LECTURE II-16: BRST QUANTIZATION OF GAUGE THEORIES

Edward Witten

Notes by Pavel Etingof and David Kazhdan

In this lecture we will discuss quantization of gauge theory by using BRST coho-
mology. This approach is an improvement of the original Faddeev-Popov approach.
An advantage of the BRST approach as opposed to the Faddeev-Popov method is
that BRST makes explicit the independence of quantization of the choice of the
gauge xing procedure.
A similar approach can be used in gravity (see D'Hoker's lectures).
16.1. The general setup.
We start with a general setup, and then consider examples. In the general setup,
we have a compact gauge group G with Lie algebra g, and the group G^ , which is
the group of gauge transformations of a principal G-bundle E over a spacetime M .
Formally, we want to compute the path integral
1 Z
(16.1) ^ DADe L(A;) ;
V ol(G)
where L(A; ) is a gauge invariant Lagrangian with a gauge eld A and matter
elds .
The diculty with a perturbative treatment of this path integral is that its ki-
netic term FA2 for the gauge eld is degenerate along orbits of G^ . One way to deal
with this diculty is to replace the integrand in (16.1) by some expression that
integrates to 1 on orbits of G^ { then (16.1) would equal to the integral of this
expression (at least if everything were nite dimensional). For example, this ex-
pression could be the delta-function of some gauge, i.e. of some submanifold in the
space of connections and matter elds which is a cross-section for the G^ -action (this
procedure is called gauge xing). As we know (see Kazhdan's lectures on gauge
theory and Faddeev's lectures), this introduces a determinant under the integral
(the Faddeev-Popov determinant). The determinant is a nonlocal expression, so
in order to work only with local expressions, one should replace this determinant
with a Gaussian integral over the space of elds times two copies of the odd space
of sections of the coadjoint bundle of E . Thus, one has to introduce additional
fermionic elds c; c with values in ad(E ). These fermions are called ghosts, since
they do not correspond to any physical particles and violate spin-statistics. Af-
ter the introduction of ghosts the path integral can be treated as usual, e.g. by
perturbation theory techniques.
16.2. The BRST dierential.
Of course, any gauge xing procedure by denition destroys gauge invariance.
Therefore, in order to obtain a sensible quantization, we must make sure that in the
A S
Typeset by -T X
nal result the gauge symmetry is restored. In particular, we must explain what
replaces the gauge symmetry in the ghost setting of the previous section.
It turns out that what replaces the gauge symmetry is a certain odd derivation
of the algebra of local functionals, which we will now construct.
We will rst consider the classical theory. Let us look what elds our theory
has after introduction of ghosts. The basic elds are the connection A, the matter
elds , and the ghosts c; c, which are sections of the adjoint bundle to E . We will
add an auxiliary scalar bosonic eld h, whose signicance will be seen below.
Let R be the algebra of local functionals. We want to dene an odd derivation
 : R ! R such that 2 = 0 (the BRST dierential).
~ , where R~ is the
Recall that the algebra of local functionals R is the quotient R=I
algebra of local expressions in the elds and I is the dierential ideal generated by
eld equations.
We rst dene a derivation  : R~ ! R~, and then make sure that the eld
equations are respected.
Dene  on generators by
(16.2) c = 12 [c; c];  = c; A = dAc; c = h; h = 0;
where c  means the variation of  along the innitesimal gauge transformation c.
It is easy to check that 2 = 0.
Recall from Kazhdan's and D'Hoker's lectures that the Lagrangian with ghosts
for our theory is
2
(16.3) L~ = L(A; ) (c(h=2 + )) = L(A; ) (h + h2 c):
where  = (A; ) is a non-gauge invariant local expression (the gauge xing
function). This Lagrangian is of course equivalent to
L^ = L(A; ) + ( 21 2 + c):
by eliminating h.
We want the Lagrangian to be invariant under the derivation . Since L(A; )
is already invariant2 (because it is gauge invariant), it is enough to check that the
expression h + h2 c is closed under , which follows from 2 = 0.
Since  preserves the Lagrangian, it preserves the set of its critical points and
hence indeed denes a derivation of the algebra of local functionals.
As an example consider the case when E is the trivial bundle (and there is no
matter elds). One of the possible gauge xing conditions is the Feynman gauge
condition d A = 0. So we take  = d A. Then one has  = ddA c, so after
elimination of h the Lagrangian for pure gauge theory is
1 Z 2  2 
(16.4) L = 4e2 F + ((d A) cd dAc):
We see that this Lagrangian is nondegenerate, so one can do perturbation theory
with it as usual. The derivation  on R in this case is dened by
(16.5) c = 12 [c; c]; A = dAc; c = dA:
2
 16.3. The properties of the BRST derivation.
Thus, we have constructed a derivation . The main properties of  are:
1. 2 = 0.
2.  is dened on R~ apriori, without the use of the Lagrangian L(A; ) and
the gauge xing term . It preserves the Lagrangian with ghosts L~ and therefore
descends to R.
Now let us turn to quantum theory. In this case local functionals are replaced by
local operators. It can be shown that there exists a renormalization procedure under
which  can be dened as above, and properties 1 and 2 hold. This is discussed
below.
However, in order to use the BRST method for quantization, we will need an-
other, purely quantum, property of . Namely, denote by Leff the eective La-
grangian, i.e. the Lagrangian for which the classical theory is equivalent to the
quantum theory for L. This Lagrangian is of course nonlocal. The third property
of  that we need is
3.  preserves the eective action Leff . That is, Leff = 0.
This property may fail, and if it fails then one says that the theory is anomalous.
Remark. Although Leff is nonlocal, it can be shown that Leff is always the
integral of a local expression. It can be shown that the obstruction to making Leff
zero by adding an auxiliary term to the Lagrangian (in a way that does not change
the physics) is given by a 1-loop calculation. Thus, anomalies arise in the 1-loop
order of perturbation theory, and don't have higher order corrections. We will see
this at the end of the lecture.
16.4. Operators in gauge theory and BRST cohomology.
Assume properties 1-3 hold. Consider the path integral Z given by Lagrangian
L~ :
Z
(16.6) Z = DADcDcDhDe L~
(possibly with some gauge-invariant insertions). Properties 1-3 imply that Z is
independent on the gauge xing condition . Indeed, for any local expression X
we have
Z L~ X = 
Z
(16.7) DADcDcDhDe DADcDcDhDe L~ X = 0;
which implies the independence of Z on .
Remark. In (16.7), we used that  preserves the measure of integration DADcDcDhD.
It is easy to see that this is equivalent to Property 3 (absence of anomalies).
The statement that Z is independent of  holds for operators (insertions) which
are annihilated by ; for example, for any gauge-invariant insertions into Z , de-
pending only on A; . On the other hand, if O = O0 then the integral of O is zero
by (16.7). Thus, the space of \physical" quantum operators in our theory is the
cohomology of . This cohomology is called the BRST cohomology.
The BRST cohomology comes with a natural Z-grading. Namely, we have a
grading in the space of local operators, in which gauge and matter elds have
degree 0, c has degree 1 and c degree 1. This degree is called the ghost number.
It is easy to see that  raises the degree by 1. This allows to introduce a grading
in cohomology: we denote by Hq the cohomology in degree q.
3
 The properties of this cohomology, which usually hold in this situation are:
1. Hq vanishes for q < 0.
2. H0 is the space of gauge invariant local operators depending only on A and
.
This shows that  plays the role of the gauge symmetry which was broken when
ghosts were introduced. Thus, we have established a setting for gauge theory which
works well in perturbation theory and in which the gauge symmetry does not die
but rather appears in the form of .
16.5. Renormalization and BRST dierential.
Now let us discuss the renormalizability and renormalization group equation in
the BRST approach. We will restrict ourselves to 4 dimensions and pure nonabelian
gauge theory.
Recall that in order for a theory to be renormalizable, all interactions have to
have nonnegative dimension. To nd out whether it is so for the Lagrangian with
ghosts, we will compute the dimensions of elds (assuming that  preserves the
scaling dimensions). It is easy to see that the dimensions are as follows: [c] =
0; [c] = 2; [A] = 1. This shows that all interactions in the Lagrangian with ghosts
are renormalizable.
Remark. In this theory, dimensions of c and c are not uniquely determined;
the only thing that is determined canonically is [cc], which equals 2. This does
not lead to a contradiction, since all operators of nonzero ghost number have zero
expectation value, and so their scaling dimension has no intrinsic meaning. This is
why we needed to make an additional assumption that  preserves dimensions to
x precise values of the dimensions. If we had assumed that  raises dimension by
k we would get a dierent answer, which would be equally good for our purposes.
Now let us look for critical couplings which will be renormalized. In the set-
ting without ghosts, the usual thing to do is to write down renormalizable (non-
gauge invariant) operators of dimension 4, which correspond to critical couplings:
[A; A]2 ; [A; A]dA; (dA)2 , and then argue that there is only one gauge invariant com-
bination of these operators, so that the only coupling which is to be renormalized
is the charge e. However, in the setting with ghosts, we also have to include oper-
ators of degree 4 involving ghosts: ccA2; cd dc; :::. The gauge invariance condition
is now replaced by the condition that  is a symmetry, so we need to renormalize
only delta-invariant interactions. This cuts down the number of operators to be
renormalized, but still leaves us with two renormalizable couplings: the charge e
and the2 gauge xing parameter a, corresponding to the scaling of the gauge xing
term 2 + c. Thus the renormalization group vector eld looks like
(16.7) @ +  (e) @ + ~(e; a) @ :
W =  @ @e @a
Here  is the beta-function of the theory, and ~ is the ghost beta-function. The beta-
function of the theory depends only on e and is physically meaningful; for example,
the negativity of its leading term insures asymptotic freedom. However, the ghost
beta-function ~ has no physical meaning: it only matters for renormalization of
operators and correlators containing c and c, which don't make sense physically.
16.6. The Hamiltonian approach.
So far we have considered BRST method from the Lagrangian point of view.
Now let us consider the connection of the BRST method with the Hamiltonian
formalism.
4
 Since ghosts violate spin-statistics (being scalar fermions), the \Hilbert space"
of the theory with ghosts cannot be an actual Hilbert space. Namely, it is possible
to construct a certain space H~ with a Hermitian form, which is analogous to the
Hilbert space in actual physical theories, but the form will not be positive denite.
However, on this space we have local quantum operators, obtained by quantization
of classical operators in the usual way. In particular, we have the global charges
{ the Hamiltonian H as well as the BRST charge Q, obtained from ghosts as
explained in D'Hokers lecture. We also have a grading of H~ by ghost number,
obtained naturally from the quantization procedure.
The operator Q has the property [Q; O] = O for any operator O (not necessarily
-closed) in the theory with ghosts. Also, Q
= 0, where
is the vacuum, and the
ghost number of Q is 1.
The operator Q has properties analogous to those of :
1. Q2 = 0. This can be conrmed by a direct computation.
2. Q is dened apriori, without the use of L and  (by an explicit formula
as in D'Hoker's lectures). In particular, if H~ is an irreducible representation of
the operator algebra, then Q is completely determined by  and the properties
[Q; O] = O, Q
= 0.
3. If there is no anomalies, the element Q commutes with the Hamiltonian and
with all gauge invariant local operators which involve no ghosts.
Remark. As in the Lagrangian setting, here the explicit expression for Q is
independent on  only if one uses the operator h corresponding to the auxiliary
eld in the Lagrangian. This operator can be expressed via other operators in the
theory, in a way which depends on : h = . If one makes this substitution, the
obtained formula for Q will involve . Thus, property 2 should be understood as
follows: there exists a formula for Q in terms of the elds (including h!) which is
independent on L and  but depends only on eld conguration.
Let HQq be the cohomology of Q on H~ , graded by ghost number.
The properties of Q which usually hold are
1. HQq vanishes if q < 0.
2. The Hermitian form is degenerate on the kernel of Q in H~0 (operators of
ghost number zero); the kernel of this form is the image of Q. The induced form
on HQ0 is positive denite.
The space HQ0 plays the role of the physical Hilbert space of the theory, so we
denote it by Hphys .
In the space Hphys , we have an action of the Hamiltonian H and \physical"
local operators O 2 H0 . These operators no longer involve ghosts and correspond
to actual observables of the theory.
Let us now compare the BRST and the \traditional" approaches to quantization
of gauge theory. For simplicity, we consider pure gauge theory. Traditionally,
a scheme of quantization would be as follows. Suppose that the space part of
the spacetime is compact. In this case we have seen that classically the space of
solutions to the equations of motion can be realized as T  A, where A is the space
of connections on a space cycle modulo gauge transformations. Therefore we would
dene the Hilbert space as L2(A) (with respect to some measure). We call this
Hilbert space the traditional Hilbert space.
We claim that these approaches give the same result, i.e. Hphys is isomorphic to
Htrad as a representation of the operator algebra.
5
 First of all, Hphys does not depend on the gauge xing term , and the Hamilton-
ian and the quantum operators in Hphys don't depend on it either. This follows from
the fact that when  is varied, operators in the pseudo-Hilbert space H~ are changed
by adding a -exact expression, so their action on -closed vectors is changed by a
-exact expression.
To identify Hphys with Htrad we can use a convenient gauge xing term . It is
enough to do it for one such term, but we will do it for two { just for fun.
Set  = ud A + vA0 , where A0 is a time component of the connection (this
uses the splitting of spacetime into space and time). Then we get a sensible theory
unless both u = 0 and v = 0. Even u = 0; v 6= 0 gives a nice theory { this gauge
xing term is called \temporal gauge".
We rst consider the case u = 0. Then the Lagrangian is
Z
^L = ( 12 F 2 + 1 v2 A20 vcDc );
4e 2 Dt
where D denotes covariant derivative. Replacing c with vc, and tending v to
innity (using the fact that nothing depends on v), we see that the path integral is
localized to the hyperplane A0 = 0, and in the limit we get a Lagrangian
Z 1 2 dc
(16.8) L^ = ( 4e2 F + c dt );
Since A0 is now zero, we get usual quantum mechanics where dynamical variables
are a spacial connection, its time derivative, and the ghosts. Thus, H~ has the form
O(A~)
^g , where A~ is the space of connections on the space cycle, O(A~) is the
space of functions on A~, g~ is the Lie algebra of the group of gauge transformations
on the space cycle, and ~g is the space of functions of c 2 ~g. Moreover, the
Hamiltonian for the ghosts vanishes since there is no nontrivial evolution on the
space of classical solutions (the Euler-Lagrange equations for ghosts are simply
dc = dc = 0). Thus the Hamiltonian in our theory is the usual gauge theory
dt dt
Hamiltonian
1 Z 3 2 e2 2
(16.9) H = 2e2 d xFA + 2 rA; A 2 A~
acting on the rst component of the tensor product.
It is easy to see that the space H~ with the grading by ghost number is nothing
but the space of the standard complex of the Lie algebra ~g with coecients in the
module O(A~). Moreover, from the explicit formula for Q one gets that in this case Q
is exactly the dierential in the standard complex. Thus, the physical Hilbert space
Hphys which is by denition the 0-th cohomology of Q, is the 0-th the cohomology
of g~ with coecients in O(A~). This is just the space of g~-invariants in O(A~), i.e.
the space of functions on A~=G, which is by denition the traditional Hilbert space
Htrad .
This shows that BRST cohomology is an innite dimensional generalization of
Lie algebra cohomology.
Now consider another gauge obtained by setting v = 0; u = 1:  = d A. Let us
try to see the isomorphism between Hphys and Htrad using this gauge. We have,
1 Z 2 1  2 
(16.10) L^ = 4e2 F + ( 2 (d A) cD dA c):
6
In this case the equations of motion for ghosts are nontrivial and of second order,
so the Hilbert space H~ consists of functions of As ; c; c; A0 , where As is a connection
on the space cycle. In this case, one nds
(16.11) Q = QLie algebra cohomology + Q0 ;
R
where Q0 = cA0 , and c, A0 are the momentum operators for c, A0 .
It is easy to check directly that the two summands in (16.11) anticommute, and
that (Q0 )2 = 0. It can also be checked that Q0 is acyclic except in 0-th degree,
where it has a 1-dimensional cohomology. Thus, by Kunneth formula, we again get
Hphys = Htrad .
16.7. Anomalies.
Now let us recall conditions 1,2,3 which were necessary for the BRST construc-
tion, and analyze when they are satised. These conditions are
1. Lagrangian: 2 = 0.
Hamiltonian: Q2 = 0.
2. Lagrangian:  is independent on  and L~.
Hamiltonian: Q is independent of ; L~.
3. Lagrangian: Leff = 0.
Hamiltonian: [Q; H ] = 0.
As we mentioned, properties 1 and 2 can always be attained.
However, as we also mentioned, Property 3 may fail if anomalies are present. So
let us consider anomalies more closely.
Consider a 4-dimensional gauge theory with a chiral spinor with values in a
complex representation  of the gauge group G and antichiral spinor  with values
in the dual representation . The basic gauge-invariant Lagrangian for such elds
is
1 Z 2 Z
(16.12) L = 4e2 F + DA :
R
In quantum theory we are interested in the path integral e L . Integrating out
in this path integral, we get
Z
(16.13) det(DA )e L(A)DA;

where DA : (S+) ! (S ) is the covariant Dirac operator and L(A) the La-
grangian of the pure gauge theory.
Integral (16.13) may not have a gauge invariant regularization. What is worse,
it may not even have a non-gauge-invariant regularization for which the gauge
invariance is restored as the cuto goes to innity. In this case the gauge theory we
are considering does not make sense quantum mechanically, even in perturbation
theory, because gauge symmetry cannot be restored. This phenomenon is called an
anomaly.
A geometric reason for an anomaly is that although the operator DA is gauge
invariant, its determinant det(DA ), in general, fails to be gauge invariant. In other
words, this determinant is not a function on the space of gauge classes of connections
but rather a section of some line bundle over this space, called the determinant line
7
bundle; this bundle comes with a canonical connection. If this canonical connection
does not trivialize the bundle, this \function" cannot be sensibly integrated.
It is useful to distinguish two types of anomalies.
1. Local anomaly. The canonical connection has a nonzero curvature. In this case
for suitable spacetime manifolds this curvature may represent a nontrivial second
cohomology class, so that the determinant bundle is not trivial topologically.
2. Global anomaly. The canonical connection is at but has a nontrivial mon-
odromy (and possibly the bundle is not trivial).
Thus, both local and global anomalies can produce topological anomalies, but
only the rst one can be seen in perturbation theory (by computing of the curva-
ture).
Here we will consider only local anomalies.
Remark 1. To analyze when we can expect local anomalies, one may consider
the situation from the topological point of view. We assume that our spacetime M
is compact and orientable (e.g. S d), with a specied point 1, and will consider
bundles, connections, and gauge transformations which are trivial at innity. In
this case the space of gauge classes of connections can be regarded as a classifying
space BG^ for the group of gauge transformations G^ . Nontrivial line bundles on BG^
are classied by H 2 (BG^ ).
Now, if M is compact and orientable, we have the transgression map  : H2(BG^ ) !
Hd+2(BG) dened as follows: given a two-dimensional homology class, pick a sur-
face S in BG^ which represents it, and take the corresponding principal G^ -bundle
on S . Its transition functions can be considered as transition functions of a G-
bundle on the 6-dimensional manifold S  M , which denes an element  ([S ]) of
Hd+2(BG). Consider the dual map   : H d+2 (BG) ! H 2 (BG^ ). It can be shown
that the Chern class of our line bundle is  (C ), where C is a xed d +2-dimensional
characteristic class which does not depend on M , and is computed locally from the
curvature. This class is exactly the local anomaly.
Thus for d = 4 local anomalies live in H 6 (BG), or (S 3g)g , where g is the Lie
algebra of G.
For example, in the standard model the gauge group is SU (3)  SU (2)  U (1),
and thus the space of anomalies (S 3g)g is 4-dimensional: it equals to the sum of
four subspaces Sinv3 (su(3)); S 3 (u(1)); S 2 (su(2))
u(1); S 2 (su(3))
u(1), which
inv inv inv
are 1-dimensional (here \inv" denotes that we are taking invariant symmetric poly-
nomials).
This discussion illustrates why anomalies don't arise in the case when all fermions
are in a real representation of the gauge group. Indeed, in this case, the determinant
bundle is real, and thus its Chern class must be zero.
Remark 2. Although the local anomaly can be considered from the above
topological point of view, one should remember that it has a purely local nature,
and has nothing to do with the macrostructure of the spacetime. If there is a local
anomaly, the quantum theory will not make sense on any spacetime, even on Rd.
The problem is that even if the determinant bundle is topologically trivial, it will
not have a at connection dened in a local way: otherwise this at connection
would have been good for any simply connected spacetime, and no topological
anomaly would arise. Thus, path integral (16.13) is not sensible even on Rd.
Remark 3. In the standard model, the gauge group is SU (3)  SU (2)  U (1). In
particular, there is a possibility for local anomalies, and they do appear in reality.
8
However, one can check that the anomalies coming from the dierent matter elds
of the standard model miraculously cancel, in all four components of the space of
anomalies. An explanation of this is that the representation of the gauge group in
the standard model extends (after adding some insignicant summands) to a spinor
representation of Spin10, for which H 6 (BG) vanishes.
Let us show how to analyse anomaly in perturbation theory. Our goal is to
explain why, after possibly enlarging the space of elds, properties 1. and 2. of
section 16.3 can always be assumed to hold (that is, 2 = 0 and  is dened
independently of the choice of a particular Lagrangian) but one cannot assume
that the eective Lagrangian is delta-invariant.
First, let us just try to make sense of integral (16.3) perturbatively. When we
write down Feynman diagrams, we will nd divergences in the 1-loop order which
we cannot remove in a gauge invariant fashion. To x the 1-loop order, we will
regularize the path integral by adding another, very heavy matter eld  such that
its determinant bundle is inverse to that for . In favorable cases, our original
theory should be recovered from this theory in the limit when the mass m of  goes
to innity. In other cases, the procedure will exhibit why there is an anomaly.
To satisfy this condition, the matter eld  can be taken to be a bosonic eld
(+ ;  ) with values in (S+  S )
. In this case the complex conjugate eld
 is with values in (S+  S )
, where S+; S are the spin representations of
the Poincare (recall that both S+ and S are self-dual and self-complex-conjugate
in Euclidean signature). It is of course needless to say that these elds violate
spin-statistics and therefore, like ghosts, don't make physical sense.
The natural Lagrangian term for the elds  would be
Z
(16.14) L0 (A;  ) = d4x((+ ; DA  )+( ; DA +)+ m(+; + )+ m( ;  ))
(Here the Dirac operator is skewselfadjoint).
Remark. The 's are called Pauli-Villars regulator elds.
If we add expression (16.14) to the Lagrangian, we will get the squared absloute
value of the determinant rather than the determinant itself, and will not x the
anomaly. Thus, we modify (16.14) in a way that breaks the gauge invariance: we
let A0 be a xed connection and set
(16.15)
00
Z 4
L (A;  ) = d x((+ ; DA  ) + ( ; DA0 +) + m(+; +) + m( ;  ))
Now consider the theory with the Lagrangian L + L00 . Integrating out the  elds,
we will get a factor det(DA DA0 m2) 1 . For m = 0 this factor is gauge invariant
up to a multiplicative constant, and cancels the determinant in the numerator,
which is caused by anomaly. This shows that in this theory, we don't have a
topological anomaly for any nite m (i.e. the appropriate determinant bundle is
trivial). However, for m > 0, the gauge invariance fails. So we have to study the
limit m ! 1 (which is supposed to recover our original theory) and see whether
the gauge symmetry reappears.
Dierentiating the determinant ratio det(DA DA0 m2)=det(DA ) in the direction
R t 2 g^, we obtain (using the path integral interpretation)
of a gauge trasformation
that it is equal to mh [(+ ; t+) + ( ; t )]i, where hX i denotes the expectation
value of X . This expectation value has a decomposition in powers of 1=m.
9
 To see whether the failure of gauge invariance persists for m ! 1, let us consider
the two-point function of the curvature operator F . It is easy to see that the
leading contribution (in 1=m) to the derivative of this function in the direction of
t is from the 1-loop diagram with a  loop having the t operator inside and two
P dabctaFFb-edges.
outgoing This contribution is of the 0-th order in 1=m, and has the form
F c , where dabc is some tensor. So if dabc 6= 0, the gauge-noninvariance
remains in the limit.
Remark. In case the original fermions were in a real representation, dabc is zero
and the regularization in (16.14) is completely satisfactory. The problem arises
when the original representation is complex. Then a regularization as in (16.14)
doesn't work unless one gives up gauge invariance.
Remark. When dabc 6= 0, one can choose a regularization scheme to remove all
loop contributions to the non-gauge invariance except 1 loop.
Now let us consider anomalies from the BRST veiwpoint. If a local anomaly is
present, we will have U = Leff (A) 6= 0 (here Leff (A) is the eective Lagrangian
for A, with the ghosts integrated out). However, since the anomaly is local U must
be the integral of a local expression of A and c which is linear in c. It is also
clear that U = 0. Furthermore, one can show that U involves only A and its rst
derivatives (and no matter elds).
On the other hand, if U is N where N is the integral of some local expression
of A then we can arrange Leff = 0 by redening the Lagrangian as L ! L + N .
Thus, anomalies lie in the cohomology of  on local functionals of degree  1 of A
and c (linear in c) modulo complete derivatives.
Now let us show that such cocycles are in fact related to invariant symmetric
tensors on the Lie algebra g (or equivalently, the cohomology of g).
Let C be a G-invariant element in S n+1g. To this element there corresponds a
2n + 2-dimensional characteristic class of G-bundles, namely C (F n+1), where F is
the curvature. The Chern-Simons form CSC (A) corresponding to C is the local
n+1
2n + 1 form such that ~C(~FA ) = CSC (A) ^ ~A modulo dierentials of local forms
(here ~ denotes the variation to distinguish from the BRST dierential ).
The main property of the Chern-Simons form is the following. Although this form
is not gauge invariant, its Lie derivative along an innitesimal gauge transformation
is a dierential of a local form.
Now let M 2n be our spacetime. LetR A be a connection on M 2n . We want to
dene a functional of the form U (A) = W (A), where W (A) is a 2n-form on M 2n
which is local in A but not gauge invariant, and such that W = 0.
Let X 2n+1 is a smooth manifold with boundary equal to M 2n . Choose an ex-
tension of the connection A to X 2n+1 in any way (for simplicity we assume that
there is no topological obstruction to the choices of XRand the extension of A; this
assumption is in fact inessential). Now set V (A) = X CSC (A). This functional
depends on the extension of A to X . However, by the main property of CS , the
functional V (A) = W (A; c) (where  is the BRST dierential) does not depend
on the extension and therefore is a local functional in A and c linear in c. One can
show that it represents a nontrivial cohomology class in the local -cohomology.
Thus, we get an injective map S n+1(g)g ! H1;local. For 4-dimensional theories
n = 2 and the cocycles come from (S 3g)g = H 6 (BG) as we expected.

10