p Kp 2 p 2 p
Z Z
Z_ p2 _ U_ (2)
º ² ¡p ² p +VU + ² ¡p ² p : (51)
2 p Kp 2 p
By means of the relation k@k G = ¡ 2Kp0 =k 2 where Kp0 = dK (z)=dz, z = p2 =k 2 the right
hand side gives
( )
R ³ 2
´ R h ³ í
(1) p 1 r¢(p¡r) p¢r p¢(p¡r)
¡ p
¯ p;0 U +² p U (2) + (Z ¡ 1) K p
+ 2 r
² r² p¡r U (3) ¡ Z (1) Kp¡ r
¡ Kr
¡ Kp
( )
³ ´ R h ³ í
K0 (1) (2) p2 1 p¢s (p+ s)¢s p¢(p+ s)
£ k2p ¯ p;0 U +² ¡p U + (Z ¡ 1) K p
+ 2 s
² ¡p¡s ² s U (3) ¡ Z (1)
Kp
¡ Ks
¡ K¡ p¡ s
( )
R ³ ´ R h ³ 2 í
Kp0 p2 1 p r2
+·h p k2
V U (2) + (Z ¡ 1) K p
+ 2 r
² ¡r ² r U (4)
+ Z (2)
Kp + Kr
· ³ ´2 ³ ´¸
K 0 R K00 + K00 0 K0 r2 =(k2 Kr ) r2 K00 r2
º¡ V U (1)2 k20 ¡ r ² ¡r ² r k2 U (2)
+ (Z ¡ 1) K r
+ k2 U
(1)
U (3)
+ Z (1) K r
( )
R ³ ´ R h ³ 2 í
Kp0 p2
1 p r2
+·h p k2
V U (2) + (Z ¡ 1) K p
+ 2 r
² ¡r ² r U (4) + Z (2) K p
+ Kr (52)
This equation is consistent since the terms O (² 2 ) are comparable thanks to the smooth-
ness of the cut-o®. We compare the coe±cients of the same ² -dependent expressions and
¯nd
Z Ã !
p2
k 2 U_ = ¡ U (1)2 K00 + h
¹ K p0 U (2) + (Z ¡ 1) ;
p Kp
Z
K 000 K0 (2)2
k 2 Z_ 0 (2) 0 (1) (1)
= ¡ 4K0 U (Z ¡ 1) ¡ 2K 0 U Z ¡ 2 2 U + ¹hZ (2)
K p0 ; (53)
k p
and Z Ã !
p2
k 2 U_ (2) = ¡ 2K00 U (2)2 ¡ 2K 00 U (1) U (3) +h
¹ Kp0 U (4) + Z (2) (54)
p Kp
We have two independent equations for two functions because the Eq. (54) follows from
the ¯rst equation of (53).
It is instructive to check the ¯rst equation in the local potential approximation Z = 1,
Z
U (1)2 k@k K p ¹hU (2) k@k Kp
U_ = ¡ : (55)
2 p2 jp=0 2 p p2
The ¯rst term describes the tree-level adjustment of the bare action in keeping the physics
cut-o® independent and some of the corresponding graphs are depicted in Fig. 1(a). The
present form, characteristic of the gradient expansion, is non-vanishing when the change
of the cut-o® at p = k is felt at the base point of the gradient expansion, p = 0. The
second term is the leading order perturbative contribution to the WH equation (14) when
the mass term is treated perturbatively, as expected and the beta functions,
Z
¹ U (n+2)
h k@k K p
n (¿ ) = ¡ ; (56)
2 p p2
for n > 2 correspond to the leading order ¯rst column in Eqs. (23) with g 2 = 0. The
independent mode approximation reproduces the leading order renormalization of the
coupling constants and the numerical integration resums the perturbation expansion in
the approximation where the e®ective vertices are replaced by their values at p = 0. The
leading order perturbative contributions, retained in the RG equation are linear in the
coupling constants and the classical, non-linear terms drop out in the linearized evolution
around the Gaussian ¯xed point therefore the usual critical exponents are reproduced.
2.3 Composite operator renormalization
The renormalization of composite operators [11] seems to be a highly technical and subtle
issue, dealing with the removal of the U.V. divergences from Green functions where com-
posite operators are inserted. But it becomes more elementary and general [12] as soon
as we are ready to give up the perturbative approach and can reformulate the procedure
in a general, non-perturbative manner. In order to see this let us go back to the remark
made at Eq. (6). It is not enough to impose the RG invariance of the partition function
only since it yields a single equation for in¯nitely many coupling constants. Instead one
should require the RG invariance of the generator functional for Green functions because
any observable can be reconstructed from this functional. This turns out to be rather
cumbersome due to the source dependence generated for the blocked action [13]. We
took another direction by imposing the RG invariance of another functional, the blocked
action, cf Eqs. (7) and (45). Instead of this strategy we shall now use the matching of
observables computed at di®erent values of the cut-o® to rederive the RG equation. This
is a non-perturbative, blocking inspired generalization of the multiplicative RG schemes
mentioned in the Introduction in what the scale dependence of the observables is followed
instead of those of non-physical bare coupling constants.
We start with a toy model to demonstrate the natural relation between blocking and
composite operator renormalization and the ¯eld theoretical application follows next by
generalizing the source term in the generator functional. This term which usually involves
the elementary ¯eld only is extended here for any local operator with inhomogeneous
source and we consider it as part of the action which now contains inhomogeneous cou-
pling constants. Our main point is that the beta functions corresponding to this extended
action provide a bridge between the blocking and the composite operator renormalization.
2.3.1 Toy model
Consider the two dimensional integral

Z
Z= dxdye¡S(x;y) (57)
over x and y which play the role of low- and high-frequency variables, respectively, where
the bare action is given by the expression
X1
1 1
S(x; y) = sx x2 + sy y 2 + g n (x + y)n (58)
2 2 n=0
in terms of bare coupling constants g n . We use h

¹ = 1 in this section. The blocked action
is de¯ned as Z
e¡S(x) = dye¡S(x;y): (59)
The elementary, bare operators are (x + y)n with n = 0; 1; 2; : : : and their expectation
values are R R
dxdy(x + y)n e¡S(x;y) dx @S(x)
@gn
e¡S(x)
R = R : (60)
dxdye¡S(x;y) dxe¡S(x)
The right hand side of this equation expresses the expectation value of the bare operator
in terms of an operator of the blocked, ’thinner’ system. The action S(x) can be expanded
in the terms of the operators xn ,
X1
1 0 m
S(x) = sx x2 + gm x ; (61)
2 m=0
giving rise the composite operators

X1
@S(x)
fxn g = = xm Sm;n (62)
@gn m=0
where 0
@gm
Sm;n = : (63)
@gn
We have chosen a basis for operators in the bare theory and searched for operators in
the blocked theory. The opposite question, the construction of the composite operators
[(x + y)n ] of the bare theory which reproduce the expectation values of the blocked
operators gives R R
dxdy[(x + y)n ]e¡S(x;y) dxxn e¡S(x)
R = R ; (64)
dxdye¡S(x;y) dxe¡S(x)
with 1
X
[(x + y)n ] = (x + y)m (S ¡1 )m;n : (65)
m=0
It is instructive to generalize the toy model for three variables,

1
1 2 1 2 1 2 X
S(x; y; z) = sx x + sy y + sz z + g n (x + y + z)n ; (66)
2 2 2 n=0
with
Z
¡S(x;y)
e = dze¡S(x;y;z) ;
Z
e¡S(x) = dye¡S(x;y); (67)
where the operator mixing looks like

@S(x; y; z)
fxn gz = = (x + y + z)n ;
@gn
R
n @S(x; y)
@
@gn
dze¡S(x;y;z)
fx gy = =¡ R ;
@g n dze¡S(x;y;z)
R
n @S(x)
@
@g
dydze¡S(x;y;z)
fx gx = =¡ Rn
: (68)
@gn dydze¡S(x;y;z)
These relations yield
R
n dzfxn gz e¡S(x;y;z)
fx gy = R ;
dze¡S(x;y;z)
R
n dyfxn gy e¡S(x;y)
fx gx = R ; (69)
dye¡S(x;y)
indicating that the evolution of the operators comes from the elimination of the ¯eld
variable in their de¯nition. We can compute the expectation value of fxn g at any level,
R R R
dxdydzfxn gz e¡S(x;y;z) dxdyfxn gy e¡S(x;y) dxfxn gx e¡S(x)
R = R = R : (70)
dxdydze¡S(x;y;z) dxdye¡S(x;y) dxe¡S(x)
The lesson of this toy model is twofold. First, one sees that the formal de¯nition of the
renormalized operator at two di®erent value of the cut-o® di®ers because these operators
are supposed to reproduce the same averages by means of di®erent number of degrees of
freedom. Second, it shows that the cut-o® dependence of the renormalized operators can
be obtained in a natural and simple manner by considering the bare coupling constants
at one value of the cut-o® as the functions of the coupling constants given at another
value of the cut-o®.
2.3.2 Quantum Field Theory
We generalize the operator mixing of the toy model for the scalar ¯eld theory given by
the bare action
Z X X
S¤ [¿ ] = Gn;x (¤)On (¿ x) = Gn~ (¤)On~ (¿ x ); (71)
x n n
~
where On (¿ x) represents a complete set of local operators (functions of ¿ x and its space-
time derivatives) and G n;x (¤) denotes the coupling constant. We simplify the expressions
by introducing a single index n~ for the pair (n; x) labeling the basis elements for the local
P P R
operators and n~ = n x . The decomposition ¿ x = ¿ k;x + ¿ ~k;x of the scalar ¯eld into a
low- and high-frequency parts is carried out as in Eqs. (7) and (39).
The Kadano®-Wilson blocking

Z
[Ák+Á~k]
e ¡Sk[Ák]
= D[¿ ~k ]e¡S¤ (72)
for the action X

Sk [¿ k] = G n~ (k)O n~ (¿ k) (73)
n
~
generates the RG °ow

k@k Gn~ (k) = n~ (k; G): (74)
The blocked operators are de¯ned by
R ~
¯ Sk [¿ k ] D[¿ ~k ]O n~ (¿ k + ¿ ~k )e¡S¤ [Ák+Ák]
fOn~ (¿ k )gk = = R (75)
¯ Gn~ (¤) D[¿ ~k ]e¡S¤ [Ák+Á~k]
in agreement with Eq. (68). They satisfy the equation

R ±Sk[Ák] ¡Sk[Ák] R ~
D[¿ k ] ±Gn
~ (¤)
e D[¿ ~
k ]D[¿ k ]O n
~ (¿ k + ¿ ~k )e¡S¤ [Ák+Ák]
R = R ; (76)
D[¿ k ]e
¡Sk[Ák]
D[¿ k ]D[¿ ~k ]e¡S¤ [Ák+Á~k]
c.f. Eq. (70), showing that they are represented in the e®ective theory by the functional
derivative of the blocked action with respect to the microscopical coupling constants. The
form (73) of the action gives the operator mixing
X ¯ Gm
~ (k) ¯ S k [¿ k ] X
fO n~ (¿ k )gk = = O m~ (¿ k )S m~
~ n (k; ¤); (77)
m
~ ¯ Gn~ (¤) ¯ Gm~ (k) m
~
in a manner similar to Eq. (69). We introduced here the sensitivity matrix
¯ Gm~ (k)
Sm;~
~ n (k; ¤) = : (78)
¯ G n~ (¤)
It is the measure of the sensitivity of an e®ective strength of interaction on the initial

condition of the RG trajectory, imposed in the U.V. domain, c.f. Eq. (153). The com-
posite operator fO n~ (¿ k )gk introduced in Eq. (77) replaces the bare operator fO n~ (¿ )g¤
in the Green functions of the e®ective theory for the I.R. modes ¿ k .
The di®erential equation generating the operator mixing (77) is obtained in the fol-
lowing manner. The relation
¢k
G n~ (k ¡ ¢k) = Gn~ (k) ¡ n~ (k; G) (79)
k
is used to arrive at the sensitivity matrix,
¯ Gm~ (k ¡ ¢k) X ¯ Gm
~ (k ¡ ¢k) ¯ G `~(k)
~ n (k ¡
Sm;~ ¢k; ¤) = =
¯ G n~ (¤) `~
¯ G `~(k) ¯ G n~ (¤)
" #
X ¢k ¯ m~ (k; G)
= ¯ ~ `~ ¡
m S`;~
~ n (k) (80)
`~
k ¯ G`~(k)
and to ¯nd
X¯ m
~ (k; G)
k@k Sm;~
~ n (k; ¤) = S`;~
~ n (k; ¤): (81)
¯ G `~(k)
`~
The scale dependence of the operator mixing matrix is governed by
1 ¯ n~ (k; G)
® n
~m~ (k) = : (82)
k ¯ G m~ (k)
Summary: Eq. (74) represents the link between Kadano®-Wilson blocking and com-
posite operator renormalization. On the one hand, n~ describes the evolution of the
action in the traditional blocking scheme, Eq. (72). On the other hand, interpreting the
coupling constants in the action as sources coupled to composite operators n~ determines
the mixing of composite operators in Eq. (82). Therefore all conepts and results of the
blocking, e.g. ¯xed point, universality, etc. has a counterpart in composite operator
renormalization.
2.3.3 Parallel transport
The operator mixing discussed above has a nice geometrical interpretation, a parallel
transport of operators along the RG trajectory [14] with the connection ¡n~ ;m
~ (k). The
independence of the expectation value from the cut-o®,
R R
D[¿ ~
k0]D[¿ k 0]O n
~ ¡Sk0[Ák0+Á~k0]
~ (¿ k0 + ¿ k 0)e D[¿ ~
k ]D[¿ k ]O n
~ (¿ k
~
+ ¿ ~k )e¡Sk[Ák+Ák]
R = R (83)
D[¿ k0]D[¿ ~k0]e¡Sk0[Ák0+Á~k0] D[¿ k ]D[¿ ~k ]e¡Sk[Ák+Á~k]
de¯nes the parallel transport of composite operators. The linearity of the mixing (77)
assures that this parallel transport is indeed linear and can therefore be characterized by
a covariant derivative
D k O k = (@k ¡ ¡)Ok (84)
¡1
where ¡m;~
~ n = (S ¢ ® ¢ S)n~ ;m~ in such a manner that D k O k = 0 along the RG °ow.
The dynamical origin of the connection comes from the fact that there are two di®erent
sources the scale dependence of hOk i comes from: from the explicit k-dependence of the
operator and from the implicit k-dependence due to the cut-o® in the path integration.
The operator mixing is to balance them. In fact, the covariant derivative could have been
introduced by the relation
@k hO k i = hD k Ok i; (85)
requiring that the operator mixing generated by the connection is to make up the implicit
k-dependence of the expectation value coming from the cut-o®.
It is obvious that ¡ is vanishing in the basis fOn~ gk ,
X X X
@k h cn~ (k)fO n~ gk i = @k cn~ (k)hfOn~ gk i = h@k cn~ (k)fOn~ gk i: (86)
n
~ n
~ n
~
The connection can in principle be found in any other basis by simple computation.
2.3.4 Asymptotical scaling
We linearize the beta-functions around a ¯xed point G ¤m~ ,

X
n~ º ¡¤n~ m~ (G m~ ¡ G ¤m
~ ); (87)
m
~
and write the scaling coupling constant, the left eigenvectors of ¡

X
¤
cleft
n ~ ¡ m;~
~ ;m ~ r = ¬ n
left
~ cn r;
~ ;~ (88)
m
~
as
X
Gsc
~ =
n cleft
n ~ ¡
~ (G m
~ ;m G ¤m~ ) : (89)
m
~
They display the scale dependence
G sc
~ ¹
n k n~ : (90)
Furthermore let us ¯x the overall scale of the local operators. This can be done by
using the decomposition
X
O(¿ x ) = bn On (¿ x) (91)
n
m
where Oqn (¿ (x)) is the product of the terms @¹1 ¢ ¢ ¢ @¹` ¿ (x) with coe±cient 1. The norm
P 2
jjOjj = n bn is introduced with the notation
O
O= ; (92)
jjOjj
for the operators of unit norm. We shall use the convention that the coupling constants
Gn~ (¤) always multiply operators of unit norm in the action.
The scaling operators
X right
O nsc
~ = cn~ m~ O m~ (93)
m
~
are obtained by means of the right eigenvectors of ¡¤ ,

X right right
¡¤r~m
~ cm~
~n = ¬ ~ cr~n
n ~ ; (94)
m
~
and they satisfying the conditions of completeness cright ¢ cleft = 1 and orthonormality
cleft ¢ cright = 1. The coupling constants of the action
X
Sk = Gn~ (k)Onsc
~ (¿ k) (95)
n
~
obviously follow (90). The operator

X
O= bn~ fOnsc
~ gk ; (96)
n
~
written at scale k in this basis yields the parallel transport trajectory

Ã !
X X k0 n
~
fOgk 0 = bn~ ffO nsc

~ gk gk 0 = bn~ fO nsc
~ gk (97)
n
~ n
~ k
in the vicinity of the ¯xed point.

The beta function introduced by Eqs. (72)-(74) agrees with the usual one. One ¯nds
that only relevant operators have non-vanishing parallel transport °ow. Universality
manifests itself at a given scaling regime in the suppression of the parallel transported
irrelevant operators.
The blocked action Sk [¿ k;x; G n~ (k); G n~ (¤)] possesses two interpretations:
° The value of the coupling constant at the running cut-o®, G n~ (k), re°ects the scale
dependence of the physical parameters.
° The dependence on the bare coupling constants, G n~ (¤), the initial condition of the
RG °ow provides us the generator functional for composite operators.
Finally, the composite operator renormalization represents an alternative way to arrive
at the RG equation. In fact, the beta functions arising from the blocking (72) are obtained
in this scheme by the parallel transport, the matching (76) of the expectation values.
2.3.5 Perturbative treatment
Let us ¯nally compare the matching of the observables, described above with the tra-
ditional method of perturbative composite operator renormalization. The inversion of
Eq. (77) gives X
[On~ (¿ k + ¿ ~k )]k = Om~ (¿ k + ¿ ~k )(S ¡1 (k; ¤))m;~
~ n: (98)
m
~
It is not di±cult to see that the operator [On~ (¿ k + ¿ ~k )]k of the bare theory which cor-
responds to the operator fOn~ (¿ k )gk of the e®ective theory agrees with the result of the
usual composite operator renormalization. In fact, the bare action (71) can be split into
the sum of the renormalized part and the counterterms the framework of the renormalized
perturbation expansion, giving G n~ B = G n~ R + Gn~ CT , G n~ (¤) = Gn~ B and G n~ (k) = Gn~ R .
The counterterms are introduced just to render certain Green functions with composite
operator insertions cut-o® independent [13]. The composite operator corresponding to
the renormalized (ie blocked) operator On~ (¿ k ) in this perturbative framework is
X ¯ G mB
~
Om~ (¿ k + ¿ ~k ) (99)
m
~ ¯ Gn~ R
which agrees with (98).

It is worthwhile recalling that the parameters and operators of a bare theory corre-
spond to the cut-o® scale in a natural manner. Therefore fOn~ (¿ k )gk represents the bare
operator [On~ (¿ k + ¿ ~k )]k in the e®ective description at the scale k and reproduces the ob-
servational scale dependence by construction in a manner similar to the scale dependence
of the hadronic structure functions obtained in the framework of the composite operator
renormalization convey the same information [15].
2.4 Continuous evolution

The RG equations obtained so far deal with the evolution of the bare action during the
gradual lowering of the U.V. cut-o®. There are two di®erent reasons to look for alternative
schemes where the evolution of the Green functions rather than the bare action is followed.
One reason is usefulness. In the traditional RG strategy followed so far the renor-
malized trajectories give the bare coupling constants as functions of the running cut-o®.
Though one can extract number of useful information from the trajectories they remain
somehow qualitative because the bare parameters of the theory with the running cut-o®
are not physical quantities. Though this remark does not apply to the multiplicative RG
scheme where the e®ective parameters are constructed by means of Green functions this
scheme is seriously limited as mentioned in the Introduction. Returning to blocking, one
can say at most that the di®erence between bare parameters and physical quantities arises
due to the °uctuations in the path integral and this latter decreases with the number of
degrees of freedom as we approach the IR end point of the trajectories. Therefore it is
the blocked action with very low cut-o® only which is supposed to be directly related to
physical quantities (in the absence of IR instability).
Another more formal point of view is to construct a strictly non-perturbative scheme.
The RG equations (10), (48) represent the complete resummation of the perturbation
expansion but they are not really non-perturbative equations. This situation is remi-
niscent of the Schwinger-Dyson equations. They were ¯rst obtained by resumming the
perturbation expansion and only later by a genuine non-perturbative method, by the
in¯nitesimal shift of the integral variable in the path integral formalism. The fact that
the naive result derived by assuming the convergence of the perturbation expansion is
correct is presumably related to the unique analytic continuation of the Green functions
in the coupling constants. Can we ¯nd in a similar manner the truly non-perturbative
RG equations? If possible, it will come by a di®erent route: by relating two path inte-
gral averages instead of computing them by brute force. The quantities in question are
one-particle irreducible (1PI) amplitudes and we shall give up to follow the evolution of
the bare coupling constants, the parameters in the path integral. This will allows us to
avoid any reference to the perturbation expansion. It remains to be seen if (10) and (48)
can be derived in a similar manner or they remain valid for strongly coupled models.
We begin this program by writing the generator functional for the connected Green
functions in the form Z
1 1
e h
· Wk[j]
= D[¿ ]e¡ h· (SB[Á]+Ck[Á]¡j¢Á); (100)
where the term Ck [¿ ] is introduced to suppress °uctuations. What we require is that

(i) for k = 1 the °uctuations be suppressed, C1 [¿ ] = 1, (ii) the original generator
functional be recovered for k = 0, C0 [¿ ] = 0 and (iii) the °uctuations are suppressed only,
ie Ck [¿ 0 ] = 0 for some con¯gurations close to the vacuum expectation value, ¿ 0 º h¿ i.
The simplest choice for models without condensate is a quadratic functional,
1
Ck [¿ ] = ¿ ¢ Ck ¢ ¿ (101)
2
but in certain cases higher order terms in the ¯eld variables are necessary in the suppres-
sion. We distinguish two kinds of suppression. The length scale of the modes ’released’
`(k) is well de¯ned when
(
¯ 2 Ck >> 1 jpj < 1=`(k),
M2k (p) = (102)
¯ ¿ ¡p¯ ¿ p jÁ=hÁi º0 jpj > 1=`(k).
The evolution generated by such suppressions shows the scale dependence in the spirit
of the Kadano®-Wilson blocking, the contribution of modes with a given length scale to
the dynamics. Examples are,
8
2 f(p)
> ap 1¡f (p)
>
<
[16]
³ ´
Ck (p) = > a k22 b [17] (103)
>
: p
2 2 2 2
a(k ¡ p )£(k ¡ p ) [18]
2 2 c
where f (p) = e¡b(p =k ) and a; b; c > 0. Another kind of suppression for which M2k (p)
shows no clear structure will be considered in section 2.5 below.
The evolution equation for W [j] is easy to obtain,
Z
1 1
¡h
@k W k [j] = ¡ e ·
Wk[j]
D[¿ ]@k Ck [¿ ]e¡ h· (Ck [Á]+SB[Á]¡j¢Á)
" #
1
¡h ¯ 1
=¡ e · W k[j] @k Ck ¹h e h· Wk[j] : (104)
¯ j
This is already a closed functional di®erential equation but of little use. The reason is
that it is di±cult to truncate W [j] being a highly non-local functional. In order to arrive
at a more local functional we shall make a Legendre transformation and introduce the
e®ective action ¡k [¿ ] as
¯ W [j]
¡k [¿ ] + W k [j] = j ¢ ¿ ; ¿ = ; (105)
¯ j
with
¯ W [j]
@k ¡k [¿ ] = ¡ @k W k [j] ¡ @k j + @k j¿ = ¡ @k W k [j]: (106)
¯ j
It is advantageous to separate the auxiliary suppression term Ck [¿ ] o® the e®ective action
by the replacement ¡[¿ ] ! ¡[¿ ] + Ck [¿ ] resulting in
" #
1
¡h Wk[j] ¯ 1
@k ¡k [¿ ] = e · @k C k h
¹ e h· Wk[j] ¡ @k Ck [¿ ]: (107)
¯ j
This relation takes a particularly simple form for the quadratic suppressions (101). Since
the quadratic part of the e®ective action is the inverse connected propagator we ¯nd [16],
[10], [19], [20], [21],
2 3
h
¹ h
¹ 1
@k ¡k [¿ ] = tr [@k Ck ¢ h¿ ¿ iconn] = tr 4@k Ck ¢ ± 2 ¡k
5: (108)
2 2 Ck + ±Á±Á
Two remarks are in order in comparing the evolution equation obtained here with the
pervious RG equations. First, the Kadano®-Wilson blocking is constructed to preserve
the generator functional W [j] for the thinner system and this is realized by eliminating
modes at the U.V. side, by changing Sk [¿ ]. One is not aiming at keeping anything ¯xed in
the continuous evolution scheme. The 1PI generator functional does depend on k which
appears as the lowest momentum of modes considered in the theory whose e®ective action
is ¡k [¿ ]. The second remark is that the right hand side is proportional to ¹h. But the
apparent absence of the³ mixing ´ of tree and loop-levels is misleading since the propagator
0
may be the sum of O ¹ tree-level and O (¹h) °uctuation contributions. The explicit
h
tree-level term is missing because the modi¯cation of the bare action was carried out
in the part O (¿ 2 ) and this correction to the free propagator was removed by the step
¡[¿ ] ! ¡[¿ ] + Ck [¿ ], the substraction of the tree-level suppression term from the e®ective
action.
It is illuminating to compare the evolution and the Schwinger-Dyson (SD) equations.
Though the formal aspects appear similar the content of the equations di®ers. A similarity
appearing immediately is that both set of equations determine the Green functions in a
hierarchical manner. The derivation of the equations shows some similarity, as well. In
fact, the crucial step in the derivation of the evolution equation is the ¯rst equation
in (104). This step, the expression of the derivative of the functional integral formally,
without its actual evaluation is the hallmark of the non-perturbative demonstration of the
SD equations. Both the SD and the evolution equations are genuinely non-perturbative
because we actually do not evaluate the functional integral, instead we relate to another
one by bringing a derivative inside the path integral. In both schemes on compares two
functional integrals which di®er slightly, either in the in¯nitesimal shift of the integral
variables or in the in¯nitesimal change in the action. Let us write the ¯rst equation in
(104) for ¯nite ¢k,
Z ³ ´
1 1
D[¿ ] e¡ h· (Ck[Á]+SB[Á]¡j¢Á) ¡ e¡ h· (Ck¡¢ k[Á]+SB[Á]¡j¢Á)
Z ³ ³ ´´
1 1 1
= D[¿ ] ¢k@k Ck [¿ ]e¡ h· Ck[Á] + O (¢k)2 e¡ h· (SB[Á]¡j¢Á): (109)
¹h
The small parameter ¢k is used to suppress the insertion of more ¯eld variables in the
evolution equation, to cut o® the higher order Green functions from the evolution, the
strategy common with the non-perturbative proof of the SD equation. The higher order
contributions in ¢k bring in higher order Green functions and we suppress them by the
smallness of the step in ’turning on’ the °uctuations. This is how the RG idea, the
replacement of the higher loops by running e®ective coupling constants and the dealing
with small number of modes at each step, is realized in the evolution equation scheme.
The obvious di®erence between the evolution and the SD equations is that the latter
express the invariance of the functional integral in a manner similar to the RG equations
but the former is simply an expression of the derivative of the functional integral with
respect to a parameter. The similarity of the evolution equation and the RG strategy
leads to another di®erence. The subsequent elimination of modes and taking into account
their dynamics by the introduction/modi¯cation of e®ective vertices generates a ’univer-

sal’ evolution equations which does not depend on the theory in question. In fact, the
evolution equation (108) and the functional RG equations Eqs. (6), (45) do not contain
any reference to the model considered. The e®ective or the bare action of the model ap-
pear in the initial condition only. The SD equation is based on a careful application of
the equation of motion within the expectation values and therefore contains the action
of the model in an obvious manner.
The evolution equation method seems to be better suited to numerical approximations
then the SD equations. This is because the ’dressing’, the summing up the interactions is
achieved by integrating out di®erential equations rather then coupling the Green functions
in the SD hierarchy, a numerical problem we can control easier.
2.5 Blocking in the internal space
We present now a generalization the RG method. The traditional RG strategy is aiming

at the scale dependence of observables and consequently follows the cut-o® dependence
in the theories. The results are obtained by the successive modi¯cation of the cut-o® and
the accumulation of the resulting e®ects. One can generalize this method by replacing
the cut-o® by any continuous parameters of the theory on which the observables depend
upon in a di®erentiable manner. Such a generalization replaces the RG equation by an
evolution equation corresponding to the parameter or coupling constant in question.
One gains and looses in the same time during such a generalization. One looses
intuition of and insight into the dynamics since the trajectories generated by the evolution
equation have no relation to scale dependence. To make things even more complicated the
trajectories do not correspond anymore to a ¯xed physical content, instead they trace the
dependence of the dynamics in the parameter considered. But we gain in °exibility. In
fact, the parameter we select to evolve can be the Plank constant or a coupling constant
and the integration of the resulting evolution resums the semiclassical or the perturbation
expansion. We may apply this method in models with non-trivial saddle point structure
or where the cut-o® would break important symmetries, e.g. in gauge theories.
A more technical aspect of this generalization touches the U.V. divergences in the
theory. The traditional RG procedure is based on the tacit assumption that the U.V.
divergences are properly regulated by the blocking, the moving cut-o®. This seems to be
a natural requirement if the RG °ow is interpreted as scale dependence. In fact, an U.V.
divergence left-over by the blocking would indicate the importance of the processes at
the regulator and would lead to the appearance of a second scale. In the generalization
of the RG procedure we introduce below there may not be a well de¯ned scale related to
the evolution and we have to regulate the U.V. divergences.
There are two di®erent kind of physical spaces in Field Theory. The space-time or
the momentum-energy space where the ’events’ are taking place will be called external
space. The ’events’ are characterized by further quantities, the ¯eld amplitudes. The
space where the ¯eld amplitudes belong will be called internal space. Therefore the ¯eld
con¯guration ¿ x realizes a map ¿ : external space ! internal space. The traditional

Kadano®-Wilson blocking orders the degrees of freedom to be eliminated according to
their scale in the external space. We may realize a similar scheme by ordering the modes
according to their scale in the internal space, their amplitude [22]. We present two
di®erent implementations of this idea, one for Wilsonian and another for the e®ective
action.
2.5.1 Wilsonian action
We shall use the °exibility of the method of deriving the Polchinksi equation in section
2.2 to separate o® and eliminate the °uctuation modes with the largest amplitude in the
path integral. Since the °uctuation amplitude is controlled by the mass this leads to the
description of the dynamics by highly massive modes, small °uctuations at the ¯nal point
of the evolution.
Let us consider the model (38) where the role of the parameter k is played by the
mass M and G¡1 2 2 2
M 2 = p + M . We shall follow the evolution from M = 0 to M >> ¤
2 2
where ¤ is the U.V. cut-o®. The decomposition
~ M 2 ; © = ¿ + ¿ ~;
G M 2 = GM 2+¢M 2 + G (110)
will be used with

1 ³ 2 ´ 1 (M 2 ¡ h )2 ~ 1 ³ 2 ´ 1 ~ ~ ¡1 ~
¿ M + ¢M 2 ¡ h ¿ + ¿~ ¿ = © M ¡ h ©+ © G D ©: (111)
2 2 ¢M 2 2 2
We replace the degrees of freedom © ~ by the ¯eld ¿ ~ with in¯nitesimal °uctuations. After
eliminating ¿ ~ the remaining ¯eld ¿ has smaller °uctuations than ©. We use the ansatz
Z · ¸
I 1
SM [¿ ] = (Z M (¿ x) ¡ 1)@¹ ¿ x @¹ ¿ x + U M (¿ x) (112)
x 2
where the U.V. regulator with smooth cut-o® is not shown explicitly. We follow the steps
outlined in section 2.2.2 which gives rise the evolution equation with left hand side
½ Z h i ¾
1
@M 2 ² ¡p ² p (Z ¡ 1)p2 + U (2) + V U : (113)
2 p
The right hand side reads as

Z ( h i Z h ³ í
)
1 (1) 2 (2) 1 (3) (1) 2 2
¡ ¯ p;0 U +² p (Z ¡ 1)p + U + ² r² p¡r U ¡ Z r¢p¡ r +p
2 p 2 r
1
£
(p2 + M 2 )2
( Z )
h i 1 h ³ í
(1) 2 (2) (3) (1) 2 2
£ ¯ p;0 U +² ¡p (Z ¡ 1)p + U + ² ¡p¡s ² s U +Z s +p¢s+p
2 s
Z ( Z )
h
¹ 1 h i 1 h i
+ V (Z ¡ 1)p2 + U (2) + ² ¡r ² r U (4) + Z (2) (p2 + r 2 )
2 p (p2 + M 2 )2 2 r
8h i2 9
Z >
< (Z ¡ 1)r2 + U (2) >
=
1 1 1 (1) (3)
º ¡ V U (1)2 ¡ ² ¡r ² r + U (U + Z (1) r2 )>
2M 4 2 r >
: (r 2 + M 2 )2 M 4
;
Z ( Z )
h
¹ 1 h i 1 h i
2 (2) (4) (2) 2 2
+ V (Z ¡ 1)p + U + ² ¡r ² r U +Z (p + r ) (114)
2 p (p2 + M 2 )2 2 r
The system of evolution equation projected onto the di®erent ² -dependent terms is
Z
(Z ¡ 1)p2 + U (2)
2M 4 @M 2 U = ¡ U (1)2 + ¹hM 4
p (p2 + M 2 )2
Z
U (2)2 1
2M 4 @M 2 Z = ¡ 4U (2) (Z ¡ 1) + 4 2 ¡ 2U (1) Z (1) + h ¹ M 4 Z (2)
M p (p 2 + M 2 )2
Z (2) 2 (4)
Z p +U
2M 4 @M 2 U (2) = ¡ U (2)2 ¡ U (1) U (3) + ¹hM 4 : (115)
p (p 2 + M 2 )2
³ ´
The O ¹h0 terms on the right hand side are to keep the tree-level observables M -invariant
within the gradient expansion ansatz. As far as the loop corrections are concerned, the
beta functions Z
2 (n+2) 1
n = M @M g n = h
¹M U ; (116)
p (p + M 2 )2
2
obtained from the ¯rst equation with Z = 1 agree with the leading order contribution to
the one-loop renormalized potential,
Z
1¡loop ¹
h (2)
UM = UM + ln[p2 + M 2 + U M ]; (117)
2 p
except their sign. This is because we intend to keep the partition function unchanged as
opposed to Eq. (117) where the mass is evolving together with the dynamics.
2.5.2 E®ective action
Let us consider for the sake of simplicity again the scalar model given by the bare action
Z " #
1 m 2B
2 2
SB = (@¹ ¿ x ) + ¿ x + U B (¿ x) (118)
x 2 2
and the suppression with no structure in the external space [22],
M2
Ck (p) = (119)
2
which acts as a ’smooth cut-o®’ for the amplitude of the °uctuations. The corresponding
evolution equation is
" #¡1
2
h
¹ ¯ ¡ M [¿ ]
@M 2 ¡M [¿ ] = tr M 2 + : (120)
2 ¯ ¿ ¯ ¿
The e®ective action of the theory is obtained by integrating this equation from the initial
condition ¡M0 [¿ ] = SB [¿ ] imposed at M 2 = M02 >> ¤2 >> m 2B to M = 0.
Let us consider how this scheme looks like for the ansatz
Z · ¸
1 2
¡M [¿ ] = Z M (¿ x )(@¹ ¿ x ) + U M (¿ x) : (121)
x 2
One ¯nds
Z
h
¹ 1
@M 2 U =
2p Zp2 + M 2 + U (2)
³ ´
Z " p2 (1)
Z + 2 Z (1) 2
p + U (3)
h
¹ d Z (2)
@M 2 Z = 2Z (1) 3 ¡ 2
2 p (Zp2 + M 2 + U (2) ) (Zp2 + M 2 + U (2) )
³ ´2 ³ ´
Z (1) p2 + U (3) 8p2 Z (1) p2 + U (3)
¡ 2Z 4 ¡ ZZ (1) 4
(Zp2 + M 2 + U (2) ) d (Zp2 + M 2 + U (2) )
³ ´2 #
(1) 2 (3)
8p2 2 Z p + U
+ Z 5 : (122)
d (Zp2 + M 2 + U (2) )
The asymptotic form of the ¯rst equation for M 2 >> U (2) in the local potential approx-
imation,
Z
(2) M2
M @M U M º ¡ U M (123)
p (p 2 + M 2 )2
up to ¯eld independent constant and it agrees with asymptotic form of the ¯rst equation
in (115) except the sign. The integrand in (123) corresponds to transformation rule
(4) of the propagator under the in¯nitesimal change of the mass for the ¯rst, leading
order graphs in Figs. 3. The numerical integration resums the higher orders in the
loop expansion. In short, we see the Callan-Symanzik scheme at work in the functional
formalism.
The suppression (119) freezes out modes with momentum p if p2 < M 2 and generates
a characteristic scale pcr where (Zp2 + M 2 + U (2) )=(Zp2 + U (2) ) deviates from 1. So long
this scale is far from the intrinsic scales of the model, M 2 >> U (2) in the p U.V. scaling
regime, such an internal space blocking generates the external scale pcr º M= Z. In other
words, the universal part of the beta functions obtained in the U.V. regime by means
of the Callan-Symanzik scheme should agree with the same part of the beta functions
coming from other schemes. This agreement has been observed in the framework of the
scalar model [23]. We shall check this quickly in the local potential approximation. By
comparing Eq. (123) with the asymptotic form of the WH equation (14),
(2)
¹ U k d k d¡2
h
k@k U k = ¡ ; (124)
2(2º )d
we ¯nd Z ¤=k
dk 2 y d¡1 dy
=2 : (125)
dM 2 0 (y 2 + M 2 =k 2 )2
Below the upper critical dimension, d < 4, the right hand side is ¯nite and the external
(WH) and the internal (CS) scales are proportional. At the critical dimension the M
dependence of the right hand side is through the proportionality factor ln(1 + ¤2 =M 2 ).
The two scales become approximately proportional only for much higher value of the
cut-o® ¤, since the RG °ow spends more ’time’ close to the non-universal short distance
regime due to the tree-level marginality of g4 . There is no agreement beyond d = 4 where
the cut-o® scale is always important. The agreement between the two schemes is violated
by higher order terms in the perturbation expansion since they represent the insertion of
irrelevant e®ective vertices and by leaving the asymptotical U.V. regimes. This explains
that similar argument does not hold when Z displays important M -dependence.
It is worthwhile noting that the non-trivial wave function renormalization Z and the
corresponding anomalous dimension ² of the ¯eld variable can be thought as the re°ection
of the mismatch between the scale dependence in the external and internal spaces around
the U.V. ¯xed point. In fact, the phenomenological form
h¿ x¿ y i º cjx ¡ yj2¡d¡´ (126)
cf Eq. (128) below connects the fundamental dimensional objects of the internal and
external spaces.
The °exibility of choosing the suppression functional Ck [¿ ] may be important for
certain models. By choosing
k
Ck [¿ ] = S[¿ ] (127)
¤
where ¤ is the U.V. cut-o® the evolution in k resums the loop expansion since the e®ective
Planck-constant, ¹h¡1 (k) = ¹h¡1 + k=¤ evolves from 0 to h ¹ . This scheme is advantageous
for models with inhomogeneous saddle point eg solitons or instantons because their space-
time structure is ’RG invariant’, being independent of the gradual control of the amplitude
of the °uctuations. Another advantage o®ered by the °exibility in choosing the suppres-
sion functional is the possibility of preserving symmetries, the point considered in section
3.5 below.
3 Applications
We shall brie°y review a few applications of the functional evolution equations. An
incomplete list of the developments not followed due to limitation in time is the following.
The exciting competition for the ’best’ critical exponents have led to several works using
this method [24], [53], [17], [50], [51], [25], [57], [26]. Such kind of application opens
up the issue of understanding the impact of truncation on the blocking [27], [28], [29]
[25], [30], [31] and looking for optimization [32], [33], [18]. The phase structure and
the nature of phase transitions are natural subjects [21], [34], [35]. The incorporation
of fermions is essential to arrive at realistic models in High Energy [36] and Condensed
Matter Physics [37]. The computation of the quenched average of Green functions, one of
the essential obstacle of progress in Condensed Matter Physics can be approached in a new
fashion by applying the internal space renormalization group method [38]. A promising
application is in general relativity [39]. Finally, the infamous problem of bound states
can be reconsidered in this framework [40], [41].
Much more to be found in review articles [42], conference proceedings [43] and PhD
thesis [44].
3.1 Fixed points

3.1.1 Rescaling
`The original form of the RG procedure consists of two steps, the blocking, followed by a
rescaling. The latter can be omitted if the RG strategy is used to solve models only. But
it becomes important when we try to understand the scale dependence of the theories.
Our physical intuition is based on classical physics and particles. After losing much of the
substance during the quantization procedure there is a chance to recover some clarity by
the introduction of quasiparticles, localized excitations with weak residual interaction.
The rescaling step of the RG scheme is to check a given quasiparticle assumption by
removing the corresponding scale dependence. The remaining scale dependence is a mea-
sure of the quality of the quasiparticle picture. The deviation from this non-interactive
system is parametrized in terms of anomalous dimensions.
The particle content of a theory is usually ¯xed by the quadratic part of its action.
Therefore the rescaling is de¯ned in such a manner that this part of the action stays
invariant during the blocking. For an O(d) invariant Euclidean or relativistically invariant
real-time system the resulting rescaling re°ects the classical dimension of the dynamical
variables and coupling constants. For non-relativistic systems the ¯xed point scaling is
an arti¯cial device only to identify the non-interacting part of the action. When smooth
cut-o® is used its details may in°uence of the scaling properties of the quadratic action,
as well, and may induce deviations from classical dimensions..
Consider the scale transformation
p ! p0 = (1 + ° )p; x ! x0 = (1 ¡ ° )x; ¿ ((1 + ° )x0 ) ! (1 ¡ ° d Á )¿ (x) (128)
in an O(d) invariant scalar model where dÁ = (d ¡ 2 + ² )=2. The parameter ² re°ects the
deviation of from classical dimensional analysis. The e®ect of this rescaling on the action
XZ
S[¿ ] = up1;¢¢¢;pn ¯ p1+¢¢¢+pn;0 ¿ p1 ¢¢¢¿ pn ; (129)
n p1;¢¢¢;pn
can be found in the following manner [45]:

° The momentum integral measure changes as dd p ! (1 ¡ d° )dd p0 , giving
R
S ! [1 ¡ ° d p ¿ p ±Á± p ]S.
° The coupling constants change as up1;¢¢¢;pn ! u(1¡²)p01;¢¢¢;(1¡²)p0n . This can be written
R
as the transformation S ! [1 ¡ ° p ¿ p p ¢ @p0 ±Á± p ]S of the action where the prime on
the gradient @p indicates that the derivative acts on the coupling constants only and
not on the Dirac-deltas.
° The Dirac-deltas change as ¯ p1 +¢¢¢+pn;0 ! ¯ (1¡²)p01+¢¢¢+(1¡²)p0n;0 , amounting to the trans-
formation S ! [1 + d° ]S of the action.
° Finally, the ¯eld transforms as ¿ p ! [1 ¡ ° (d Á ¡ d)]¿ p, inducing

R
S ! [1 ¡ ° (d Á ¡ d)] p ¿ p ±Á± p S.
Adding up these contributions we ¯nd the rescaling generator
Z " #
¯ ¯
G =¡ ¿ pp ¢ @p0 + d Á¿ p : (130)
p ¯ ¿ p ¯ ¿ p
The RG equation for the rescaled action can be obtained by adding G acting on the
(e®ective) action to the right hand side, eg the evolution (49) turns out to be
1 I 1 I
k@k e¡ h· Sk[Á] = (G + ¹hB) e¡ h· Sk [Á]: (131)
The rescaling of the ¯eld may be viewed as a transformation of the action without
changing the physics at the ¯xed point. The terms generated on the ¯xed point action
by the rescaling, G S ¤ [¿ ] are called redundant at the ¯xed point in question.
3.1.2 Reparametrizing
We considered linear rescaling of the ¯eld variable but one can easily generalize the rescal-
ing to non-linear reparametrization. The in¯nitesimal change ¿ x ! ¿ 0x = ¿ x + ° ª[¿ ; x]
performed inside the path integral gives
Z Z Z " Z #
¡S[Á] 0 ¡S[Á0] ¯ ª[¿ ; x] ¡S[Á]¡² R ª[Á;x] ¯ ¯ S[¿¿ ]
D[¿ ]e ! D[¿ ]e = D[¿ ] 1 + ° e x x : (132)
x ¯ ¿ x
The reparametrization invariance of the integral assures that the modi¯cation
S[¿ ] ! S[¿ ] + ° G ª S[¿ ] (133)
of the action where Z " #

¯ S[¿ ] ¯ ª[¿ ; x]
G ª = ª[¿ ; x] ¡ (134)
x ¯ ¿ x ¯ ¿ x
is an invariance of the partition function. It has furthermore been noted [46] that some
of the in¯nitesimal blocking relations for the (e®ective) action can be written in the form
(133).
Though it is certainly very interesting to ¯nd a common structure for the functional
RG equations the connection with reparametrization invariance is not clear. The point is
that non-linear reparametrizations can not be carried out inside of the path integral as in
ordinary integrals. The problem arises from the fact that the typical con¯gurations are
rather singular in the path integration. Calculus known from classical analysis is replaced
by Ito-calculus in Quantum Mechanics due to the nowhere-di®erentiable nature of the
quantum trajectories [47]. The quantum ¯eld con¯gurations are even more singular.
Consider a free massless ¯eld ¿ in d dimensions and its partition function
YZ d¡2 P YZ P
¡a (¢· Áx)2 ¡ 12 (¢· Á~x)2
Z= d¿ xe 2 x = d¿ xe x (135)
x x
where ¢¹ ¿ x = ¿ x ¡ ¿ x¡¹ and the dimension of the ¯eld was removed in the second
equation, ¿ ~ = ad=2¡1 ¿ . The typical con¯gurations
³
~ 0
´ have ¢¿ = O (a ), ie the discontinuity
of the original ¯eld variable is ¢¿ = O a1¡d=2 .
This simple scaling, the basis of the usual U.V. divergences in quantum ¯eld theory,
gives non-di®erentiability in Quantum Mechanics, for d = 1, ¯nite discontinuity in d = 2
and diverging discontinuities for d > 2. The non-di®erentiability can be represented
by e®ective potential vertices in Quantum Mechanics, [48], a re°ection of the unusual
quantization rules in polar coordinates, sensitivity for operator ordering and quantum
anomalies [49]. In two dimensions the discontinuity is ¯nite and the continuous structure
can either be preserved or destroyed by quantum °uctuations, cf. section 3.4. In higher
dimensions the singularities remain always present. It is this singular nature of the
trajectories which requires that one goes to unusually high order in the small parameter
° characterizing the in¯nitesimal change of variables in the path integral. This problem,
composite operator renormalization, renders the non-linear change of variables a poorly
controlled subject in quantum ¯eld theories.
3.1.3 Local potential approximation
We shall search for the ¯xed points of the Polchinski equation in the local potential
approximation, [50] [51]. We introduce dimensionless quantities ¿ = k 1¡d=2 © and
uk (¿ ) = k d U k (©). The RG equation with rescaling for u is
² +d¡ 2 ¹ 0 u(2)
u_ = ¡ du + ¿ u(1) ¡ u(1)2 K 00 + ¹hK (136)
2
R
where K¹ 0 = k ¡d K 0 and we set ² = 0 in the approximation Z = 1. Notice the explicit
p p
dependence on the cut-o® function K . This re°ects the fact that the RG °ow depends
on the choice blocking transformation. Only the qualitative, topological features of the
renormalized trajectory and the critical exponents around a ¯xed point are invariant
under non-singular rede¯nitions of the action.
Gaussian ¯xed point: The ¯xed point equation, u_ ¤ = 0 has two trivial solutions,
u¤ (¿ ) = 0; (137)
and
¿ 2 ¹0
K
u¤ (¿ ) = + h
¹ ; (138)
2jK00 j dK00
¹ 0 < 0. For any other solution u = O (¿ 2 ) as ¿ ! 1.
where K00 ; K
In order to ¯nd the scaling operators we introduce a perturbation around the ¯xed
point by writing u = u¤ + ° k ¡¸ v(¿ ) where ° is in¯nitesimal and solve the linearized
eigenvalue equation
Ã !
¹ 0 v (2) = (d ¡ 2¡ d
h
¹K ¶ )v + ¿ + 2u¤(1) K00 v(1) : (139)
2
Having a second order di®erential equation one can construct a one-parameter family of
solution after having imposed say v (1) (0) = 1 in a theory with the symmetry ¿ ! ¡ ¿ .
But the polynomial solutions
n
X
2`
vn (¿ ) = vn;` ¿ ; (140)
`=1
which are parametrized by a discrete index n correspond to discrete spectrum [52].

The critical exponents are identi¯ed by comparing the terms O (¿ 2n ) in Eq. (139),
¶ n = d + (2 ¡ d)n, and ¶ n = d¡ (d+2)n, for the ¯xed points (137) and (138), respectively
[53]. The leading order critical exponents are given entirely by the tree-level contribu-
tions. The dimensionful coupling constants are cut-o® independent in this case and we
have ¶ n = [g 2n ] for the ¯xed point (137) where the dimension of the coupling constant [g]
is given by Eq. (24). The dominant, largest exponent ¶ T and the corresponding coupling
constant play distinguished role in the scaling laws. The dimensionless coupling constant
can be identi¯ed with the ’reduced temperature’ t. According to the de¯nition ¹ º t¡º of
the critical exponent ¸ we have the mean-¯eld exponent, ¸ = 1=¶ T = 1=2. The mass is
relevant at the point (137) as expected, as shown by the beta function 2 = ¡ 2K 00 g~22 ¡ 2~
g2 ,
according to Eq. (55). The scaling potentials with exponential growth for large ¿ [54]
correspond to the continuous spectrum and their physical interpretation is unclear.
All exponents are negative at (138), this is an IR ¯xed point.
Wilson-Fischer ¯xed point: There are non-Gaussian ¯xed point solutions for Eq. (136)
when 2 µ d < 4. One can construct a one-parameter family of solutions but the ¯xed
point potentials which remain non-singular for arbitrary ¿ correspond to a discrete set
[55], [52]: the Wilson-Fischer ¯xed point for 3 < d < 4 and as many ¯xed points as
relevant when 2 < d < 3. The perturbation around these ¯xed points is a one-parameter
family of scaling potentials with continuous spectrum of critical exponents. The restric-
tion for solution which are ¯nite and non-singular everywhere produces a discrete spec-
trum [56] in good agreement with other methods of determining the critical exponents
[53], [57].
The truncation of the ¯xed point solution at a ¯nite order of ¿ introduces error and
spurious solutions [27], [28] which can partially be eliminated by expanding the potential
along the cut-o® dependent minimum [29].
3.1.4 Anomalous dimension
The RG equation for the wave function renormalization constant z(¿ ) = Z(©) is
K 000 K 0 (2)2
z_ = ¡ 4K00 u(2) (z ¡ 1) ¡ 2K00 u(1) z(1) ¡ 2 u
k2
¹0 ¡ d+² ¡ 2
hz (2) K
+¹ ² (z ¡ 1) + ¿ z(1) : (141)
2
The lesson of the case Z = 1 is that the requirement of the existence and ¯niteness of
u¤ (¿ ) introduces discrete number of ¯xed point solutions. Let us try to follow the same
strategy again. The dominant terms of such ¯xed point solutions u_ = z_ = 0 of Eqs. (136)
and (141) are

2¡ ² d¡2+²
u¤ (¿ ) º ¿ 4
+ A¿ d+2¡ ² ; z ¤ (¿ ) º B: (142)
4
Together with the conditions u(1) (0) = z (1) (0) = 0 imposed for the ¿ ! ¡ ¿ symmetrical
models the solution are well determined in terms of A and B. The problem is that such
kind of argument does not ¯x the value of ² .
There is another condition to be ful¯lled by the ¯xed points, the critical exponents
should be invariant under rescaling. This is su±cient to determine ² . Unfortunately the
rescaling invariance of the ¯xed point is lost unless sharp or specially chosen polynomial
smooth cut-o® is used [56], [58]. Furthermore the truncation of the gradient expansion
contributes to the violation of this invariance, too. An approximation to ¯nd the ’best’
solution when the rescaling invariance is not respected is the following [51]: Introduce
a further condition which violates rescaling invariance, say ¯x the value of z(0). This
allows the determination of ² which would be unique if rescaling could be used to relax
our last condition. We are as close as possible to the invariant situation within our
parametrized problem when the dependence of ² on z(0) is the slowest. Therefore the
condition d² =dz(0) = 0 selects the ’best’ estimate of ² .
The numerical error due to the truncation of the gradient expansion has been the
subject of extensive studies in the framework of the scalar model, c.f. Ref. [42]. The
general trend is that the truncation of the gradient expansion at the local potential
approximation (zeroth-order) or at the wave function renormalization constant level (¯rst
order) yields approximately 10% or 3% di®erence in the critical exponents compared with
Monte-Carlo simulations, seven-loop computations in ¯xed dimensions or the ¯fth-order
results of the expilon-espansion.
3.2 Global RG
The usual application of the RG method can be called local because as in the determi-
nation of the critical exponent it is performed around a point in the space of coupling
constants. Models with more than one scale may visit several scaling regimes as the ob-
servational scale changes between the U.V. and the IR ¯xed points. The determination
of the ’important’ coupling constants of such models which parametrize the physics goes
beyond the one-by-one, local analysis of the scaling regimes. It requires the careful study
of crossovers, the overlap between the scaling laws of di®erent ¯xed points, a problem
considered in this section.
3.2.1 RG in Statistical and High Energy Physics
It is important to realize the similarity the way RG is used in Statistical and High Energy
Physics despite the super¯cial di®erences. The most obvious di®erence is that while the
running coupling constants are introduced in Particle Physics by means of Green functions
or scattering amplitudes at a certain scale, the parameters of Solid State Physics models
are de¯ned at the cut-o®, the latter being a ¯nite scale parameter, eg lattice spacing.
Since the bare coupling constants characterize the strength of the physical processes at
the scale of the cut-o®, the two ways of de¯ning the scale dependence are qualitatively
similar and reproduce the same universal scaling laws.
The U.V. ¯xed point where the correlation length diverges in units of the lattice spac-
ing in Statistical Mechanics corresponds to renormalized theory in High Energy Physics
where the U.V. cut-o® is removed.
There are two classi¯cation schemes of operators, one comes from Statistical and the
other from High Energy Physics: if the coupling constant of an operator increases, stays
constant or decreases along the renormalized trajectory towards the IR the operator is
called relevant, marginal or irrelevant in Statistical Physics. The coupling constants which
can be renormalized as the U.V. cut-o® is removed in such a manner that observables
converge are called renormalizable in High Energy Physics. The important point is these
classi¯cation schemes are equivalent, in particular the set of irrelevant operators at an
U.V. ¯xed point agrees with the set of non-renormalized ones.
One can easily give a simple quantitative argument in the leading order of the per-
turbation expansion. First let us establish the power counting argument about renormal-
izability. Suppose for the sake of simplicity that there is only one coupling constant, g,
playing the role of small parameter and a physical quantity is obtained as
X
hOi = g n In (143)
n
where In is sum of loop integrals. This series gives rise the expression
[In ] = [O] ¡ n[g] (144)
for the (mass) dimension of the loop integral. Let us recall that the degree of the overall
(U.V.) divergence of a loop integral is given by its dimension. We distinguish the following
cases:
° [g] < 0 : The higher powers of g decrease the dimension in (143) which is compen-
sated for by increasing the degree of the overall divergence of the loop integrals,
cf Eq. (144). Graphs with arbitrary high degree of divergence appear in the pertur-
bation expansion and g is called non-renormalizable.
° g[g] > 0 : The higher order loop integrals are less divergent, there are ¯nite number of
divergent graphs in the perturbation series of any observable. The coupling constant
is super-renormalizable.
° [g] = 0 : The maximal degree of (U.V.) divergence is ¯nite for any observable but
there are in¯nitely many U.V. divergent graphs. g is a renormalizable coupling
constants.
Since the IR degree of divergence of massless loop integrals is just ¡ [In ] the perturbation
expansion of a super-renormalizable or non-renormalizable massless model is IR unstable
(divergent) or stable (¯nite), respectively. The compromise between the U.V. and the IR
behaviors is attained by dimensionless, renormalizable coupling constants. The IR sta-
bility can be realized by a partial resummation of the perturbation expansion in massless
super-renormalizable models [59].
It is rather cumbersome but possible to prove by induction that the de¯nition of

renormalizability outlined above according to the overall divergence of the loop integrals
remains valid in every order of the perturbation expansion.
In order to separate o® the trivial scale dependence one usually removes the classical
dimension of the coupling constants by means of the cut-o® in Statistical Physics. On
the tree-level, in the absence of °uctuations classical dimensional analysis applies giving
the relation g = k [g]g~ between the dimensional and dimensionless coupling constants, g
and g~, respectively. Assuming that there is no evolution on the tree-level we have the
scaling law
Ã !¡[g]+O(¹h)
k
g~(k) = g~(¤) (145)
¤
for the dimensionless coupling constants, showing that the non-renormalizability of a
coupling constant, [g] < 0 is indeed equivalent to irrelevance, the decrease of the cou-
pling constant towards the IR direction. Higher loop contributions do not change this
conclusion so long as the loop expansion is convergent and the anomalous dimension, the
contribution O (¹h) in the exponent can not overturn the sign of the classical dimension
[g]. In case of a marginal coupling constant, [g] = 0 on has to carry on with the loop
expansion until the ¯rst non-vanishing contribution to the anomalous dimension.
One can construct a much simpler and powerful argument for the equivalence of the
irrelevant and the non-renormalizable set of coupling constants in the following manner.
Consider the U.V. ¯xed point P and the region around it where the blocking relations can
be diagonalized, as depicted in Fig. 5. The solid line shows a renormalized trajectory of
a model which contains relevant operator only at the cut-o®. The Lagrangian of another
model whose trajectory is shown by the dotted line contains an irrelevant operator, as
well. The di®erence between the two models becomes small as we move towards the IR
direction and the physics around the end of the U.V. scaling regime is more independent
on the initial value of the irrelevant coupling constants z longer the U.V. regime is. This
is what is called universality of the long range, low energy phenomena. By looking
"backwards" and increasing the cut-o® energy, as done in Particle Physics, the non-
vanishing irrelevant coupling constants explode and the trajectory is de°ected from the
¯xed point. As a result we cannot maneuver ourselves into the U.V. ¯xed point in the
presence of irrelevant operators in the theory. Since the in¯nite value of the cut-o®
corresponds to the renormalized theory, represented by the ¯xed point, the irrelevant
operators are non-renormalizable. Despite its simplicity, this argument is valid without
invoking any small parameter to use in the construction of the perturbation expansion.
Renormalizable models were thought to be essential in High Energy Physics. These
models in which a su±ciently large class of observables can be made convergent are
distinguished by their simplicity, their interactions exist without any cut-o®. Examples
are Yang-Mills models in four dimensions and the sine-Gordon, Thirring and the X-Y
models in two dimensions. The non-renormalizable models are those whose existence
requires a ¯nite value of the U.V. cut-o®, like QED, the ¿ 4 model in four dimensions, the
z
More precisely all the irrelevant coupling constants can modify is an overall scale.
Fig. 5 An U.V. xed point and its vicinity. The x and the y axis correspond a relevant and
an irrelevant operator, respectively. The circle denotes the vicinity of the xed point where the
blocking relation is linearizable.
Standard Model and models in Solid State Physics. It is believed that asymptotically
free models are renormalizable only.
The traditional reason to discard non-renormalizable models was their weak predictive
power due to the in¯nitely many renormalization conditions they require. Since we can
never be sure what kind of Lagrangian to extrapolate up to high energies we need another
point of view. According to the universality scenario the non-renormalizable models can
be excluded because they cannot produce anything di®erent from renormalizable theories.
The cut-o® independent physics of the latter is parametrized by means of renormalized
coupling constants. But the subtle point to study below is that this reasoning assumes
the presence of a single scaling regime with non-trivial scaling laws in the theory which
is a rather unrealistic feature [7].
Thus renormalizability is the requirement of a simple extrapolation to the U.V. regime
without encountering "new physics". The evolution of High Energy Physics shows that
this is a rather unrealistic assumption, any model with such a feature can be an ap-
proximation at best, to study a given interaction and to sacrify the rest for the sake
of simplicity. The goals are less ambitious in Statistical Physics and apart of the case
of second order phase transitions the renormalizability of models is required seldom, for
convenience only, not to carry the regulator through the computation. All models in
Solid State Physics are e®ective ones given with a physically motivated cut-o®.
Let us start the discussion of the possible e®ect of the co-existence of several scaling
regimes along the RG °ow with the simplest case, a model with a gap in the excitation
spectrum above the ground state, ie with ¯nite correlation length, ¹ < 1. Suppose that
the RG °ow of the model starts in the vicinity of an U.V. ¯xed point and reaches an IR
¯xed point region as shown in Fig. 6. Universality, the parametrizability of the physics
beyond the U.V. scaling regime by the relevant coupling constants of the U.V. scaling
laws, tacitly assumes the absence of any new relevant coupling constants as we move along
the renormalized trajectory towards the IR, as shown in Fig. 5. This assumption is in fact
correct for massive models whose IR scaling laws are trivial, meaning that that only the
Gaussian mass term is relevant. To see why let us consider the RG °ow in the IR side of
the crossover, when the U.V. cut-o®, say the lattice spacing a >> ¹ and make a blocking
step, a ! a0 which generates the change g n (a) ! gn (a0 ) in the coupling constants. The
UV IR
Fig. 6 The RG ®ow of a model with massive particles. The renormalized trajectory connects
the U.V. and IR scaling regimes, indicated by circles which are separated by a crossover at the
intrinsic scale of the model, p º 1=¹ .
change ¢gn (a) = gn (a0 ) ¡ g n (a) is due to °uctuation modes whose characteristic length
is a < ` < a0 . Owing to the inequality ¹ << a < ` these °uctuations are suppressed
by e¡a=» and the °ow slows down, gn (a) º gn (a0 ). This implies the absence of run-away
trajectories, the absence of relevant non-Gaussian operator. But more realistic models
with massless particles or with dynamical or spontaneous symmetry breaking occurring
at ¯nite or in¯nite length scales, respectively, or with condensate in the ground state
the IR scaling may generate new relevant operators and universality, as stated in the
introduction, referring to a single scaling regime is not a useful concept any more.
3.2.2 Scalar model
In order to understand better the generic case shown in Fig. 6 we return to its simplest
realization, the scalar model in the local potential approximation, Eq. (14), [7]. One
does not expect anything unusual here since the model possesses a gap and the IR scaling
regime is trivial in either the symmetrical or symmetry broken phase. But we shall see
that the model has soft large amplitude °uctuations and the true vacuum is just there
where this non-perturbative and the well known perturbative domains join. The second
derivative of the local potential is discontinuous there, depends on the direction from
which the vacuum is approached. We shall arrive at some understanding of the soft
modes in two step, by separating the loop contributions from the tree-level structure. We
start with the more traditional loop contributions.
To facilitate the dealing with soft modes we introduce an external constraint which
R
controls the expectation value x h¿ x ij = V ©. The usual realization of this constraint
is the introduction of an external source as in Eq. (5). We shall consider the model
R
in the symmetry broken phase where x h¿ x i0 = V ©vac 6= 0 and choose homogeneous
external source jx = J which can destabilize the naive symmetry broken vacuum whenever
J©vac < 0 and can induce j©j < ©vac . Note that such a control of the ¯eld expectation
value by an external source preserves the locality of the action.
The beta-functions, as given by Eq. (23), depend strongly on the propagator
G(k 2 ) = 1=(k 2 + g2 (k)), represented by the the internal lines in Figs. 3. The qualitative
features of the function G¡1 (k 2 ) are shown in Fig. 7 for © = 0. According to the loop
expansion
(
(2) (2)
¡1 2 2 (2) k 2 + U ¤ (©) + O (¹h) U k << k 2 ,
G (k ) = k + U k (©) = (2) (2) (146)
k 2 + U 0 (©) + O (¹h) k 2 << U k ,
the bending of the lines in the ¯gure is due to radiative corrections, as far as the per-
turbative region, G ¡1 >> 1 is concerned. The dashed line corresponds to the massless
Coleman-Weinberg case [23]. The °ow above or below this separatrix is in the symmet-
rical or symmetry broken phase, respectively. This qualitative picture is valid so long
j©j < ©vac .
It is important to recall that the argument of the logarithmic function in the WH
equation in Eq. (14), G¡1 (k 2 ), is the curvature of the action at the cut-o® and serves
as a measure of the restoring force driving °uctuations back to their trivial equilibrium
position. Thus perturbative treatment is reliable and the loop contributions are calculable
so long G ~ ¡1 (k 2 ) = G ¡1 (k 2 )=k 2 >> g~4 .
In order to have a qualitative ides if what is happening let us consider the model where
U ¤ (©) = m 2 ¿ 2 =2 + g¿ 4 =4! at the cut-o®, k = ¤. By simply ignoring the loop corrections
we have g 2 (k) = m2 , g4 (k) = g and gn (k) = 0 for n > 4 on the tree-level, ie m 2 and g can
be identi¯ed with renormalized parameters and m2 < 0. For an isolated system (j = 0)
where j©j = ©vac = ¡ 6m2 =g and the use of the perturbation expansion appears to be
justi¯ed for G ~ ¡1 (k 2 ) = 1 ¡ m2 =k 2 > g4 k d¡4 . Spontaneous symmetry breaking, or the
appearance of a condensate in the ground state in general, is characterized by instability
of °uctuations around the trivial, vanishing saddle point. When an external source is cou-
pled to ¿ x to dial j©j < ©vac in the symmetry broken phase then non-perturbative e®ects
set in as k decreases. We can see this easily when the ¯eld expectation value is squeezed in
the concave part of the potential, j©j < ©in° = ¡ 2m2 =g = ©vacq=3. In fact, G¡1 approaches
zero and the amplitude of the °uctuations explodes at k = ¡ m 2 ¡ g©2 =2. Below this
non-perturbative regime the plane wave modes become unstable since G ¡1 (k 2 ) < 0 and a
coherent state is formed, re°ected q in the appearance of a non-homogeneous saddle point
with characteristic scale 1= ¡ m ¡ g©2 =2.2
The mechanism responsible of spreading the instability over the whole region j©j < ©vac
comes from the tree-level structure and will be discussed later, in section 3.3.
We identify at this point four di®erent scaling regimes:
A: In the U.V. scaling regime k 2 ¾ g2 (k): Gn º g n =k 2 , the mass term is negligible.
B: The explicit scale dependence disappears in the IR regime k 2 ½ g2 (k) of the sym-
metrical phase, where G n º gn =g 2 . This is a trivial scaling, there are no interactive
relevant operators.
C: The onset of the condensation is at G ¡1 (k 2 ) = k 2 + g 2 (k) º 0 and the higher loop
e®ects start to dominate.
D: The spinodal unstable region where the homogeneous ground state becomes unstable
against in¯nitesimal °uctuations, k 2 +g 2 (k) < 0. The leading scaling laws are coming
from tree-level, classical physics.
There are two crossover regimes. One is between the scaling regimes A and B at
2
k º g2 (k). Another one is between A and C when j©j < ©in° . This will be extended to
A
G 1
C p2
Fig. 7 The inverse propagator G¡1 (p2 ) of the scalar model as the function of the momentum
square p2 for © = 0. The dotted and solid lines correspond to the symmetrical and the symmetry
broken phase, respectively. The massless case is shown by dashed line. One can distinguish four
di¬erent scaling regimes, namely A: U.V., B: symmetrical massive, C: precursor of condensation
and D: spinodial phase separation regime.
the region j©j < ©vac by tree-level contributions. We shall look into the crossover of the
symmetry broken phase.
The best is to follow the evolution of the local potential the plane (©; k 2 ), as shown
in Fig. 8. The potential U k (©) should be imagined as a surface above the plane
(©; k 2 ). The RG equations for the dimensionless coupling constants, g~_ n = ~n where
~n is given by Eq. (25) are integrated in the direction of the arrows. The initial condi-
tion U~¤ (¿ ~) = m 2 ¿ ~2 =2k 2 + gk d¡4 ¿ ~4 =4!, set at k = ¤, along the horizontal line on the top.
As the running cut-o® k is decreased and the RG equations are integrated the potential
becomes known along horizontal line intercepting the ordinate at k 2 . The spontaneous
breaking of the symmetry in the vacuum implies that for su±ciently small j©j the prop-
agator explodes by approaching a singularity at k = k cr (©) as k decreases. This happens
¯rst at © = 0 when the running horizontal line touches the two curves shown in the Figure.
These curves correspond to the generalization of the minimum and the in°ection point
of the potential for k 2 > 0 and are given by the relations k 2 + m 2B + g©2min =6 + O (¹ h) = 0
2 2 2
and k + m B + g©in°=2 + O (¹ h ) = 0, respectively. We shall argue in section 3.3 that the
singularity lies close to the extension of the minimum,
k cr (©min (k)) º k: (147)
The numerical integration of the WH equation requires a truncation of the summation

in Eq. (20) for the local potential. By using nmax µ 22 the singular line can be located
with reasonable accuracy. Close to the singular line on the plane (©; k 2 ) the coupling
constants increase rapidly and the truncation of the potential is not acceptable. In the
U.V. scaling regime (denoted by A in Fig. 7) the propagator is small, G = O (k 2 ), and
the ¯rst term on the right hand side of Eq. (23) is dominant. As we approach kcr the
propagator starts to increase, all term contribute equally in the beta functions and we
enter into a new scaling regime (C in Fig. 7).
In order to connect the two scaling regimes we shall consider
fm;n (k=¤) = @ ~m;k (¿ )=@ g~¤;n , the dependence of the beta functions at k on the micro-
2
k
2 2
k =L
Fig. 8 The (©; k 2 ) plane where the local potential Uk (©) is de ned. The potential should be
imagined as a surface above the this plan. The initial condition gives U¤ (©) on the horizontal
dashed line and the RG equation is integrated from this line in the direction of the arrows. The
extension of the minimum and the in®ection point of the potential from k 2 = 0 to k 2 > 0 is
shown by solid and dotted lines, respectively.
scopical initial condition, imposed at k = ¤. The simplest way to estimate fm;n (x) is to
start with fm;n (1) º 1 which can be inferred from the form (23). The scaling law (145)
allows us to write
Ã ! Ã !¡[gn] 0Ã !¡[gn] 1
k @ ~m (¿ ) @ ~m (¿ ) k k
fm;n = = = O@ A (148)
¤ @ g~n (¤) @ g~n (k) ¤ ¤
up to corrections O (¹ h), con¯rming the universality of the U.V. scaling regime. In fact,
initial, microscopical value of the the irrelevant, non-renormalizable coupling constants
([g] < 0) leave vanishing trace on the dynamics as ¤=k ! 1.
The numerical determination of the function fm;n (x) in d = 4 supports the prediction
(148) in the U.V. scaling regime, namely the impact of a non-renormalizable coupling
constants g n (¤) on the beta functions weakens as (k=¤)¡[gn] for x º 1. But this trend
changes for x < 1 as the critical value of the cut-o®, kcr , found to be in agreement
with Eq. (147), is approached in what f starts to increase. The potential must not be
truncated close to the singularity and the numerical results appear inconclusive. But the
increase of fm;n (x) at the crossover actually starts already far enough from kcr when all
couplings are small, g~n << 1 and the truncation is safe. The lesson of this numerical result
is that universality as used around the U.V. scaling regime alone is not valid anymore.
The instability at the onset of the condensation at k º k cr introduces divergences which
are strong enough to overwrite the U.V. scaling laws and generate new relevant operators.
Such an instability of the RG °ow enhances the sensitivity of the physics at ¯nite length
scales on the microscopical parameters.
The need of generalizing universality in a global manner has been shown for j©j < ©vac .
It remains to see if such a phenomenon can be observed at the highly singular point
j©j = ©vac , in the true vacuum.
A similar phenomenon has already been noticed in connection with the BCS ground
state. It has been pointed out that kinematical factors turn the four electron operator
which is irrelevant according to the power counting into marginal for processes close to
the Fermi level [60], [61] [62]. The collinear divergences at the Fermi level drive the
instability of the non-condensed vacuum and generate new scaling laws. Another similar
mechanism is the origin of the strong long range correlations in the vacuum state of
Yang-Mills models. They appear as a consequence of non-renormalizable, U.V. irrelevant
Haar-measure term of the path integral [63].
We note ¯nally that the view of the RG °ow as a parallel transport, mentioned in
section 2.3 is particularly well suited for the studies of crossovers. This is because the
sensitivity matrix (78) is a global quantity displaying clearly the sensitivity of the IR
physics on the U.V. parameters by construction.
3.2.3 RG microscope
We now embark on a rather speculative subject, playing with the possibility of matching
di®erent scaling laws. Let us start with the generic case, a model with U.V. and IR scaling
regimes, separated by a crossover, as sketched in Fig. 6. We have a classi¯cation of the
very same operator algebra at both scaling regimes. One can write any local operator A
P
as a linear superposition of scaling operators, A = n cA;n On , the latter being the eigen-
operators of the linearized blocking relations of a scaling regime, On (k) º (¤=k)ºn O(¤)n .
Let us denote by ¸ A the largest scaling dimension of scaling operators which occur in the
a linear decomposition of A, ¸ A = maxn ¸ n . We shall simplify matters by calling a local
A operator relevant, marginal or irrelevant if ¸ A > 0, ¸ A = 0 or ¸ A < 0, respectively.
We ignore marginal case by assuming that radiative corrections always generate non-
vanishing scaling dimensions.
One can distinguish between the following four cases:
(a) rU:V: rIR : relevant in both regimes,
(b) rU:V: iIR : relevant in the U.V., irrelevant in the IR,
(c) iU:V:rIR : irrelevant in the U.V., relevant in the IR,
(d) iU:V:iIR : irrelevant in both regimes.
The qualitative dependence of the corresponding coupling constants is shown in Figs. 9.
Let us consider for example QED containing electrons, muons and ions with metallic
density. There is an U.V. scaling regime for energies well above the electron mass. Sup-
pose that the ground state is a BCS type superconductor where the IR scaling laws are
driven by long range phonon mediated interactions. The electron mass is an example of
the class (a). The muon mass belongs to class (b) owing to the decoupling of muons at
low energy. By assuming that radiative corrections turn the four fermion vertex relevant
it serves as an example of class (c). Finally the six fermion vertex is doubly irrelevant
and belongs to class (d).
The interesting class is (c). Suppose that we modify the values of the bare coupling
constants at the initial condition and ¯nd the dashed line. Due to the universality of the
U.V. scaling regime the two renormalized trajectories converge towards each other for
0 < x < xuv and run very close between the two scaling regimes, xuv < x < xir . But in
the IR scaling regime where the coupling constant is relevant the two trajectories start to
g g
x uv x ir x uv x ir
(a) (b)
g g
xuv x ir x uv x ir
(c) (d)
Fig. 9 The qualitative dependence of the four types of coupling constants on the observational
length scale. The U.V. and the IR scaling regimes are 0 < x < xuv and xir < x, respectively.
The dashed line in (c) corresponds to a di¬erent initial condition for the RG ®ow.
diverge from each other. The question is whether the di®erence of the trajectories in the
U.V. scaling regime has any impact on the di®erence found at the IR. It may happen that
the IR phenomena is independent of the U.V. regime and the crossover smear completely
out the small e®ects of the U.V. initial condition of the trajectory. But one can imagine
that the extremely small di®erences left at the crossover remain important and lead to
an initial condition dependent divergence at the IR. The answer is model dependent and
can be given by the detailed analysis of the set of coupled di®erential equations.
In order to make this point clearer let us imagine a model with two parameters, say a
mass m and a coupling constant g, both expressed in a dimensionless manner by means
of the U.V. cut-o®, the lattice spacing a. Suppose the following rather simple scaling
laws: The beta function of the mass is approximately constant,
a @m
® (m; g) = º m >0 (149)
m @a
and of the coupling constant g can be written as
· µ ¶ µ ¶ ¸
@g m mcr
(m; g) = a = À ¸ U:V: + À ¸ I:R: g (150)
@a mcr m
where À (z) is interpolating smoothly between À (0) = 1 and À (1) = 0. The running mass
is given by µ ¶ ¯m
a
m(a) = m(a0 ) ; (151)
a0
and the length scale of the crossover between the U.V. and the I.R. scaling regimes is at
acr º aU:V:(m cr =mU:V: )1=¯m. The asymptotical scaling for the coupling constant is
³ ´¡ºU:V:
aU:V:
g(aU:V: ) º cU:V: g(acr ) acr
; aU:V: ½ acr ;
³ ´ºI:R:
aI:R:
g(aI:R: ) º cI:R: g(acr ) acr
; aI:R: ¾ acr : (152)
Let us furthermore assume that ¸ U:V: < 0 and ¸ I:R: > 0, i.e. this coupling constant
belongs to the class (c). The sensitivity of the coupling constant in the asymptotical
regions, g(aU:V:) and g(aI:R:) on g(acr ) is
µ ¶
@g(aU:V: ) aU:V: ¡ºU:V:
º cU:V: ;
@g(acr ) acr
µ ¶
@g(aI:R: ) aI:R: ºI:R:
º cI:R: (153)
@g(acr ) acr
therefore
@g(aI:R: ) cI:R: aºI:R:
I:R: ºU:V:
aU:V:
º º +º
: (154)
@g(aU:V: ) cU:V: acrI:R: U:V:
If the IR scaling regime is long enough then the di®erence between renormalized trajec-
tories with di®erent initial conditions imposed in the U.V. regime can be as large or even
larger than at the initial condition. When the U.V. and the I.R. observational scales are
related by
µ ¶
acr ¡ºI:R: =ºU:V:
aU:V: º acr (155)
aI:R:
then (154) is O (a0 ) cI:R:=cU:V: and despite the ’focusing’ of the universality in the U.V.
regime the ’divergence’ of the trajectories in the I.R. can amplify the extremely weak
dependence on the initial condition at ¯nite scales. The increase of aI:R: in this case
enhances the sensitivity on the initial condition and we can ’see’ the initial value of the
non-renormalizable coupling constants at smaller distance a.
The lesson of the numerical results mentioned above is that the scalar theory in the
unstable regime realizes such a RG ’microscope’. The loss of the U.V. based universality
as we approach the critical line on the plane (©; k 2 ) suggests that there is at least one
relevant operator at the new scaling laws which is irrelevant in the U.V.. We do not
know the eigen-operators of the linearized blocking relations in this region but this new
relevant operator must contain the local monomials ¿ nx with n > 4. To make it more
di±cult, it is not obvious that this is a local operator as opposed to the scaling operators
of an U.V. ¯xed point.
In order to parametrize the physics of the scalar model we have to use this new
coupling constant at the U.V. cut-o®, as an additional free parameter. This is rather
un-practical not only because we do not know the operator in question but mainly due
to the smallness of this coupling constants in the U.V. and the crossover region. It seems
more reasonable to use a mixed parametrization, consisting of the renormalizable coupling
constants at the U.V. cut-o® and the new coupling constant taken close to the singularity
where it has large enough value, ie k º k cr . In the scalar model kcr ! 0 as the external
source is turned o®. We shall call the new coupling constants appearing in this manner
hidden parameters.
What operator corresponds to the hidden parameter of the scalar model? It seems
reasonable to assume that the only observables whose value is determined at the onset of
the condensate is just the magnitude of the condensate. Therefore the conjecture is that
the strength of the condensate, ©vac , is dynamically independent of the renormalizable
parameters g 2 , g 4 in d = 4. We have therefore the following possibilities in parametrizing

the scalar model with spontaneously broken symmetry:
° We identify a non-renormalizable coupling constant, gnr , which in°uences the value
of the condensate and use the bare values of g2 (¤); g4 (¤); gnr (¤) at the U.V cut-o®.
° One uses the renormalizable bare coupling constants from the U.V. end and the
strength of the condensate from the I.R. regime, g 2 (¤); g 4 (¤); ©vac .
° One may use I.R. quantities only, say g 2 (0); g 4 (0); ©vac .
Our point is that the scalar model with spontaneously broken symmetry has three free
parameters instead of two. This rather surprising conjecture can be neither supported
nor excluded by perturbation expansion because one can not connect the U.V. and the
I.R. regimes in a reliable manner. The numerical studies of the scalar model carried out
so far are inconclusive, as well, because they were constrained to the quartic potential
and the possible importance of higher order vertices in forming the condensate was not
considered.
The modi¯cation of the scaling laws of the scalar model, necessary to generate a
hidden parameter comes from the onset of a condensation. As discussed below in Section
3.3 there are non-trivial tree-level scaling laws in a condensate which overwrite the loop-
generated beta functions and may provide the dynamical origin of the hidden parameter.
In a similar manner one can speculate about the role of the Cooper-pair condensate in the
BCS vacuum. If there turns out to be a hidden parameter as in the one-component scalar
model then the non-renormalizable e®ective coupling constants of the Standard Model
would in°uence the supercurrent density in ordinary metals! Another possible example
is the Higgs sector of the Standard Model, where the hidden parameter, if exists, would
be an additional free parameter of the model.
One may go further and inquire if dynamical symmetry breaking can modify the
scaling law in a similar manner. The bound state formation, responsible to the generation
of large anomalous dimensions in strong coupling QED [64], in strong extended technicolor
scenario [65, 66], in the Nambu-Jona-Lasinio model [67] or in the top-quark condensate
mechanism [68, 66] may be a source of hidden parameters, as well.
Hidden parameters represent an unexpected ’coupling’ between phenomena with very
di®erent scales and question of our traditional strategy to understand a complex system
by analyzing its constituents ¯rst. The axial [70] or scale [71] anomalies represent a
well know problem of this sort, except that the regulator independence of the anomaly
suggests that for each classical IR ¯xed point there is a single ’anomalous’ one without
continuous ¯ne tuning.
Universality is expressed in the context of High Energy Physics by the decoupling
theorem [69]. Let us start with a renormalizable model containing a heavy and a light
particle and consider the e®ective theory for the light particle obtained by eliminating
the heavy one. There are two possible classi¯cation schemes for the e®ective vertices for
the light particle, generated by the elimination process. One is according to the U.V.
scaling laws in the e®ective theory, i.e. there are renormalizable and non-renormalizable
vertices. Another scheme is based on the strength of the e®ective interactions, i.e. there
are vertices which are stay constant or diverge when the ratio of the light and the heavy
particle mass tends to zero and there are vertices whose coe±cients tend to zero in the
same limit. The decoupling theorem asserts that these two classi¯cation schemes are
equivalent, i.e. the non-renormalizable e®ective coupling constants are vanishing when
the heavy mass diverges. The existence of the hidden parameter is, at the ¯nal count is
a violation of this theorem.
3.2.4 The Theory of Everything
The appearance of a hidden coupling constant renders our goal of understanding the
hierarchy of interactions in Nature extremely complicated. In fact, the RG °ow of the
Theory of Everything must be imagined in the space of all coupling constants in Physics,
including the parameters of the Grand Uni¯ed Models down to di®erent models in Solid
State Physics or hydrodynamics in the classical regime, as indicated in Fig. 10. Di®erent
¯xed points are approached by the trajectory when the scale dependence is dominated
by a single interaction. But all ¯xed point except the U.V. one is avoided due to the
non-renormalizable, irrelevant vertices [69] generated by the dynamics of the next ¯xed
point in the U.V. direction.
Such a wandering in the space of coupling constants echoes the age old disagreement
between High Energy and Solid State Physics. It is usually taken for granted in the High
Energy Physics community that the su±ciently precise determination of the microscopical
parameters of the Theory of Everything would ’¯x’ the physics at lower energies. The
obvious di±culties to extract the high energy parameters from experiments which render
each new experiment expensive, long and large scale operation indicate that this might
not be a practical direction to follow. In other words, the relevant coupling constants have
positive Lyapunov exponent and render the trajectory extremely sensitive on the initial
conditions. Therefore the characterization of the RG °ow of The Theory of Everything
by the initial condition, though being possible mathematically, is not practical due to the
¯nite resolution of the measurements. The other side of the coin, the physical parameters
of a ¯xed point are clearly the relevant coupling constants of the given scaling regime.
They are the object of Solid State Physics, as far as the scaling regimes QED, CM and IR
are concerned in Fig. 10. One looses sight of the fundamental laws but gains predictive
power by restricting oneself to local studies of the RG °ow.
3.3 Instability induced renormalization
In the traditional applications of the RG method which are based on the perturbation
expansion the RG trajectory is sought in the vicinity of the Gaussian ¯xed point. One can
describe in this manner crossovers which connect scaling regimes which share the same
small parameter. We can enlarge the classes of accessible crossovers by relying on the
saddle point expansion, the only systematical non-perturbative approximation scheme.
The generic way to induce new crossover in this scheme is by passing a condensation at
a ¯nite length scale [72].
GUT QCD
IR’
CM
EW
TOE QED
IR
Fig. 10 The renormalization group ®ow of the Theory of Everything. The branching drawn at
an IR scale is a phase transition driven by the environment, such as particle or heat reservoirs.
One may object that there is nothing surprising or new in ¯nding that the physics is
fundamentally changed by a condensation mechanism. What additional knowledge can
then be gained by looking into condensation by the method of the RG? The point is that
the classi¯cation of operators around a ¯xed point which was achieved in the framework of
the perturbation expansion changes essentially by the condensation, ie the appearance of
a non-trivial saddle point. To see how this happens we return to the strategy followed at
Eq. (2). Let us consider an observable A in a model with a single coupling constant ³g for ´
¹ F 1 (g; ¤) + O ¹h2 ,
simplicity, computed in the saddle point expansion, hAi = F 0 (g; ¤) + h
where F ` (g; ¤) denotes the `-loop contribution. We can obtain the beta function by
taking the derivative of this equation with respect to cut-o® ¤,
³ ´ · µ ¶¸ ³ ´
k@k (F 0 + h ¹ F1) k@k F 0 @k F 1 @g F 1
= k@k g = ¡ ¹2 = ¡
+O h 1+¹
h ¡ ¹ 2 : (156)
+O h
@g (F 0 + ¹hF 1 ) @g F 0 @k F 0 @g F 0
The loop corrections F` (g; ¤), ` ¶ 1 are polynomials of the coupling constants g. But
the leading order, tree-level piece usually has stronger dependence on g and may become
singular as g ! 0. The tree-level contribution may induce qualitatively new scaling laws
with new set of relevant operators. Due to the singularity at g = 0 the condensation
actually realizes the dangerous irrelevant variable scenario [73].
At the onset of an instability, region C in Fig. 7 certain modes experience strong
nonlinearity and develop large amplitude °uctuations. But as we enter in the unstable
regime one may hope to recover quasiparticles with weak residual interaction after having
settled the dynamics of the unstable modes. At least this is what happens within the
framework of the semiclassical, or loop expansion where the °uctuations remain small
after the proper condensate is found. There are models where the vacuum state is just at
the edge of the instability, like the scalar model discussed above in the symmetry broken
phase.
The tree-level contributions to the blocking transformation, if exist, are more impor-
tant
³ then
´ the loop corrections. This is the reason that we shall consider the tree-level,
O h¹ 0 RG °ow. There is no evolution in this order so long the system is stable and
the saddle point is trivial. But as soon as we arrive at the unstable region the RG °ow
becomes non-trivial.
3.3.1 Unstable e®ective potential
Before turning to the actual tree-level blocking let us review brie°y the kind of instabilities
one expects. As we enter into the unstable regime k 2 + g2 (k) < 0 the trivial saddle point,
¿ k = 0 becomes unstable in the blocking (7). To understand this instability better we
introduce the constrained functional integral,
Z µ Z ¶
¡SB[Á] 1
Z© = D[¿ ]e ¯ ¿ x ¡ © (157)
V x
corresponding to a conserved order parameter. By following the RG trajectory until

the IR end point one eliminates all but the homogeneous mode by keeping the partition
function unchanged,
Z © = e¡Sk=0[©] = e¡V Uk=0(©) : (158)
To ¯nd another local representation for this constrained partition function we consider the
generator functional density w[j] = W [j]=V de¯ned by means of Eq. (5) and compute its
Legendre transform, V e® (©) in the mean ¯eld approximation. For this end we set jx = J
and write
¡ V e® (©) = w[J ] ¡ J © = min[¡ U k=0 (¿ ) + J(¿ ¡ ©)]: (159)
J;Á
The minimization with respect ¿ and J yields J = dw(J )=dJ and and ¿ = ©, respectively.
Finally we have V e® (©) = U k=0 (©), the result announced after Eq. (16) so long the mean
¯eld approximation is reliable, namely in the thermodynamical limit and in the absence
of large amplitude °uctuations. It is further known that Veff (©) is convex.
The instabilities of the kind mentioned above are well known in the case of ¯rst order
phase transitions. But similar instabilities may appear at higher order phase transitions,
as well. In fact, the magnetization of the Ising model as the function of the external
magnetic ¯eld shows discontinuous behavior below the critical temperature. Therefore
the free energy constrained into a sector with a given magnetization displays such kind
of instabilities.
Suppose that the U k=0 (©) determined perturbatively has degenerate minima and a
concave part. Then it is advantageous to introduce two curves on the (©; k) plane of
(1) (2)
Fig. 8, ©in°(k) and ©min (k) de¯ned by k 2 ©min + U k (©min ) = 0, and k 2 + U k (©in° ) = 0.
The stable region in the mean-¯eld approximation is j©j > ©min (0) = ©vac .
For ©in° (0) < j©j < ©min (0) there are two minima in ¿ for the last equation in (159).
One of them is metastable, ie is unstable against su±ciently large amplitude modes.
The spinodal phase separation, the instability against in¯nitesimally small amplitude
°uctuations occurs when some eigenvalues of the second functional derivative of the action
becomes negative. This is region D in Fig. 7, j©j < ©in° (0). It is unreachable by the mean-
¯eld treatment because no local minimum is found in ¿ which would satisfy j¿ j < ©in° (0).
To understand better the nature of these unstable regions one has to go beyond the
mean-¯eld approximation and to follow the dynamics of the growing inhomogeneous
instabilities. We shall use the tree-level WH equation to deal with such large amplitude,
inhomogeneous °uctuations.
3.3.2 Tree-level WH equation
The construction of the condensate within the unstable region implies the minimization
of the action with respect to a large number of modes. The RG strategy o®ers an
approximation for such a rather involved problem, it deals with the modes one-by-one
during the minimization [75], [74].
The negative curvature of the potential makes the saddle point in (9) non-trivial,
Sk¡¢k [¿ ] = min Sk [¿ + ¿ ~0 ] = Sk [¿ + ¿ ~cl[¿ ]] 6= Sk [¿ ]: (160)

Á~0
Since the saddle point depends in general on the background ¯eld, ¿ ~cl = ¿ ~cl[¿ ], the action
is modi¯ed during the blocking and we ¯nd a non-trivial RG °ow [74]. The local potential
approximation to the tree-level blocking is
Z · ¸
1
U k¡¢k (©) = min (@¹ ¿ ~cl;x)2 + U k (© + ¿ ~cl;x;) : (161)
Á~cl;x x 2
One encounters here a conceptual problem. The k-dependence of the tree-level RG
°ow might well be singular since there is no obvious reason that the saddle points which
are usually rather singular functions of the parameters evolve smoothly. This problem
has already been noticed as a possible ’¯rst order phase transition’ in the blocking which
induces discontinuous RG °ow [76], [77], [78]. It was later shown by rigorous methods that
the RG °ow is either continuous or the blocked action is non-local [79]. The resolution of
this apparent paradox is that the saddle point actually develops in a continuous manner
as we shall shown below. As the RG °ow approaches the onset of the condensate, k ! k cr
then the new scaling laws generate such an action that the saddle point turns out to be
continuous in k.
3.3.3 Plane wave saddle points
The saddle points, the minima in (160) satisfy the highly non-linear Euler-Lagrange
equations whose solutions are di±cult to ¯nd. The use of sharp cut-o® slightly simpli¯es
the problem since it reduces the functional space in which the minimum is sought to
¿ ~ 2 Fk nFk¡¢k . We shall retain the the plane wave saddle points only,
» k h iµk i
¿ ~cl;p = e ¯ p;kek + e¡iµk ¯ p;¡kek ; ¿ ~cl;x = » k cos(kek ¢ x + ³ k ): (162)
2
The parameter ³ k and the unit vector ek correspond to zero modes, they control the
breakdown of translational and rotational symmetries.
This is a key to what happens at the unstable line: Despite the discreteness of
the internal symmetry ¿ ! ¡ ¿ there are soft modes because the inhomogeneous sad-
dle points break the continuous external symmetries. The condensation of particles with
non-vanishing momenta automatically generates Goldstone modes. This phenomenon is
well known for solids where the saddle point is a crystal of solitons which breaks external
symmetries and there is no integration over the zero mode to restore the symmetrical
ground state. Our point is that the soft modes make their appearance even if the sym-
metry of the ground state is restored by the integration over the zero modes.
The amplitude » k is determined by minimization,
µ Z ¶
1 2 2 1 ¼
U k¡¢k (©) = min k » k + dyU k (© + » k cos y) ; (163)
½k 4 º 0
in the local potential approximation. The numerical implementation of this iterative
procedure to ¯nd the RG °ow with an initial condition imposed on the potential at
k = ¤ gave the following results [74]:
(1) The recursive blocking relation, Eq. (163) is not a ¯nite di®erence equation. Despite
of this the °ow converges as ¢k ! 0.
(2) The amplitude of the saddle point satis¯es the equation © + » (k) = ©min (k).
(3) The potential obtained by the tree-level blocking is U k (©) = ¡ 12 k 2 ©2 .
(4) The tree-level results above hold independently of the choice of the potential at the
cut-o®.
The key is point 3, point 1 follows immediately from it. The lesson of this result is that
the action is degenerate for the modes at the cut-o®, the kinetic and potential energies
cancel. The ’best’ e®ective theory for a given plane wave mode is the one whose cut-o®
k is slightly above the wave vector of the mode. This result suggests the degeneracy of
the action within the whole unstable region. Conversely, if we can show that the action
is degenerate at the cut-o® within the unstable regime we established this potential.
We prove by induction in the number of steps k ! k ¡ ¢k that the variation of the
action density within the unstable regime is O (¢k). This result protects the consistency
of the saddle point expansion for d > 1 since ¢k ³ ¶ 2º´ =L where L is the size of the system
and therefore the variation of the action is O Ld¡1 . Recall that k cr denotes the cut-o®
where the kinetic snd the potential energies cancel.
First step: Let us denote by k 0 the value of the cut-o® at the ¯rst occurrence of non-
trivial saddle point in the numerical implementation of Eq. (163). It obviously satis¯es
the inequality kcr ¡ ¢k < k 0 < k cr . The trivial saddle point ¿ ~cs = 0 becomes unstable ³p for
´
blockings with j©j < ©in° (k ) in which case j© + ¿ ~cs;x j < ©in° (k ), ie j¿ ~cs;x j = O
0 0
¢k .
³ ´
The term O ¿ ~ is canceled in the action on a homogeneous background ¯eld so the
³ ´
O ¿ ~2cs contribution gives O (¢k) variation.
Induction: Suppose that the variation of the action is O (¢k) in the unstable region
and we lower the cut-o®, k ! k ¡ ¢k. At the new cut-o® the balance between the kinetic
and the potential energy is lost by an amount of O (¢k) since the potential energy is
still the given by k but the kinetic energy corresponds to the lowered cut-o®, k ¡ ¢k.
Thus the negative curvature potential energy wins and the action bends downward in the
unstable region as the function of the amplitude of the plane wave. By assuming that
the amplitude is stabilized at an O (¢k 0 ) value (will be checked later) the variation of
the action density O (¢k) in the unstable region.
Point 2 can be understood in the following manner: During the minimization of the
action » k slides down on the O (¢k) slope until the potential starts to increase again. We
can ¯nd where this happens by equating the slope of the original, bare potential with that
stated in point 3. One would expect that the deepest point is reached at » k = ©in° (k).
But the O (¿ 4 ) term in U k (¿ ) is not yet strong enough to make the potential increase
strong enough here. The two slopes agree just at » k = ©min (k).
Result 4 which follows from 3, as well, can be called super universality since it re°ects
scaling laws where all coupling constants are irrelevant. Note that the k-dependence is
continuous through the whole RG trajectory.
Point 3 is a generalized Maxwell construction. It reduces to the traditional Maxwell
construction for k = 0, to the degeneracy of U k=0(©) for j©j < ©min (0) = ©vac . The naive
Maxwell cut, applied for the concave part of the potential would give the degeneracy
for j©j < ©in° only. The problem with this argument is that it produces an e®ective
potential which is convex everywhere except at j©j = ©in° (0). The second derivative
of the potential is ill de¯ned and the ¯rst derivative is discontinuous at this point. By
placing the system into a ¯nite box the singularity is rounded o® and the second derivative
becomes ¯nite, but turns out to be negative in a vicinity of j©j = ©in° (0). Convexity
regained only if the cut is extended between the minima, j©j < ©min (0).
The generalized Maxwell construction in the mixed phase of a ¯rst order phase tran-
sition can be understood by the dynamics of the domain walls. The °atness of certain
thermodynamical potentials in the mixed phase re°ects the presence of zero-modes, the
location of domain walls. Such a rather simple kinematical mechanism which is inde-
pendent of microscopic details is the source of the `super universality’, point 4 above.
The role of the domain walls is played by saddle points in our computation, the cosine
function in (162) realizes in¯nitely many equally spaced, parallel domain walls. The inte-
gration over the zero modes ³ k and ek according to the rules of the saddle point expansion
restores translational and rotational symmetries of the ground state and reproduces the
mixed phase.
One might object our independent treatment of the plane wave saddle points. The
RG equation (163) handles the plane waves in the consecutive momentum space shells
independently which seems as a vast oversimpli¯cation. But one should recall at this
point that according to the general strategy of the RG method the dynamics of the modes
eliminated during the blocking is retained by the modi¯cation of the e®ective coupling
constants. This is not always possible since there are more modes than coupling constants,
the problem mentioned at the end of section 2.1.1. More precisely, the general framework
of the blocking what is in principle always applicable is turned into a powerful scheme
when an approximation is made. One truncates the e®ective action and assumes that the
solution of this over-determined problem exists. This is the main, and so far unproven,
assumption of the RG method. Accepting this point we can determine the blocked
action by means of a homogeneous background ¯eld and use it for non-homogeneous ¯eld
con¯gurations at the next blocking step as written in Eq. (163).
3.3.4 Correlation functions
One can compute the tree-level contributions to the correlation functions in a combination
of the mean-¯eld and the saddle point approximation [80]. Let us split the complete
functional integral into the sum of contributions which come from the unstable and
the stable regions. The °uctuations from the stable regions are governed by a non-
trivial action and they are taken into account as loop corrections. The °uctuations in
the unstable region experience a °at action and their contributions must be taken into
account on the tree-level. The complication of integrating up these contributions is the
determination of the region where the action is °at. Our approximation consists of
estimating this region for each plane wave independently, ie extending the integration
over the amplitude rp of the degenerate plane wave
¿ ~cl (x) = rp cos(pep x + ³ p ) (164)
over the interval ¡ » p < r < » p , where » p is given by (162). Let us denote the integration
over the resulting domain by D© [r] and write the correlation function in momentum space
in this single-mode determined °at region as
·Z ¸
¡1 Z
1
G©
tree (p; q) = D[³ ]D[e]D© [r] D[³ ]D[e]D © [r]rprq
4³ ´³ ´
£ eiµp ¯ p;kep + e¡iµp ¯ p;¡kep eiµq ¯ q;keq + e¡iµq ¯ q;¡keq : (165)
Due to the °atness of the action these integrals are purely kinematical and can easily
be carried out. All integration whose variable does not show up in the integrand drops
out. The integration over the shift of the plane waves, the phase angle ³ , restores the
translation invariance,
·Z Z ¸¡1 Z Z
G©
tree (p; q) = ¯ p+q;0 de drp de drp rp2
2(2º )dd (©min (p) ¡ ©)2
=¯ p+q;0 ; (166)
3 d ©¡1
min (©)
where ©¡1 ¡1
min (©) is de¯ned as ©min (©min (©)) = ©. The Fourier transform of this propaga-
¡1
tor describes a di®raction type oscillation with characteristic length scale ¹ = kcr , the
characteristic feature of domains in a homogeneous state.
3.3.5 Condensation as crossover
The scaling laws of the scalar model change already approaching the condensation in the
stable region C of Fig. 7. Inside the instable region D the action with the potential given
by point 3 of section 3.3.3 becomes invariant under blocking, ie the whole region D is an
IR ¯xed point.
The vacuum of the scalar model is a single, homogeneous coherent state consisting
of zero momentum particles when j©j > ©vac . When j©j < ©vac we encounter the sin-
gularities as seen above. This singularity may not be real, it might be smoothened out
by higher order vertices and a saddle points appearing for k < kcr . But the vacuum is
the superposition of inhomogeneous coherent states with characteristic scales ` > 1=kcr .
Each of them breaks space-time symmetries but their sum remains symmetrical.
One may wonder if any of the discussion applies to the true vacuum j©j = ©vac . For a
weakly coupled system the °uctuations are small and one may hope that the °uctuations
around ©vac are stable and un-in°uenced by what is happening with the unstable modes
¿ x < ©vac . But the answer to this question is rather uncertain and is open because the
true vacuum is just at the border of the instabilities and the typical °uctuations around
the true vacuum (¿ x ¡ ©vac ) º g 2 (0)=g 4 (0) penetrate into the unstable region.
The issue of the modi¯cation of the scale dependence by a condensate [64] is rather
general and points far beyond the simple scalar model.
The vacuum of asymptotically free models supports strong correlations at long dis-
tance. In the case of 4 dimensional Yang-Mills models the naive, perturbative vacuum is
destabilized by the one-loop level e®ective action which predicts that the vacuum is made
up by a coherent state of gluons [82]. But this state cannot be the true vacuum. The
problem is not only that the ¯eld generated by the condensate is strong and spoils the
saddle point expansion but it turns out that there is an unstable mode. The true vacuum
is supposed to be found at even lower energy densities where the long range °uctuations
restore the the external and color symmetry broken by a homogeneous condensate of the
charged spin one gluons [83] and the vacuum is thought to contain domains of homoge-
neous ¯eld in a stochastic manner [84]. This scenario is close to the view of the mixed
phase of the scalar model developed here with the di®erence that the instability comes
from the loop or the tree-level renormalized action in the Yang-Mills or scalar model,
respectively. Furthermore the instability in the Yang-Mills model can be avoided by ex-
treme environment only, by immersing the system into strong external ¯eld or bringing
into contact with heat or particle reservoir.
Similar instability occurs in vacua containing condensate of bound states. The BCS
vacuum is made homogeneous in a non-trivial manner when represented in terms of
the electrons making up the Cooper pairs. The spontaneous breakdown of the chiral
invariance in QCD manifests itself in the condensate of quark-anti quark pairs which is
homogeneous after integrating out the instanton zero modes only [85]. The crossover
from the U.V. scaling regime to the instability takes place at k º 1=` where ` is the size
of the bound states. Higher order derivative terms appearing as e®ective vertices may
generate similar crossover, as well [72], [86].
Finally it is worthwhile mentioning the tunneling phenomena, the dynamical extension
of the instabilities considered above. The interesting feature of the dynamical realization
of the tunneling by means of time-dependent, tree-level instabilities is that they take
place in conservative systems but the long time distribution agrees with the equilibrium
predictions coming from the canonical ensemble [81].
3.4 Sine-Gordon model
It is not unusual that one arrives at periodic or variables in the construction of e®ec-
tive theories, cf non-linear sigma models or non-Abelian gauge theories. Such variables
represent a double challenge. One problem which can be considered as local in the in-
ternal space is that the perturbation expansion around a minima of the potential should
keep in¯nitely many vertices in order to preserve the periodicity. Another, more di±cult,
global problem is a con°ict between the two requirements for the e®ective potential for
periodic variables, namely periodicity and convexity. One can ful¯ll both requirements
in a trivial manner only, by constant e®ective potential. If true, this conclusion has far
reaching consequence for the phenomenological description of such systems. As a case
study we shall consider the simplest model with periodic variable, the two dimensional
sine-Gordon model [87], [88].
3.4.1 Zoology of the sine-Gordon model
The sine-Gordon model

1
LSG = (@¹ ¿ )2 + u cos ¿ (167)
2
has been shown to be equivalent with the X-Y model [89], [90], [87], the Thirring model
[91], [92] and a Coulomb gas [93]. The methods of Refs. [90], [91] and [92], approximate
duality transformation, bosonization and semiclassical approximation are valid in certain
regions of the coupling constant space. The maps used in Refs. [87] and [93] are exact.
The mapping [87] Ã x = e¯Áx transforms the model (167) into the compacti¯ed sine-Gordon
model
1 u
LCSG = 2
@¹ Ã ¤ @¹ Ã + (Ã + Ã ¤ ): (168)
2 2
The models (167) and (168) are equivalent in any order of the perturbation expansion in
continuous space-time where the con¯gurations are assumed to be regular.
The X-Y model which is characterized by the action
2 3
14 X X X
SXY =¡ cos(³ x ¡ ³ x0) +h cos(³ x )5 ¡ ln z m2x ; (169)
T <x;x0> x x
where T = 2 , h=T = u, z is the vortex fugacity and mx denotes the vortex density. The
RG equation of the X-Y model obtained in the dilute vortex gas limit [89], [90]
µ ¶ µ ¶
dT 3 2 2 2 dh T dz º
a = 4º z ¡ º T h ; a = 2¡ h; a = 2¡ z: (170)
da da 4º dz T
The X-Y model appears as a generalization of the compacti¯ed sine-Gordon model
since the lattice regularization transforms the Lagrangian (168) into (169) with z = 1.
What is surprising here is that the vortex fugacity is a relevant operator in the high
temperature phase of the X-Y model but it is entirely missed by the lattice regularization!
Does that mean that the sine-Gordon model Lagrangians (167) and (168) are incomplete?
The vortex term could be missed by the regulator because the plane z = 0 is RG
invariant. Therefore it is consistent to exclude vortices from the path integral in the
sine-Gordon model given in continuous space-time. When we recast the model in lattice
regularization without ever thinking about vortices our lattice action neither suppresses,
nor enhances the singular con¯gurations and sets z = 1. But once the vortices are not
suppressed, z 6= 0 the renormalized trajectories moves away from from the plane of ¯xed
z. There is still no problem in the low temperature phase because all what happens is
that an irrelevant coupling constant became ¯xed. This is how regulators work. But the
problem is more serious in the high temperature phase. The Lagrangians (167) or (168)
can not give account of this phase because they lack a relevant, renormalizable coupling
constant. The perturbative continuum theory without vortices is consistent but once
vortices are available kinematically their density evolves in a non-trivial manner.
There is no problem with the lattice regulated X-Y model because the vortex fugacity
gets renormalized due to the cut-o® dependence of the vortex action. In other words, we
can move along the RG °ow by adjusting T and h only, the renormalization of z is carried
out ’automatically’ in the partition function. The renormalization of z is a problem in
the continuous formalism only where the vortices are introduced formally as point-like
charges.
This is an unexpected mechanism which brings the singular nature of the ¯eld con-
¯gurations into play during renormalization and may plague any quantum ¯eld theory.
The knowledge of the classical action in continuous space-time leaves open the possibility
that we have of adjust the fugacity of certain localized singularities or topological defects
not considered in the continuum.
It is interesting to speculate about similar phenomenon in QCD. Lattice QCD is con-
structed with ln z = 0 where z is the fugacity of some topological defects, like instantons,
monopoles, merons etc. These topological defects and singularities are supposed to play
an important role in the con¯ning vacuum, as vortices do in the high temperature phase
of the X-Y model. The renormalization of the fugacities of these objects should ¯rst be
studied in lattice QCD in order to construct a continuum description of the vacuum.
The sine-Gordon model possesses a topological current, j¹;x = ° ¹º @º ¿ x =2º which
is obviously conserved when the path integral is saturated by ¯eld con¯gurations with
analytic space-time dependence. Its °ux, the vorticity, gives the soliton number. The
world lines of the sine-Gordon solitons end at the X-Y model vortices, cf Fig. 11. The
distance between the bound vortex-anti vortex pairs shrinks with the lattice spacing in
the low temperature continuum limit. Any measurement with ¯nite resolution loose sight
of the instability of solitons in the renormalized theory. The average distance between
vortices stays ¯nite, cut-o® independent in the continuum limit of the high temperature
phase and the soliton decay can be observed. One expects the breakdown or at least
important modi¯cation of the bosonization transformation in this phase. In fact, the non-
conservation of the topological current requires fermion number non-conserving terms in
the fermionic representation, a fundamental violation of the rules inferred from the weak
coupling expansion.
Fig. 11 The world line of a soliton starts at a vortex (V) and ends at an anti-vortex (A) in the
two dimensional space-time.
3.4.2 E®ective potential
In order to understand better the dynamics of the long wavelength modes we compute the
e®ective potential in the sine-Gordon model. We follow the WH method truncated to the
local potential approximation which picks up the tree level evolution, too. It is obvious
that period length of the potential 2º = remains RG invariant in this approximation.
The tree ands the loop level RG equations are
" Z 1 #
k2 2
U k¡¢k (¿ ) = min » + duU k (¿ + » cos(º u)) ; (171)
½ 4 ¡1
and
k2 (2)
kU k¡¢k (¿ ) = kU k (¿ ) + ¢k ln[k 2 + U k (¿ )] (172)
4º
in the plane z = 0.
We use the Fourier expanded form for the potential
1
X
U k (¿ ) = un (k) cos (n ¿ ) (173)
n=0
and compute the leading order contribution to the WH equation when expanded in the
potential, Ã !
2 n2
k@k u
~n = ¡ 2 u~n ; (174)
4º
~n = un =k 2 . This agrees with the
in terms of the the dimensionless coupling constants u
second equation in (170). The solution of (174) is
Ã ! 2n2 ¡2
k 4º
u
~n (k) = u
~n (¤) : (175)
¤
The coupling constants un are irrelevant in the disordered phase T = 2 > 8º and
the e®ective potential obtained for k = 0 is °at. The coupling constants n < 8º =T are
relevant in the ordered phase and the e®ective potential is non-trivial. At this pointe one
suspects that Maxwell construction interferes with the evolution because a non-trivial
periodic functions necessarily has concave regions. To settle this question one has to
rely on the numerical solution of the evolution equation (171). We followed the loops-
generated evolution (172) from the initial condition set at k = ¤ until the stability is lost
at k = k cr for background ¯eld values which lie at local maxima of the periodic potential.
For k < kcr equation (171) was used. The result is that the coupling constants approach
zero as k ! 0 due to the plane wave saddle points. The e®ective potential is trivial in
both phases.
3.4.3 Breakdown of the fundamental group symmetry
The local potential U k (¿ ) °attens without developing singularities in the high temperature
disordered phase. The transformation
2º
¿ x !¿ x + (176)

is a discrete symmetry of the action and is preserved in the vacuum since the potential
barrier between the minima is vanishing for long range modes. Since we are at the lower
critical dimension, d = 2, the large amplitude long range modes realize the ’tunneling’
between the minima of the periodic potential. On the contrary to this situation, the
potential develops discontinuous second derivatives in the ordered, low temperature phase
and the instability driven °attening of the potential U k (¿ ) re°ects the survival of barriers
between the minima of the potential. All this looks like a spontaneous symmetry breaking,
therefore our conclusion is that the transformation (176) is not a symmetry of the vacuum
in the ordered, low temperature phase.
The usual circumstance under which such a phenomenon arises is the multiple con-
nectedness of the internal space. In the present context the non-linear U (1) ¼ -model
parametrization, Eq. (168), is based on the internal space U (1), with the fundamental
group Z generated by the transformation (176). The dynamical breakdown of the funda-
mental group symmetry is a genuine quantum e®ect. In fact, the path integral formally
extends over all homotopy classes, this is the symmetrical phase. When the dynamics de-
velops su±ciently high barriers between the homotopy classes the path integral becomes
restricted to a single homotopy class.
It is important to recall that the time evolution described by the Schrodinger equation
can be derived from the path integral by performing in¯nitesimal variations on the end
point of the trajectory. Therefore the consistency of the dynamics can be maintained
when the path integration is restricted into any functional subspace which is closed under
continuous deformation of the trajectories. The loss of the interference between homotopy
classes is the characteristic feature of the symmetry broken phase.
Such symmetry breaking is the key to understanding the way quarks become decon-
¯ned at high temperature or the droplet phase is formed for quantum liquids [94]. The
con¯guration space for global gauge transformations is SU (3)=Z3 for gluons, with the
fundamental group Z 3 . The quark propagator vanishes in the symmetrical phase due to
the destructive interference between the three homotopy classes. In the Z3 symmetry
broken decon¯ned phase there is no interference and quarks can be observed. Quantum
liquids in the ¯rst quantized formalism display similar symmetry breaking pattern. The
coordinate space for N particles is R 3N =SN where SN consists of permutations of the
particles. The absence of the overlap among states belonging to di®erent droplets sup-
presses the (anti)symmetrization of the states and reduces the exchange symmetry of the
ground state.
How can we recognize the dynamical breakdown of the fundamental group symmetry?
The simplest strategy, to look for the minima of the e®ective potential fails because
Maxwell construction hides any structure in a periodic potential. The answer to this
question lies in the topological structure of excitations and the order parameter will be
a topological susceptibility.
This is rather natural since the phase transition is the restriction of the path integral
into a single homotopy class, the imposition of a topological constraint. Let us assume
that the analysis sketched above for the plane z = 0 remains qualitatively valid for
z 6= 0, ie the e®ective potential is always °at, due to the loop-generated evolution or the
Maxwell construction in the disordered or ordered phase, respectively. The topological
invariant characterizing the homotopy classes is the soliton number. In the weak coupling
expansion (ordered phase) the path integral is constrained into a single homotopy class
hence the soliton number is conserved, its susceptibility is vanishing in the continuum
limit. The stability of the soliton, based on the continuity of the time evolution is lost in
the disorder phase because the short distance, large amplitude °uctuations of the typical
¯eld con¯guration extends the path integral all soliton number sector. Therefore the
susceptibility of the soliton number is non-vanishing in this phase.
3.4.4 Lower critical dimension
It is worthwhile noting the manner the Mermin-Wagner-Coleman theorem [95] appears

in the framework of the local potential approximation. By setting g2 = 0 for d = 2
the linear part of Pn in Eqs. (23) provides a ¯nite evolution but the higher order terms
make the renormalized trajectory to diverge as k ! 0 unless the coupling constants
approach zero. We can construct interactive, massless models in two dimensions so long
the running coupling constants approach zero su±ciently fast in the IR regime as in the
case of the sine-Gordon model, presented above. The retaining of the terms O (@ 2 ) in
the gradient expansion provides another mechanism to suppress the IR divergences by
generating singular wave function renormalization constant, Zk as k ! 0 [96].
In both cases the way to avoid the non-interactive system predicted by the Mermin-
Wagner-Coleman theorem is to use the RG improved perturbation expansion. In the
ordinary perturbation series we have the contributions like
Z
n
g fp1;¢¢¢;pk (177)
p1;¢¢¢;pk
and the integral is IR divergent in d = 2. One has running coupling constants, g ! gP ,

in the RG improved perturbation expansion where the analogous contribution is
Z
g Pn (p1;¢¢¢;pk)fp1;¢¢¢;pk : (178)
p1;¢¢¢;pk
When the IR ¯xed point is Gaussian, limP !0 g P = 0, then the theory remains interactive
and the running coupling constants suppress the IR divergences.
3.5 Gauge models

The blocking step of the RG scheme, being scale dependent, violates local symmetries
which mix modes with di®erent length scales.
The most natural way to deal with this breakdown of gauge invariance is to cancel
the gauge non-invariant terms by ¯ne tuning non-invariant counterterms [97]. But this
is possible in perturbation expansion only. Another, better suited strategy to the RG
method is to argue that the way Ward identities are violated by blocking shows that
the BRST invariance is broken by the running cut-o® only and the BRST symmetry is
recovered at the IR ¯xed point when the cut-o® is removed [20], [98]. This were certainly
a good procedure if it could be implemented without further truncation. In a weakly
coupled model such as QED one could construct approximations with controllable errors.
In asymptotically free models like QCD non-perturbative long range correlations make
any scheme which imposes gauge invariance in an approximate manner unreliable [63].
The demonstrate this point consider the static force law between two test charges in a
Yang-Mills theory in a scheme where the gauge invariance, Gauss’ law is implemented
approximately. The gauge non-invariant components of the vacuum state appear charged,
by de¯nition. Thus we have an uncontrollable color charge distribution around the test
charges. When the distance between the test charges are large enough then it will be
energetically favorable for the uncontrollable charges to break the °ux tube. We loose
the string tension in a manner similar how it happens in the true QCD vacuum due to
the virtual quark-anti quark polarizations. In more formal words, an arbitrarily weak
gauge-dependent perturbation can change the long range features of the vacuum as in
the ferromagnetic phase of the Ising model.
The most natural way to guarantee gauge invariance is to achieve independence on the
choice of gauge by using gauge invariant quantities only. The proposal of Ref. [99] goes
along this line but the use of loop variables renders the computation rather involved.
We discuss now a version of the RG scheme in the internal space which at least for
Abelian models produces the e®ective action without any gauge ¯xing. The electron
mass combined with chiral transformation has already been used in generating evolution
equation in Ref. [100]. We shall follow a simpler and more general by avoiding gauge
dependence in an explicit manner for the photon Green functions [101]. The gauge
invariant electron composite operators are not di±cult to include [41].
3.5.1 Evolution equation
Let us consider the generator functional

Z R
1 ¹ 1
[¡ 4e12 F · ¸ ¹ =¡m)Ã+¹
¡ ¬ 2 (@· A· )2 +Ã(iD ¹ · ¢A
e · W [j;j;J]
h = ¹
D[Ã]D[Ã]D[A]e h
· x
F· ¸ j¢Ã+Ã¢j+J · ]
; (179)
where D¹ = @¹ + iA¹ and the dimensional regularization is used to render W ¯nite. We

control the quantum °uctuations by modifying the action, S ! S + S¸ with
Z
¶
S¸ = 2 A ¹;x h T ¹º Aº (x); (180)
4e x
where T ¹º = g ¹º ¡ @¹ @º =h . The in¯nitesimal change of the control parameter ¶ ! ¶ +¢¶

modi¯es the free photon propagator
Z ³ ´
¹º ¹º
D (x ¡ y) ! D (x ¡ y) + ¢¶ dzD ¹½ (x ¡ z)h z T½· D
·º
(z ¡ y) + O (¢¶ )2 ; (181)
in a manner reminiscent of the Callan-Symanzik scheme.

The evolution equation (108) with the present suppression takes the form
2 Ã !¡1 3
2~
h
¹ ¯ ¡[A]
@¸ ¡¸ [A] = 2
tr 4h T ¹º 5 (182)
2e ¯ A ¹ ¯ Aº
~ ¸ [A] = ¡¸ [A] + S¸ [A]. We project the evolution

in terms of the photon e®ective action ¡
equation on the functional space given by the ansatz
Z
~ 1
¡[A] = A¹;x D ¡1 (i@)T ¹º A º;x + C [A] (183)
2 x
where
1+¶
D ¡1¹º = ¡ h T
¹º
¡ ¬ h L¹º ; (184)
e2
L¹º = ¯ ¹º ¡ T ¹º and C [A] is a gauge invariant functional.
The control parameter ¶ ’turns on’ the °uctuations of the photon ¯eld. Therefore the
electrons loop contributions to the e®ective action must already be present at the initial
condition which is chosen to be C [A] = ¡ itr ln(iD= ¡ m) at ¶ = ¶ 0 .
The second functional derivative matrix is written as
~
¯ 2 ¡[A] ¯ 2 C [A]
= D ¡1¹º + (185)
¯ A¹ ¯ A º ¯ A¹¯ Aº
and the inversion is carried out by expanding in the non-diagonal pieces to write the
evolution equation as
" 1
Ã !n #
h
¹ ¹º
X
n ¯ 2 C [A]
@¸ C [A] = 2 tr h T D (¡ 1) D : (186)
2e n=0 ¯ A¹ ¯ A º
3.5.2 Gauge invariance
We show now that the limit ¬ ! 0 can be taken in the evolution equation without hitting
any singularity. The gauge ¯xing parameter ¬ enters through the photon propagator
D in the Neuman-expansion of the right hand side of Eq. (186). The ¬ -dependent
longitudinal contributions of the ¯rst and the last D factor are suppressed by the gauge
invariance of the suppression term, represented by the transverse projection T ¹º in (186).
The longitudinal photon contributions of the internal propagators are suppressed by the
gauge invariance of the e®ective action,
¯ C [A]
@¹ = 0: (187)
¯ A¹
According to this equation ¯ 2 C [A]=¯ A ¹ ¯ A º does not mix the longitudinal and the trans-
verse modes, ie non-longitudinal contributions appear in the evolution equation.
We choose a gauge in the argument above what was relaxed at the end of the compu-
tation. But the steps followed remain well de¯ned even if we start with ¬ = 0, without
any gauge ¯xing. Our argument about the decoupling of the longitudinal and transverse
contributions to the evolution equation still applies but it is not clear if the longitudinal
part was well de¯ned. This subtle issue is settled by Feynman’s ° parameter, devised to
lift in¯nitesimally the degeneracies of the action. It is easy to see that it gives a weak
variation of the action along gauge orbits by breaking gauge invariance. As long as our
truncation of the evolution equation, the functional C [A] is explicitly gauge invariant °
plays the role of an in¯nitesimally external magnetic ¯eld in the Ising model, ie helps the
breakdown of the gauge symmetry only if it is really broken in the true vacuum.
4 What has been achieved
The traditional implementation of the RG idea has proven to be essential in Statistical and
High Energy Physics, starting with the understanding of critical phenomena [102], ¯nite
size scaling [103], ° -expansion [104], dynamical processes with di®erent time scales [105],
continuing with partial resummation of the perturbation expansion [4], parametrizing
the scale dependence at high energies [106] and ending with the construction of e®ective
theories in Particle Physics [107, 108, 109]. This is an inexcusable short and incomplete
list, its role is to demonstrate variability and the importance of the method only. Our
main concern here was the functional form of the RG method and its generalizations.
From this point of view one may distinguish conceptual and more technical achievements
as the power of the functional formalism is more exploited.
We should consider the RG method as ’meta-theory’, or in more practical terms
as a language, as Feynman graphs are used in particle physics. But this language is not
bound by small parameters and can provide us a general, non-perturbative approximation
method beyond the semiclassical expansion and numerical simulations.
The path integration was ¯rst viewed as a powerful book-keeping device for perturba-
tion expansion and the truly non-perturbative application came later, after having gained
some experience with the formalism. This is similar to the development of the functional
RG method where the RG equation refers to the generator function of the e®ective cou-
pling constants and provides us with a simple procedure to keep track of the Feynman
graphs. The steps in functional calculus are more cumbersome because they deal with
generator function(al)s, with in¯nitely many coupling constants. The ultimate goal is to
go beyond this level and to use this formalism in a genuinely non-perturbative manner.
The semiclassical expansion is our ¯rst step in this direction. It will be important to
check how the RG continues to be applicable where the saddle point expansion ceases to
be reliable.
Another ¯eld the RG scheme might be compared with is lattice regulated ¯eld theory.
Both are general purpose tools to deal with non-perturbative systems. The bottle-neck of
the numerical simulations on the lattice is the need to send the U.V. and the IR cut-o®s
su±ciently far from each others and the restriction to Euclidean space-time. The RG
strategy is set up in continuous, Minkowski space time and there is no particular problem
with keeping the U.V. and IR cut-o®s far from each other. But the drawback of the RG
strategy is that it is rather lengthy to extend the space of (e®ective) action functionals
used in the computation. Since the limitations of the two methods are quite di®erent
they might be used in a complementary manner.
The functional formalism is promising because of the possibility of following the mix-
ing of a much larger number of operators as in the traditional strategy. This feature gives
the hope of extending the applicability of the method from a single scaling regime to the
whole range of scales covered by the theory. Such an extension may provide us valuable
information about the competition of interactions in realistic theories.
Acknowledgments
I thank Jean Alexandre and Kornel Sailer for several useful discussions during our col-
laboration.
References
[1] N.N. Bogoliubov, D.V. Shirkov, Introduction to the Theory of Quantized Fields, 3rd
ed., John Wiley and Sons, New York, 1973.
[2] K.G. Wilson, J. Kogut, Phys. Rep. C 12, 77 (1974); K.G. Wilson, Rev. Mod. Phys.
47, 773 (1975); Rev. Mod. Phys. 55, 583 (1983)
[3] P.W. Anderson, G.J. Yuval, J. Phys. C 4, 607 (1971).
[4] M. Gellman, F. Low, Phys. Rev. 95, 1300 (1954).
[5] N. Goldenfeld, O. Martin, V. Oono, and F. Liu, Phys. Rev. Lett. 64, 1361 (1990); T.
Kunihiro, Progr. of Theor. Phys. 95, 503 (1995).
[6] C. G. Callan, Phys. Rev. D 2, 1541 (1970); K. Symanzik, Comm. Math. Phys. 18,
227 (1970).
[7] J. Alexandre, V. Branchina, J. Polonyi, Phys. Rev. D 58, 16002 (1988).
[8] F.J.Wegner, A.Houghton, Phys. Rev. A 8, 401 (1973).
[9] J. Polchinski, Nucl. Phys. B 231, 269 (1984).
[10] T. Morris, Int. J. Mod. Phys. A9 (1994) 2411.
[11] J. Collins, Renormalization, Cambridge University Press, 1984.
[12] J. Polonyi, K. Sailer, Phys. Rev. D 63, 105006 (2001).
[13] J. Hughes and J. Liu, Nucl. Phys. B 307, 183 (1988); J Hughes, Nucl. Phys. B 312,
125 (1989); G.M. Shore, Nucl. Phys. B 362, 85 (1991); G. Keller and C. Kopper,
Comm. Math. Phys. 148, 445 (1992); Comm. Math. Phys. 153, 245 (1993); C. Becchi,
On the Construction of Renormalized Gauge Theories Using Renormalization Group
Techniques, hep-th/9607188.
[14] B.P. Dolan, Int. J. of Mod. Phys. 10, 2439 (1995); Int. J. of Mod. Phys. 10, 2703
(1995);Int. J. of Mod. Phys. 12, 2413 (1995); D.C. Brody, A. Ritz, Nucl. Phys. B
522, 588 (1998); B.P. Dolan, A. Lewis, Phys. Lett. B 460, 302 (1999).
[15] G. Altarelli, G. Parisi, Nucl. Phys. B 126, 278 (1977); A. H. Mueller, Phys. Rep. 73,
237 (1981).
[16] C. Wetterich, Phys. Lett. B 301, 90 (1993).
[17] T. Morris, Phys. Lett. B 329, 241 (1944).
[18] D. Litim, Optimized renormalization group °ow, hep-th/0103195.
[19] M. Reuter and C. Wetterich, Nucl. Phys. B 391, 147 (1993); N. Tetradis and D.F.
Litim, Nucl. Phys. B 464, 492 (1996).
[20] U. Ellwanger, Phys. Lett. B 335, 364 (1994).
[21] J. Adams, J. Berges, S. Bornholdt, F. Freire, N. Tetradis, C. Wetterich, Mod. Phys.
Lett. A 10, 2367 (1995).
[22] J. Alexandre, J. Polonyi, Ann. Phys. 288, 37 (2001).
[23] S. Coleman, E. Weinberg, Phys. Rev. D 7, 1888 (1973).
[24] A. Parola, L. Reatto, Phys. Rev. Lett. 53, 2417 (1984); Phys. Rev. A 31, 3309 (1985);
V. I. Tikar, Phys. Lett. A 104, 135 (1984); J. Comellas, A. Travesset, Nucl. Phys. B
498, 539 (1997).
[25] K. I. Aoki et al, Progr. of Theor. Phys. 95, 409 (1996).
[26] S. B. Liao, M. Strickland, Nucl. Phys. B 497, 611 (1997).
¶
[27] A. Margaritis, G. Odor, A. Patkos, Z. Phys. C 39, 109 (1998)
[28] P. E. Haagensen, Y. Kubyshin, J. I. Latorre, E. Moreno, Phys. Lett. B 323, 330
(1994)
[30] N. Tetradis, C. Wetterich, Nucl. Phys. B 422, 541 (1994).
[31] M. Alford, Phys. Lett. B 336, 237 (1994).
[32] S. Liao, J. Polonyi, M. Strickland, Nucl. Phys. B 567, 493 (2000).
[33] D. F. Litim, Phys. Lett. B 486, 92 (2000); Int. J. of Mod. Phys. 16, 2081 (2001).
[34] S. Seide, C. Wetterich, Nucl. Phys. B 562, 524 (1999).
[35] J. Berges, C. Wetterich, Nucl. Phys. B 487, 675 (1997).
[36] M. Maggiore, Z. Phys. C 41, 687 (1989); T. E. Clarke, B. Haeri, S. T. Love, Nucl.
Phys. B 402, 628 (1993); M. D’Attanasio, T. Morris, Phys. Lett. B 409, 363 (1997);
J. Berges, D. U. Jungnickel, C. Wetterich, Phys. Rev. D 59, 034010 (1999);
[37] D. Zanchi, H. J. Schultz, Phys. Rev. B 54, 9509 (1996); Phys. Rev. B 61, 13609
(2000); C. J. Halboth, W. Metzner, Phys. Rev. B 61, 4364 (2000); C. Honerkamp,
M. Salmhofer, N. Furukawa, T. M. Rrice, Phys. Rev. B 63][45114, 2001 (;) M.
Salmhofer. C. Honerkamp, Progr. of Theor. Phys. 105, 1 (2001); S. Correia, J.
Polonyi, J. Richert, Functional Callan-Symanzik Equation for the Coulomb Gas,
cond-mat/0108048.
[38] J. Polonyi, Current-density functional for disordered systems, cond-mat/0203090.

[39] A. Bonanno, M. Reuter, Phys. Rev. D 62, 043008 (2000).
[40] H. Gies, C. Wetterich, Renormalization group °ow of bound states, hep-th/0107221.
[41] J. Polonyi, K. Sailer, E®ective actions for the density functional theory, cond-
mat/0108179.
[42] C. Bagnuls, C. Bervillier, hep-th/0002034; Berges, N. Tetradis, C. Wetterich, hep-
th/0005122.
[43] Proceedings of Workshop on the Exact Renormalization Group, Faro, Portugal, 1998,
ed. Krasnitz et al. (World Scienti¯c,Singapore, 1999); 2nd Conference on Exact RG,
Rome 2000, ed. S. Arnone, at al. Int. J. of Mod. Phys. 16, 1809 (2001).
[44] C. Harvey-Fros, PhD Thesis, university of Southampton, hep-th/0108018; M. D.
Turner, PhD Thesis, University of Southampton, hep-th/0108031.
[45] F. Wegner, J. Phys. C 7, 2098 (1974).
[46] J. I. Latorre, T. Morris, Int. J. of Mod. Phys. 16, 2071 (2001).
[47] L. Schulmann, Techniques and Applications of Path Integrations, John Wiley and
Sons, New York, 1981.
[48] D. McLaughlin, L. Schulman, J. Math. Phys. 12, 2520 (1971).
[49] J. Polonyi, Ann. Phys. 252, 300 (1996).
[50] R. D. Ball, P. E. Haagensen, J. I. Latorre, E. Moreno, Phys. Lett. B 347, 80 (1995);
[51] J. Comellas, Phys. Lett. B 509, 662 (1998).
[52] T. Morris, Nucl. Phys. B 495, 477 (1997); Phys. Rev. Lett. 77, 1658 (1996).
[53] A. Hasenfratz, P. Hasenfratz, Nucl. Phys. B 270, 687 (1986).
[54] K. Halpern. K. Huang, Phys. Rev. D 35, 3252 (1996); Phys. Rev. Lett. 77, 1659
(1996); V. Periwal, Mod. Phys. Lett. A 11, 2915 (1996); K. Halpern, Phys. Rev. D
57, 6337 (1998).
[55] G. Felder, Comm. Math. Phys. 111, 101 (1987); A. E. Filippov, S. A. Breus, Phys.
Lett. A 158, 300 (1981); S. A. Breus, A. E. Filippov, Physica A 192, 486 (1993); T.
Morris, Phys. Rev. Lett. 1658, 1996 (.)
[57] C. Bagnulus, C. Bervillier, Phys. Rev. B 41, 402 (1990).
[58] T. Morris, Phys. Lett. B 345, 139 (1995);
[59] R. Jackiw, S. Templeton, Phys. Rev. D 23, 2291 (1981).
[60] J. Polchinski, in Proceedings of the 1992 TASI Elementary Particle Physics, Eds. J.
Polchinski and J. Harvey, World Sci., 1992.
[61] S. Weinberg, Nucl. Phys. B 413, 567 (1994).
[62] R. Shankar, Rev. Mod. Phys. 66, 129 (1994).
[63] J. Polonyi, Nucl. Phys. (Proc. Suppl.) B20 (1991) 32; Con¯nement and
Renormalization, hep-ph/9511243, in the Proceedings of the International School of
Physics, Enrico Fermi, Course CXXX, on Selected Topics in Nonperturbative QCD,

Soc. It. di Fisica; K. Johnson, L. Lellouch, J. Polonyi, Nucl. Phys. B 367, 675 (1991).
[64] V. A. Miransky, Nuovo Cimento A 90, 149 (1985); P.I. Fomin, V. P. Gusynin, V. A.
Miransky, Yu. A. Sitenko, Riv. Nuovo Cimento 6, 1 (1983); C. N. Leung, S. T. Love,
W. A. Bardeen, Nucl. Phys. B 273, 649 (1986); W. A. Bardeen, S. T. Love, C. N.
Leung, Nucl. Phys. B 323, 493 (1989).
[65] T. Appelquist, T. Takeuchi, M. B. Einhorn, L. C. R. Wijewardhana, Phys. Lett. B
220, 223 (1989); K. Matumoto, Progr. of Theor. Phys. 81, 277 (1989); T. Takeuchi,
Phys. Rev. D 40, 2697 (1989).
[66] V. A. Miransky, M. Tanabashi, K. Yamawaki, Phys. Lett. B 221, 177 (1989); Mod.
Phys. Lett. A 4, 1043 (1989).
[67] V. A. Miransky, Int. J. of Mod. Phys. 6, 1641 (1991).
[68] Y. Nambu, Proceedings of the 1989 Workshop on Dynamical Symmetry Breaking,
eds. T. Muta, K. Yamawaki (Nagoya University, 1990); W. J. marciano, Phys. Rev.
Lett. 62, 2793 (1989);Phys. Rev. D 41, 1647 (1990); W. A. Bardeen, C. T. Hill, M.
Lindner, Phys. Rev. D 41, 1647 (1990).
[69] V. A. Alessandrini, H. de Vega, F. Schaposnik, Phys. Rev. B 10, 3906 (1974); Phys.
Rev. B 12, 5034 (1975); T. Appelquist, J. Carazzone, Phys. Rev. D 11, 2865 (1975);
T. Appelquist, J. Carazzone, H. Kluberg-Stern, M. Roth, Phys. Rev. Lett. 36, 768
(1976), Phys. Rev. Lett. 36, 1161 (1976); T. Appelquist, J. Carazzone, Nucl. Phys.
B 120, 77 (1977).
[70] J. S. Bell, R. Jackiw, Nuovo Cimento A 60, 47 (1969); S. Adler, Phys. Rev. 177,
2426 (1969).
[71] S. Adler, J. Collins, A. Duncan, Phys. Rev. D 15, 1712 (1977); J. Collins, A. Duncan,
S. Joeglekar, Phys. Rev. D 16, 438 (1977); N. K. Nielsen, Nucl. Phys. B 120, 212
(1977).
[72] V. Branchina, J. Polonyi, Nucl. Phys. B 433, 99 (1995).
[73] M. E. Fischer, in Proceedings of the Conference Renormalization Group in Critical
Phenomena and Quantum Field Theory, edt. J. D. Gunton, M. S. Green, Temple
University, 1973.
[74] J. Alexandre, V. Branchina, J. Polonyi, Phys. Lett. B 445, 351 (1999).
[75] A. Ringwald, C. Wetterich, Nucl. Phys. B 334, 506 (1990); N. Tetradis, C. Wetterich,
Nucl. Phys. B 383, 197 (1992).
[76] R. B. Gri±th, P. A. Pierce, Phys. Rev. Lett. 41, 917 (1978); J. Stat. Physics. 20, 499
(1979); O. Hud¶ak, Phys. Lett. A 73, 273 (1979); C. Rebbi, R. H. Swendsen, Phys.
Rev. B 21, 4094 (1980); R. B. Gri±th, Physica A 106, 59 (1981); A. Stella, Physica
A 108, 211 (1981).
[77] A. Hasenfratz, P. Hasenfratz, Nucl. Phys. B 295, 1 (1988)
[78] H. W. BlÄote, R. H. Swendsen, Phys. Rev. Lett. 43, 799 (1979); C. Lang, Nucl. Phys.
B 280, 255 (1987); K. Decker, A. Hasenfratz, P. Hasenfratz, Nucl. Phys. B 295, 21
(1988); A. Gonzales-Arroyo, M. Okawa, Y. Shimizu, Phys. Rev. Lett. 60, 487 (1988);
F. Igloi, C. Vanderzande, Physica A 135, 347 (1986).
[79] R. B. Israel, in Random Fields (Esztergom, 1979), J. Fritz, J. L. Lebowitz, D. Sz¶asz

eds. (North-Holland, Amsterdam, 1981); A. C. D. van Enter, R. Fern¶andez, A. Sokal,
J. Stat. Physics. 72, 879 (1993).
[80] J. Alexandre, J. Polonyi, Tree-level renormalization, hep-th/9906017.
[81] Sz. Borsanyi, A. Patkos, J. Polonyi, Zs. Szep, Phys. Rev. D 62, 085013 (2000).
[82] S. Matinyan, G. Savvidy,Nucl. Phys. B 134, 539 (1978); G. Savvidy, Phys. Lett. B
71, 133 (1977); M. Du®, R. Ram¶on-Medrano, Phys. Rev. D 12, 3357 (1976); H.
Pagels, E. Tomboulis, Nucl. Phys. B 143, 485 (1978).
[83] N. K. Nielsen, P. Olesen, Nucl. Phys. B 144, 376 (1978); Phys. Lett. B 79, 304 (1978);
J. Amjorn, N. K. Nielsen, P. Olesen, Nucl. Phys. B 152, 75 (1979); H. B. Nielsen, P.
Olesen, Nucl. Phys. B 160, 380 (1979); H. B. Nielsen, M. Ninomiya, Nucl. Phys. B
156, 1 (1979); J. Amjorn, P. Olesen, Nucl. Phys. B 60, 1980 (.)
[84] G. Dosch, Phys. Lett. B 190, 555 (1987); G. Dosch, Y. Simonov, Phys. Lett. B 205,
339 (1988); Y. Simonov, Yad. Fiz. 48, 1381 (1988); Yad. Fiz. 50, 213 (1988).
[85] G. ’tHooft, Phys. Rev. Lett. 37, 8 (1976).
[86] J. Fingberg, J. Polonyi, Nucl. Phys. B 486, 315 (1997); V. Branchina and H.
Mohrbach, J. Polonyi, Phys. Rev. D 60, 45006 (1999); Phys. Rev. D 60, 45007
(1999); M. Dufour Fournier, J. Polonyi, Phys. Rev. D 61, 065008 (2000).
[87] K. Huang, J. Polonyi, Int. J. of Mod. Phys. 6, 409 (1991).
[88] I. N¶andori, J. Polonyi, K. Sailer, Phys. Rev. D 63, 045022 (2001).
[89] J. M. Kosterlitz, D. J. Thouless, JournalJ. Phys. CC61181973; J. M. Kosterlitz,
JournalJ. Phys. CC710461974.
[90] J. V. Jose, L. P. Kadano®, S. Kirkpatrick, D. R. Nelson, Phys. Rev. B 16, 1217
(1977).
[91] S. Coleman, Phys. Rev. D 11, 3424 (1975).
[92] S. Mandelstam, Phys. Rev. D 11, 3026 (1975).
[93] S. Samuel, Phys. Rev. D 18, 1916 (1978).
[94] J. Polonyi, The con¯nement-decon¯nement mechanism, in Quark Gluon Plasma,
World Scienti¯c, 1990, R. Hwa, edt.
[95] N. D. Mermin, H. Wagner, Phys. Rev. Lett. 17, 1133 (1966); S. Coleman, Comm.
Math. Phys. 31, 259 (1973).
[96] G. V. Gersdor®, C. Wetterich, Nonperturbative Renormalization Flow and Essential
Scaling for the Kosterlitz-Thouless Transition, hep-th/0008114.
[97] C. Becchi, in Elementary Particles, Field Theory and Statistical Mechanics, Eds.
M. Bonini, G. Marchesini, E. Onofri, Parma University 1993, hep-th/9607188; M.
Bonini, M. D’Attanasio, G. Marchesini, Nucl. Phys. B 409, 1993 (441); Nucl. Phys.
B 418, 1994 (81); Nucl. Phys. B 421, 1994 (429); Nucl. Phys. B 437, 1995 (163); M.
Reuter, C. Wetterich, Nucl. Phys. B 417, 181 (1994); Nucl. Phys. B 427, 291 (1994);
M. D’Attanasio, T. R. Morris, Phys. Lett. B 378, 213 (1996).
[98] D. F. Litim, J. M. Pawlowski, Phys. Lett. B 435, 181 (1998); F. Freire, D. F. Litim,
J. M. Pawlowski,Phys. Lett. B 495, 256 (2000).
[99] T. R. Morris, Nucl. Phys. B 573, 97 (2000), J. High of Energy Phys. 0012, 012
(2000).
[100] M. Simionato, Int. J. of Mod. Phys. 15, 2121 (2000); Int. J. of Mod. Phys. 15, 2153
(2000).
[101] J.Alexandre, J. Polonyi, K. Sailer, in preparation.
[102] K. Wilson, Rev. Mod. Phys. 55, 583 (1983).
[103] M. E. Fischer, in the Proceedings of Enrico Fermi International School of Physics,
ed. M. S. Green (Academic Press, New York, 1971); M. P. Nightingale, Physica A
83, 561 (1976).
[104] K. Wilson, M. E. Fischer, Phys. Rev. Lett. 28, 248 (1972); K. Wilson, Phys. Rev.
D 7, 2911 (1973).
[105] P. C. Hohenberg, B. I. Halperin, Rev. Mod. Phys. 49, 435 (1977).
[106] G. ’tHooft, Proceedings of Colloquim on Renormalization of Yang-Mills Fields and
Application to Particle Physics, ed. C. P. Korthes-Altes, (Marseilles, 1972); D. Gross,
F. Wilcek, Phys. Rev. Lett. 30, 1343 (1973); H. D. Politzer, Phys. Rev. Lett. 30, 1346
(1973); S. Coleman, D. Gross, Phys. Rev. Lett. 31, 851 (1973).
[107] S. Weinberg, Physica A 96, 327 (1979); J. Gasser, H. Leutwyler, Ann. Phys. 158,
142 (1984); Nucl. Phys. B 250, 465 (1985); Phys. Rep. 87, 77 (1982); D. B. Kaplan,
A. V. Manohar, Phys. Rev. Lett. 56, 2004 (1986).
[108] S. Weinberg, Phys. Lett. B 91, 51 (1980) I. J. R. Aitchison, Acta Phys. Polonica B
18, 191 (1987).
[109] N. Isgur, M. B. Wise, Phys. Lett. B 232, 113 (1989); Phys. Lett. B 237, 527 (1990).

Lectures On The Functional Renormalization Group Method: Janos Polonyi

Uploaded by

Copyright:

Available Formats

Lectures On The Functional Renormalization Group Method: Janos Polonyi

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lectures On The Functional Renormalization Group Method: Janos Polonyi

Uploaded by

Copyright:

Available Formats

CEJP 1 (2003) 1{71

Lectures on the functional renormalization

Received 20 June 2002; revised 7 December 2002

Abstract: These introductory notes are about functional renormalization group

Keywords: strongly coupled theories, non-perturbative approximation, universality

2.1 Resumming the loop expansion

2.1.1 Blocking in continuous space-time

where ¿ 2 Fk¡¢k and ¿ ~ 2 Fk nFk¡¢k .

¿ x = ©, ¿ ~cl = 0 with the ansatz

2.1.2 Local potential approximation

In order to relate the WH equation to conventional perturbation expansion we shall use

and write the beta functions corresponding to ¿ 0 = h¿ i as

It is easy to see that ­ n is an n-th order polynomial

of the expression G n = gn =(k 2 + g2 ), eg

where G~ n = g~n =(1 + g~2 ).

Fig. 4 A two-loop contribution to g2 .

2.1.3 Gradient expansion

of the action are

where Z (n) = Z (n) (©), U (n) = U (n) (©) and K s = K (s2 ).

including an integration over the direction of k, the unit vector e.

c.f. Eq. (14). The terms O (² 2 r 2 ) give

2.2 Resumming the perturbation expansion

2.2.1 Polchinski equation

Let us start with the partition function

(3) Integrate out ¿ ~.

2.2.2 Gradient expansion

2.3 Composite operator renormalization

2.3.1 Toy model

Consider the two dimensional integral

in terms of bare coupling constants g n . We use h

giving rise the composite operators

It is instructive to generalize the toy model for three variables,

where the operator mixing looks like

2.3.2 Quantum Field Theory

The Kadano®-Wilson blocking

for the action X

generates the RG °ow

in agreement with Eq. (68). They satisfy the equation

in a manner similar to Eq. (69). We introduced here the sensitivity matrix

It is the measure of the sensitivity of an e®ective strength of interaction on the initial

The scale dependence of the operator mixing matrix is governed by

2.3.3 Parallel transport

2.3.4 Asymptotical scaling

We linearize the beta-functions around a ¯xed point G ¤m~ ,

and write the scaling coupling constant, the left eigenvectors of ¡

They display the scale dependence

are obtained by means of the right eigenvectors of ¡¤ ,

obviously follow (90). The operator

written at scale k in this basis yields the parallel transport trajectory

fOgk 0 = bn~ ffO nsc

in the vicinity of the ¯xed point.

2.3.5 Perturbative treatment

which agrees with (98).

2.4 Continuous evolution

where the term Ck [¿ ] is introduced to suppress °uctuations. What we require is that

their dynamics by the introduction/modi¯cation of e®ective vertices generates a ’univer-

2.5 Blocking in the internal space

We present now a generalization the RG method. The traditional RG strategy is aiming

con¯guration ¿ x realizes a map ¿ : external space ! internal space. The traditional

It is easy to see that n is an n-th order polynomial