Lecture 2: Perturbative renormalization

(continued)
Edward Witten1
October 1996

2.1. Renormalizability of quantum eld theories.

Last week we considered the 3 theory and found that its renormalizability
(i.e. how many graphs are divergent, and how badly divergent they are)
depends on whether n < 6, n = 6, or n > 6. Now we will see how to nd
that the critical value of n in the 3 theory is 6 without considering any
graphs at all.
In order to nd the critical value of n, we will dene and compute the
homogeneity degree of all terms in the Lagrangian. We recall that the
Lagrangian is
Z 1
(2:1) ( 2 (r)2 + 12 m22 + 3!g 3 )dn x;
(we are considering the Euclidean theory; in the Minkowski version, the
story is the same). Consider the dilation x ! t 1 x, t 2 R, x 2 V . Under
this transformation dn x ! t n dn x, r ! tr. If we want the integral (2.1)
to be invariant under this dilation, we must impose the following three
transformation laws: (r)2 ! tnn (2r)2, m2 2 ! tn m2 2 , g3 ! tn g3.
From the rst law6 n
we get  ! t 2 , from the second m ! tm, and from
the third g ! t 2 g . In the future we will write these scaling laws like this:
(2:2) [x] = 1; [dnx] = n; [r] = 1; [] = n 2 2 ; [m] = 1; [g ] = 6 2 n :
In other words, any homogeneous quantity a scales as a ! t[a] a under the
dilation, and the number [a] is called the dimension of a.
Remark. It is not dicult to dene the notion of dimension completely
formally (using the tensor calculus on the tangent bundle of V ), but it is
more illuminating to illustrate it on examples.
We see that the number 6 appears in the scaling law for g . More precisely,
we see that the behavior of the 3 theory is determined by whether [g ] is
positive, zero, or negative. Namely, we have three cases.
1
Notes by Pasha Etingof and David Kazhdan, TeXnical editing Misha
Verbitsky

1
 1. [g ] > 0 (n < 6): there is a nite number of supercially divergent
graphs (by graphs we always mean connected graphs).
2. [g ] = 0 (n = 6): the number of supercially divergent graphs is
innite, but div ( ) is bounded from above.
3. [g ] > 0 (n > 6): there are innitely many graphs with any number of
external vertices and arbitrarily high div ( ).
Now we will explain why this method works, and prove that it can be
used to compute the critical value of the dimension of the spacetime in a
general eld theory. In general, we can consider the following setup. Suppose
we have a quantum eld theory with elds 1 ; :::; N , and the Lagrangian
Z X X
(2:3) L = ( Qi(i) + gk Ik (1; :::; N ))dnx;
i k
where Qi is a free (quadratic) part, and Ik are interaction (coupling) terms
(dierential monomials in elds, cubic and higher). Each interaction term
comes with a small parameter gk , which is called the coupling constant.
Denition An interaction Ik is called subcritical if [gk ] > 0, and critical
if [gk ] = 0. A eld theory is called superrenormalizable if all interaction
terms in the Lagrangian are subcritical, and is called renormalizable, or
critical, if all interaction terms are critical or subcritical, but not all of them
are subcritical. Otherwise, the theory is called non-renormalizable.
For example, the 3 theory is superrenormalizable for n < 6, renormal-
izable for n = 6, and non-renormalizable for n > 6.
We will assume that our theory is a perturbation of a free theory, (for
the denition of a free theory, see Kazhdan's lectures). In a free theory, one
can easily see that the dimension of bosonic elds is n 2 2 , and of fermionic
elds is n 2 1 . Since dimensions of elds are determined from the quadratic
part of the Lagrangian, these dimensions will be the same in the perturbed
(classical) theory as well.
We will also assume that n  2 (In the quantum mechanical case n = 1
renormalization theory is not necessary). Then [i ]  0.
Theorem 2.1 (i) If a theory is superrenormalizable, there is a nite
number of supercially divergent graphs in its Feynman diagram expansion.
(ii) If a theory is renormalizable then the number of supercially diver-
gent graphs is innite, but div ( ) is bounded from above.
(iii). If a theory is non-renormalizable then there are innitely many
graphs with any number of external vertices and arbitrarily high div ( ).
Proof Let be a graph in the Feynman diagram expansion of our theory.
Types of internal vertices of such a graph correspond to interaction terms

2
in the Lagrangian, and types of its external vertices and edges correspond
to elds i . Let ei be the number of external vertices of type i , and vk be
the number of internal vertices of type Ik . Then it is easy to show that
X X
(2:4) div ( ) = n ei[i ] vk [gk ]:
i k
Statements (i)-(iii) of the theorem follows immediately from (2.4).
Denition A theory of the form (2.3) is called classically scale invari-
ant if the n-form under the integral in the Lagrangian is invariant under
dilations.
For example, the theory of a free scalar eld is scale invariant i it is
massless.
It is clear that a massless theory is scale invariant if and only if it is
purely critical, i.e. [gk ] = 0 for all k.
Remark. Scale invariant theories are always conformal. Indeed, the La-
grangian is always written naturally in terms of the metric on the spacetime,
and depends only on the elds and their rst derivatives. This implies that
a scale invariant Lagrangian has to be invariant under conformal changes of
metric.
2.2. Critical dimensions of some eld theories.
Now we will compute the critical dimension for several important eld the-
ories.
Example 1. Sigma-models.
Let M be a Riemannian manifold. The Lagrangian of the sigma-model
on Rn with the target space M is
Z
(2:5) L() = dn xgij ()ri
rj :
Since this theory is conformal in two dimensions, 2 must be the critical
dimension. Let us show by a direct computation that this theory is not
renormalizable for n > 2.
Let i be coordinates on M near some point. If the metric gij is constant
in these coordinates, the theory is free. Consider a nonconstant metric of
the form
(2:6) gij () = ij + aijk k + rijklk l + :::
(If we chose normal coordinates, we could get rid of a, but not of r, as r
in normal coordinates is the Riemann curvature tensor). Substituting this
3
into the Lagrangian, we nd that [r] = 2 n. This shows that sigma-model
is not renormalizable beyond n  2, unless the metric is at.
Example 2. Gravity.
p the spacetime is the space Rn with a metric of
In the theory of gravity
the form gij = ij + hij G, where G is the Newton's constant (for us, it is
just a formal parameter). The Lagrangian of the theory is
1 Z
(2:7) L(g) = R(g )dnx;
G
where R(g ) is the scalar curvature of the metric. In terms of h, this La-
grangian can be rewritten as
Z p
(2:8) L = dnx((rh)2 + Gh(rh)2 + :::);
so we get [h] = n 2 2 , [G] = 2 n. Thus, as in the previous example, the
theory is non-renormalizable for n > 2 and critical for n = 2.
Remark. Here h(rh)2 stands for an expression which is linear in h and
quadratic in rh. It is easy to compute what it is exactly, but it does not
matter to us, since we are only interested in the dimension. So we will use
such sloppy notation.
Example 3. Gauge theory.
In gauge theory elds are conections in a xed principal G-bundle on the
space Rn, where G is a compact Lie group. The Lagrangian has the form
Z
(2:9) L(A~) = 1 Tr(F ^ F );
e2 A~ A~

where FA~ is the curvature of the connection A~, and Tr is an invariant non-
degenerate bilinear form on the Lie algebra g of G.
In the computation of dimension, we will assume that our G-bundle is
trivial, so a connection is represented by a 1-form A: A~ = d + A. Then
F = dA + A ^ A.
Consider the eld B = A=e. In terms of B , the Lagrangian takes the
form Z
L = ((rB)2 + eB2rB + e2B4 )dnx
We have [B ] = n 2 2 , so [B 2 rB ] = 32n 2, and [e] = 4 2 n . Thus, if the
group G is noncommutative, the theory is superrenormalizable for n < 4,
renormalizable for n = 4, and non-renormalizable for n > 4. In n = 4, the
theory is conformal (If G is commutative, the theory is free).

4
 Remark. Dimensions of elds in our computation agree with geometric
dimensions. Indeed, in geometry, since d ! td under the dilation x ! t 1 x,
we have [d] = 1, so we must have [A] = 1. This coincides with our result:
[A] = [Be] = [B ] + [e] = n 2 2 + 4 2 n = 1:
Example 4. Gauge theory with a scalar bosonic eld or with a
fermionic eld.
Consider the setting of Example 3, and let V be a nite-dimensional
representation of G. Let  be a section of the corresponding vector bundle
on Rn (associated with the above G-bundle). The Lagrangian of the gauge
theory with  is
Z
(2:10) 0 ~ ~
L (A; ) = L(A) + (rA~)2dnx;
where L(A~) is the Lagrangian (2.9). Writing A~ in the form A~ = d + A,
A = eB , we get
Z
L0 = ((rB)2 + eB2 rB + e2B4 + (r)2 + 2e(r; B) + e2(B)2)dnx
For this formula we get [B ] = [] = n 2 2 , and (even when G is commutative!)
[e] = 4 2 n . Thus, the theory is superrenormalizable in n < 4, renormalizable
in n = 4, and nonrenormalizable for n > 4 (regardless of the commutativity
of G).
The same answer applies if we have a fermionic eld. Let be a section
of V
S , where S is the spin bundle over the spacetime. The Lagrangian
of the gauge theory with is
Z
(2:11) L00(A;~ ) = L(A~) + ( ; DA~ )dnx;
where DA~ is (i times) the Dirac operator along the connection A~. The
critical dimension in this theory, as before, is 4, for any nontrivial compact
group G.
Remark. If G = U (1) and V is the standard 1-dimensional representa-
tion of G, this theory is the quantum electrodynamics (QED).
Example 5. Theory of a scalar bosonic eld.
Consider the Lagrangian
Z 1
L() = ( 2 (r)2 + m2 2 + Q())dnx;
2

P
where Q() = gk k . We already considered this type of Lagrangian in
Lecture 1. In particular, the 3 -theory is a special case of this situation,
5
when Q is a cubic polynomial. We have [gk ] = n k n 2 2 . Thus, for n = 2
all terms in the Taylor expansion of Q() are subcritical. For n = 3, the
term k is subcritical for k < 6, critical for k = 6, and non-renormalizable
for k > 6. For n = 4, the term k is subcritical for k < 4, critical for k = 4,
and non-renormalizable for k > 4. For n = 5; 6, the terms k , k  4 are
non-renormalizable, and the term 3 is subcritical for n = 5 and critical for
n = 6.
In particular, the following theories are critical: the 3 theory in 6 di-
mensions, the 4 -theory in 4 dimensions, and the 6 theory in 3 dimensions.
Example 6. Yukawa interaction.
Consider a theory with a scalar bosonic eld  and a fermionic eld .
We will consider the Lagrangian
Z
(2:12) L(; ) = ((r)2 + ( ; D ) + g[ ; ])dnx
( takes values in a vector bundle W which is a direct sum of several copies
of the spin bundle; (,),[,] are a symmetric and a skew-symmetric form on W
which are invariant under gauge transformations). The cubic term [ ; ]
is called the Yukawa interaction. Let us compute the dimension of its coef-
cient g .
We have [] = n 2 2 , [ ] = n 2 1 , so [g ] = 4 2 n . So this theory is critical in
dimension 4, superrenormalizable in dimension < 4, and non-renormalizable
in dimension > 4.
We observe that in dimension 4 all interactions except 3 , 4 and [ ; ]
are \bad" (non-renormalizable).
Example 7. Standard model.
Let us now try to write down the most general renormalizable theory
that lives in our 4-dimensional physical world. According to the above exam-
ples, we cannot include gravity or -model, but we can include connections,
bosons with terms up to degree 4, and fermions with Yukawa interaction If
we only take these elds, these are the only renormalizable terms we can
write. Thus the most general Lagrangian we can write in dimension 4 giving
a renormalizable theory is
(2:13) Z
L(A; ; ) = (e 2FA2 +( ; D )+(r)2 + g14 + g2 2 +lower terms)d4x:
The Standard Model is a theory which belongs to this family, with the
group G containing SU (3)  SU (2)  U (1).

6
2.3. Perturbative renormalization of critical theories.
From now on we will consider only critical theories. As a model example we
can consider 4 -theory in 4 dimensions, which has one critical interaction
4 , or its extension containing fermions, which has an additional critical
interaction  2 (the Yukawa interaction).
Consider the Lagrangian of 4-theory:
Z 1 m 2 g 
(2:14) L = 2 (r) + 2  + 4!  dnx:
2 2 4

As before, we want to study the Schwinger function h(x1):::(xN )i (it

is enough to consider only even N , as the Schwinger function vanishes for
odd N ). As usual, it is more convenient to consider its Fourier transform.
This Fourier transform has the form GN (k1; :::; kN ) (k1 + ::: + kN ), where
GN is a function on the hyperplane k1 + ::: + kN = 0.
Consider the Feynman diagram expansion of the function GN . In 4
theory we have to sum over graphs whose internal vertices have 4 edges.
As usual, we can restrict ourselves to connected graphs, as the sum over
disconnected graphs expresses via Schwinger functions with fewer insertion
points. In fact, as shown in Kazhdan's lectures, we can always restrict to
1-particle irreducible graphs (i.e. those which cannot be split by cutting one
edge). So from now on we only consider connected, 1-particle irreducible
graphs.
From formula (2.4) we get that the supercial divergence index of any
graph equals 4 E , where E is the number of external edges. Thus, any
graph with 2 external edges is quadratically divergent, any graph with 4
external edges is logarithmically divergent, and any graph with > 4 external
edges is supercially convergent.
Now we will explain how to renormalize all N -point functions at all
orders in g .
First of all, we can exclude graphs with 2 external edges which connect
to the same internal vertex, e.g.

and all those which contain such a loop-like graph as a subgraph. Indeed,
these graphs produce a constant function of k2 , so they can be removed by
7
renormalization of mass. More precisely, there exists a function P (; g ) =
gP1() + g 2P2 () + :::, such that the sum of terms corresponding to all
graphs for the theory with mass m2 , computed with the cuto propagator
(see lecture 1) equals the sum of terms corresponding only to graphs without
loop-like subgraphs, but with mass M () such that M 2 = m2 + P (; g ). So
we can assume from the beginning that we have a theory with mass M ()
and not worry about loop-like graphs.
Now we have no divergent graphs with one internal vertex, so we have
no corrections to make in the rst order in g . Let us look at the second
order in g . In this case we have the following bad graphs:

Denote the rst graph by 2 and the second by 4 .

Thus, in order g 2 we have a problem in the 2-point and the 4-point func-
tions. The problem in the 2-point function is created by 2 . The graph 2
diverges quadratically. Therefore, the term corresponding to 2 computed
using the cuto propagator is a function of k2 ;  of the form g 22 (k2; ),
where
(2:15)
2 (k2; ) = Ak2 ln(=) + B () + O(1); B () = B0 2(1 + o(1));  ! +1
This asymptotics follows from the fact that the rst derivative of (2.15) with
respect to k2 is logarithmically divergent, and the second one is convergent.
Another problem we have is in the 4-point function, created by the graph
4 This graph is logarithmically divergent. This means, if we compute the
.
term corresponding to this graph using the cuto propagator, we will get a
function of ki ;  of the form g 22 (k1; k2; k3; ), where
(2:16) 2 (k1; k2; k3; ) = C ln(=) + O(1);  ! +1
This asymptotics follows the fact that the derivative of the integral corre-
sponding to 4 with respect to ki is convergent.
In order to renormalize the 2- and 4- point functions in order g 2, we
choose renormalized functions R2 (k2), R2 (k1; k2; k3). Here R2 is an arbi-
trary function of k2 whose second
 derivative is given by the (convergent!)
2
integral obtained by applying dkd2 to the quadratically divergent integral

8
corresponding to 2 . Analogously, R2 is an arbitrary function of k1; k2; k3
whose derivatives are given by the convergent integrals obtained by dier-
entiating the logarithmically divergent integral corresponding to 4 . It is
clear that the function R2 is dened uniquely up to addition of a function
of k2 of the form ak2 + b, and the function R2 is dened uniquely up to
addition of a constant.
Now we will make second order corrections to the coecients of (r)2,
 and 4 in the Lagrangian. Namely, we will consider a new Lagrangian of
2
the form
(2:17)Z  
L = 12 (r)2 + M2 2 + 4!g 4 + () 2 +  () 2 + () 4 dn x;
2
2 (r  ) 2 4!
Let 02(k2 ; ; ; ; ), 02 (k1; k2; k3; ; ; ; ) be the functions 2 , 2 for
this Lagrangian and the cuto propagator. We will choose the functions
; ; in such a way that
lim 0 (k2 ; ; m; ();  (); ()) = R2 (k2 );
!1 2
(2:18)
lim 0 (k ; k ; k ; ; m; ();  (); ()) = R2 (k1; k2; k3);
!1 2 1 2 3
As we are working modulo g 3, we can choose ; ; in the form  = g 22 ,
 = g22 , = g 2 2, where 2 ; 2; 2 are independent of g . It is easy to
check that in order for (2.18) to hold, the functions 2 ; 2; 2 should have
the following asymptotics:
(2:19)
2 = g 2(A ln(=m) + D1 ) + o(1); 2 = g 2( B() + D2) + o(1);
2 = g 2(C ln(=m) + D) + o(1);  ! +1;
where D1; D2 depend on the choice of R2 , and D depends on the choice of
R2 . Of course, there are many ways to choose such functions, but they are
unique up to adding terms o(1),  ! 1.
Thus, we have renormalized the graphs 2 , 4 . This removes divergence
in all correlation functions modulo g 3. Thus, all correlation functions of our
theory are now dened modulo g 3.
Now we proceed inductively in the order of g . Suppose we have removed
divergences and dened all correlation functions modulo g K . Consider the
2N -point function (for the deformed Lagrangian and the cuto propagator)
modulo g K +1 :
KX1
(2:20) F2N = g j F2RN;j + g K F2N;K
j =0

9
(the superscript R means that the corresponding coecient has already been
renormalized). The term F2N;K is represented by the sum over all graphs
with K internal vertices. This sum has supercial divergence index 4 2N .
Therefore, the second derivative of F2 by k2, the rst partial derivatives
of F4 , and F2N , N  3, are supercially convergent. The crucial fact for
renormalization theory, which follows from the \Strong Weinberg Theorem"
(see Lecture 1), is
Proposition 2.2 There exists nite limits, as  ! 12, of the functions
F2N;K (k1; :::; k2N 1; ), rk F4;K (k1; k2; k3; ), and dkd2 F2;K (k2 ; ) N 
3.
Remark. For 4 theory, this proposition holds for the term correspond-
ing to each particular graph, but in general (for example, for theories with
gauge elds) this is not the case: the sum over all graphs may have a meaning
while each individual graph does not. However, an analogue of Proposition
2.2 (for the sum over all graphs) holds in any renormalizable theory.
Proposition 2.2 allows us to full the induction step. It shows that the
function F2 (in the limit) is dened up to adding ak2 + b, the function F4
is dened up to adding a constant, and F2N , N  3, is dened uniquely. So
one can choose renormalized functions F2RN;K and make corrections in the
Lagrangian,  !  + g K K ,  !  + g K K , ! + g K K , to compensate
the divergence in F2N;K and obtain F2RN;K instead of it. This procedure is
completely analogous to the one for order g 2. In this way we will complete
the renormalization in order K .
Remark 1. At every step of our renormalization procedure we had to
choose 3 constants of integration. This may create an impression that we
get a family of theories parametrized by 3 innite sequences of constants.
However, it is easy to see that in fact we get a family of theories parametrized
by only 3 constants. This means that any 4 invariants attached to the theory
(for example, values of the 2-point function at 4 points in spacetime) are
linked by a universal functional relation.
Remark 2. Even if in the original theory certain critical or subcritical
interactions were not present, they may appear in the process of renormaliza-
tion. In general, renormalization brings in all missing critical and subcritical
terms, unless there is a symmetry which prevents it from doing so. Let us
demonstrate it by a few examples.
Example 1. In the process of renormalization of the Lagrangian (2.12)
in 4 dimensions we will be forced to introduce the subcritical term 3 and
the critical term 4 , in order to remove logarithmic divergence in the graphs

10
 ,

where wavy lines correspond to bosons and straight ones to fermions.

However, the critical term 2 r will not appear, since there is no Poincare
invariant expression of this form. In terms of graphs, this means that the
graph

whose supercial divergence is linear, in fact diverges only logarithmi-

cally, because of cancellations in the integrand caused by Poincare symme-
try; so there is no linear divergence to compensate and hence no need for
2r to appear.
Example 2. In 4 theory in 4 dimensions, the subcritical term 3 does
not appear in renormalization, since it is not preserved by the symmetry
 !  of the original theory. In the language of graphs, this is clear: there
is no graphs with 3 external edges, so there is no divergence to compensate
by 3 .
In a general eld theory, every type (in terms of external edges) of a su-
percially divergent graph corresponds to a number of critical and subcritical
terms in the Lagrangian, which should be renormalized in order compensate
the divergence in the corresponding graph. More precisely, divergent terms
which are quadratic in k correspond to terms in the Lagrangian which have
two derivatives by x, linear terms in k correspond to terms with one deriva-
tive, and constant divergencies correspond to terms without derivative.
Remark. It follows from formula (2.4) that in a renormalizable the-
ory, all divergences in N  2-point correlation function are no worse than
quadratic. So the coecient of k2 in the 2-point function if divergent at

11
most logarithmically. Therefore, if the quadratic forms Qi have to be renor-
malized (like in 4 theory), they will be multiplied by coecients which
depend on the cuto parameter  at worst logarithmically. This shows that
the dimension of Qi and hence of i survives renormalization to all nite
orders in the asymptotic expansion.

12