a

r
X
i
v
:
m
a
t
h
/
0
4
0
6
2
5
1
v
1


[
m
a
t
h
.
G
T
]


1
2

J
u
n

2
0
0
4
FEYNMAN DIAGRAMS FOR PEDESTRIANS AND
MATHEMATICIANS
MICHAEL POLYAK
Dedicated to Dennis Sullivan on the occasion of his 60th birthday
1. Introduction
1.1. About these lecture notes. For centuries physics was a potent source pro-
viding mathematics with interesting ideas and problems. In the last decades some-
thing new started to happen: physicists started to provide mathematicians also
with technical tools, methods, and solutions. This process seem to be especially
strong in geometry and low-dimensional topology. It is enough to mention the
mirror conjecture, Seiberg-Witten invariants, quantum knot invariants, etc.
Mathematicians, however, en masse failed to learn modern physics. There seem
to be two main obstructions. Firstly, there are few textbooks in modern physics
written in terms accessible for mathematicians. Mathematicians and physicists
speak two dierent languages, and a good physical-mathematical dictionary is
missing
1
. Thus, to learn something from a physical textbook, a mathematician
should start from a hard and time-consuming process of learning the physical jargon.
Secondly, mathematicians consider (and often rightly so) many physical methods
and results to be non-rigorous and do not consider them seriously. In particular,
path integrals still remain quite problematic from a mathematical point of view
(due to some usually unclear measure aspects), so mathematicians are reluctant to
accept any results obtained by using path integrals. Yet, this technique may be put
to good use, if at least as a tool to guess an answer to a mathematical problem.
In these notes I will focus on perturbative expansions of path integrals near a
critical point of the action. This can be done by a standard physical technique of
Feynman diagrams expansion, which is a useful book-keeping device for keeping
track of all terms in such perturbative series. I will give a rigorous mathematical
treatment of this technique in a nite dimensional case (when it actually belongs
more to a course of multivariable calculus than to physics), and then use a simple
dictionary to translate these results to a general innite dimensional case.
As a result, we will obtain a recipe how to write Feynman diagram expansions
for various physical theories. While in general an input of such a recipe includes
path integrals, and thus is not well-dened mathematically, it may be used purely
formally for producing Feynman diagram series with certain expected properties.
A usual trick is then to sweep under the carpet all references to the underlying
2000 Mathematics Subject Classication. Primary: 81T18, 81Q30, Secondary: 57M27, 57R56.
Key words and phrases. Feynman diagrams, gauge-xing, Chern-Simons theory, knots, con-
guration spaces.
Partially supported by the ISF grant 86/01 and the Loewengart research fund.
1
With a notable exception of [9], which is somewhat heavy.
1
2 MICHAEL POLYAK
physical theory, keeping only the resulting series. Their expected properties often
can be proved rigorously, directly from their denition.
I will illustrate these ideas on the interesting example of the Chern-Simons the-
ory, which leads to universal nite type invariants of knots and 3-manifolds.
A word of caution: during the whole treatment I will brush aside all questions
of measures, convergence, and such; see the discussion in Section 4.5.
1.2. Basics of classical and quantum eld theories. The remaining part of
this section is a brief sketch on the physical level of rigor of some basic notions
and physical jargon used in the quantum eld theory (QFT). Its purpose is to give
a basic mathematical dictionary of QFTs and a motivation for our consideration
of Gaussian-type integrals in this note. An impatient reader may skip it without
much harm and pass directly to Section 2. Good introductions to eld theories
can be found e.g. in [12], [21]; mathematical overview can be found in [9]; various
topological aspects of QFT are well-presented in [23]. Very roughly, by a eld theory
one usually means the following.
Given a space-time manifold X, one considers a space T of elds, which are
functions of some kind on X (or, more generally, sections of bundles on X). A
Lagrangian L : T R on T gives rise to the action functional S : T R dened
by
S() =
_
X
L()dx.
In classical eld theory one studies critical points of the action S (classical
trajectories of particles). These elds can be found from the variation principle
S = 0, which is simply an innite-dimensional version of a standard method for
nding the critical points of a smooth function f : R R by solving
df
dx
= 0.
In the quantum eld theory one considers instead a partition function given by
a path integral
(1) Z =
_
F
e
ikS()
T.
over the space of elds, for a constant k R and some formal measure T on T.
This is the point where mathematicians usually stop, since usually such measures
are ill-dened. But let this not disturb us.
In the quasi-classical limit k , the stationary phase method (see e.g. [8]
and also Exercise 2.5) states that under some reasonable assumptions about the
behavior of S this fast-oscillating integral localizes on the critical points of S, so
one recovers the classical case.
The expectation value f) of an observable f : T R is
f) =
1
Z
_
F
T e
ikS()
f().
For a collection f
1
, . . . , f
m
of observables their correlation function is
f
1
, . . . , f
m
) =
1
Z
_
F
T e
ikS()
n

i=1
f
i
().
By solving a theory one usually means a calculation of these integrals or their
asymptotics at k .
FEYNMAN DIAGRAMS FOR PEDESTRIANS AND MATHEMATICIANS 3
Increasingly often, due to a simpler behavior and better convergence properties,
one considers instead the Euclidean partition function, equally well encoding phys-
ical information (and related to (1) by a certain analytic continuation in the time
domain, called Euclidean, or Wick, rotation):
(2) Z =
_
F
e
kS()
T.
Since at present a general mathematical treatment of path integrals is lacking,
we will rst consider a nite dimensional case.
1.3. Finite-dimensional version of QFT. Let us take T = R
d
as the space of
elds. An action S and observables f
i
are then just functions R
d
R. For a
constant k R, consider the partition function Z =
_
R
d
dxe
kS(x)
and the correla-
tion functions f
1
, . . . , f
m
) = Z
1
_
R
d
dx e
kS(x)

i
f
i
(x). We are interested in the
behavior of Z and f
1
, . . . , f
m
) in the quasi-classical limit k .
A well-known stationary phase method states that for large k the main contri-
bution to Z and f
1
, . . . , f
m
) comes from some small neighborhoods of the points x
where S/x = 0. Thus it suces to study a behavior of Z and f
1
, . . . , f
m
) near
such a point x
0
. Considering the Taylor expansions of S and f
i
in x
0
(and noticing
that the linear terms in the expansion of S vanish), after an appropriate changes
of coordinates we arrive to the following problem: study integrals
_
R
d
dx e

1
2
x,Ax+U(x)
P(x)
for some bilinear form A, higher order terms U(x), and monomials P(x) in the
coordinates x
i
.
Further in these notes we will calculate such integrals explicitly. To keep track
of all terms appearing in these calculations, we will use Feynman diagrams as a
simple book-keeping device. See the notes of Kazhdan in [9] for a more in-depth
treatment.
2. Finite-dimensional Feynman diagrams
2.1. Gauss integrals. Recall a well-known formula for the Gauss integral (ob-
tained by calculating the square of this integral in polar coordinates):
Proposition 2.1.
_

dxe

1
2
ax
2
=
_
2
a
.
More generally, let A = (A
ij
) be a real d d positive-denite matrix, x =
(x
1
, . . . , x
d
) the Euclidean coordinates in V = R
d
, and , ) : (R
d
)

R
d
R the
standard pairing x
i
, x
j
) =
j
i
. Then
Proposition 2.2.
(3) Z
0
=
_
R
d
dx e

1
2
Ax,x
=
_
det
A
2
_

1
2
.
Indeed, by an orthogonal transformation (which does not change the integral)
we can diagonalize A and apply the previous formula in each coordinate.
4 MICHAEL POLYAK
Remark 2.3. In a more formal setting, this may be considered as an equality for a
positive-denite symmetric operator A : V V

from a d-dimensional vector space

V to its dual (and , ) : V

V R). Indeed, A induces det A :

d
V
d
V

,
so that det A (
d
V

)
2
and (det A)

1
2
[
d
V [. Hence equality (3) with R
d
changed to V still makes sense if we consider both sides as elements of [
d
V [. In a
similar way, for C-valued symmetric operator A : V V

with a positive-denite
ImA one has
_
V
dx e
i
2
Ax,x
=
_
det
A
2i
_

1
2
where now both sides belong to [
d
V [
C
.
A more general form of equation (3) is obtained by adding a linear term b, x)
with b (R
d
)

to the exponent: dene Z

b
by
(4) Z
b
=
_
dx e

1
2
Ax,x+b,x
.
Then, by a change x x A
1
b of coordinates, we obtain
Proposition 2.4.
(5) Z
b
=
_
det
A
2
_

1
2
e
1
2
b,A
1
b
= Z
0
e
1
2
b,A
1
b
.
Exercise 2.5. Verify the stationary phase method in the simplest case: use an
appropriate change of coordinates to pass
from
_

dx e
k(
ax
2
2
+bx)
to
_

dx e

x
2
2
.
What happens to a small -neighborhood of the critical point x
0
= b/a under this
change of coordinates? Conclude that in the limit k integration over a small
neighborhood of x
0
= b/a gives the same leading term in the expansion of this
integral in powers of k, as integration over the whole of R.
2.2. Correlation functions. The correlators f
1
, . . . , f
m
) of m functions f
i
:
R
d
R (also called m-point functions) are dened by plugging the product of
these functions in the integrand and normalizing:
(6) f
1
, f
2
, . . . , f
m
) =
1
Z
0
_
dx e

1
2
Ax,x
f
1
(x) . . . f
m
(x).
They may be computed using Z
b
. Indeed, notice that

b
i
_
dx e

1
2
Ax,x+b,x
=
_
dx e

1
2
Ax,x+b,x
x
i
,
hence for correlators of any (not necessary distinct) coordinate functions we have
(7) x
i1
, . . . , x
im
) =
1
Z
0

i1
. . .
im
Z
b

b=0
=
i1
. . .
im
e
1
2
b,A
1
b

b=0
where we denoted
i
= /b
i
.
In particular, 2-point functions are given by the Hessian matrix

2
b
2
(Z
b
/Z
0
)

b=0
with the matrix elements
(8) x
i
, x
j
) =
i

j
e
1
2
b,A
1
b

b=0
= (A
1
)
ij
.
FEYNMAN DIAGRAMS FOR PEDESTRIANS AND MATHEMATICIANS 5
Thus the bilinear pairing Sym
2
(V

) R given by 2-point functions is just the

pairing determined by A
1
. This explains the similarity of our notations for the
2-point functions and , ) : V

V R.
For polynomials, or more generally, formal power series f
1
, . . . , f
m
in the coordi-
nates we may apply (7) (with i
1
= i
2
= = i
n
= i for each monomial (x
i
)
n
) and
then put the series back together, noting that each x
i
should be substituted by
i
.
This yields:
Proposition 2.6.
(9) f
1
, f
2
, . . . , f
m
) = f
1
_

b
_
. . . f
m
_

b
_
e
1
2
b,A
1
b

b=0
2.3. Wicks theorem. Denote by A
ij
the matrix elements (A
1
)
ij
of A
1
. The
key ingredient of the Feynman diagrams technique is Wicks theorem (see e.g. [22])
which we state in its simplest form:
Theorem 2.7 (Wick).
(10)
i1
. . .
im
e
1
2
b,A
1
b

b=0
=
_

A
j1j2
. . . A
jm1jm
, m = 2n
0, m = 2n + 1
where the sum is over all partitions (j
1
, j
2
),. . . , (j
m1
, j
m
) in pairs of the set
i
1
,i
2
,. . . ,i
m
of indices.
Proof. For each k, the expression
i1
. . .
i
k
e
1
2
b,A
1
b
, considered as a function of
b, is always of the form P
i1...i
k
(b)e
1
2
b,A
1
b
, where P
i1...i
k
(b) is a polynomial. Each
new derivative
j
acts either on the polynomial part, or on the exponent, by the
rule

j
_
P(b)e
1
2
b,A
1
b
_
=
j
(P(b))e
1
2
b,A
1
b
+P(b)(

i
A
ji
b
i
)e
1
2
b,A
1
b
,
so the polynomial part P
i1...im
(b) may be dened recursively by P

(b) = 1 and
(11) P
i1...im
(b) = (
i1
+

i
A
i1i
b
i
)P
i2...im
(b) = . . .
= (
i1
+

i
A
i1i
b
i
) . . . (
im
+

i
A
imi
b
i
)1,
where 1 is the function identically equal to 1. We are interested in the constant
term P
i1...im
(0). Directly from (11) we can make two observations. Firstly, if m
is odd, P
i1...im
(b) contains only terms of odd degrees, in particular P
i1...im
(0) = 0.
Secondly, unless each derivative
i
k
, k < m acts on the term

i
A
i
l
i
b
i
in some
l-th, l > k, factor of the (11), the evaluation at b = 0 would give zero. Each such
pair (i
k
, i
l
) contributes a factor of A
i
l
i
k
to the constant term of P
i1...im
. These
observations prove the theorem.
It is convenient to extend (10) by linearity to arbitrary linear functions of the
coordinates, note that in this case we may dene the 2-point functions f, g) by
f, A
1
g) in view of (8), and nally combine it with (7) into the following version
of Wicks theorem:
6 MICHAEL POLYAK
Theorem 2.8 (Wick). Let f
1
(x), . . . , f
m
(x) be arbitrary linear functions of the
coordinates x
i
. Then all m-point functions vanish for odd m. For m = 2n one has
(12) f
1
, . . . , f
m
) =

f
i1
, f
i2
) . . . f
im1
, f
im
),
where the sum is over all pairings (i
1
, i
2
),. . . , (i
m1
, i
m
) of 1, . . . , m and the 2-point
functions f
j
, f
k
) are given by f
j
, A
1
f
k
).
Remark 2.9. Another idea for a proof of Theorem 2.8 is the following. Note that
both sides of (12) are symmetric functions of 1, . . . , m, so they may be considered
as functions on m-th symmetric power S
m
(V ) of V = R
d
. Thus it suces to check
(12) only for f
1
= = f
m
= f; in this case it is obvious.
Exercise 2.10. Check that the number of all pairings of 1, . . . , 2n is (2n)!/2
n
n!.
Calculate
_

dx x
m
e
x
2
/2
using integration by parts and Proposition 2.1. Calculate
d
m
dx
m
e
x
2
/2

x=0
substituting x
2
/2 instead of x in the Taylor series expansion of e
x
. Compare these
expressions and explain how are they related to the above number of pairings.
Exercise 2.11. Find formulas for the 4-point functions x
1
, x
1
, x
2
, x
3
) and x
1
, x
1
, x
1
, x
2
).
2.4. First Feynman graphs. It is convenient to represent each term
f
i1
, f
i2
) . . . f
im1
, f
im
)
in Wicks formula (12) by a simple graph. Indeed, consider m points, with the k-th
point representing f
k
. A pairing of 1, . . . , 2n gives a natural way to connect these
points by n edges, with an edge (a propagator in the physical jargon) e = (j, k)
representing A
1
e
= f
j
, A
1
f
k
). Equation (12) becomes then
(13) f
1
, . . . , f
m
) =

eedges()
A
1
e
,
where the sum is over all univalent graphs as above.
Example 2.12. An application of equation (13) for n = 2 (see Figure 1a) gives the
following:
x
1
, x
2
, x
3
, x
4
) = A
12
A
34
+A
13
A
24
+A
14
A
23
,
x
1
, x
1
, x
2
, x
2
) = A
11
A
22
+ 2A
12
A
12
,
x
1
, x
1
, x
1
, x
1
) = 3A
11
A
11
.
2.5. Adding a potential. The above computations may be further generalized by
adding a potential function U(x) (with some small parameter = k
1
) to Ax, x)
in the denition of Z
0
. Namely, dene
2
Z
U
by
(14) Z
U
=
_
dx e

1
2
Ax,x+U(x)
.
Applying (9) for f = e
U(x)
we get:
2
Again, let me remind that we ignore problems of convergence: for most U(x) this integral will
be divergent!
FEYNMAN DIAGRAMS FOR PEDESTRIANS AND MATHEMATICIANS 7
2 4
1 3 1
2 4
3 1 3
2 4
a b
Figure 1. Terms of x
1
, x
2
, x
3
, x
4
) and graphs of degree two
Proposition 2.13.
(15) Z
U
= Z
0
e
U(

b
)
e
1
2
b,A
1
b

b=0
.
Correlation functions f
1
, . . . , f
m
)
U
are dened similarly to (6):
(16) f
1
, f
2
, . . . , f
m
)
U
=
1
Z
U
_
dx e

1
2
Ax,x+U(x)
f
1
(x) . . . f
k
(x).
Using (9) once again, we get
Proposition 2.14.
(17) f
1
, f
2
, . . . , f
m
)
U
=
Z
0
Z
U
e
U(

b
)
f
1
_

b
_
. . . f
m
_

b
_
e
1
2
b,A
1
b

b=0
.
2.6. A cubic potential. Consider the important example of a cubic potential
function U(x) =

U
ijk
x
i
x
j
x
k
. Let us compute the expansion of the partition
function (14) in power series in . The coecient of
n
in the expansion of (15) is
Z
0
n!
_
_

i,j,k
U
ijk

k
_
_
n
e
1
2
b,A
1
b

b=0
.
Let us start with the lowest degrees. By Wicks theorem, the coecient of vanishes
and the coecient of
2
is given by
(18)
Z
0
2!

i,j,k

,j

,k

U
ijk
U
i

k

i

i

j

k
e
1
2
b,A
1
b

b=0
=
Z
0
2!

i,j,k

,j

,k

U
ijk
U
i

A
i1i2
A
i3i4
A
i5i6
,
where the last sum is over all pairings (i
1
, i
2
),. . . ,(i
5
, i
6
) of i, j, k, i

, j

, k

. We
may again encode these pairings by labelled graphs, connecting 6 vertices labelled
by i, j, k, i

, j

, k

by three edges (i
1
, i
2
),(i
3
, i
4
),(i
5
, i
6
) representing A
i1i2
A
i3i4
A
i5i6
.
This time, however, we have an additional factor U
ijk
U
i

k
. To represent U
ijk
graphically, let us glue the triple (i, j, k) of univalent vertices in a trivalent vertex;
to preserve the labels, we can write them on the ends of the edges meeting in this
new vertex (i.e., on the star of the vertex). Similarly, we represent U
i

k
by gluing
the remaining triple (i

, j

, k

) of univalent vertices into a second trivalent vertex.

Thus for each of the 6!/(2
3
3!) = 15 pairings of i, j, k, i

, j

, k

we end up with a
graph with two trivalent vertices; we get 6 copies of the -graph and 9 copies of
the dumbbell graph shown in Figure 1b.
Note, however, that each of these labelled graphs is considered up to its auto-
morphisms, i.e. maps of a graph onto itself, mapping edges to edges and vertices
to vertices and preserving the incidence relation. Indeed, while the application of
an automorphism changes the labels, it preserves their pairing (edges) and the way
8 MICHAEL POLYAK
1 2 1 1 2 1 2 2
Figure 2. Degree two graphs with two legs
they are united in triples (vertices), thus corresponds to the same term in the right
hand side of (18). Instead of summing over the automorphism classes of graphs,
we may sum over all labelled graphs, but divide the term corresponding to a graph
by the number [ Aut [ of its automorphisms. E.g., for the -graph of Figure 1b
[ Aut [ = 12, and twelve copies of this graph (which dier only by transpositions
of the labels) all give the same terms U
ijk
U
i

k
A
ii

A
jj

A
kk

.
Also, when summing the resulting expressions over all indices, note that the
terms corresponding to i, j, k, i

, j

, k

and to i

, j

, k

, i, j, k are the same (which will

cancel out with 1/2! in front of the sum). Hence, we may write the coecient of

2
in the following form:
Z
0

1
[ Aut [

labels

v
U
v

e
A
1
e
,
Here the sum is over all trivalent graphs with two vertices and labellings of their
edges, U
v
= U
ijk
for a vertex v with the labels i, j, k of the adjacent edges, and
A
1
e
= A
ij
for an edge e with labels i, j.
Exercise 2.15. Calculate the number of automorphisms of the dumbbell graph of
Figure 1b.
In general, for the coecient of
n
we get the same formula, but with the sum-
mation being over all labelled trivalent graphs with n vertices.
2.7. Correlators for a cubic potential. We may treat m-point functions in a
similar way. Let us rst consider the power series expansion in of Z
U
x
i1
, . . . , x
im
)
U
.
The coecient of
n
is
Z
0
n!
_
_

i,j,k
U
ijk

k
_
_
n

i1
. . .
im
e
1
2
b,A
1
b

b=0
.
Thus it may again be presented by a sum over labelled graphs, with the only
dierence being that now in addition to n trivalent vertices these graphs also have
m ordered legs (i.e. univalent vertices) labelled by i
1
, . . . , i
m
. See Figure 2 for
graphs representing the coecient of
2
in Z
U
x
1
, x
2
)
U
.
However, not all of these graphs will enter in the expression for x
i1
, . . . , x
im
)
U
,
since we should now divide this sum over graphs by Z
U
(represented by a similar
sum, but over graphs with no legs). This will remove all vacuum diagrams, i.e. all
graphs which contain some component with no legs. Indeed, the term corresponding
to a non-connected graph is a product of terms corresponding to each connected
component. Each component with no legs appears also in the expansion of Z
U
and
thus will cancel out after we divide by Z
U
. For example, the rst graph of Figure
2 contains a vacuum -graph component. But it also appears in the expansion of
Z
U
(see Figure 1b). Thus the corresponding factor cancels out after division by
Z
U
. The same happens with the second graph of Figure 2. As a result, only the
FEYNMAN DIAGRAMS FOR PEDESTRIANS AND MATHEMATICIANS 9
last two graphs of Figure 2 will contribute to the coecient of
2
in the expansion
of x
1
, x
2
)
U
.
Example 2.16 (A nite dimensional
3
-theory). Take U
ijk
=
ij

jk
, i.e. U =

i
(x
i
)
3
. Note that since all ends of edges meeting in a vertex are labelled by the
same index, we may instead label the vertices. Thus the calculation rules are quite
simple: we count uni-trivalent graphs; a vertex represents a sum

i
over its labels;
an edge with the ends labelled by i, j represents A
ij
. The coecients of
2
in Z
U
and in x
1
, x
2
)
U
are given by
Z
0

i,j
6(A
ij
)
3
+ 9A
ij
A
ii
A
jj
,

i,j
9A
1i
A
2j
A
ii
A
jj
+ 6A
1i
A
2j
(A
ij
)
2
respectively. We can identify these terms with two graphs of Figure 1b and two
last graphs of Figure 2, respectively. The rst two graphs of Figure 2 represent

i,j
6A
12
(A
ij
)
3
+ 9A
12
A
ij
A
ii
A
jj
.
These terms do appear in the
2
coecient of Z
U
x
1
, x
2
), but cancel out after we
divide it by Z
U
= Z
0
_
1 +
2

i,j
(6(A
ij
)
3
+ 9A
ij
A
ii
A
jj
) +. . .
_
.
2.8. General Feynman graphs. It is now clear how to generalize the above re-
sults to the case of a general potential U(x): the k-th degree term U
i1...i
k
x
i1
. . . x
i
k
of U will lead to an appearance of k-valent vertices representing factors U
i1...i
k
. We
will call such a vertex an internal vertex. We assume that there are no linear and
quadratic terms in the potential, so further we will always assume that all internal
vertices of any Feynman graph are of valence 3; denote their number by [[.
Denote by
0
the set of all graphs with no legs. Also, for m 1, denote by
m
the
set of all non-vacuum (i.e. such that each connected component has at least one
leg) graphs with m ordered legs.
Denoting U
v
= U
i1...i
k
for an internal vertex v with the labels i
1
, . . . , i
k
of adja-
cent edges, and A
1
e
= A
ij
for an edge e with its ends labelled by i, j, we get
Proposition 2.17.
(19) Z
U
= Z
0

||
[ Aut [

labels

v
U
v

e
A
1
e
.
Note that instead of performing the internal summation over all labellings, one
may include the summation over labels of the star of a vertex into the weight of
this vertex.
In a similar way, for m-point functions we get
Proposition 2.18. For even m,
(20) x
i1
, . . . , x
im
)
U
=

||
[ Aut [

labels

v
U
v

e
A
1
e
,
where the sum is over all labelled graphs with m legs labelled by i
1
, . . . , i
m
.
Again, we may include the summation over the labels of the star of an internal
vertex into the weight of this vertex.
10 MICHAEL POLYAK
2.9. Weights of graphs. Let us reformulate the above results using a general
notion of weights of graphs.
Let V be a vector space. A weight system is a collection (a, u
k

k=3
) of a
Sym
2
(V ) and u
k
Sym
k
(V

). A weight systemW denes a weight W

: (V

)
m

R of a graph
m
in the following way. Assign u
k
Sym
k
(V

) to each internal
vertex v of valence k, associating each copy of V

with (an end of) an edge. Also,

to the i-th leg of , i = 1, . . . , m assign some f
i
V

. Now, for each edge contract

two copies of V

associated to its ends using a Sym

2
(V ). After all copies of V

get contracted, we obtain a number W

(f
1
, . . . , f
m
) R.
In our case, a bilinear form A
1
and a potential U(x) determine a weight
system in an obvious way: set a = A
1
and let u
v
to be the degree k part of U(x).
These rules of computing the weights corresponding to a physical theory are called
Feynman rules.
Formulas (19) and (20) above can be reformulated in these terms as
(21)
Z
U
= Z
0

0
1
[ Aut [
W

,
f
1
, . . . , f
m
)
U
=

m
1
[ Aut [
W

(f
1
, . . . , f
m
).
Exercise 2.19 (Finite dimensional
4
-theory). Consider a potential U =

i
(x
i
)
4
.
Formulate the Feynman rules. Find the graphs which contribute to the coecient
of
2
of Z
U
and compute their coecients. Do the same for x
1
, x
2
)
U
. Draw the
graph representing

i
A
12
(A
ii
)
2
; does it appear in the expansion of x
1
, x
2
)
U
and
why?
2.10. Free energy: taking the logarithm. The summation in equation (21) is
over all graphs in
0
, which are plenty. Denote by
0
conn
the subset of all connected
graphs in
0
. There is a simple way to leave only a sum over graphs in
0
conn
,
namely to take the logarithm of the partition function (called the free energy in the
physical literature):
Proposition 2.20. Let W be a weight system. Then
log
_

0
1
[ Aut [
W

_
=

0
conn
1
[ Aut [
W

.
Proof. Let us compare the terms of the power series expansion for the right hand
side with the terms in the left hand side:
exp
_
_

0
conn
1
[ Aut [
W

_
_
=

1
n
1
! . . . n
k
!
W
n1
1
. . . W
n
k

k
,
where the sum is over all k, n
i
, and distinct
i

0
conn
, i = 1, . . . , k. Consider
= (
1
)
n1
. . . (
k
)
n
k

0
. Since in addition to automorphisms of each
i
there
are also automorphisms of interchanging the n
i
copies of
i
, we have [ Aut [ =
n
1
! . . . n
k
![ Aut
1
[ . . . [ Aut
k
[. Also, any weight system satises W

= W

,
hence W

= W
n1
1
. . . W
n
k

k
. The proposition follows.
Exercise 2.21. Formulate and prove a similar statement for graphs with legs.
FEYNMAN DIAGRAMS FOR PEDESTRIANS AND MATHEMATICIANS 11
Remark 2.22. It is possible to restrict the class of graphs to 1-connected (in the
physical literature usually called 1-point irreducible, or 1PI for short) graphs. A
graph is 1-connected, if it remains connected after a removal of any one of its edges.
This involves a passage to a so-called eective action, which I will not discuss here in
details. Mathematically, it simply means an application of a Legendrian transform
(a discrete version of a Fourier transform): if z(b) = log(Z
b
) is given by the sum
over all connected graphs as in Proposition 2.20, then z(x) = b, x) z(b) is given
by a similar sum over all 1PI graphs (and b(x) may be recovered as z/x).
3. Gauge theories and gauge fixing
3.1. Gauge xing. All calculations of the previous section dealt only with the case
of a non-degenerate bilinear form A; in particular, the critical points of the action
S(x) had to be isolated (see Section 1.3). However, gauge theories present a large
class of examples when it is not so. Suppose that we have an l-dimensional group
of symmetries, i.e. the Lagrangian is invariant under a (free, proper, isometric)
action of an l-dimensional Lie group G. Then instead of isolated critical points we
have critical orbits, so A has l degenerate directions and the technique of Gauss
integration can not be applied.
Let us try to calculate the partition and correlation functions without a superu-
ous integration over the orbits of G. In other words, we wish to reduce integrals of
G-invariant functions on X to integrals on the quotient space

X = X/G of G-orbits.
For this purpose, starting from a G-invariant measure on X we should desintegrate
it as the Haar measure on the orbits over some quotient measure on the base

X.
If G is compact then is the standard push-forward of . For example, if f is a
rotationally invariant function on R
2
, we can take the pair of polar coordinates (r, )
as coordinates in the quotient space

X and the orbit, respectively. The measure on

X in this case is 2r dr and we get the following elementary formula:

_
R
2
f([x[) d
2
x = 2
_

0
f(r)r dr.
If G is a locally compact group acting properly on X then can be dened by
the property
(Y ) =
_

X
[Y G( x)[ d ( x),
where G( x) is the ber over x

X and [[ is the Haar measure on it. In this case the
integral
_
X
f dx in question is innite, but it can be formally dened (regularized)
as
_

X
fd x.
A standard physical procedure for the desintegration that can be applied also
to a non-locally compact gauge group is called a gauge xing (see e.g., [23]); it
goes as follows. Suppose that f : X R is G-invariant, i.e. f(gx) = f(x) for
all x X, g G. Choose a (local) section s :

X X which intersects each
orbit of G exactly once. Suppose that it is dened by l independent equations
F
1
(x) = = F
l
(x) = 0 for some F : X R
l
. Firstly, we want to count each G-
orbit only once. This is simple to arrange by inserting an l-dimensional -function

l
(F(x)) in the integrand. Secondly, we want to take into account the volume of a
G-orbit passing through x, so we should count each orbit with a certain Jacobian
12 MICHAEL POLYAK
factor J(x) (called the Faddeev-Popov determinant). How should one dene such a
factor? We wish to have
_
X
f(x) dx =
_
X
f(x)J(x)
l
(F(x)) dx.
Rewriting the right hand side to include an additional integration over G and
noticing that both f(x) and J(x) are G-invariant, we get
_
X
f(x) dx =
_
X
f(x)J(x)
l
(F(x) dx =
=
_
X
dx
_
G
dg f(x)J(x)
l
(F(gx)) =
_
X
dx f(x)J(x)
_
G
dg
l
(F(gx)).
Thus we see that we should dene J(x) by
J(x)
_
G
dg
l
(F(gx)) = 1,
where dg is the left G-invariant measure on G. Thus, the Faddeev-Popov deter-
minant plays the role of Jacobian for a change of coordinates from x to (s( x), g).
Example in 3.3 below provides a good illustration.
Remark 3.1. A formal coordinate-free way to dene J(x) is as follows. The section
s :

X X determines a push-forward s

: T
x

X T
x
X of the tangent spaces. The
tangent space T
x
X thus decomposes as s

i : T
x

Xg T
x
X, where i : g T
x
X
is the tangent space to the orbit, generated by the Lie algebra g of G. The Jacobian
J(x) may be then dened as J(x) = det(s

i).
Remark 3.2. Equivalently, one may note that the tangent space to the ber at
x s may be identied with g, to directly set J(x) = det , where = (
F
i
g
j
) and
g
j

l
j=1
is a set of generators of the Lie algebra g of G, see e.g. [3]. I.e., J(x) is the
inverse ratio of the volume element of g and its image in R
l
under the action of G
composed with F.
Indeed, since F has a unique zero on each orbit and since (due to the presence
of the delta-function) we integrate only near the section s, we can use F as a local
coordinate in the ber over x. Making a formal change of variables from g to F we
get
J(x)
1
=
_
G
dg
l
(F(gx)) =
_
G
dF
l
(F(gx)) det
_
g
F
_
= det
_
g
F
_

F=0
.
Calculating (
F(gx)
g
)

F=0
at a point x s and identifying the tangent space to the
ber with g, we obtain (
F
g
)

F=0
= (
F
i
g
j
).
Exercise 3.3. Let us return to the simple example of a rotationally invariant function
f(x
1
, x
2
) = f([x[) on R
2
, using this time the gauge-xing procedure. The group
G = S
1
acts by rotations: x = e
i
x and the (normalized) measure on G is
1
2
d.
We should use the positive x
1
-axis for a section s, so we may take e.g. F = x
2
. A
slight complication is that the equation x
2
= 0 denes the whole x
1
-axis and not
only its positive half, so each ber of G intersects it twice and not once. This can
be taken care of, either by dividing the resulting gauge-xed integral by two, or
by restricting its domain of integration to the right half-plane R
2
+
in R
2
. In any
FEYNMAN DIAGRAMS FOR PEDESTRIANS AND MATHEMATICIANS 13
case, using x
2
instead of as a local coordinate in the ber Gx near x s we get
d = d
_
arctan(x
2
/x
1
)
_
= x
1
[x[
2
dx
2
Thus for x s we have
J(x)
1
=
_
(F(x))
1
2
d =
1
2
_
(x
2
)x
1
[x[
2
dx
2
=
1
2x
1
,
so J([x[) = 2[x[ as expected and
_
R
2
f([x[) d
2
x =
_
R
2
+
f([x[)2[x[(x
2
) dx
1
dx
2
= 2
_

0
f(r)r dr.
3.2. Faddeev-Popov ghosts. After performing the gauge-xing, we are left with
the gauged-xed partition function
Z
GF
=
_
R
d
dx e

1
2
Ax,x

l
(F(x)) det .
We would like to make it into an integral of the type we have been studying be-
fore. We have two problems: to include (F(x)) det in the exponent (i.e., in the
Lagrangian) and more importantly to make A into a non-degenerate bilinear
form.
The -function is easy to write as an exponent using the Fourier transform:

l
(F(x)) = (2)
l
_
R
l
d e
i,F(x)
.
The gauge variables (called Lagrange multipliers) supplement the variables x, and
the quadratic part of , F(x)) supplements Ax, x) so that the quadratic part A
F
of the gauge-xed Lagrangian is non-degenerate.
The det term is somewhat more complicated; it can be also represented as
a Gaussian integral, but over anti-commuting variables c = (c
1
, . . . , c
l
) and c =
( c
1
, . . . , c
l
), called Faddeev-Popov ghosts. Thus
(22) c
i
c
j
+ c
j
c
i
= c
i
c
j
+c
j
c
i
= c
i
c
j
+ c
j
c
i
= 0
There are standard rules of integration over anti-commuting variables (known to
mathematicians as the Berezin integral , see e.g. [23, Chapter 33] and [13, 16]). The
ones relevant for us are
_
c
i
dc
j
=
_
c
i
d c
j
=
ij
and
_
1 dc
j
=
_
1 d c
j
= 0.
The multiple integration (over e.g., dc = dc
l
. . . dc
1
) is dened by iteration. One
may show that this implies (see the Exercise below) that for any matrix
_
e
c,c
dc d c = det .
Exercise 3.4. Let l = 1 and dene the exponent e
cc
by the corresponding power
series. Use the commutation relations (22) to verify that only the two rst terms
of this expansion do not vanish. Now, use the integration rules to deduce that
_
e
cc
dc d c =
_
(1 + cc) dc d c = .
Thus we may rewrite Z
GF
by adding to the Lagrangian the gauge-xing term
and the ghost term:
Z
GF
=
_
dxd dc d c e

1
2
Ax,x+ c,c+i,F(x)
.
14 MICHAEL POLYAK
At this stage we may again apply the Feynman diagram expansion to the gauge-
xed Lagrangian. The Feynman rules change in an obvious fashion. The quadratic
form now consists of two parts: A
F
and , so there are two types of edges. The rst
type presents A
F
, with the labels x
i
and
i
at the ends. The second type presents
, with the labels c
i
and c
i
at the ends. Note that since is not symmetric, these
edges are directed. Also, there are new vertices, presenting all higher degree terms
of the Lagrangian (in particular some where edges of both types meet). An example
of the Chern-Simons theory will be provided in Section 5.
3.3. An example of gauge-xing. Let us illustrate the idea of gauge-xing on
an example of the standard C

-action on C
2
. In the coordinates (x
1
, x
1
, x
2
, x
2
) on
C
2
the gauge group acts by x
i
x
i
, x
i


x
i
. Let us take A =
x1
x2
x1
x2
as an
invariant function.
Of course, the orbit space CP
1
is quite simple and an appropriate measure
on CP
1
is well known; in the coordinates z = x
1
/x
2
, z = x
1
/ x
2
it is given by
dzd z/(1 +z z)
2
We are thus interested in
(23) Z
GF
=
_
dzd z
(1 +z z)
2
e

1
2
z z
.
Let us pretend, however, that we do not know this and proceed with the gauge-xing
method instead.
The invariant measure on C
2
is dx
1
d x
1
dx
2
d x
2
/(x
1
x
1
+x
2
x
2
)
2
. In a gauge F = 0
we have
Z
GF
=
_
dx
1
dx
2
d x
1
d x
2
(x
1
x
1
+x
2
x
2
)
2
e

1
2
x
1
x
2
x
1
x
2

2
(F(x, x)) det ,
where =

x
1
F
x1
+x
2
F
x2
x
1
F
x1
+ x
2
F
x2
x
1

F
x1
+x
2

F
x2
x
1

F
x1
+ x
2

F
x2

.
E.g., for F = x
2
1 we get
2
([x
2
1[) and det = x
2
x
2
.
Exercise 3.5 (Dierent gauges give the same result). Consider F = x

2
1. Show
that
2
([x

2
1[) = [x
1
2
[
2

2
([x
2
1[) and det = (x
2
x
2
)

. Check that the

dependence on in Z
GF
cancels out, thus gives the same result as F = x
2
1.
Show that it coincides with formula (23).
Finally, let us check that while the initial quadratic form A is degenerate, the
supplemented quadratic form A
F
is indeed non-degenerate. It is convenient to make
a coordinate change x

1
= x
1
, x

2
= x
2
1. Using a Fourier transform we get

2
(x
2
1) = (2)
2
_
dd

e
i(x

2
)
.
Also, we have
x1
x2
= x

1
+x

n=1
(1)
n
x

2
. We can now compute A and A
F
; in the
coordinates (x

1
, x

1
, x

2
, x

2
) and (x

1
, x

1
, x

2
, x

2
, ,

), respectively, we have:
A =

0 1 0 0
1 0 0 0
0 0 0 0
0 0 0 0

, A
F
=

0 1 0 0 0 0
1 0 0 0 0 0
0 0 0 0 i 0
0 0 0 0 0 i
0 0 i 0 0 0
0 0 0 i 0 0

.
FEYNMAN DIAGRAMS FOR PEDESTRIANS AND MATHEMATICIANS 15
4. Infinite dimensional case
4.1. The dictionary. Path integrals are generally badly dened, so instead of try-
ing to deduce the relevant results rigorously, we will just provide a basic dictionary
to translate the nite dimensional results to the innite dimensional case.
The main change is that instead of the discrete set i 1, . . . , d of indices we
now have a continuous variable x M
n
(say, in R
n
), so we have to change all
related notions accordingly. The sum over i becomes an integral over x. Vectors
x = x(i) = (x
1
, . . . , x
d
) and b = b(i) become elds = (x) and J(x). A quadratic
form A = A(i, j) becomes an integral kernel K = K(x, y). Pairings Ax, x) =

i,j
x
i
A
ij
x
j
and b, x) =

i
b
i
x
i
become K, ) =
_
dxdy (x)K(x, y)(y) and
J, ) =
_
dx J(x)(x) respectively. The partition function Z
b
dened by (4)
becomes a path integral Z
J
over the space T of elds
Z
J
=
_
T e

1
2
K,+J,
.
The inverse A
1
of A dened by

k
A
ik
A
kj
=
j
i
corresponds now to the inverse
G = K
1
of K dened by
_
dz K(x, z)G(z, y) = (x y).
Formula (5) for Z
b
then translates into
Z
J
= Z
0
e
1
2
J,GJ
.
Correlators x
i1
, . . . , x
im
) dened by (6) become now m-point functions
(x
1
), . . . , (x
m
)) =
1
Z
0
_
T e
1/2K,
(x
1
) . . . (x
m
).
4.2. Functional derivation. A counterpart of the derivatives /x
i
is given by
the functional derivatives /(x). The theory of functional derivation is well-
presented in many places (see e.g. [10]), so I will just briey recall the main notions.
Let F() be a functional. If the dierential
DF()() = lim
0
F( +) F()

can be represented as
_
(x)h(x)dx for some function h(x), then we dene
F
(x)
=
h(x). In general, the functional derivative F/(x) is the distribution representing
the dierential of F at . The reader can entertain himself by making sense of the
following formulas, which show that its properties are similar to usual derivatives:

(x)
(y) = (x y),

(x)
(F()H()) =

(x)
(F()) H() +F()

(x)
(H()).
Example 4.1.

J(y)
e
J,
=

J(y)
e
_
dxJ(x)(x)
= (y)e
_
dxJ(x)(x)
= (y)e
J,
.
16 MICHAEL POLYAK
Exercise 4.2. Consider a (symmetric) potential function
(24) U() =

n
1
n!
_
dx
1
. . . dx
n
U
n
(x
1
, . . . , x
n
)(x
1
) . . . (x
n
).
Prove that

(y)
U() =

n
1
n!
_
dx
1
. . . dx
n
U
n+1
(y, x
1
, . . . , x
n
)(x
1
) . . . (x
n
).
The inverse G(x, y) can be written as a Hessian, similarly to equation (8) for
A
1
:
G(x, y) =
1
Z
0

J(x)

J(y)
Z
J

J=0
.
More generally, for m-point functions we have, similarly to (9),
(x
1
), . . . , (x
m
)) =
1
Z
0

J(x
1
)
. . .

J(x
m
)
Z
J

J=0
.
4.3. Wicks theorem and Feynman graphs. Wicks theorem now states that,
similarly to (10),

J(x
1
)
. . .

J(x
m
)
e
1
2
J,GJ

J=0
=

G(x
i1
, x
i2
) . . . G(x
im1
, x
im
),
where the sum is over all pairings (i
1
, i
2
) . . . (i
m1
, i
m
) of 1, . . . , m. Just as in the
nite dimensional case, we may encode each pairing by a graph with m univalent
vertices labelled by 1, . . . , m, and edges connecting vertices i
1
with i
2
, . . . , and
i
m1
with i
m
presenting the factors of G.
Let us add a potential (24) to the action and dene
Z
U
=
_
T e
1/2K,+U()
.
Then, similarly to (15), we have
Z
U
= Z
0
e
U(

J
)
e
1
2
J,GJ

J=0
.
Using again the Wicks theorem, we can rewrite the latter expression in terms of
Feynman graphs to get
(25) Z
U
=

||
Aut
_
labels

v
U
v

e
G
e
,
where the integral is over all labellings of the ends of edges, U
v
= U(x
1
, . . . , x
k
) for a
k-valent vertex with the labels x
1
, . . . , x
k
of the adjacent edges, and G
e
= G(x
i
, x
j
)
for an edge with labels x
i
, x
j
. Sometimes it is convenient to include the integration
over the labels of the star of a vertex into the weight of this vertex.
4.4. An example:
4
-theory. Let us write down the Feynman rules for a poten-
tial U() =
_
dx
4
(x). Firstly, the relevant graphs have vertices of valence one
or four. Secondly, all edges adjacent to a vertex should be labelled by the same x,
so we may instead label the vertices. An edge with labels x, y represents G(x, y)
and (including the integration over the vertex labels into the weights of vertices)
an x-labelled vertex represents
_
dx. The linear term in the power series expansion
of x
1
, x
2
)
U
should correspond to non-vacuum graphs with two legs, labelled by x
1
and x
2
, and one 4-valent vertex. There is only one such graph, see Figure 3. It
FEYNMAN DIAGRAMS FOR PEDESTRIANS AND MATHEMATICIANS 17
Figure 3. Graphs appearing in the
4
-theory
represents
_
dx G(x
1
, x)G(x, x)G(x, x
2
) and should enter with the multiplicity 12
(the number of all pairings of 6 vertices x
1
, x
2
, x, x, x, x in which x
1
is not connected
to x
2
). Let us now check this directly. Indeed, the coecient of in x
1
, x
2
) is

J(x
1
)

J(x
2
)

4
J(x)
4
e
1
2
J,GJ

J=0
= 12
_
dx G(x
1
, x)G(x, x)G(x, x
2
),
where we applied Wicks theorem to obtain the desired equality.
In a similar way, the linear term in the expansion of Z
U
should correspond to
the graph with no legs and one vertex of valence four (see Figure 3), representing
_
dx G
2
(x, x) (and entering with the multiplicity 3).
Exercise 4.3 (
3
-theory). Let U() =
_
dx
3
(x). Find the Feynman rules for
this theory. Which graphs will contribute to the coecient of
2
in the power
series expansion of the 2-point function x
1
, x
2
)
U
? Write down these coecients
explicitly.
4.5. Convergence. Usually the integrals which we get by a perturbative Feynman
expansion are divergent and ill-dened in many ways. Often one has to renormalize
(i.e. to nd some way to remove divergencies in a unied and consistent manner)
the theory to improve its behavior. Until recently renormalization was consid-
ered by mathematicians more like a physical art than a technique; lately Connes
and Kreimer [7] have done some serious work to explain renormalization in purely
mathematical terms (see a paper by Kreimer in this volume).
But even in the best cases, the Green function G(x, y) usually blows up near
the diagonal x = y, which brings two problems: Firstly, the weights of graphs with
looped edges, starting and ending at the same point (so-called tadpoles) are ill-
dened and one has to get rid of them in one or another way. Secondly, all diagonals
have to be cut out from the spaces over which the integration is performed, so the
resulting conguration spaces are open and the convergence of all integrals dening
the weights has to be proved. Mathematically these convergence questions usually
boil down to the existence of a Fulton-MacPherson-type (see [14]) compactication
of conguration spaces, to which the integrand extends.
There is also a challenging problem to interpret the Feynman diagrams series
in some classical mathematical terms and to understand the way to produce them
without a detour to physics and back. In many examples this may be done in terms
of a homology theory of some grand conguration spaces glued from conguration
spaces of dierent graphs along common boundary strata.
We will see all this on an example of the Chern-Simons theory in the next section.
5. An example of QFT: Chern-Simons theory
The Chern-Simons theory has an almost topological character and as such presents
an interesting object for low-dimensional topologists. For a connection in a trivial
18 MICHAEL POLYAK
SU(N)-bundle over a 3-manifold M one may dene ([6]; see also [11]) the Chern-
Simons invariant CS() as described in Section 5.1 below. It is the action functional
of the classical Chern-Simons theory and was extensively used in mathematics for
many years to study properties of 3-manifolds (mostly due to the fact that the
classical solutions, i.e. the critical points of CS(), are at connections). But it
is the corresponding quantum theory which is of interest for us. Its mathematical
treatment started only about a decade ago, following Wittens suggestion [25] that
it leads to some interesting invariants of links and 3-manifolds, in particular, to
the Jones polynomial. While Wittens idea was based on the validity of the path
integral formulation of the quantum Chern-Simons theory, his work catalyzed much
mathematical activity. By now mathematicians more or less managed to formalize
the relevant perturbative series and exorcize from them all physical spirit, leaving
a (surprisingly rich) rigorous mathematical extract. In this section I will describe
this process in a number of iterations, starting from an intuitive and roughest de-
scription and slowly increasing the level of rigor and details. Finally, I will try to
reinterpret these Feynman series in some classical topological terms and formulate
some corollaries.
5.1. Chern-Simons theory. Further we will use the following data:
A closed orientable 3-manifold M with an oriented framed link L in M.
A compact connected Lie group G with an Ad-invariant trace Tr : g R
on the Lie algebra g of G.
A principal G-bundle T M.
To simplify the situation, we will additionally assume that G is simply connected,
since for such groups any principal G-bundle over a manifold M of dimension 3
(which is our case) is trivializable, see e.g. [11].
The appropriate notions of the Chern-Simons theory, considered as a eld theory,
are as follows. The manifold M plays the role of the space-time manifold X. Denote
by / the space of G-connections on T and let ( = Aut(T) be the gauge group.
Fields on M are G-connections on T, i.e. T = /. The Lagrangian is a functional
L : /
3
(M) dened by
L() = Tr( d +
2
3
).
Remark 5.1. This choice can be motivated as follows. Let = d + be
the curvature of . Then Tr( ) is the Chern-Weil 4-form
3
on T, associated
with Tr; this form is gauge invariant and closed. The Chern-Simons Lagrangian
CS() = Tr( +
2
3
) is an antiderivative of Tr( ) on T: it is a nice
exercise to check that d(CS()) = Tr( ).
The corresponding Chern-Simons action is a function CS : / R given by
CS() =
1
4
_
M
dxTr( d +
2
3
).
It is known that the critical points of this action correspond to at connections and
(assuming that Tr satises a certain integrality property
4
, which holds in particular
3
Chern-Weil theory states that the de Rham cohomology class of this form is a certain char-
acteristic class of P
4
Namely that the closed form
1
6
Tr( ) represents an integral class in H
3
(G, R)
FEYNMAN DIAGRAMS FOR PEDESTRIANS AND MATHEMATICIANS 19
for the trace in the fundamental representation of G) it is gauge invariant modulo
2Z.
The partition function is given by the following path integral:
(26) Z =
_
A
e
ikCS()
T.
Here the constant k N is called level of the theory; its integrality is needed for
the gauge invariance of Z.
Now, let L =
m
j=1
L
j
, j = 1, . . . m be an oriented framed m-component link in
M such that each L
j
is equipped with a representation R
j
of G. Given a connection
/, let hol
Lj
() be the holonomy
(27) hol
Lj
() = exp
_
Lj

of around L
j
. Observables in the Chern-Simons theory are so-called Wilson loops.
The Wilson loop associated with L
j
is the functional
J(L
j
, R
j
) = Tr
Rj
(hol
Lj
()).
The m-point correlation function L) = L
1
, L
2
, . . . L
m
) is dened by
(28) L) = Z
1
_
A
e
ikCS()
m

j=1
J(L
j
, R
j
)T.
Since the action is gauge invariant, extrema of the action correspond to points on
the moduli space of at connections. Near such a point the action has a quadratic
term (arising from d) and a cubic term (arising from ). We would
like to consider a perturbative expansion of this theory.
5.2. What do we expect. Which Feynman graphs do we expect to appear in the
perturbative Chern-Simons theory?
Firstly, a gauge-xing has to be performed, so the ghosts have to be introduced.
As a result, we should have two types of edges: the usual non-directed edges (cor-
responding to the inverse of the quadratic part) and the directed ghost edges.
Secondly, in addition to the quadratic term the action contains a cubic term, so
the internal vertices should be trivalent. Also, this time the cubic term is given by
an antisymmetric tensor instead of a symmetric one, so one should x a cyclic order
at each trivalent vertex, with its reversal negating the weight of a graph. Two types
of edges should lead to two types of internal vertices: usual vertices where three
usual edges meet, and ghost vertices where one usual edge meets one incoming and
one outgoing ghost edge.
Thirdly, note that the situation with legs is somewhat dierent from our earlier
considerations. Indeed, the legs (i.e. univalent ends of usual edges) of Feynman
graphs, instead of being xed at some points, should be allowed to run over the
link L, with each link component entering in L) via its holonomy (27). To reduce
this to our previous setting, we can use Chens iterated integrals to expand the
holonomy in a power series where each term is a polynomial in . In terms of a
parametrization L
j
: [0, 1] R
3
, this expansion can be written explicitly using the
pullback L

j
of to [0, 1] via L
j
:
20 MICHAEL POLYAK
hol
Lj
() = 1 +
_
0<t<1
(L

j
)(t) +
_
0<t1<t2<1
(L

j
)(t
2
) (L

j
)(t
1
)+
+
_
0<t1<<t
k
<1
(L

j
)(t
k
) (L

j
)(t
1
) + . . .
where the products are understood in the universal enveloping algebra U(g) of g.
Thus we should sum over all graphs with any number k
j
of cyclically ordered legs
on each L
j
, and integrate over the positions
(x(t
1
), . . . , x(t
kj
)) L
kj
j
, 0 < t
1
< < t
kj
< 1
of these legs.
These simple considerations turn out to be quite correct. Of course, one should
still nd an explicit formulas for the weights of such graphs. An explicit deduction
of the Feynman rules for the perturbative Chern-Simons theory is described in
details in [3, 15]. Let me skip these lengthy calculations and formulate only the
nal results. For simplicity I will consider only an expansion around the trivial
connection in M = R
3
.
5.3. Feynman rules. It turns out (see [3]) that the weight system W
CS
of the
perturbative Chern-Simons theory splits as W
CS
= W
G
W, where W
G
contains all
the relevant Lie-algebraic data of the theory (but does not depend on the location
of the vertices of a graph), and W contains only the space-time integration. Since
the whole construction should work for any Lie algebra, one may encode the anti-
symmetry and Jacobi relations already on the level of graphs, changing the weight
W
G

of a graph to a universal weight [], which is an equivalence class of

in the vector space over Q generated by abstract (since we do not care about the
location in R
3
of their vertices) graphs, modulo some simple diagrammatic anti-
symmetry and Jacobi relations, shown on Figure 4. The same relations hold for
graphs with either usual or ghost edges, so we may think that the relations include
the projection making all edges of one type.
= = =
Figure 4. Antisymmetry and Jacobi relations
The drawing conventions merit some explanation. It is assumed that the graphs
appearing in the same relation are identical outside the shown fragment. In each
trivalent vertex we x a cyclic order of edges meeting there; unless specied other-
wise, it is assumed to be counter-clockwise. The edges are shown by dashed lines,
and the link component L
j
(xing the cyclic order of the legs) by a solid line. An
important consequence of the antisymmetry relation is that for any graph with a
tadpole (a looped edge) we have [] = 0 due to the existence of a handle twisting
automorphism, rotating the looped edge. Thus from the beginning we can restrict
the class of graphs to graphs without tadpoles.
FEYNMAN DIAGRAMS FOR PEDESTRIANS AND MATHEMATICIANS 21
It remains to describe the weight W

of a graph . Roughly speaking, for each

internal vertex of we are to perform integration over its position in R
3
(for a ghost
vertex we should also take a certain derivative acting on the term corresponding
to the outgoing ghost edge); for each leg we are to perform integration over its
position in L
j
(respecting the cyclic order of legs on the same component). As for
the edges, we are to assign to each usual and ghost edge inverses of the operator
curl and of the Laplacian, respectively.
Somewhat surprisingly (see e.g. [15]) two types of edges may be neatly joined
into one combined edge, thus reducing the graphs in question to graphs with only
one type of edges (and just one type of uni- and trivalent vertices). The weight
G(x(e), y(e)) of such an edge e with the ends in (x(e), y(e)) R
3
R
3
has a nice
geometrical meaning: it is given by G(x, y) = (x y), where
(x) =
x
1
dx
2
dx
3
2[[x[[
3
+ cyclic permutations of (1,2,3)
is the uniformly distributed area form on the unit 2-sphere [x[ = 1 in the standard
coordinates in R
3
. In fact, the usual and the ghost edges (with two possible orien-
tations) give respectively the (1, 1), (2, 0), and (0, 2) parts of (x y) in terms of
its dependence on dx and dy. Abusing notation, I will depict the combined edge
again by a dashed line.
Remark 5.2. A simple explanation for an existence of such a simple unied propa-
gator escapes me. The only explanation which I know is way too complicated: it is
the existence (see [2]) of the superformulation of the gauge-xed theory, i.e. the
fact that the connection together with the ghosts may be united in a superconnec-
tion of a supertheory, which leads to an existence of a superpropagator, uniting
the usual and the ghost propagator. I believe that there is a simple explanation,
probably emanating from the scaling properties and the topological invariance of
the Chern-Simons theory, by which one should be able to predict that the combined
propagator should be dilatation- and rotation-invariant.
Remark 5.3. Note that the weight G(x, x) of a tadpole is not well-dened, so it is
quite fortunate that we got [] = 0 for any such graph.
To sum it up, we are interested in the value
(29) L) =

||
Aut
[]
where [[ is half of the total number of vertices (univalent and trivalent) of , and
the weight of is given by the integral
(30) W

=
_
C

e
G(x(e), y(e))
over the space C

of all possible positions of vertices of , such that all vertices

remain distinct. Here = (k + h

)
1
, where h

is the dual Coxeter number of G

(see [25]).
I shall describe in more details the type of graphs which appear in this formula
and their weights W

(both the conguration spaces C

, and the integrand).

22 MICHAEL POLYAK
5.4. Jacobi graphs. Let us start with the graphs. Instead of thinking about
graphs embedded in R
3
, consider abstract graphs (with just one type of edges),
such that
all vertices have valence one (legs) or three;
there are no looped edges;
all legs are partitioned into m subsets l
1
, . . . , l
m
;
legs of each subset l
j
are cyclically ordered;
each trivalent vertex is equipped with a cyclic order of three half-edges
meeting there;
for technical reasons it will be convenient to think that, in addition to the above,
all edges are ordered and directed.
We will further address the last three items simply as an orientation of a graph.
For such a graph with a total of 2n (univalent and trivalent) vertices dene the
degree of by [[ = n, and denote the set of all such graphs by

J
n
. Set

J =

J
n
. The
ordering and directions of edges of graphs in

J may be dropped by an application
of an obvious forgetful map. See Figure 5 for graphs of degree one with m = 2
and m = 1, and graphs of degree two with m = 1. Both antisymmetry and Jacobi
relations of Figure 4 preserve the degree of a graph, thus we may consider a vector
space over Q generated by graphs in

J
n
modulo forgetful, antisymmetry and Jacobi
relations. We will call it the space of Jacobi graphs of degree n and denote it by J
n
;
denote also J =
n
J
n
, and let as before [] be the class of

J in J.
a b c
Figure 5. Graphs of degree one and two
Exercise 5.4. Let m = 2. Write the relations between the equivalence classes of
degree two graphs shown in Figure 5c. What is the dimension of J
2
?
This settles the type of graphs appearing in formula (29): the summation is over
all graphs in

J, while L) J[[]]. It is somewhat simpler to study separately the
components of dierent degrees; dene
(31) L)
n
=

Jn
W

[ Aut()[
[]
5.5. Conguration spaces. Let us deal now with the weights (30) of graphs (see
[5, 18, 24] for details). The domain of integration in (30) is the conguration space
C

of embeddings of the set of vertices of to R

3
, such that the legs of each subset
l
j
lie on the corresponding component L
j
of the link L in the correct cyclic order.
It is easy to see that for a graph with k trivalent vertices and k
j
legs ending
on L
j
, j = 1, . . . , m we have C

= (R
3
)
k

j
(S
1

kj 1
) , where
k
is
a k-dimensional simplex, and is the union of all diagonals where two or more
points coincide. Indeed (forgetting for a moment about coincidences of vertices),
FEYNMAN DIAGRAMS FOR PEDESTRIANS AND MATHEMATICIANS 23
each trivalent vertex is free to run over R
3
, while k
j
legs ending on L
j
run over
S
1

k1
, where S
1
encodes the position of the rst leg, and the following legs are
encoded by their distance from the previous one.
Exercise 5.5. Show that the dimension of C

is twice the number of the edges of

.
Now, an orientation of a graph determines an orientation of C

; its idea is in
fact based on Exercise 5.5. Let me describe this construction in some local coordi-
nates. Near each trivalent vertex of there are three local coordinates (describing
its movement in R
3
); assign one of them to each of the three ends edges meeting
in this vertex using their cyclic order. Near each leg of there is only one local
coordinate (describing its movement along the link); assign it to the corresponding
end of the edge. By now the end of any edge has one coordinate assigned to it. It
remains to order them using the given ordering of all edges of and their directions.
Let us order them as (x
1
, y
1
, x
2
, y
2
, . . . , x
n
, y
n
) where (x
i
, y
i
) are the coordinates
assigned to the beginning and the end of i-th edge. This denes an orientation of
C

.
Exercise 5.6. The above construction involves a choice in each trivalent vertex since
we had only a cyclic order of the edges meeting there, while we used a total order
of these three edges. Show that a cyclic permutation of the three local coordinates
used there preserves the orientation of C

. Also, we used the orientation of ; what

happens with the orientation of C

if:
(1) The cyclic order of three half-edges in one vertex is reversed?
(2) A pair of edges is transposed in the total ordering of all edges?
(3) The direction of an edge is reversed?
5.6. Gauss-type maps of conguration spaces. To understand the integrand
in (30), consider a directed edge e. Its ends (x, y) represent a point in the square
R
3
R
3
with the diagonal = (x, y)[ x = y cut out. This cut square C =
R
3
R
3
has the homotopy type of S
2
, with the Gauss map
: (x, y)
y x
[[y x[[
providing the equivalence. The form (y x) assigned to this edge is nothing more
than a pullback of the area form on S
2
to C via the Gauss map:
(y x) =

Each edge e of a graph denes an evaluation map ev

e
: C

C, by erasing
all vertices of but for the ends of e. The composition
e
= ev
e
denes the
Gauss map corresponding to (the ends of) an edge e. The graph with an ordering
e
1
, . . . , e
n
of edges denes the product

n
i=1

ei
: C

(S
2
)
n
of Gauss maps.
Finally, the weight W

is given by integrating the pullback of the volume form

dvol =
n
i=1
on (S
2
)
n
to C

by the product Gauss map

:
(32) W

=
_
C

dvol
Exercise 5.7. Suppose that a graph has a double edge (i.e., a pair of edges both
endpoints of which coincide). Show that dim(

(C

)) dim(C

) 1. Deduce that
W

= 0.
24 MICHAEL POLYAK
The following important example shows that at least in some simple cases W

has an interesting topological meaning:

Example 5.8. Let = e be a graph with one edge with the ends on two link
components L, see Figure 5a. The conguration space C
e

= S
1
S
1
C is a
torus. It is mapped to S
2
by the Gauss map =
e
. The weight W
e
=
_
Ce

=
deg() is in this case just the degree of the map . This fact has many important
consequences. In particular W
e
takes only integer values and is preserved if we
change the uniformly distributed area form to any other volume form dvol on S
2
normalized by
_
S
2
dvol = 1. It is also preserved if we deform the link by isotopy
(since then the conguration space changes smoothly and the degree can not jump),
so is a link invariant. This invariant is easy to identify:
_
Ce

= lk(L
1
, L
2
) is
the famous Gauss integral formula for the linking number lk(L
1
, L
2
) of L
1
with L
2
.
Thus we get
Proposition 5.9. Let = e be a graph with one edge with the ends on two link
components L. Then the weight W
e
is the linking number lk(L
1
, L
2
).
1 2
+
1 2

a b
Figure 6. Signs of crossings and the south pole on S
2
Exercise 5.10. There is a simple combinatorial way to compute lk(L
1
, L
2
) from
any link diagram: count all crossings where L
1
passes over L
2
, with signs shown in
Figure 6a. Interpret this formula as a calculation of deg() by counting (with signs)
the number of preimages of a certain regular value of (hint: look at Figure 6b).
What formula would we get if we counted the preimages of the north pole?
For other graphs the situation is more complicated. For example, let = e be
the graph with one edge, both ends of which end on the same link component, see
Figure 5b. Then the conguration space C
e
is an open annulus (R
3
)
0
S
1

1
=
S
1
(0, 1) (torus cut along the diagonal). The Gauss map is badly behaved near
the diagonal, so the integrand blows up near the diagonal and we can not extend
it to the closed torus. The integral nevertheless converges; one way to see it is to
compactify C
e
, cutting out of it some small neighborhood of the diagonal. This
makes C
e
into a closed annulus C

e
= S
1
[, 1 ] (thus making the integral
convergent) and we can recover the initial integral by taking 0. But the
Gauss integral W
e
is no more a knot invariant: it may take any real value under
a knot isotopy. A detailed discussion on this subject may be found in [5]. Why
does this happen? The reason is that the compactied space C
e
is not a torus,
but an annulus, so has a boundary and the degree of the Gauss map is not well-
dened. When both ends of the edge start to collide together, the direction of the
vector connecting them (which appears in the Gauss map) tends to the (positive
or negative) tangent direction to the knot. The image of the unit tangent to the
knot under the Gauss map is a certain curve on S
2
. One of the boundary circles
FEYNMAN DIAGRAMS FOR PEDESTRIANS AND MATHEMATICIANS 25
S
1
and S
1
(1 ) of C

e
is mapped into , while the other is mapped into ,
and the weight W
e
is part of the area of S
2
covered by the annulus (C

e
) between
these curves. Unfortunately, may move on S
2
under an isotopy of L, so this area
may change.
In this particular case there is a neat way to solve this problem: let L be framed
(i.e. x a section of its normal bundle). We may think about the framing as about
a unit normal vector n(x) in each point x of a knot. This allows us to slightly
deform the Gauss map:
(
x, y) (x, y) + n(y). Now both boundary circles of
the annulus C
e
map into the same curve on S
2
(why?) and we may glue the annulus
into the torus so that the map

extends to it. It makes W

e
into an invariant of
framed knots, called the self-linking number (the same result may be obtained by
slightly pushing L o itself along the framing and considering the linking number
of the knot with its pushed-o copy).
It turns out that for other graphs there are also no divergence problems, so
all integrals W

converge, and that a collision of all vertices of a graph to one

point (so-called anomaly, see [18, 24]) is the only source of non-invariance, exactly
as for W
e
above. Thus there is a suitable normalization of the expression (31)
for L) =

n
L)
n
which gives a link invariant. To avoid a complicated explicit
description of this normalization, let me formulate this result as follows:
Theorem 5.11 ([1], [18], [24]). Let L =
m
i=1
L
i
be a link. Then L) depends only
on the isotopy class of L and on the Gauss integrals W
e
(L
i
) of each component
L
i
. In particular, an evaluation of L) at representatives of L for which W
e
(L
1
) =
= W
e
(L
m
) = 0 is a link invariant.
Remark 5.12. It is known that this is a universal invariant of nite type. In partic-
ular this means that it is stronger than both the Alexander and the Jones polyno-
mials (it contains the two-variable HOMFLY polynomial) and all other quantum
invariants. Conjecturally the anomaly vanishes and this invariant coincides with
the Kontsevich integral, see [18].
Example 5.13. Let L be a knot, and take n = 2. There are four graphs of degree
two, shown in Figure 5c. We will denote the rst of them X, and the second by
Y . By Exercise 5.7 the weight of the third graph vanishes. Also, choose a framing
of L so that the self-linking is 0; then the contribution of the last graph vanishes
(another way to achieve the same result is to add to L)
2
a certain multiple of the
self-linking number squared); we can set then [X] = [Y ]. Thus we will consider
simply
v
2
=
1
4
_
CX

X
( ) +
1
3
_
CY

Y
( ).
The rst integral is 4-dimensional, while the second is 6-dimensional; none of them
separately is a knot invariant (see [20] for a discussion); however, their sum v
2
is
(see [3], [20])! This invariant is, up to a constant, the second coecient of the
Alexander-Conway polynomial. See [20] for its detailed treatment as the degree of
a Gauss-type map.
5.7. Degrees of maps. How can we explain the result of Theorem 5.11? We may
try to repeat the reasoning of Example 5.8 in the general case. Recall that by
Exercise 5.5 the dimensions of C

and (S
2
)
n
match, so if C

would be a closed
manifold, then equation (32) would be a formula for a calculation of the degree of

. In other words, if C

would have a fundamental class, (32) would be its pairing

26 MICHAEL POLYAK
with the pullback

dvol. That would be great: we would know that W

takes only
integer values, and would be able to compute it in many ways, including a simple
counting of preimages of any regular value of .
Unfortunately, the reality is much worse: C

is an open space, so the degree

is not well-dened and even the convergence is unclear. To guarantee the conver-
gence, we should construct a compactication

C

of C

to which

extends. This
however will cause new problems: the space

C

will have many boundary strata.

There are various way to deal with them: we can relativize some of them (i.e.
consider a relative version of the theory), cap-o some others (i.e. glue to them
some new auxiliary conguration spaces), or zip them up (gluing a stratum to itself
by an involution). But in general, some boundary strata will remain; indeed, in
Example 5.13 we have seen that none of W
X
or W
Y
separately can be made into
a knot invariant. The remedy would be to glue together the conguration spaces
for dierent graphs in

J
n
along the common boundary strata. This tedious work
can be done indeed [18, 24] and (up to a certain anomaly correction) one can in-
terpret L) as the degree of a certain map
n
from a grand conguration space
(
n
to (S
2
)
n
. Too many technicalities are involved to describe this construction in
necessary details, so I refer the interested reader to [18, 24] and will present only a
brief sketch of this construction.
The rst problem is that initially the dimensions of C

for various

J
n
do not
match. E.g., in Example 5.13, for n = 2 the spaces C
X
and C
Y
have dimensions
4 and 6 respectively. This can be xed by considering a product C

(S
2
)
k
of
C

with enough spheres to make the maps

(id)
k
to have the same target
space (S
2
)
N
for all . Now one should do the gluings. When two endpoints of an
edge e of a graph collide, the corresponding boundary stratum of C

looks like
C
G
S
2
for G = /e. Thus we can glue together such strata for all pairs (, e)
with isomorphic G = /e. Some more, so-called hidden, strata remain after these
main gluings. Fortunately, each of them can be zipped-up (i.e. glued to itself by a
certain involution). The only codimension one boundary strata which remains after
all these gluings are the anomaly strata, where all vertices of a graph

J collide
together. These problematic anomaly strata can be glued [18] to a new auxiliary
space. One ends up with a grand conguration space (
n
endowed with a map
: (
n
(S
2
)
N
(glued from the corresponding maps

:

C

(S
2
)
N
). One may
show that the cohomology H
2N
((
n
) of this space projects surjectively to J
n
(see [17]
for a similar case of 3-manifold invariants). Then L)
0
n
= L)
n
+anomaly correction
can be interpreted as the degree of
n
, or more exactly, the image in J
n
of the
fundamental class [(S
2
)
N
] under the induced composite map

: H
2N
(S
2
)
N

H
2N
((
n
) J
n
.
5.8. Final remarks. There remain many questions: which compactication should
we take, why do the antisymmetry and Jacobi relations appear in the cohomology
of the grand conguration space, etc. Each of them is quite lengthy and is out of
the scope of this note. We refer the interested reader to [5, 18, 24]. A mathematical
treatment of invariants of 3-manifolds arising from the Chern-Simons theory was
done in [2, 4]; I especially recommend [17]. While I do not know whether similar
Feynman series arising in other topological problems always have a reformulation
in terms of degrees of maps of some grand conguration space, it seems quite plau-
sible. There are at least some other notable examples, see e.g. [19] for a similar
interpretation of Kontsevichs quantization of Poisson structures.
FEYNMAN DIAGRAMS FOR PEDESTRIANS AND MATHEMATICIANS 27
References
[1] D. Altschler, L. Freidel, On universal Vassiliev invariants, Comm. Math. Phys. 170 (1995)
4162.
[2] S. Axelrod, I. M. Singer, Chern-Simons perturbation theory, Proc. XX DGM Conf. (New-
York, 1991) (S. Catto and A. Rocha, eds.) World Scientic, 1992, 345; Chern-Simons per-
turbation theory II, J. Di. Geom. 39 (1994) 173213.
[3] D. Bar-Nathan, Perturbative aspects of the Chern-Simons topological quantum eld theory,
Ph.D. thesis, Princeton Univ. 1991; Perturbative Chern-Simons theory, J. Knot Theory and
Ramif. 4 (1995) 503548.
[4] R. Bott, A. Cattaneo, Integral invariants of 3-manifolds I, II, J. Di. Geom. 48 (1998),
91133, and J. Di. Geom. 53 (1999), no. 1, 113.
[5] R. Bott, C. Taubes, On the self-linking of knots, J. Math. Phys. 35 (1994) 52475287.
[6] S. S. Chern, J. Simons, Some cohomology classes in principal ber bundles and their appli-
cation to riemannian geometry, Proc. Nat. Acad. Sci. U.S.A. 68 (1971) 791794.
[7] A. Connes, D. Kreimer, Renormalization in quantum eld theory and the Riemann-Hilbert
problem I, II, Commun.Math.Phys. 210 (2000) 249273, Commun.Math.Phys. 216 (2001)
215241.
[8] E. T. Copson, Asymptotic expansions, Cambridge Tracts in Math. and Math. Phys. 55, 1967.
[9] Quantum elds and strings: a course for mathematicians, vol 1-2, (P. Deligne et al, eds.)
AMS 1999.
[10] A. Dubrovin, A. Fomenko, S. Novikov, Modern geometry methods and applications. Part
II. The geometry and topology of manifolds, Graduate Texts in Mathematics, 104, Springer,
1985.
[11] D. Freed, Classical Chern-Simons theory I, II, Adv. Math. 113 (1995) 237303, Houston J.
Math. 28 (2002) 293310.
[12] J.-M. Droue, C. Itzykson, Statistical eld theory, Cambridge Univ.Press, 1989.
[13] L. D. Faddeev, V. N. Slavnov, Gauge elds, introduction to quantum theory, Ben-
jamin/Cummings, Reading, 1980.
[14] W. Fulton, R. D. MacPherson, A compactication of conguration spaces., Annals of Math.
(2) 139 (1994), 183225.
[15] E. Guadagnini, M. Martinelli, M. Mintchev, Perturbative aspects of the Chern-Simons eld
theory, Phys. Let. B277 (1989) 111; Chern-Simons eld theory and link invariants, Nucl.
Phys. B330 (1990) 575607.
[16] C. Itzykson, J. Zuber, Quantum eld theory, McGraw-Hill, New-York, 1985.
[17] G. Kuperberg, D. Thurston, Perturbative 3-manifold invariants by cut-and-paste topology,
math.GT/9912167.
[18] S. Poirier, The Conguration space integral for links in R
3
, Algebr. Geom. Topol. 2 (2002)
10011050.
[19] M. Polyak, Quantization of linear Poisson structures and degrees of maps, Let. Math. Phys.
60 (2003) 1535.
[20] M. Polyak, O. Viro, On the Casson knot invariant, J. Knot Theory and Ramif. 10 (2001)
711738.
[21] A. Polyakov, Gauge elds and strings, Harwood academic publishers, 1987.
[22] S. Schweber, An introduction to relativistic quantum eld theory, Row, Peterson 1961.
[23] A. S. Schwarz, Quantum eld theory and topology, Springer, 1993.
[24] D. Thurston, Integral expressions for the Vassiliev knot invariants, M.A. thesis, Harvard
Univ. 1995, math.QA/9901110.
[25] E. Witten, Quantum eld theory and the Jones polynomial, Comm. Math. Phys. 121 (1989)
351399.
Department of Mathematics, The Technion, 32000 Haifa, Israel
E-mail address: polyak@math.technion.ac.il