Lecture 5.

Feynman graphs
David Kazhdan

5.1Feynman graph expansion. The technique of Feynman graphs allows one to write
down an asymptotic series for functional integral of a QFT in a neighborhood of a free
QFT (see Witten's lecture 1). Nevertheless, the technique itself can be applied to a purely
nite dimensional integral. In this lecture we will discuss this application only.
Let V be a nite dimensional real vector space, let V 0 be its dual. We will view elements
of symmetric algebra Sym(V 0 ) as polynomial functions on V . We x a nondegenerate
positive denite quadratic form b 2 Sym2(V 0 ). Let b 1 2 Sym2(V ) be the corresponding
quadratic form on V 0 .
Let 0 = e b 1=2 dv be the Gaussian measure on V 0 ; we have F(0) = e b=2 .
Let P be a polynomial function on V 0 . We want to study the \perturbed" measure
 = e"P 0 . This is a well-dened measure if " > 0, and P is negative.
We are going to write down an asymptotic expansion for the Fourier transform F().
More precisely, we will obtain a formal Taylor series e 2 C [[V 0 ; "]] in the following way.
Let D(P ) be the dierential operator with constant coecients on V corresponding to P .
More precisely, D is a homomorphism of the algebra of polynomial functions on V 0 to the
algebra of dierential operators with constant coecients on V , such that D(x) = i@=@x
for x 2 V .
We write:
1
X
"P
F ( e 0 ) = e "D (P ) (e ) = n1! "nD(P )n (e b=2 ) :
b=2
n=0
This is obviously an element of C [[V 0 ; "]].
This formal power series is connected with the original analytical problem via the notion
of Borel summability. However in this lecture we will be interested in the formal expression
only.
1
 It is notationally convenient to perform \Wick rotation". So we consider the series
1
X
e(v ) = F(e"P 0 )(iv ) =
 1 n n b=2
n! " Pb (e ) :
n=0

where Pb = D(P ( ix)), so that xb = @=@x.

The series e is dened for arbitrary nondegenerate quadratic form b, and a polynomial
function P over any eld of characteristic 0. (The condition on b to be positive denite is
of course irrelevant in this setting.)
There exists a combinatorial way to \compute" e. To describe it we rst x some
notations.
By a graph we will mean the data consisting of: two sets 1 (edges) and 0 (vertices)
and a map j : 1 ! 0  0. We also assume an involution  of 1 is xed such that
j ( ) = (x; y ) , j ( ( )) = (y; x), and  has no xed points. We will only consider graphs
with no isolated vertices (i.e. pr1  j is surjective). An edge of a graph is an element of
1 = . For a vertex 0 2 0 we dene its \star" as ( 0 ) = f 1 2 1 j pr2  j ( 1) = 0 g. The
set of external vertices is ex = f 0 2 0j#( ( 0 )) = 1g, and the set of inner vertices is
in = 0 e ( ) =
L( 0 ); L( ) =
L( ).
ex . We put: L( 0 ) = Sym#( ( 0)) (V ); L
0 2 0 2 in
Assume that for any 0 2 0 we are given an element s( 0) 2 L( 0). We then take `s =

s( 0) 2 Le( ). For any map s~: 0 ! V we dene ss~( 0 ) = (~s( 0))#( ( 0)) 2 L( 0 ),
0 2 0
and denote `s~ = `s~.
s

Lemma 1. There exists a unique linear function eb on Le( ) such that for any map
s~: 0 ! V we have:
Y
1
eb (`s~) = #Aut( s(pr1  j ( 1 )); s~(pr2  j ( 1 ))) :
b (~
)
1 2 1 =

Proof is clear.
We can view eb as a map b : L( ) ! Sym# ex (V 0 ).
Example. n is a disjoint union of n copies of the graph:   . (The nonoriented
edge on the picture corresponds to two elements of 1 permuted by .) Then L( ) = C
2
canonically, because in is empty. We have:
b n (1) = 2bn n! :
n

1
P
Note the equality b (1) = eb=2 where 0 is the empty graph.
n

n=0 P
Let P , e be as above. We write P = Pn =n! where Pn is homogeneous of degree n
n2
(thus we assume for convenience that P contains no constant or linear term).
Theorem. We have an equality of formal series:
X 
e=
 "# in b (sP ) eb=2
2g
Here g is the set of isomorphism classes of graphs and sP ( 0) = P# ( 0).
Proof. [Another proof appears in x1.3 of Witten's lecture 1].
The idea of the proof is as follows. We want to prove an equality between two expres-
sions, which are power series in " and in Pn's. We will show that both the left hand side
and the right hand side satisfy the same system of linear dierential equations in Pn 's as
variables, and then we will check that the initial conditions coincide.
So, let be a graph and let   ex , be such that jj = n. We denote by  the \cone
over ". This means that  is a graph, such that 0 = 0 [ fg  and 1 = 1 fall
edges of with an end in g[fall edges from  to vertices connected with g. (We add all
edges connecting  to a vertex in , and then erase the vertices belonging to .) Assume
now that n > 1. Then  is an internal vertex of .
For any d 2 S n(V ) and s 2 L( ) we dene sd = d
s 2 L( ).
Lemma 2. For any d 2 S n(V ), 2 g and s 2 L( ) we have
d^ X
(b( )(s)) = b (  )(sd )
n!
 ex ;jj=n
(Here d^ is the dierential operator with constant coecients on V corresponding
to d.)
Now we can prove the theorem. Let l(P ) denote the left hand side and let r(P ) be the
right hand side of the equality. We want to write down a certain system of dierential
3
equations in Pn's as variables, which will be satised by both sides. It is easy to see from
the denitions that l(P ) satises the system of equations
@l(P )
@Pn
(Q) = ("Q=n
b !)l(P )

where Q is an arbitrary polynomial of degree n on V 0 and Qb is the corresponding dierential

operator on V ; in the left hand side Q is considered as a tangent vector to the space of
polynomials of degree n.
But now Lemma 2 tells us precisely that the right hand side of this expression is equal
to @r@P(P ) (Q).
n

Therefore l(P ) and r(P ) satisfy the same system of dierential equations. Let's now
check the initial conditions. By the denition, l P =0 = eb=2 . On the other hand, one also
has r P =0 = eb=2 by the above Example. Hence we have l(P ) = r(P ) for any P , so the
theorem is proved.
Remark 1. The theorem and its proof along with all the statements below can be immedi-
ately generalized in the following way. We can take P to be an element of C [V ][["] rather
than a polynomial, i.e. P = 1
P m
m=0 " P(m) is a Taylor series in " with polynomial functions
on V 0 as coecients.
Remark 2. This way of writing an asymptotic expansion for a measure uses Fourier trans-
form, and hence the linear structure of V heavily. For example when one applies the
Feynman graph expansion to the gauge theory, V being the space of connections modulo
gauge transformation, then V has no natural linear structure. In this case one still can
write an asymptotic expansion for functional integral, but an individual term assigned to
a particular graph is not canonically dened.
P
We denote Z = 2g b (sP )"# in .
One of the diculties in using this expansion is that there are \too many" terms; so we
P
try to reduce it somehow. We denote F = 2gcon b (sP )"# in , where gcon is the set of
connected nonempty graphs.
Note that F (and all series that will be introduced below) lie in a subring C [V 0 ][["] 
C [ V 0 ; "] .
4
Claim 1. F = log Z .
Proof. Decomposing a graph into union of its irreducible components one can identify the
two combinatorial expressions:

Z =
X
b =
X 1 b 1 : : : b = exp(F )
k

2g 1 ;::; 2gcon ;n1 ;::;n

k k
n 1 ! n2 ! ::nk !

where the summation is over all sets of distinct elements i 2 gcon.

We next want to study the \quasi-classical approximation" to our integral. So we
introduce another variable ~ and consider the expression:
 
F e(b =2+"P )=~ dv (~v) :
1
F~ (v ) = log

As an immediate corollary to the Theorem, we get:

P
Lemma 3. F~ = ~ 1 2gcon ~h1 ( )b (sP )"# in where h1( ) = dim(H 1 ( )) = # 1=
# 0.
Proof. Apply the same formal procedure to the quadratic form ~b = ~b and polynomial Pe =
~ 1 P , and note that for a connected graph we have ~b (sP=~ ) (~v ) = ~h ( ) 1 b (sP )(v ).
1

5.2 Quasi-classical (low-loop) approximations. Recall that the \classical" approx-

imation to Fourier transform is the Legendre transform. An appropriate version of the
denition of Legendre transform is as follows.
If f is a function on V 0 , then we can view its dierential df as a map df : V 0 ! V .
If df is an isomorphism then the Legendre transform of f is dened by



L(f )(v ) = v; (df ) 1 (v ) f (df ) 1 (v)

i.e. L(f )(v) is the critical value of the function v f where v is considered as a linear
function on V 0 .
Let now G 2 C [ V; "] be a formal power series. Then its dierential dG (in the V 0 -
direction) is an element of Hom(V 0 ; V )
C [ V; "] .
5
 Assume that G = g0 + g1 + "G1 where g0 is a nondegenerate quadratic form on V 0 , and
g1 contains no terms of degree less then 3 in V . Then dG is invertible, i.e. there exists
H = (dG) 1 2 Hom(V; V 0 )
C [ V; "] with no constant term such that dG  H = id and
H  dG = id. In this situation the Legendre transform L(G) 2 C [[V 0 ; "]] can be dened by
the same formula.
To describe the next term of the asymptotics we dene for any function f on V 0 as


above a new function H (f ) on V as follows. Let Hessf (p) = det @v@ @vf jp be the Hessian
2
i j

of f (we assume that a constant volume form on V is xed so that the determinant of the
quadratic form is taken with respect to that volume form).


We put: H (f )(v) = Hess (df ) 1 (v) .
For G = g0 + g1 + "G1 2 C [[V 0 ; "]] as above we can dene Hess(G) and H (G) 2 C [[V 0 ; "]]
by the same formula.
Claim 2. a) Let F0 = P 2T b (sP )"
# in , where T is the set of (nonempty) trees. Then
is the Legendre transform of b 2 "P .
1
F0
P
b) Let F1 = b (sP )"# in where g1 is the set of one-loop connected graphs (i.e.
2g1
graphs with h0( ) = h1( ) = 1. Then F1 = log H b 2 1 "P 1 . (We assume that

the background volume form is such that det( b 2 ) = 1. Then H ( b 21 "P ) 2 1 +

1

"C [ V 0 ; "] + V 0 C [ V 0 ; "] , and log(H ( b 2 "P )) is a well-dened element of C [ V 0 ; "] ).
1

Proof of the Claim. From Lemma 3 we see that F~(v) = F0 (v)=~ + F1(v) + O(~). Thus
the claim follows from the stationary phase approximation applied to the integral
Z
F(f )(iv) = 1d=2 e(v;p)+("P b =2)(p)dp
1

(2)
V0

One can also directly show the equality of the two combinatorial expressions:
 
b 1 1 X X Pn
F0 d(
2 "P )jp = b 2 (p) "
Pn
1)! (p) + " (p)
n (n n n!
b 1   b 1 
(5.1) = hp; d 2 "P jp i (b 1 =2 "P ) d(
2 "P )
6
and
 b 1  b 1 
(5.2) F1 d( "P )jp = log Hess 2 "P
2


Here (5.1) is equivalent to F0 = L b 2 1 "P and (2) is equivalent to F1 = log H b 2 1

1

"P .
Let us sketch the combinatorial proof of (5.1) and (5.2).


Let us identify V and V 0 by means of b and use an equality "# in b (sP )  d b 2 1 "P =
P
( 1)c "# in b (sP ) where the graph i is obtained from the graph by adding ci vertices
i
i i

of arbitrary valencies, each one of which is connected with an external vertex of in such
a way that any external vertex of is connected with no more then one new vertex.



Now it is not hard to check that in the expansion of F0  d b 2 1 "P all terms with
"i for i  2 cancel, and identify the sum of remaining terms with the RHS of (5.1).



Likewise in the expansion of F1  d b 2 1 "P only the terms with one-particle irre-
ducible one-loop graphs (\circles") do not cancel; the sum of these terms is identied with
P1 i n
n=1 " tr(dP ) =n = log det(Id dP ) which coincides with the RHS of (5.2).
5.3 Eective potential. Let us call a nonempty graph 1-particle irreducible (or just 1-
irreducible) if it is connected and remains connected after removal of any internal edge.
To any connected graph there corresponds a unique tree with a 1-particle irreducible
graph assigned to each vertex together with an identication of the set of edges coming
to a vertex of a tree with the set of external vertices of the corresponding graph. Thus
\computation" of F0 can be reduced to summation over trees and over 1-particle irreducible
P
graphs separately. Denote F1 irr = 2g1 irr "# in b (sP ) where g1 irr is the set of 1-
particle irreducible graphs. (Notice that the graph   is 1-particle irreducible by the
denition).
Claim 3. =b
F1 irr b (L(F )) where b is viewed as a map from V to V 0 and L is the
Legendre transform.
Proof. By Remark 1 after the proof of the theorem we can apply claim 2 to P 0 =
" 1 (b 1 ) (F1 irr b=2) 2 C [V 0 ][["] .
7
 From the above combinatorial observation we see that F = FP = (F0 )P 0 . Hence claim


2 implies that F = L(b 1 =2 "P 0 ) = L (b 1 ) (b F1 irr) . Since Legendre transform is
involutive we get the claim.
F1 irr is called the eective potential of the theory.