LECTURE 2: RENORMALIZATION GROUPS (CONTINUED)

David Gross

Notes by P.Etingof and D.Kazhdan

2.1. Finite renormalization.
As we saw in the previous lecture and in lectures of Witten and Gawedzki,
renormalizable quantum eld theories always come in families (which may even be
innitely-parametric in 2 dimensions), and we need to specify certain quantities in
order to pinpoint a specic theory among such a family. The only ways to do this
are to measure these quantities or to set their values articially. But when we know
them, we can (in principle) compute all other quantities in the theory as functions
of them.
The procedure of picking a theory by setting its parameters to the results of
measurements (or articially) is called nite renormalization, as opposed to \in-
nite" renormalization which we discussed above. While innite renormalization
deals with removing divergences (innities), nite renormalization adjusts the -
nite part of the coecients in the Lagrangian (as functions of the cuto ), which
innite renormalization leaves ambiguous.
Thus, a prescription of nite renormalization is a way to dene a system of
coordinates (or a parametrization) of the space of renormalizable theories of a
given type.
It is clear that the number of quantities to be specied (i.e. the dimension of
the space of renormalizable theories) equals to the number of coecients of the
Lagrangian of nonnegative dimension. However, these coecients themselves are,
in general, not suitable as coordinates on the space of theories, as they depend on
the cuto  and are innite at  ! 1. Therefore, people use other systems of
coordinates.
For example, for the space Y of renormalizable theories of a scalar eld, we could
use the following parametrizations:
1. T ! (G2(0); @@pG (0); G4 (0; 0; 0)), where Gi are connected correlation functions.
2
2

2. T ! (G2(2 ); @@pG (2 ); G4(p1 ; p2 ; p3)), where  2 M is some scale, and

2
2

p1; p2 ; p3 ; p4 = p1 p2 p3 are vertices of a regular tetrahedron such that p2i = 2.

Notice that the the rst recipe denes one xed parametrization of the space of
theories, while the second one is in fact a family of such parametrizations, depending
on a scale  2 M . In the future, such families of parametrizations will be more
important to us than xed parametrizations.
Remark. The coordinates in both of the above parametrizations are not mea-
surable quantities, since momenta of all particles which can be generated in a lab
satisfy the condition p2 = m2, where m is the mass of the particle (people say
Typeset by AMS-TEX
that \p is on the mass shell"). In the case of 4 theory, it is possible to make up
a system of coordinates purely out of measurable quantities. (Of course, here we
make the (invalid) assumption that the 4 theory describes a real physical process.)
Let us explain how to do this.
Suppose that we have an elementary particle  which obeys a renormalizable 4-
theory. We have seen that renormalizable 4-theories live in a 3-parametric family
Y . For T 2 Y , let T be the residue of the 2-point function of T (in momentum
space) at its pole. The dimensionless quantity T is called the scale of  in T .
Measurable quantities are masses of particles and scattering amplitudes, which
are scale-of--independent objects. So let us be interested only in scale-of--
independent characteristics of a given theory. In this case we can restrict our
attention to the 2-parameter subfamily of theories Y1  Y in which the scale of 
equals 1. This causes no loss of generality, since any theory T 2 Y can be nor-
malized to a theory T1 2 Y1 by a transformation  ! , in such a way that all
scale-of--independent quantities are unchanged.
Now our space of theories is 2-dimensional, so we need to measure 2 things in
order to specify a concrete theory. One of the possible prescriptions is the following.
The two measured things are
(i) The mass m of the eld . Mathematically, it is determined from the condition
that m2 is the pole of the 2-point function of the theory, regarded as a function
of the square of momentum k2 (in the Euclidean picture).
(ii) The eective coupling g { the scattering amplitude for two P particles at the
some point (p1; p2 ; p3 ; p4) \on shell", i.e. with p2i = m2 and pi = 0. One of
the possible choices is to take pi to be vertices of a regular tetrahedron lying on
some 3-dimensional subspace of the dual spacetime V  (it is clear that all such
congurations can be obtained from each other by rotation). The eective coupling
measures how strong the interaction is.
However, as we will see in the future, one often needs to consider parametriza-
tions of the space of theories in which parameters are not necessarily measurable
or have any physical meaning. An example of such physically meaningless, but
very useful parametrization is the dimensional regularization prescription, which is
described in the next section.
2.2. The dimensional regularization prescription of nite renormal-
ization.
Now let us describe a prescription of nite renormalization which uses dimen-
sional regularization.
Recall from Lecture 1 that given a renormalizable 4-Lagrangian, we can com-
pute its correlation functions using dimensional regularization and minimal sub-
traction. At rst sight it seems that the answer is completely canonical and does
not depend on anything. However, after careful consideration it becomes obvious
that the answer depends on the choice of unit of length. This happens because the
dimensions of couplings in the Lagrangian, in general, depend on the dimension D
of the spacetime, so that they should be regarded not as functions but rather as
sections of some line bundle over the space of values of D. In order to work with
functions, we should trivialize this bundle.
The correct way of computing correlation functions with dimensional regulariza-
tion is, in general, the following. We should choose a scale of momentum , and
2
for any coupling (or mass)  in the Lagrangian nd the dimensionless quantity d
such that  = ddim(;d), where dim(; D) is the dimension of  when dimension
of spacetime is D (this is some function of the form aD + b), and d is the physical
dimension of the spacetime. Now we should replace all masses and couplings i
in the Lagrangian with the expressions i(D) = iddim(;D), and perform dimen-
sional regularization { minimal subtraction with this new Lagrangian, as described
in Section 1.6 of Lecture 1.
Remark. One should remember that the Feynman integrals used to dene
correlation functions will have d-dependent dimensions, so they have to be mul-
tiplied by appropriate powers of  to make their dimensions constant, before the
subtraction procedure is applied.
It is clear that the answer (correlation functions) heavily depend on the choice
of . The reason is that some couplings  will nontrivially depend on D (for xed
d), and their rst derivatives by D (which contain ln ) will in uence the regular
part of the correlation functions at D = d.
For example, in the case of 4-theory, we have a Lagrangian of the form
Z 2
(2.1) L = d4x( a2 (r)2 + m2 2 + 4!g 4);
and m(D) = md; a(D) = ad; g(D) = gd4 D .
Now, the dimensional regularization prescription of nite renormalization is: for
any  2 M , T ! (m; a; g), where (m; a; g) are the parameters of the 4-Lagrangian
of the form 2.1 which gives the correlation functions of T if renormalized with the
help of dimensional regularization, using the scale .
The parameters m; a; g have the meaning of \the eective mass at scale ", \the
inverse of the eective scale of  at ", and \the eective coupling at " (we will
later explain why).
2.3. Scale-dependence of nite renormalization prescriptions.
In physics we are interested in the behavior of a system at dierent scales, and
how it changes when going from scale to scale. Therefore, we are interested in
scale-dependent parametrizations of the space of theories, such that at each scale
 the corresponding parametrization represents in some way the behaviour of the
system at . Among the prescriptions 1,2 given in Section 2.1, only 2 satises this
property, while 1 is scale-independent.
The dimensional regularization prescription also satises this property. That is,
given a theory T 2 Y , the parameters of the Lagrangian L (T ) which reproduces
the correlation functions of T if renormalized by dimensional regularization using
 represent the behavior of the theory at scale . This is not immediately obvious,
since the dimesional regularization procedure is formal and has no physical meaning.
The reason is that for large  the mass can be neglected, and thus the function 4
is a function of pi =, which does not explicitly depend on . Therefore, the coupling
g(), which is formally obtained from the dimensional regularization procedure,
characterizes the magnitude of 4 at pi  .
Let us illustrate this by considering an example.
Given a 4-Lagrangian L of the form (2.1), its 1-particle irreducible 4-point
function, computed by dimensional regularization using the scale , in the 1-loop
3
approximation equals
2
(2.2) 4 (pi ; ; m; a; g ) = g + ag2 [f ((p1 + p2)2 ) + f ((p1 + p3 )2) + f ((p1 + p4 )2)];
where f (k2 ) is the amplitude of the corresponding 1-loop diagram. This amplitude
was computed in Gawedzki lecture 3 (page 12; the formula for I^4;ren(k)), and the
answer is

(2.3) f (k 2 ) = A Z 1 ln((1 ) k2 + m2 )d + const;

6 0 2 a2
where A is some positive numerical constant.
For large  the term m2=a2 can be neglected, and we get

(2.4) f (k 2 )  A Z 1 ln((1 ) k2 )d + const = A ln(k2=2) + const;

6 0 2 6
In particular, if p1; p2 ; p3; p4 are vertices of a regular terrahedron such that p2i = 2
(a point representing the scale ), then for large  in the 1-loop approximation we
get
g2
4 = g + a2 C;
where C is some numerical constant independent of . In particular, the scale-of-
-independent quantity a has the form
4
2

4 = g + g2C; g = g=a2
a 2

If g is small, 4 =a2  g . So the \eective coupling constant" g does indeed

characterize the intensity of interaction at the scale .
2.4. The renormalization group ow corresponding to a scale-dependent
renormalization prescription.
Given a scale-dependent nite renormalization prescription, we can dene the
renormalization group ow in the space of parameters, which is given by transition
functions from the coordinates at one scale to coordinates at another scale. This is
done as follows.
Let Y denote the space of renormalizable theories of a single scalar eld . For
any scale  2 M we have an injective map  : Y ! P~ , where P~ is a xed (-
independent) space of parameters. We will assume that the image P = Im is
independent of . Denote this image by P .
Now we can x a theory T and look at the dynamics of  (T ) as  changes. To
do this, for any 1; 2 2 M dene a transformation R  : P ! P by R  =
   1 . This allows us to dene a ow on P  M : t(z; ) = (R;t z; t). The
1 2 1 2

2
owlines of this ow are of the form ( (T ); ), where T 2 Y is a xed theory.
1

This ow is called the renormalization group ow associated to the corresponding

family  of renormalization prescriptions.
4
 It is convenient to represent the renormalization group ow as a vector eld. Let
 (T ) = (m (T ); a (T ); g (T )), where m = m ; a = a ; g = g are the eective
mass, inverse scale, and coupling at  for some renormalization prescription. Then
the vector eld of renormalization group looks like
(2.5) W =  @ @ + ~m @ + ~ @ + a ~ @;
@m @g @a
where ~; ~; ~ are functions on P  M .
It is clear that the renormalization group commutes with the group G1 of rescal-
ings of : a ! ta; m ! t1=2m; g ! t2g;  ! , as well as with the group G2 of
changing the units of measurement: a ! a; m ! p tm; g ! g;  ! t. The invariants
of these actions are g = g=a2 and m = m= a. This shows that ~ = ^(g ; m ),
~ = ^(g ;  )a2 , and ~ = (g ; m )a.
It is convenient to rewrite W in the coordinates ; g ; m ; a. The result is
(2.6) W =  @ @ + ( 1)m @ +  @ + a @ ;
 @m @g @a

where := ^ =2, and  = ^ 2g .
Using this formula, we can write down a dierential equation for correlation func-
tions. Let j (pi ; ; m ; a; g ) be the j -point function of the theory with parameters
m ; a; g computed at scale . Since the \engineering" (i.e. classical) dimension of
j is 4 j , we have
 
@ +s @ +j 4
 @
(2.7) @s j (spi ; ; m ; a; g ) = 0:

Subtracting from (2.7) the equation W j = 0, we get

 
@ @ @ @
(2.8) s @s ( 1)m @m  @g a @a + j 4 j (spi ; ; m ; a; g ) = 0:
 
This equation allows one to study the dynamics of j under the scaling of external
momenta.
Now consider the renormalization group ow in terms of the coordinates g; m ; a.
According to formula (2.6), integral curves of this ow are given by the formulas
g = g (); m = m(); a = a(), where
(2.9) g0 =  (g; m ); m0 = ( (g ; m) 1)m ; a0 = (g ; m)a:
and prime denotes the derivative with respect to ln . From these formulas it is
clear that a expresses trivially through g ;  :
R
(2.10) a() = a0 e 0
(g (s);m (s))d ln s :

However, equations for g and m cannot be easily solved, even if  and are
known, unless  and have some special form.
5
 However, suppose that the vector eld W has a third one-parameter symmetry
group:  ! ; a ! a; m ! tm ; g ! g (mass rescaling). In this case we will call
our renormalization prescription mass independent. Then ; ,  are independent
on m , and equations (2.9) are easily solved. Namely, g () is represented implicitly
by
 R g  dg

0 = e
g   g ;
( )
(2.11) ( 0) ( )

and m is expressed via g by

m () = 0 m (0)e  (g (s))d ln s :
R
(2.12) 0

Unfortunately, most renormalization prescriptions are mass-dependent. This is

the case, for example, for prescription 2 from Section 2.1. However, this prescrip-
tion is \asymptotically mass-independent", i.e. almost mass independent at high
momenta. Indeed, the correlation functions, which are used as parameters in this
prescription, do not signicantly depend on m at high momenta. Therefore, al-
though  and depend on m , there exist limits  (g) := limm !0  (g ; m ),
(g ) := limm !0 (g ; m ), so that formulas (2.11), (2.12) become valid for very
large 0 ; , and, in particular, are applicable for studying the asymptotics of solu-
tions at  ! 1.
This \asymptotic mass independence" is characteristic of all prescriptions 
which are based on correlation functions at . But all of them are mass dependent
at nite .
Nevertheless, mass independent renormalization prescriptions exist. The most
convenient of them is the dimensional regularization prescription.
Let us explain why the dimensional regularization prescription is mass-independent.
We will work in the spacetime dimension D (a generic complex number), with
 = 4 D. For a renormalizable Lagrangian L of the form (2.1) with parame-
ters m; a; g, let i(pj ; m; a; g) be the correlation functions of L in dimension D,
computed by plain integration in D dimensions as described in Lecture 1, and
( ;sub
i (pj ; ; m; a; g ) be the correlation functions in dimension D, computed using

scale , with pole parts subtracted. The functions i are singular at  = 0, while
;sub are regular at  = 0, and the functions ;sub are the true
i i := lim!0 i
correlation functions of the theory, computed at scale  as in Section 2.2.
Let mb; ab; gb be the bare parameters, dened by the condition
 (pj ; mb ; ab ; g b ) = ;sub(pj ; ; m; a; g )
i i
(b stands for \bare"). These bare parameters are functions of renormalized param-
eters, , and , of the form
mb = mb(m; a; g; ; ); ab = ab(m; a; g; ; ); gb = gb(m; a; g; ; ):
From dimensional analysis (using the groups G1 and G2 ) it follows that the func-
tions mb, ab , gb are dened by the equations
ab = aZa (m ; g ; ); mb(ab ) 1=2 = mZm (m ; g ; ); gb(ab ) 2 = g Zg (m ; g ; );
6
where X
Zi (m ; g ; ) = 1 + Zik (g ; m ) k :
k1
(no positive powers of , as we used minimal subtraction).
The key point now is that Zi do not actually depend on m . Indeed, the func-
tions Zi are sums of counterterms introduced in the process of renormalization.
These counterterms at each step of the process are just sums of pole parts (in )
of amplitudes of all diagrams with a given number of loops. These pole parts are
determined purely by the asymptotic expansion of the amplitudes at jpij ! 1.
From this one can deduce that these pole parts are independent on m. Thus,
Zi = Zi (g ; ).
Let W be the vector eld of the form
W =  @@ + ( 1)m @ +  @ +  a @ ;
  @m  @g  @a
 
which preserves functions ;sub
i (pj ; ; m; a; g ). It is clear that such W exists and is
unique, and lim"!0 W" = W . On the other hand, using the denition of i , we ;sub
get that W" is dened by the condition Wmb = 0; W ab = 0, Wgb = 0. Writing
these equations in components, we get
0
Zm0 + Zm = 0; Za0 + Za = 0; g +  (gg ZZg ) = 0
 g
(prime denotes the derivative by g ). From these equations we nd:
"g2 Zg Z 0
m Z a
0
 = (g Z )0 ;  =  Z ;  =  Z :
 g m a
Using the facts that Zi has a Laurent expansion in inverse powers of , and the fact
that ; ;  are by denition analytic at " = 0, we get the following formulas:
(g ) = g + g2(Zg1 )0 ; (g ) = g(Zm1 )0 ; (g ) = g (Za1)0
In the limit " ! 0, these formulas give
 (g ) = g2(Zg1)0 ; (g ) = g(Zm1 )0 ; (g ) = g(Za1 )0
The last formula shows that ; ;  depend only on g, and gives a simple and useful
rule of evaluating them.
In particular we see that although counterterms in dimensional regularization
can have poles of arbitrary order, the contribution to the renormalization group
vector eld comes only from residues of these poles. Thus, in principle the renor-
malization group vector eld should be more elementary than the complete system
of counterterms. This is one of the reasons why it is useful to write down the
renormalization group equation.
We have seen that the asymptotic properties of our eld theory at  ! 1 depend
on the properties of the function  (g), since this function determines the dynamics
7
of the coupling g(). This function is called the beta-function. Later we will see
that the sign of this function near g = 0 is especially important.
2.5. Computation of the renormalization group ow in the 1-loop
approximation.
In this section we will compute the vector eld W in the 1-loop approximation.
Let m = m (), g = g(), a = a() be an integral curve of the ow. We will
express the derivatives m0; g0 ; a0 from the equations  @@ 2 (p2 ; m ; a; g ; ) = 0,
 @@ 4(pi ; m ; a; g ; ) = 0.
First of all, the 1-loop correction to the 2-point function can be absorbed in mass
renormalization: m() = m0(1 + b ag ln(=0 )) + O(g2 ), where b is a numerical
2

constant. Thus, a0 = 0; m0 = m (1 + bg ).

In order to compute g0 , we dierentiate equation (2.2) with respect to ln . This
yields, modulo g3, the following expression:
(2.13) g0 = Ag2:
Thus, we get in the one-loop approximation:  (g ) = Ag2; A > 0; (g ) =
bg ; (g ) = 0.
Remark. A more careful calculation gives: A = 163 ; b = 321 .
2 2

Now consider the renormalization group equation (2.13). The general solution
of this equation is
(2.14) g = 1 Ag0gln( 0 :
 =0 )
Thus, we have found an approximation to the function g() based on the 1-loop
approximation to the  -function.
Now we face the following fundamental questions. In what respect is expression
(2.14), obtained from the 1-loop approximation to the beta-function, any better
than the expression
(2.15) g = g0 + A(g0)2 ln(=0);
obtained directly from (2.2)? (Nothe that (2.15) coincides with (2.14) modulo
(g0)3 !) In what sense does (2.14) better represent \reality"than (2.15)? In other
words, why did we care to introduce renormalization group, beta-function, and so
on, instead of just getting (2.15) directly?
The answer is the following. Assume that 4 theory exists in the non-perturbative
setting (it is not believed to be the case, but never mind; you can think in-
stead of another eld theory). Then formula (2.15) has a chance to be valid only
when g0 ln(=0 ) is small, since the series for g will contain terms of the form
(g0)k+1 ln(=0 )k . On the other hand, formula (2.13) is valid for small g, since it
is the beginning of the Taylor expansion of  . So in the region where g is small
but g0 ln(=0 ) is not (for example, Ag0 ln(=0) = 1=2; g0 << 1), formula (2.14)
is reliable but (2.15) is not.
2.6. Asymptotic freedom.
So far we have always worked with quantum eld theories in the perturbative
setting, primarily because this was the only setting in which we could dene and
8
study them. However, physics lives over R rather than R[[ ]], so we should investi-
gate when we can use perturbative results to judge about the actual process. One
necessary condition is clear: all couplings should be small, so that expansions in
formal series with respect to them have a chance to make sense.
In particular, we can ask whether we can trust the rules of perturbative renor-
malization, which were discussed in the previous lecture and in Witten's lectures.
As you remember, these rules tell how to handle divergences arising at high mo-
menta (ultraviolet divergences). Thus, in order to trust the theory of perturbative
renormalization we should be sure that interactions are small at high momenta.
A (Euclidean) quantum eld theory (in the non-perturbative sense; for example,
in the sense of Wightman axioms) is called asymptotically free if its interactions are
small at high momenta. That is, a theory is asymptotically free if its correlation
functions in momentum space are asymptotically (at high momenta) close to those
of a free theory. Thus, in order to use perturbative renormalization legitimately,
we should be sure that our theory is asymptotically free.
Of course, in the perturbative setting we cannot even show that a quantum eld
theory exists, much less that it is asymptotically free. It is therefore a pleasant
surprise that if we assume that a theory with a given perturbation expansion exists,
its asymptotic freedom can be checked perturbatively, usually already in the 1-
loop approximation. Later we will discuss it in more detail, but now we will just
demonstrate it for 4-theory.
We will only consider renormalizable theories, i.e. points of the space Y . It is
clear that such a theory T 2 Y is free (i.e. quadratic) if the eective coupling
g vanishes. Thus, the condition for asymptotic freedom is that the coupling g
vanishes asymptotically, at high momenta. That means, lim!1 g () = 0.
It is clear that we cannot see whether g () approaches zero or not by looking at
g () in the perturbation expansion. This is clear from the example of the function
f () = 1+ ln(1= ) , which vanishes at innity for any nite  , but diverges at
0
innity if reduced modulo  N for each N . What we can see, however, is whether,
as  increases, the value of g () increases or decreases.
Let us do this for 4 theory. In this case, the physically meaningful region of
values of g is g > 0. So we should nd out whether g0 is positive or negative
when g is in this region.
Now it is very easy for us to answer this question. Namely, if g is small, we
can trust formula (2.13): g0 = Ag2 . This means that g increases, i.e. gets away
from zero. Of course, we cannot claim on the basis of (2.13) that g increases
all the time (remember that (2.13) is only valid for small g ), but what we can
denitely guarantee on the basis of (2.13) is that lim!1 g () 6= 0. Thus the 4
theory is not asymptotically free, and its perturbation expansion is not valid
at high momenta. This is a non-perturbative result which we obtained using only
perturbation theory.
Thus, renormalization group gives us a subtle way of partial summation of the
perturbation series, allowing us to extend the region of validity of perturbation the-
ory, and thus pushing us closer toward the world of non-perturbative eld theories.