Lecture 3: Perturbative Renormalization (Continued) : 3.1. Local Functionals in A Classical Eld Theory
Lecture 3: Perturbative Renormalization (Continued) : 3.1. Local Functionals in A Classical Eld Theory
(continued)
Edward Witten1
October 1996
1
so the left and the right hand sides of (3.1) evaluated at x 2 V represent
the same local functionals of . Thus, local functionals are all possible
dierential expressions in modulo the classical eld equations.
Remark 1. One should be careful to distinguish between two dierent
notions of eld dimension which arise in eld theory. The rst is \the
engineering dimension" and says in what units the eld is measured (if the
units are cm d then the engineering dimension is d). The second is \the
scaling dimension", which is the dimension we have been talking about.
These two dimensions are not always the same. For example, in the theory
of one eld dened by (3.1) the engineering dimension of the eld m2 is
3 (it is measured in cm 3 ), while the scaling dimension is the same as that
of , i.e. 1.
It is the basic principle of physics that all meaningful expressions and
equations are homogeneous with respect to the engineering dimension. In
particular, it is true for the eld equations in eld theory (e.g. (3.1)), which
implies that engineering dimension denes a grading on the space of local
functionals. On the other hand, whenever a eld theory is not scale in-
variant, its eld equations are not homogeneous with respect to the scaling
dimension, so scaling dimension denes a ltration, rather than a grading,
on the space of local functionals. This means, the space of all local function-
als is a union of subspaces of functionals of dimension d, and a functional
is said to be of dimension d if it is of dimension d but not d 1. On the
other hand, if the theory happens to be scale invariant, scaling dimension
does dene a ltration on the space of local functionals.
In this lecture, we will not use engineering dimension, and the word
\dimension" will always mean scaling dimension.
Remark 2. From a localR functional O one can obtain more general
functionals on X of the form O(; x)d(x), where d(x) is a (generalized)
density on V . Using this operation, one can obtain from any local functional
all derivatives of thus functional (it is enough to take for d(x) all possible
densities supported at x). Therefore, without loss of information we could
consider local functionals modulo the image of derivatives. However, for the
purposes of this lecture this is not necessary.
3.2. Quantization of local functionals in a free theory
We have seen (see Bernstein's Lectures and Witten's problem sets) that
the space X of classical solutions of a meaningful classical eld theory always
carries a natural closed 2-form
. If this form is nondegenerate, X is a
symplectic manifold. In this case, suitable functions on X form a Poisson
algebra.
In the quantum theory, the space X should be quantized, and the Poisson
2
algebra of functions on X should become the algebra of operators (observ-
ables) in some Hilbert space of states. In particular, we should be able to
assign an operator to every local functional.
If V is a Minkowski space, and the eld theory satises the Wightman
axioms (for example, the free theory), the Hilbert space of states H is con-
structed as described in Kazhdan's lectures. In this case, Wightman elds
(x) are distributions on V with values in the space of operators on the sub-
space D H of muptiparticle states. This means, for any Schwarz function
f on V there is an honest operator '(f ) on D.
However, we would like to have more general operators of the form N (x),
(r)2(x), etc. which correspond to local functionals in the classical theory.
That is, for any Schwarz function f we want to have operators 'N (f ),
(r')2(f ), etc.
Unfortunately, such operators are not automatically dened. For ex-
ample, even '2 (f ) does not, in general, make sense. Indeed, let " =
e j xj 2 ="
(2=")n=2 be the smooth approximation to the -function, and let
'
' be the operator-valued distribution on V V given by '
'(f1
f2 ) =
'(f1 )'(f2). The natural denition of '2(f ) would be that '2(f ) is the limit,
as " ! 0, of the operator '2" (f ) := '
'(f (x)" (x y )) (in the sense of
convergence of matrix elements). However, it is easy to see that this limit
does not usually exist, as Wightman functions usually have singularities on
the diagonals.
One way to deal with this problem is to say that an \operator" A is just
a collection of its matrix elements, i.e. a collection of correlation functions
h(y1):::(yl)jAj(z1):::(zr)i. If we accept this point of view, we might as
well forget about the Hilbert space of states, i.e. perform a Wick rotation
from Minkowski space to Euclidean space, and consider Schwinger functions
instead of Wightman functions.
So from now on we will consider only the Euclidean situation, in which
we will mean by an operator A a collection of functions
h(y1):::(yl)jAj(z1):::(zr)i
with certain properties. Roughly, an operator is just a symbol which can be
inserted in a correlation function.
Remark. Of course we should remember at all times that A is not really
an operator and does not act in any Hilbert space. However, the information
we get from studying A in the Euclidean situation can be used (after Wick
rotation) for studying the Minkowski situation.
Consider now the problem of quantization of local functionals. By the
denition, in order to quantize the local functional O(x), we should say
3
what the functions h(y1):::(yl)jO(x)j(z1):::(zr )i are. As our theory is
Euclidean, the order of factors in the correlation function does not matter
(any two distinct points are space-like separated). So, in order to dene
O(x), it is enough to dene the correlation functions h(y1):::(yr)O(x)i for
all r.
Consider the theory of a scalar bosonic eld in n > 2 dimensions, with
the Lagrangian
Z 2
(3:2) L() = ( 21 (r)2 + m2 2 + Q())dnx;
We will give all denitions and constructions for this example. In other eld
theories, everything is done in a similar manner.
We rst consider the case when the theory is free (Q = 0). In a free
theory, quantization is done with the help of normal ordering, as follows.
Consider a local functional O of degree J in . According to Feynman
rules, in order to compute the correlation function h(y1):::(yr )O(x)i, we
should consider all graphs with r external vertices and only one internal
vertex v , which has J edges. We should put a certain function at the vertex
v , and compute the term (amplitude) corresponding to this graph as usual
in Feynman calculus. If J > 1, we will run into trouble: we will get some
graphs with loops going from v to itself, and the integration over the loops is
divergent. The easiest remedy is to ignore all such graphs. Then we obtain
certain correlation functions, which dene some operator. This operator is
denoted by : O(x) : and called the normal ordering of O(x). Apriori, O(x)
does not make sense as an operator, while : O(x) : does. We call operators
of the form : O(x) : composite operators. When no confusion is possible, we
will drop the dots and write O(x).
Thus, we have assigned canonically to each dierential polynomial O in
an operator : O : in the free theory. However, recall that two dierent
polynomials O1 ; O2 might dene the same local functional. So does the map
O !: O : actually dene canonically a quantization of the local functional
represented by O? The answer turns out to be positive: whenever polyno-
mials O1 , O2 dene the same local functional, the operators : O1 :; : O2 :
coincide. To check this is an easy exercise. For example, the polynomials
O1 = , O2 = m22 dene the same local functional, and the operators
: :, : m22 : coincide.
3.3. Multiplication of composite operators.
It is easy to multiply composite operators supported at dierent points.
That means, given two composite operators O, O0 , and points x; x0 2 V ,
x 6= x0 , it is easy to dene an operator O(x)O0(x0 ), using the Feynman
4
diagram expansion. In this case we will have two special vertices v; v 0, and
we should sum over all ways of connecting them which involve no loops
from a vertex to itself. However, multiplication of two composite operators
O; O0 at the same point x cannot be dened in the same way: on each edge
connecting v with v 0, we will have to evaluate the Green's function at (x; x),
where it is singular.
To avoid this problem,
R one can try to smear the operators, i.e. consider
the operator O(f ) = f (x)O(x)dnx, where f is a compactly supported
smooth function on V . Such smeared operators can sometimes be multiplied:
for example, we can dene the product (f1 ):::(fs). (see Kazhdan lectures).
However, most composite operators cannot be multiplied, even after
smearing. Indeed, it can be deduced from re
ection positivity that for any
complex operator O hO(x)O(x0 )i jx x0j 2[O] , x ! x0 where [O] is the
dimension of [O], and O is the conjugate of O. On the other hand, for
a real operator O, in order for the product O(f )O(f 0 ) to be dened, it is
necessary and sucient that the functions h(y1 ):::(yr)O(x)O(x0 )i be dis-
tributions. The function jxj l denes a distribution (without regularization)
if and only if it is locally L1 , i.e. i l < n. Thus, the product O(f )O(f 0 ) is
automatically dened i [O] < n=2.
Remark. If the Minkowski situation, the product O(f )O(f 0) exists (in
the sense of correlation functions) if and only if the \operator" O(x) is an
operator-valued distribution, so that O(f ) is actually an honest operator
(i.e. matrix elements of O(f ) are matrix elements of some operator on D).
Indeed, the \if" part is clear. The \only if" part: for any f1 ; :::; fN we can
regard the product X = O(f )'(f1):::'(fN ) as a linear functional on D. The
fact that O(f )2 is dened implies that this functional has nite L2 -norm, as
jjX jj2 = h'(fN ):::'(f1)O(f )2'(f1):::'(fN )i. Therefore, X corresponds to a
vector in H, Q.E.D.
3.4. Operator product expansion (OPE) in the free theory
Composite operators in a reasonable eld theory have an important al-
gebraic property, which is called the operator product expansion (OPE).
This is a property of algebraic nature, saying roughly that composite oper-
ators form an algebra of a certain kind. It does not follow obviously from
Wightman axioms, but on the other hand is often satised when Wightman
axioms fail. This property is actually useful in practice.
We will now dene and compute the OPE for the free theory of a scalar
bosonic eld. The idea is to formalize the behavior of the product O(x)O0 (x0)
as x ! x0 . In the classical theory, this is accomplished by the usual Taylor
5
series. Indeed, we have
(3:3) O(x)O0(x0) = O(x0)O0(x0) + (x x0)@O(x0)O0(x0) + :::
Thus, for any two local functionals O; O0 the product O(x)O0 (x0 ) decom-
poses in a series in products of powers of (x x0 ) whose coecients are
other local functionals.
In the quantum theory, the product : O(x) :: O0 (x0) :, as we have seen,
is singular at x = x0 . Thus, we have to compute the asymptotic expansion
of this product near x = x0.
We will assume that O; O0 are represented by some monomials in and
its derivatives. According to Feynman rules, we should assign to both O; O0
a vertex, say v and v 0. The numbers of edges at v; v 0 are J; J 0 , where J; J 0
are the degrees of O; O0 as monomials of . At both v; v 0, we have to place
symmetric distributions in J; J 0 variables corresponding to O; O0. Now we
have to sum amplitudes over all ways of connecting some edges going from
v with some edges going from v 0. The amplitude of each particular graph
is computed as follows: at each edge going from v to v 0 we put the function
D(y z), the two-point Schwinger function of the free theory (the Green's
function for the Helmholtz equation). On each remaining external edge we
put (y ). Then we compute formally the product of things on vertices and
edges, and put it under the normal ordering sign. Since for each number
K of connecting edges, we have only one possible graph, we should just
compute the amplitude of this graph, and then multiply it by the number
of ways to get this graph, i.e. by J !J 0 !=K !(J K )!(J 0 K )!.
This will give us a formula of the form
L
X
(3:4) : O(x) :: O0 (x0 ) := : O~k (x)O~k0 (x0) : Ek (x x0 );
k=1
where : O~k :; : O~k0 : are some local functionals, and Ek (y ) are polynomial
dierential operators in D(y ). Here : O(x)O0 (x0) : is the amplitude of the
graph in which the vertices v and v 0 are not connected.
Now we can use the Taylor formula (3.3) inside of the normal ordering,
to get the following (already innite) expansion:
X
(3:5) : O(x) :: O(x0) := : Os (x0) : Ds (x x0);
s
where Os is local functional, and Ds (y ) is a product of the function Ek (y )
for some k and a polynomial in y .
6
Denition. Expansion (3.5) is called the operator product expansion
(OPE).
Now recall that the Green's function D(x) has the following behavior
at x = 0: D(x) jxj2 n . Therefore, all functions dk (x) have the property
jdk (x)j Ckjxjrk for small x, where the numbers rk are almost all positive.
This proves an important property of the OPE: only nitely many terms are
singular. However, these singular terms carry the most interesting informa-
tion in many situations.
Remark. We see that the operator product expansion (3.5) is in fact the
quantum analogue of the Taylor expansion (3.3). Indeed, it is an expansion
of the form similar to (3.3), but also involving nitely many singular terms
at x = x0 .
Let us now consider examples of OPE (we will always write the nite
formula (3.4)). Let us rst compute (x)(x0). Using the above rules, we
get two graphs:
, ,
: J (x) :: J 0
(x0 ) := 0
(3:8) K =0 K !(J K )!(J K )!
D(x x0)K : J K (x)J K (x0) :
0 0
7
The last formula can be written more nicely using generating functions:
(3:9) : e(x) :: e(x ) := eD(x x ) : e(x)e(x ) :;
0 0 0
8
l is suciently big. Now all the loop integrals converge, and for any local
functional O of degree d we can consider the cuto correlation function
h(y1):::(yr)O(x)i. If we take ! 1, we will of course nd that the
limit does not exist. However, it is not diucult to prove the following.
Let Ad be the space of local functionals of dimension d.
Proposition 3.1. There exists a -dependent linear map R : Ad !
Ad , strictly triangular with respect to the ltration of Ad by dimension, such
that for any local functional O 2 Ad there exists a limit
(3:10) lim h(y1):::(yr )RO(x)i ;
!1
Thus, Proposition 3.1 allows us to assign to every local functional O an
operator O~ , which is, by denition, the operator whose matrix elements are
the limits of the corresponding matrix elements of R O. However, given
O, the operator O~ is dened apriori noncanonically, as the map R is not
unique: it is dened up to left multiplication with a -independent strictly
triangular map.
That is, O~ is dened uniquely up to adding composite operators of lower
dimension. This shows that the space of composite operators is naturally a
ltered object (by dimension), and not a graded object.
Remark. Of course, if the theory is not classically scale invariant, we
saw that there is no grading on functionals already at the classical level. The
statement here is that even for a classically scale-invariant theory, where the
space of classical functionals automatically has a grading by scaling dimen-
sion, the grading is usually lost in the process of quantization. An exception
is a free scale-invariant theory, where there is a canonical quantization by
normal ordering, and therefore the grading survives quantization.
Thus, in general we may be able to quantize naturally the space of local
functionals of dimension d, but not every functional separately.
3.6. Composite operators in an interacting critical theory.
Now consider an interacting renormalizable eld theory. As a model
example we consider the Lagrangian (3.2) and take Q() = 4!g 4 . Let O be
any local functional represented by a monomial. To quantize O, we proceed
as in the free case, but we will formulate everything in a slightly dierent
language.
We want to consider correlation functions h(y1):::(yr )O(z )i. These
correlation functions can be viewed as the "-coecient in the usual Schwinger
functions for a perturbed Lagrangian, of the form
(3:11) L"() = L() + "O(; z);
9
where " is a formal variable such that "2 = 0. In the language of Feynman
diagrams, this means that we are introducing an additional vertex v corre-
sponding to O, and summing over all graphs which contain exactly one such
vertex (with no self-loops at this vertex) and are otherwise as usual.
Remark. One should remember that there is no momentum conserva-
tion at the vertex v .
In general, such an alteration will worsen the divergence properties of
the Feynman graphs. More precisely, now the supercial divergence of a
graph with E external edges is given by div ( ) = [O] E . However, we can
renormalize these divergences, using the cuto propagator considered in the
previous section. Then, analogously to Proposition 3.1, one can prove the
following.
Proposition 3.2. There exists a -dependent linear map R : Ad !
Ad , triangular (in general, not strictly) with respect to the ltration of Ad
by dimension, such that for any local functional O 2 Ad there exists a limit
(3:12) lim h(y1):::(yr )RO(x)i ;
!1
As in the free theory, this Proposition allows to quantize the space of
local functionals of dimension d, but in general there is no canonical
quantum analogue for each classical local functional. This non-uniqueness
is not only due to the non-uniqueness of renormalization, but also due to
the non-uniqueness of representation of a given functional by a dierential
polynomial.
Example. Let O = 2=2. Let us compute the renormalization of O of
order g . The only divergent graphs we have in this order are
Let us call the rst graph by 0 and the second by 2 . The graph 0 is
quadratically divergent. If we replace the usual propagator with the cuto
propagator, the integral will converge to a -dependent constant of the form
gC , where C, which grows quadratically in .
Now consider the graph 2 . As usual, it is more convenient to work
in the momentum space, i.e. consider the Fourier transform of the term
corresponding to this graph. Let k1; k2; k be the corresponding momentum
variables. Then the Fourier transform of the term corresponding to 2 is of
10
the form F2 (k1 + k2 + k), where F2 is a function on the plane k1 + k2 + k = 0.
Set k1 = r, then k2 = r k (on this plane). Thus, F2 = F2 (r; k), and its
order g correction is the amplitude of the Feynman diagram
q q+k
r -r-k
and thus
(x)(x0) = 2R(x0 )(1 gA ln j x x0 j) + D(x x0)
(3:20) 2
+ regular part + O(g ): 2
14
Now consider a more complicated operator product, for example
2R(x)2R(x0):
In this case the expansion is of the form
2R(x)2R(x0 ) = 4R(x0)f1 (x x0 ) + (r2R)(x0)f2 (x x0 )
(3:21) + 2R (x0)f3 (x x0 ) + f4 (x x0 )
+ regular part ;
where the subscript R means \renormalized composite operator". Of course
the functions f1 ; :::; f4 will depend on the choice of the renormalization, but
they are well dened up to an upper triangular linear transformation.
For the sake of brevity we will only compute the expansion modulo O(jx
0
x j 1 ). Thus, the functions f1; f2 can be ignored, and we only have to
compute f3; f4 modulo O(jx x0 j 1 ).
The function f4 is dened canonically up to multiplication by a scalar.
As before, this function is computed using the 0-point function: f4 (x x0 ) =
h2R(x)2R(x0)i. The only graph that contributes to the 1-st order in g of f4
is
k k
16
transform of the function C ln jq=j=q 2 modulo O(jx x0 j) 1 and O(g 2). As
before, the constant depends on the choice of renormalization; changing
of this constant is equivalent to adding a multilple of D(z ) to f31 (z ), which
is the same as multiplying f3 (z ) by a scalar of the form 1 + cg .
Remark: Operator product expansion in conformal eld theory.
In a 2-dimensional conformal eld theory, the OPE of composite oper-
ators has especially simple form. In this case, V = C , and the space of
functions on V n 0 has a \bigrading": the function z a zb has bidegree (a; b)
(a b 2 Z). The space of composite operators also has a bigrading: to any
homogeneous operator O one assigns two numbers { the holomorphic di-
mension d(O) andPantiholomorphic dimension d(O0 ) (d d 2 Z). Therefore,
if O(z )O0 (z 0) = k Dk (z z 0 )Ok (z 0 ), then Dk (z z 0 ) has degree (dk ; dk ),
where dk = d(O) + d(O 0) d(Ok ), dk = d(O) + d(O0 ) d(Ok ). This implies
d
d
that Dk (z ) = Ck z k z k . Also, the action of the Virasoro algebra allows to
reduce the problem of computing OPE of arbitrary operators to the problem
of computing OPE for primary elds only, i.e. for elds which are highest
weight vectors for the Virasoro algebra. In a rational conformal eld theory,
one has a chiral algebra of symmetries (for example, an ane Lie algebra)
which is so big that there are only nitely many elds which are primary with
respect to this algebra. This fact allows to treat a rational 2-dimensional
conformal eld theory in a purely algebraic setting, and reduce many of its
problems to problems in algebraic geometry.
17