This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
represents an inner product of I L%> and I#>; the totality of the states spans a Hilbert space. t
From (66) and (67), we obtain
('V(x), V(y)1 = f d4k (sgn k(°) 6 (m2 
k2)e_ik(xy)
(69)
where J d4k stands for the integration over all 4momentum space of k = (k(0), k(l), kit>, 1ct3>), and sgn k(O) = kt°>/I kto>I . [For the meaning of
the 6function used here, see Appendix A1.] Since the righthand side of (69) vanishes for xt0> = yi0> with x + y and is Lorentz invariant, we have
[v(x), ,(y)] = 0 for (x  y)2 < 0
(610)
Such a relation as (610) is called the microcausality condition. It is closely related to Einstein's relativity principle, which states that any action cannot propagate faster than the light speed (c = 1).
Now, we consider field operators which are not free. Since various elementary particles interact each other with some regularity, we have to specify in what ways various field operators are mutually related; it is called dynamics. To specify dynamics, it is the orthodox way to consider a Lagrangian L of the system; L is a Lorentz invariant, herlnitian, operatorvalued functional of various fields. It consists of a free Lagrangian Lfree and and interaction Lagrangian L;,,i. The Euler equations yielded by Lfree are introduces additional terms to them; for equations for free fields, while example, the Euler equation yielded by L for a spinless particle is [m2
+ (alax)2] ve(x) = Q(x)
(611)
where 9(x) is nonlinear with respect to fields and represents interactions. The field operator which satisfies an equation containing an interaction term such as (611) is called a Heisenberg operator. It is prohibitive to describe its operator properties completely, but it is usually supposed to satisfy the microcausality condition such as [p(x), zp(y)] = 0 for (x  y)2 < 0
(612)
Because of the translational invariance of the theory, there exist four generators of translations, which are denoted by P = (P(0>, Ptl>, p(2), p(3)) t The above description of operators and states is not mathematically rigorous. We here intend to explain the theory intuitively. 4
Nakanisbi
50
GRAPH THEORY AND FEYNMAN INTEGRALS
Their simultaneous eigenstates form a complete set of states. The eigenvalues of P(O) and p2 are supposed to be real and nonnegative because they
represent the energy and the invariant square of the 4momentum of the system, respectively. From physical consideration, the existence region of the eigenvalues of P2 is usually further restricted. This restriction is called a spectral condition. The axiomatic field theory [9] is a theory which starts from postulates such as the invariance under the Poincare group, the Hilbertspace properties of states, and microcausality and spectral conditions. In the
axiomatic field theory, neither the existence of a Lagrangian nor detailed operator properties of fields are assumed. Therefore this theory cannot describe the system completely, but from it we can obtain certain amounts of information about the general framework of the quantum field theory. For example, with one additional postulate, one can show that the Smatrix exists and is unitary for the physical values of 4momenta (this fact represents the conservation of total probability). In order to find the Smatrix explicitly, we have to specify not only the Lagrangian but also the operator properties of fields completely. To do this, it is convenient to make a (formal) unitary transformation of states in such a way that equations of fields reduce to those of the corresponding free fields. The transformed states are called the states in the interaction re
presentation, which are timedependent. We can describe those states Lorentzcovariantly by using a certain device, but we here omit details. In the interaction representation, we know the commutation relations of fields completely such as (69), and therefore the theory is fully specified. Given a state at a particular time, we can in principle calculate a state in any time by means of the Hamiltonian formalism. The state at xt0)   oo and that at
x(0)  + oo are called the initial and the final state, respectively. The particles in those states are supposed to be free. The inner product of the initial state and the final one is nothing but the Smatrix element. After a lot of analysis [10], which is known as the perturbation theory, one finds that the Smatrix is given by S  9/, where 1M
M
E M! i=i 11 (J d4xj T[2snt(xi) ... M=O
2'1nt(x,f)]
(613)
Here
Lint = f d4x .int(x)
(614)
and the symbol Tindicates that the succeeding operators should be arranged
in the order of times xj(0). The series expansion of S, which is called a
FEYNMANPARAMETRIC FORMULA
51
perturbation series, is only formal because, as we see later, it is not welldefined mathematically in several respects. Nevertheless, it is very important to analyze (613) in detail because it is the only known explicit expression for the nontrivial Smatrix in the quantum field theory. The Smatrix element is dependent on certain factors related purely to the
quantummechanical states of the particles present in the initial and final states. It is convenient, therefore, to drop those "kinematical" factors from the Smatrix element because they are not essential. Furthermore, when we work in the momentum space, that is, when we take the fourier transform, it is customary to omit a fourdimensional Sfunction expressing the overall conservation of the 4momentum, which is due to the translational invariance of the theory. In this way, we obtain the essential part of the Smatrix element; we call it a transition amplitude.
The transition amplitude is most conveniently calculated by means of Feynman integrals. They are easily written down according to the Feynman rules [10, 11, 12] if the corresponding Feynman graphs are given. A Feynman graph is constructed as follows. The number of its external lines is equal to the total number of the particles present in the initial and final states. The
number M of vertices corresponds to the order of the term in (613). If 2in,(x) is a monomial of degree k in fields (its coefficient is called a coupling constant), the Fdegree of each vertex is equal to k. The lines incident with a vertex should be labeled according to the names of the fields involved in
2int(x). If 2i.,(x) is a polynomial, each vertex is specified by one of its terms in the abovementioned way, that is, there are various types of vertices. Each of external and internal lines should be incident with a vertex or a pair of vertices so as to be consistent with the abovementioned labeling. For an Mth order term of (613), it is possible to construct many Feynman graphs. It equals a sum of Feynman integrals (apart from a fourdimensional 6function) over all possible Feynman graphs which involve no vacuum polarization parts (this restriction comes out from being divided by in the definition of S). Let p and m be a 4momentum and a mass of a particle, respectively. If p satisfies the relation p2 = m2, then it is said to be on the mass shell; otherwise it is said to be off the mass shell. In the transition amplitude, all the
4momenta of external particles (i.e., particles associated with external lines) are on the mass shell. It is possible and convenient, however, to extend them to the ones being off the mass shell straightforwardly. The offthemassshell extension of the transition amplitude is called the extended 4*
52
GRAPH THEORY AND FEYNMAN INTEGRALS
transition amplitude. Correspondingly, Feynman integrals also are extended to the quantities off the mass shell. Hereafter, we always use the terminology of the Feynman integral in this extended sense without writing so explicitly. Let n be the total number of external particles. The transition amplitude is physically meaningful only for n >_ 4 because at least two particles are present in each of the initial and the final state. The transition amplitudes for n = 4 and for n >_ 5 are called the scattering amplitude and the production amplitude,respectively. On the other hand, the extended transition amplitude 2. The extended transition amplitudes for n = 2 and is meaningful for n
for n = 3 are called the selfenergy and the vertex function, respectively. Those for n = 4 and for n z 5 are called the extended scattering amplitude and the extended production amplitude, respectively. 62
Definition of the Feynman Integral
We first consider the simplest case in which the interaction Lagrangian density 2;,,,(x) is a polynomial in spinlessparticle fields alone and involves no derivatives of them. It is convenient to work in the momentum space, that is, we take the fourdimensional transform. Then the Feynman propagator for a spinless particle labeled by 1 is given by
i/(mI  ki  iei)
(615)
where m, and k, are its mass and 4momentum, respectively, and e, is a positive constant which should tend to zero at the final stage of the calculation. We consider a connected Feynman graph G, which has N internal lines and M vertices. The Feynman integral associated with G is a formal multiple integral
\ fi 64(I [b : l] k, + pbl FG(p) = 'EIGI
f
d4k !1 1
bEV(G)
11
1
Gj
(ml  kl  iEI)
/
(616)
IEIGI
that is, it is an integral of a product of the Feynman propagators corresponding to all internal lines of G (apart from a numerical factor) such that the 4momentum is conserved at every vertex of G. In (616), the integration range goes over a 4Ndimensional real euclidean space, and 64(k) indicates 3
a fourdimensional dfunction, i.e. fi 6(k'J)); Pb is an algebraic sum of J=O
external momenta outgoing from the vertex b, where an external momentum
FEYNMANPARAMETRIC FORMULA
53
is a 4momentum of a particle associated with an external line. j We call k1 an internal momentum and likewise ml an internal mass. From (25), we have
Z ( E [b :1] kl + Pb) _bcv(G) Y Pb
bev(G)
(617)
1EjGj
Hence (616) is proportional to S4(
Pb), that is, we have the overall
11 bV(G)
conservation of the 4momentum. The number of independent fourdimen
sional 6functions involving integration variables is rank sad = M  1. After carrying out those M  1 trivial integrations, the number of the remaining fourdimensional integrations is N  (M  1) = It, the connectivity of G. It is convenient to decompose k, into an integration momentum k, and a constant one q1 in such a way that
k1 = k1 + qt (l a IGI) [b : l] q1 + pb = 0 (b c v(G)  a)
(618) (619)
IcIGI
where a is a particular vertex. Then 64 bev(G)
(
\ IEIGI
[b :1] k1 + Pb = 64( beu(G)
Pb)
k' S4( [b : 1] k1
b*a
(620)
\ lCIGI
Because of (318) and (37), we have N
L S4( E [b: l] k1) = 1I S4( , [Si :1] k1) \IEIG1 b*a i=/s+1 l1cJGj
(621)
where {Sµ+1, ..., SN} is the fset of cutsets corresponding to a tree T = {lµ+1, ..., IN} in the sense of Theorem 229. Furthermore, Theorem 35 implies
IEI
[Si : 1]
k, = k,,

14
[C1 : 1i] kt,
i
(j = ,u + 1, ..., N)
(622)
where {Cl , ..., Cj is the fset of circuits corresponding to a chord set T* = {11, ..., lµ}. By means of (618)(622), (616) becomes FG(P) = 64
(
bev(G)
Pb) jj7j
I'ET*
fr
J
d4k1,)
z
fI [ml  (kl + q1)2 IEtGl (623)
where the momenta k, for 1 e T are redefined now by !t
k, = Y_ 1C': I] k,, I=1
t We also call Pb an external momentum later.
(624)
54
GRAPH THEORY AND FEYNMAN INTEGRALS
The ,u momenta fC, for 1 e T* are called basic momenta; (624) becomes a trivial identity for any one of basic momenta. We note that the constant momenta q, are not uniquely defined by (618) and (619). Indeed, for any chord set T*, we can arbitrarily choose q, for any 1 E T*, because then the M  1 simultaneous equations
E [b: 1] q, = Pb  E [b: 11 !ET
(b + a)
q,
(625)
LET*
can be solved with respect to q, (1 e T) owing to det sV' + 0. The Feynman integral (623) will, in general, not well defined mathematically owing to the following four possible reasons.
(1) Since Feynman propagators should be regarded as distributions (see Appendix A), their product as in (623) is not necessarily well defined. This difficulty can be bypassed by regarding s, to be finite temporarily. (2) The integral is at best conditionally convergent because (k, + q,)2 can remain small for very large fCj")(v = 0, 1, 2, 3). Therefore the way of carrying out the integrations has to be specified explicitly. The most satisfactory way is to convert it to a Feynmanparametric integral as done in the next section. (3) Even if the difficulty (2) is removed, (623) may be divergent owing to the contribution from very large k,(°). This divergence is known as ultraviolet divergence. This difficulty is removed by renormalization as explained in Subsection 102. At present, we bypass it by replacing (623) by
FG(p, 2) = S4( I Pb) fl (f dkk,,) 11 [mi / L'ET* bEV(G)
(k, + q1)2
IcIGI
(626)
where Re 21 > 0 should be chosen in such a way that (626) is free from ultraviolet divergence.
(4) When some masses are zero, (623) may be divergent owing to the contribution from vanishing k, = lc, ± q, . This divergence is known as infrared divergence, which cannot be removed in the transition amplitude but automatically cancels out in the transition probability. We avoid this difficulty here by supposing m, > 0 for all 1 E IGI
.
Finally, we briefly mention the case in which ',,,,(x) involves nonzerospin
particle fields and/or derivatives of fields. The Feynman propagator for a Dirac particle is given by (615) multiplied by m, + yk,. [Any line corresponding to a fermion such as a Dirac particle has to be given an intrinsic
FEYNMANPARAMETRIC FORMULA
55
direction. The Feynman propagator is defined according to this direction.] Likewise, the Feynman propagator for a higherspin particle can be obtained from (615) by multiplying a certain (matrix) polynomial in kj. If Pint(x) involves nonfield factors such as constant matrices and differentials (8/8x), then for each vertex b we should introduce a matrix factor or a polynomial in k1(1 e S[b]). We need some additional rules for treating matrices, but we omit them. After all, we should multiply the integrand of (616) or (623) by a certain matrix polynomial, P(k), in k1 = k, + qi (1 e IGI), which may also depend on external momenta. It is difficult to compute a product of matrices involved in P(k) generally, but recently a systematic method of calculating a product of y matrices has been proposed [13].
§7 FEYNMANPARAMETRIC INTEGRAL 71
Derivation of the FeynmanParametric Integral
It is important from both theoretical and practical points of view to convert the Feynman integral defined in Subsection 62 into a Feynmanparametric integral, whose integration region is compact. The Feynmanparametric integral in the general form is derived on the basis of the following generalized Feynman identity. THEOREM 71. n
rj Axi'
J=1
U I'(A )
i=1
d(1x
n
i=1
a;l )
Ijafi1
i=1
n
lAD
a;AA/ n
where A; + 0 and Re A. > 0 for all j and AO =
1
= (n  1)!
JdcJ)
jl J=1
i=1
8 (1
2j. Especially,
 ;=1 ai n
% A;I
o
rjA,
(72)
;=1
J=1
/
In the above, we may replace the upper limit, oo, of each integration by 1 because of the afunction.
Proof We can prove (71) by mathematical induction with respect to n by using Euler's formula 1
as1(1 f da J 0
_
01)b1
[aA + (1  x) B]'+'
with Re a > 0 and Re b > 0. q.e.d. 56
I'(a) I'(b)
l(a + b)
1
1
A°Bb
(73)
FEYNMANPARAMETRIC FORMULA
57
Our procedure is first to apply (71) to (626) or (72) to (623), second to interchange the order of integrations,[ and finally to carry out the momentum
integrations explicitly. After the first and second steps mentioned above, (626) becomes
FG(P, 2) = 64(I Pb) [ II 1'(AI)]1 f dl' X rj ai`1 H(q, x,2) b
IEIG!
J
1EIGI
A
(74)
Here
f dN1 a
H
(f
IEIG ` 0
A
dal I d(1 J \
al)
(75)
H(q, a, A) = 1(20) 11 (f d4k1) [O(k, q, a)  k]'O
(76)
icT*
where
20 = Y 21
(77)
IeJGj
E = Y X181
(78)
IeJGj
and 0(%, q, a) =
a1[mI  (i + q,)2]
a1fC2
IEIGI
2
c 1ql
l + L(q, X)
I
I
(79)
with
L(q, a) =
1E GI
o,(mi  qi)
(710)
To proceed to the third step, we substitute (624) for kI in (79): 1=1 J=1
where
[C': 1] [Ci : 1] IEIGI
al)
qc =
2
gck,., +
Jot
L
(711) (712)
[C: 1] a1g1 IEIGI
It is convenient to reexpress (711) in the matrix notation. For this purpose, hereafter we introduce the vector notation. We denote the column vectors and (gct, ..., gc,,)t by kT. and qT., respectively. Then (711) is rewritten as
0 _ (kT.)t QlT*kT 
kT. + L
(713)
where °11T. is defined by (323). t We do not prove the adequacy of this interchange because it should rather be regarded as the definition of the Feynman integral (see Section 6).
58
GRAPH THEORY AND FEYNMAN INTEGRALS
All the eigenvalues of the symmetric matrix 411T. are
THEOREM 72
positive for at > 0 (1 e IGI).
Proof
Any principal minor of °11T. is expressed by a derivative of
U = det °11T. with respect to at for 1 E I (I c T*), which is positive definite because of (325). q.e.d. [We can also prove this theorem by noting that Ya1k i
is a positivedefinite quadratic form if we suppose k: to be real numbers.] Let 77 be a very small positive number and
f
a
ac", dN1 7

1E IG i
(
j
(714)
da,ll
6(1  `E
a,)
Then (°11T.)1 exists for a, z 77 (1 c I GI). Let
kT. = kT.  (glT.)_1 4T
(715)
then
_ (kT*)t QIT.k'. + (4T*)' (°&T*)1 QT. + L
(716)
The Jacobian of the transformation (715) is of course unity. Next, we transform &T. into a diagonal matrix 2T.: kT. = YT.kT« (717)
WTO 1 O&T+YT* = 2T with (27.)t = (2T.)1. Then (716) becomes
_ Here
qT*kT. + (qT.)t
qT. + L
det 2T. = det °11T. = U(a)
(718) (719)
because of (330). THEOREM 73
f
d4k
2i
1
(720)
(2  1) (A  2) J (a  bk2)1 where Im a < 0, b > 0, and Re A > 2, provided that the k(0) integration is carried out first. b2a2_2
Proof Since the integrand is holomorphic in the first and third quadrants in the k(0) plane, we can rotate the contour into the imaginary axis by means of Cauchy's theorem. We then have an integral involving a fourdimensional euclidean vector, which can be easily calculated by considering it in polar coordinates. q.e.d.
59
FEYNMANPARAMETRIC FORMULA
By means of (720), we can straightforwardly carry out the integrations over ki'(1 e T*); (76) becomes (721)
H(4, a, A) = F(20  21u) (n2i)'/U2(a) [V(4, a)  is]x0 2
where from (718) with (710) we have
V(4, a) ° E al(mi  q) + (4T*)` ('&T*)1 4T*
(722)
ZEIGI
The above result can be summarized as follows. THEOREM 74
FG(P, A) = 64 6
Pb) (n2i)lJa(P, A)
(723)
with r(710  2,u) f,(P 2)
II leIGI
Y acx,1
lim r dv1 a U2(V a+0 J leIGI
(724) i8)102'
If we take Re 71 > 2, then as seen in Subsection 101, the 77  + 0 limit exists, that is, we may replace lim f by The genuine Feynman '1,+0
df )
f
integral FG(p) is formally FG(p, 1) for N > 21u, where 2 = 1 means 21 = 1 for all 1 E IGI. We write fG(p)  fG(p, 1). In deriving (724) we have employed a particular chord set T*. Since the original Feynman integral (616) is, however, independent of T*, (724), and
therefore V, should be independent of it. We can check this property directly.
Let kJ be a column vector formed by It independent integration momenta,
so that kT. = fk, with det 0. Then °1 T. and qT. in (716) are replaced by QIJ  /IGIIT./ and qJ = ftgT., respectively; the kJ integrations may be done by the same manipulations as used above. We observe that the extra factor (det &,)2/(det U?CT.)2 in the denominator exactly cancels the
Jacobian (det /)4 and that (4J)`
72
4J = (4T*)'
(GIIT.)1 qT.
(725)
Topological Formulas for the V Function
The central importance in this chapter is to derive topological formulas for V defined by (722).
GRAPH THEORY AND FEYNMAN INTEGRALS
60
THEOREM 75 For any values of q{ (1 E I GI) which may not satisfy (619), V is expressed as follows:
(a) Circuit representation [3]
V=
{EGI
a,({ni  qi) + U1
Uc(Y_ [C: 1] a1g{)2 CEC
(726)
{EIGI
(b) Cutset representation [7, 8]
V=
a{mI
WS('E [S: 1] q{)2
 U1
(727)
/
\ {EIGI
SEs
{EIGI
Proof of (726) [3] Comparing (722) with (726) together with (712), we see that we have only to prove k
u ['&T*]{.J 4C,4C,
{=1 J=1
(728)
=Z Ucgc CEC
with [gIT.]"J = [°11T.]C''CJ. In the following we prove the identity
EE
{=1 J=1
_
[q1T.]{,J ( Y
[C{ :1] g{)
I
2
[cC:1'] q{') _ Y_ Uc( CEC
[C:1141
{
(729)
for arbitrary q{ (1 e IGI); (728) follows from (729) by setting q{ = a{q{.
Now, we take an arbitrary circuit C and put a(C) = 0 (i.e., a, = 0 for all 1 e C). Since the lefthand side of (729) is independent of T* as shown in Subsection 71, we may choose an fset of circuits {C1, ..., C,j which contains C, say C1 = C. Then Theorem 313 shows while
[q'T]a(c)=o = UC [°I1T]a'k)=o
= 0 unless i = j = 1
(730) (731)
because all elements of its first row or column vanish. Therefore the lefthand
side of (729) reduces to UC ( [C: 1]q,) 2 for a(C) = 0.
Next, for any C' 4 C we prove
U. = 0 for a(C) = 0
(732)
Since Uc' is a chordset product sum in G/C', any term of it corresponds to a chord set T'* in G/C'. Since C C', C  C' contains a circuit in G/C' because of Theorem 22 (c). Hence T'* n C + 0, from which (732) follows. Let R be the difference between the lefthand and the righthand side of
(729). From the above arguments, R vanishes when a(C) = 0 for any
61
FEYNMANPARAMETRIC FORMULA
circuit C. However, since R is a polynomial in at (1 E IGI) of degree /Z  1, u1
if R * 0 then R contains a term of the form k fJ a,, (lc + 0) for some ,s1
J=1
lines 11i ..., 1,i_1, which may not be all distinct. Let I  U {1;}; then 1=1
N(I*) z M, where I* IGI  I. Hence 1* includes a circuit C because of Theorems 26 (b) and 25 (d), so that I n C = 0. The existence of such a term is inconsistent with the property of R derived above. Therefore R  0. q.e.d. Before proving (727), we prove the following identities. THEOREM 76
a U/Ba, = E Uc
(a)
(733)
Ca l
al  Ul,° _ YSalWS
(b)
(734)
[C: i] [C: 1'] alas, U
(c) cec
[8: 1] [S : 1'] WS
for 1 + /'
SeS
(735)
Proof (a). It is convenient to translate (733) into the inverseFeynmanparametric form: (736) Z, f I Nk L j 1 A, = Y, Tot kcT+1 C31 TOC keT°
where T is a tree, T° being a pseudotree. Given a tree TO 1, T'  T + 1 is a pseudotree such that 1 E C c T° where C is a circuit. Conversely, given a pseudotree T° C 3 1, T  To  I is a tree which does not contain l (see Theorem 26 (b)). Thus (736) holds identically. (b). We rewrite (734) as Y_ 11 1'k Tel keT1
=E
E
fl
Sal T2eT2(S) kcT2
(737)
1'k
where T and S denote a tree and a cutset, respectively. Given a tree T a 1, there exists a unique cutset S such that T n S = {I} because of Theorem 229.
Then the 2tree T2  T  I is disjoint from S. Conversely, given a 2tree TZ E T2(S) with S 31, T = T2 + 1 is a tree because of Theorem 26 (b). (c). The equality which we have to prove is
Y_ [C: 1] [C: 1] E C
fl
To.C kcTo11'
F'k = Y_ [S: 1] [S: 1'] S
for 1 + 1'
fl
T2eT2(S) kcT2
Nk
(738)
62
GRAPH THEORY AND FEYNMAN INTEGRALS
Given a 2tree T2 e T2(S) with S a 1, 1', TO = T2 + 1 + 1' is a pseudotree because T2 + 1 is a tree. Therefore, TO includes a unique circuit C, and 1, 1' E C because both TO  1 and TO  1' are trees. Hence S n T2 = 0 implies C n S = {l, 1'}, whence (212) reduces to [C: 1] [S: 1] + [C: 1'] [S: 1'] = 0
(739)
[C: 1] [C: 1'] =  [S: 1] [S : 1]
(740)
that is, Conversely, given a pseudotree TO C a 1, 1', since TO  1 is a tree, from Theorem 229 there exists a unique cutset S such that (T°  1) n S = {I'}.
Hence T2 = To  1  1' is a 2tree, and 1, 1'e S because N(S n C) is even. Thus TO and T2 e T2(S) are in onetoone correspondence. q.e.d.
Proof of (727) Since
U = 8U aI + Ula'_o for any 1 8xl
we have
UIEIGI E x,qI =
IGIG
(
Cal
Uc)
aI ql +
IEIGI
t , (Ul a, =o) qi
(741)
(742)
with the aid of (733). Hence the coefficient of q2 in (726) is  U1 E aI Ula,=o
(743)
IcIGI
and that of glgl. (1 + 1') is U
(744)
[C: 1] [C: 1] 0,10,1, UC C
On account of (734) and (735), therefore, we immediately obtain (727). q.e.d. THEOREM 77
In terms of external momenta V is expressed as follows:
(a) Vertex representation [3, 14, 15]
V=
IeIGI
c mi + U1 Z W (b1 c) PbPC (b.c}c9
(745)
(b) Modified cutset representation [4, 5, 6]
V= ieIGI
!XIm2 I
 U 1 E W(h19h) (h I9h)
bEh
Pb 2
(746)
where g is the set of all external vertices (i.e., Pb = 0 for any b 0 g) and the
summation E runs over all nontrivial divisions of g. (hl9h)
FEYNMANPARAMETRIC FORMULA
63
Proof From (727), (746) immediately follows by using the momentum conservation
E [S: Z] ql = ± E 1EIGI
for S E S(h l g  h)
Pb
(747)
bEh
and (343) (without tildas). Here (747) is derived by means of (210) and (619).
The equivalence between (745) and (746) are shown as follows: (hlgh)
{b,c}cg
W (bl `) PbPc =
{b,c}cg
= (h1gh)
PnPc
W
h9b
gh3c W(hlg1'
(
(
g
ceh
bh be h
_  E W(hlgh) ( E pb)2 (hIgh)
(748)
bEh
with the aid of (346) (without tildas) in the first step. q.e.d. We can also present a direct proof of (745) in the following way. We introduce the following matrix Y"^T+ of order ,a + 1: 2
a 1q1 T* =
(qT )t
1
I
qTS
(749)
&T
Then from the definition (722) of V we have U(
almi
 V) = U I
algi  1,1'eT*
[?tT.]1'1 Qtg1' = det YET.
(750)
As remarked in Subsection 62, we may put q, = 0 for all 1 E T*, and then the remainders q, (I E T) are fixed as linear combinations of external momenta pb. Furthermore, by means of the overall momentum conservation we have
Pb = Pb Y_ Pc
(751)
c*b
Therefore the quantity in (750) can be expanded into
Y
{b,c}cv(G)
AbcPbPC =
E AbcPbPc
(752)
{b,c}cg
where the coefficients Abc are functions of a, (I E IGI). Comparing (752) and (745), we have only to show
 Abc = W (b l c)
(753)
64
GRAPH THEORY AND FEYNMAN INTEGRALS
Since we can choose external momenta arbitrarily as long as they satisfy the conservation law, we put
Pb = pa, Pa = 0 unless a = b or
c
(754)
then (752) reduces to AbCpb . Accordingly, our task is to show that det YET, with (754) is equal to W(bl`)pb
Now, the tree T includes a unique path P e P(bc) because of TheoG(b`)' which is obtained from G rem 25 (c). We consider a Feynman graph by linking b and c by a new line l0 . Let {C1, ... , Cu} be the fset of circuits in G corresponding to the chord set T* = {/,, ... , 1,1; then since T* n P = 0, {C,, C1, ..., Cu} is an fset of circuits in GOO) % where Co  P + 10 . We define the orientation of C0 by setting [C0 : 1] = [b : l0], and put
q1 = [C0: 1] Pb
for all 1 e IGI
(755)
Then (755) satisfies the conservation law at each vertex and q1 = 0 for all
1 e T* as required. On substituting (755) in (712), we find that (749) becomes
det Yl r = pb det
[Cf : 1] [C f, : 11'X1 I id = 0, 1, ..., ,u)
(756)
\1EIGI
The matrix in the righthand side of (756) is exactly the 071 matrix in with a,. = 0, whence its determinant equals the chordset product sum in G°' because of Theorems 310 and 319. Therefore
G(bc)'
det i' T = W("Ic)Pb
(757)
because W(11`) is the U function in G('°). q.e.d. 73
Remarks
In this subsection, we make some remarks on applications of the Feynmanparametric formula. We consider it at the physical value of A, that is, we consider FG(p) = FG(p, 1). THEOREM 78 For every Feynman integral FG(p), there always exists an interaction Lagrangian density in1(x) such that the nonvanishing lowestorder contribution to a certain extended transition amplitude coincides with FG(p) apart from a constant coefficient.
Proof We define £° (x) by
E
bev(G)
7
f,j
1eS[b]
01(x)
(758)
FEYNMANPARAMETRIC FORMULA
65
Here lpb(x) (b e g) and f1(x) (l e I GI) are field operators corresponding to distinct spinless particles, and pb(x)  1 for b 0 g; the mass of ?P,(x) is equal to mi, and gb (b e v(G)) denotes a coupling constant. It is easy to show by
means of the standard analysis [11] that the nonvanishing lowestorder extended transition amplitude for the process specified by g is exactly given by FG(p) multiplied by a certain constant. q.e.d.
The proportionality constant is given by i M(bf gb) (27c) 41 [ t/(27t)4]N
(759)
EV(G)
Here the first factor comes out from the coefficient in (613) (1/M! is canceled by the M! permutationst of xl, ..., xM), the second factor is evident from (758), the third one is due to the 4Mdimensional fourier transform,
and the last one comes out from (615) together with the inverse fourier transform. Hence, with the aid of (723) with 2 = 1, the contribution, TG(p), from G to the extended transition amplitude is given by TG(p) = i(2 r)4 (47L)_2P (f gb) .fG(P)
(760)
As remarked in Subsection 61, the S matrix is unitary for the physical values of momenta, that is,
SSt = StS = 1
(761)
On setting S = 1 + iR, (761) becomes
RRt = RtR = i(Rt  R)
(762)
or, for two (orthogonal) states Ia> and I#>, Y = i
(763)
V
One half of the righthand side of (763) is called the absorptive part of . If we consider the nonvanishing lowestorder term of «I R I x>
as in Theorem 78, it is given by only one Feynman integral. Then its absorptive part is proportional to Im fG(p) (s ). + 0). The lefthand side in the lowest order is a sum of products of simpler Feynman integrals. Each term of those products is expressed as follows. Let h and g  h be the sets
of the external lines of the particles present in the states Ia> and Ii9>, t If G is symmetric under some permutations of internal vertices, a factor 11r remains, where r denotes the number of the permutations leaving G invariant. 5
Nakanishi
66
GRAPH THEORY AND FEYNMAN INTEGRALS
respectively. For each S e S(h I g  h), we consider an integral F's(p), which is obtained from (623) with s  + 0 by replacing the propagator (615) by]'
2n 6(ki°) 6(mi  ki)
(764)
for all 1 e S. Then the unitarity relation (763) implies
y f s)(P)
2 Im J G(P) =
(765)
SES(blsh)
with FG )(P) = 64(Y Pb) (
21)KfGS)(P).
b
From (724) together with (A17), ImfG(p) (s + + 0) is formally expressed as follows [3, 14] :
ImfG(p) = at f dN1 a U2 a(N2p1) (V)
(766)
4
where for N 2 bev(G)g
/
WbCxbXC  2 1 Ybxb + > NL(xi  YL) bEV(G) g
(831)
LEIGI
where .
(832)
b = >2 [b:1]NLYL LEIGI
We can carry out the integrations over Xb (b e v(G)  g) in quite an analogous way to the calculation made in Subsection 71. Thus we obtain
PG(x, 0,2) = (70i)m [P(20  2m)/ fl I'(2L)]
X
f
dN1'3Q
J
11
fx,1 L
(833)
U2(17 + ie)102
A
where m = M  n and n stands for the number of external vertices. The topological formulas for U and V are found in the following way. We introduce a tadpole graph G', which is obtained from G by identifying all external vertices; G' has m + 1 vertices and N internal lines. We may denote the sole external vertex of G' by g without confusion. The quantities defined with respect to G' are denoted by affixing a prime such as T', U', etc. Hence
U' = E
>2 f1 NL 11 YL = TET"(g7 LET
TeT' LET
(834)
73
FEYNMANPARAMETRIC FORMULA
where T°[g] is the set of all ntrees in G such that any two vertices of g are not linked by the lines of Tin G. We may call U' an ntree product sunz ; it is a homogeneous polynomial of degree m. Since the matrix K' of G' can be constructed from #' by summing up all rows of g to make a single row and all columns of g to make a single column, we have K'g = (w,,C I b, c E v(G)  g)
(835)
Hence (831) implies
U = det /V'g = U'
(836)
with the aid of (329). Thus CJ is an ntree product sum.
[16] For arbitrary values of yl, V is expressed as follows:
THEOREM 81
(a) Cutset representation V
U'1
_ IEIGI Y I(xi  Yi) +
WS (
SEs'
lejGl
IS: 1]
PIYI) 2
(837)
(b) Circuit representation V
= IEIGI Y #IxI 
UC(E [C:1]Yl)2
U'1
CEC'
\ IEIGI
(838)
Proof From (831), f/ is defined by V= I
Y [/V'g]b.Cj3 j'c I (xi  yl) + U'1 b,cev(G) g
(839)
From Theorems 316 and 318, we have Ig]b.c
[
= g '(glb)
for b = c
=
for b
W,(glbc)
(840)
c
Hence, with the aid of (346), (839) is rewritten as V
= I NI(xi  Yi) + U'1 ucv(G')g L W'(glg*) I
beu
Yb)2
(841)
where u* = v(G')  u. It is straigthforward to show that (841) is equal to (837), by using (832), (210), and (344). To prove (838), we make use of Theorem 75. Since (726) is identically equal to (727), we have
 1Y2 U + I Uc(> [C: CEC
I
_
WS(E [S: l] c S
1]YI)2
I
1 y' )
(842)
GRAPH THEORY AND FEYNMAN INTEGRALS
74
by setting q1 = ytIai formally. Putting al = 1/f1 and considering G' instead of G, we obtain SES'
cI Uc
[C: 1] y,)2
(843)
This identity shows that (837) equals (838). q.e.d.
In terms of xb (b Eg), i has the following vertex
THEOREM 82 [17, 16]
representation :
Y
#1x1
UbC(xb  xc2
(8A4)
{b,c}cg
1EIGI
where Ubc is a sum of 11 Yt over (n 1)trees T in G such that b and c are IET
linked by the lines of T but any vertices of g  c are not linked by them.
Proof We now use (830); then
E [C: 1] Yi = I (E [a: 1] [C: 1]) xa aE9
LEIGI
(845)
lEIGI
Any circuit C in G' is either a circuit or a path between two external vertices in G. In the former case, the orthogonality relation (28) implies that (845) vanishes. In the latter case, if C E PG(bc) then we evidently have
(xb  x,,)
[C: 1] Y,
(846)
Thus (838) directly implies (844) by writing
Ub, ° q.e.d.
Y_
Uc
CEPc (bc)
(847)
§9 FEYNMANPARAMETRIC FUNCTIONS 91
Properties of U and V
In this subsection, we summarize some properties of U and V as functions of Feynman parameters x, for fixed values of momenta. From the definitions, U is a homogeneous polynomial of degree p and is linear in each x,, and V is a homogeneous rational function of first order. Furthermore, for all x, >_ 0, U > 0 and U vanishes if and only if x(C) = 0 for a circuit C. THEOREM 91 [18]
Let H be a connected subgraph of G and R = GII HI
Then
U = U$UR + U'
(91)
V= VR + V'
(92)
where UH (or UR) and VR denote the chordset product sum in H (or R) and the V function in R, respectively; U' is a sum of terms whose degree with
respect to a, (I e IHI) is at least u(H) + 1, and V' is of (at least) first order with respect to x, (I e IHI).
Proof An arbitrary term of U corresponds to a chord set T*. Since its complement T is a tree, T n IHI includes no circuit, whence Theorem 25 (d) and the maximality of a tree imply N(T n IHI) S M(H)  1 = N(H)  ,u(H)
(93)
Therefore
(94) N(T* n IHI) _> ,u(H) If the equality does not hold in (94), the term corresponding to T* belongs 
to U'. Hence we have only to consider the equality case. Then, since N(T n IHI) = M(H)  1 and T n IHI includes no circuit, T n IHI is a tree in H, that is, T* n IHI is a chord set in H. Since
N(T n I R I) = N(T)  N(T n I H I) = M  M(H) M(R)  1 (95) and since T n IRI is connected in R and passes through all vertices of R, T n IRI is a tree in R because of Theorem 26(c), that is, T* n IRI is a chord
set in R. Thus any term of U is contained in the righthand side of (91). Conversely, let T$ and TR be arbitrary trees in H and R, respectively. Then T = TR u TR is a tree in G because it cannot include a circuit and N(T)
=M(H)+M(R)2=M1.
Next, we prove (92) in the vertex representation (745) of V. [Actually, it holds for arbitrary values of q,, but the proof is more complicated.] We 75
76
GRAPH THEORY AND FEYNMAN INTEGRALS
have only to discuss the second term of (745). Since W"' is the U function in G"bc) = G(b`)'/lo (see the proof of Theorem 77), (91) implies W(blc) = UHWR(blc)
+ W'
(96)
where W( bIc) is the W(11c) in R and Wis of degree (at least),C(H) + 1 with respect to a1(l e JHI). On substituting (96) and (91) in (745), we easily obtain (92) by taking it into account that WR I c) = 0 and W, (alb) = W;Ic>
if b and c are not distinct in R. q.e.d. If JHI consists of only one line, (91) reduces to (741). If JHI equals a circuit C, (91) becomes
U = (y ai) Uc + U' lrc
(97)
If JHI = IGI  1 and if 1 is not a cutline in G, then (92) reduces to
V = almi + V'
(98)
REMARK 92 [18] If we define the value of V at a(I) = 0 for any I c IG1 by VG/I, then V is a continuous function in al > 0 for all 1 e IGI. Therefore V assumes the maximum and the minimum value in the compact region d. [As shown by concrete examples with ,u z 2, however, 8V/8a, is bounded
but indefinite at x(C) = 0 if 1 E CeC.] Furthermore, (98) implies that if at least one noncutline 1 such that m! > 0 exists then V assumes positive values in some subregion of d for any values of external momenta. THEOREM 93 [3]
0 < U < \N) (1/N)" ind
(99)
where (r) = N!/,u!(N ,u)!. Proof From the definition of U we have
0 r1. implies r1 > 2r1., and the symbol Sym means to take the arithmetic mean over all permutations of the relative magnitudes of r1 (1 E IGI).
The above prescription is not only equivalent to the conventional renormalization procedure apart from a finite renormalization in the renorm
alizable field theories but also applicable to the unrenormalizable field theories. Unfortunately, however, one cannot regard (1031) as a physically significant method of calculating all ultravioletdivergent quantities without cutoff.
Example For the N = 2 selfenergy graph G (the two internalline particles are spinless and have mass m), we have I
r(21 + 22  2) fG(P; 21, 22) = ra( I) r ( a2)
0CAl1 (1
da

)12I
[m2  all  OG) p2  ic] Z,+222
0
Because of the symmetry of the integrand, we may first carry out the contour integration over 22 which is equivalent to set 22 = 1. Then I
fG(P; 2, 1) = (2  1)
1
J da WI {[m2  x(l  a)p2  is]'+1 _ 0
+ (2 
1)IAI
1}
87
FEYNMANPARAMETRIC FORMULA
Since the first term is holomorphic at 2 = 1, we find
G(p) f da log [m  a(l  a) p2  ie]  1 0
Thus the result is dependent on the choice of mass unit.
The second method of obtaining the renormalized Feynma.nparametric integral is due to Lam [26] and Appelquist [27]. It is more closely along the
line of the conventional renormalization theory [20] and based on the formula for the remainder of a Taylor series:
a1
n
1
F(xI, ., Xk)  Y_ 
ae
=o j!
...,xx)
/I
g=o
I
f d (1 0
n
)a
In+1
F(exl, ..., xk)
(1032)
Let I be a nonempty union of circuits such that
d(I) = 41z(I) + v(I)  2N(I) z 0
(1033)
and H be a subFeynmangraph such that JHI = I. From the subFeynmanintegral corresponding to H we subtract its Taylor expansion to order d(I) around the point at which all external momenta of H are zero, and make use of (1032) (we denote the integration parameter by I). After doing this for all possible circuitunions I such that d(I) z 0 and H is nonseparable, we
convert the subtracted Feynman integral into the Feynmanparametric form. The resulting expression becomes particularly simple when all particles
are spinless (v(I) = 0). By putting i = CI, the renormalized Feynmanparametric integral explicitly reads [cf. (771)] I
m
f
f dCI  dI I
dai
0
eivE
(1034)
Ua
aCr
0
where dI = 21z(I)  N(I) z 0, and U and V are defined as follows:
fl (ai 1] U ° T*ET* lcT*
I)I [II
aims + U 1
17 LeIGl
ICI)
JV(blc)pbpc {b,c)cg
(1035)
(1036)
GRAPH THEORY AND FEYNMAN INTEGRALS
88
with yy(bl0 being the U function in G(bc). As seen from (91), V is a polynomial not only in oci but also in Cr; when Cr = 0 and , = 1 for all J + I, U reduces to UH UR , where I HI = I and R = G/I. In order to see the meaning of (1034), we consider the simplest case in which only one I is involved in (1034) and dr = 0. Then we have i app
d,,
{EI 0
dCr 0
lim
exp ( i V  s) U2
ac,
daI
exp ( i V  s) U2
exp [  i VH(0)  sHI U2
exp ( i VR  R)
x IEIRI
(1037)
U2R
with the aid of Theorem 91, where VH(0)
oclmi and sH + SR = s (ER
0, SR >_ 0). The second term of (1037) is the subtraction term, which is an
infinite constant times the Feynmanpararnetric integral corresponding to R.
References (1) R. P. Feynman, Phys. Rev. 76, 769 (1949).
(2) R. Chisholm, Proc. Camb. Phil. Soc. 48, 300 (1952); Erratum, 48, 518 (1952). (3) N. Nakanishi, Progr. Theoret. Phys. 17, 401 (1957). (4) Y. Nambu, Nuovo Cim. 6, 1064 (1957). (5) K. Symanzik, Progr. Theoret. Phys. 20, 690 (1958). (6) N. Nakanishi, Progr. Theoret. Phys. 26, 337 (1961.). (7) Y. Shimamoto, Nuovo Cim. 25, 1292 (1962). (8) T. Kinoshita, J. Math. Phys. 3, 650 (1962). (9) R. Jost, General Theory of Quantized Fields (American Math. Soc. Publications, 1963).
(10) H. Umezawa, Quantum Field Theory (North Holland, Amsterdam, 1956). (11) S. S. Schweber, An Introduction to Quantum Field Theory (Harper & Row, New York, 1961).
FEYNMANPARAMETRIC FORMULA
89
(12) J. D. Bjorken and S. D. Drell, Relativistic Quantum Fields (McGrawHill, New York, 1965).
(13) J. Kahane, J. Math. Phys. 9, 1732 (1968).
(14) N. Nakanashi, Progr. Theoret. Phys. Suppl. 18, 1 (1961); Errata, Progr. Theoret. Phys. 26, 806 (1961). (15) T. T. Wu, Phys. Rev. 123, 678 (1961). (16) N. Nakanashi, Progr. Theoret. Phys. 42, 966 (1969). (17) C. S. Lam and J. P. Lebrun, Nuovo Cim. 59A, 397 (1969). (18) N. Nakanishi, Progr. Theoret. Phys. 22, 128 (1959). (19) R. Karplus and N. M. Kroll, Phys. Rev. 77, 536 (1950). (20) F. J. Dyson, Phys. Rev. 75, 1736 (1949). (21) N. Nakanishi, J. Math. Phys. 4, 1385 (1963). (22) N. N. Bogoliubov and 0. S. Parasuik, Acta Math. 97, 227 (1957). (23) K. Hepp, Commun. Math. Phys. 2, 301 (1966). (24) E. R. Speer, J. Math. Phys. 9, 1404 (1968). (25) M. Riesz, Acta Math. 81, 1 (1949). (26) C. S. Lam, Nuovo Cim. 59A, 422 (1969). (27) T. Appelquist, Ann. of Phys. 54, 27 (1969).
CHAPTER 3
Singularities of the Feynman Integral The Feynman integral can be regarded as a boundary value of an analytic function of complex momenta, which is called a Feynman function. It is a very interesting and important problem to investigate its analytic properties, that is, to find the locations of its singularities and their natures. Necessary conditions for the existence of a singularity are expressed in terms of certain equations for momenta and Feynman parameters, which are usually called Landau equations. Some general properties of the singularity can be discussed by means of the Landau equations together with the Feynmanparametric formula. More detailed study of the singularity structures is, however, extremely difficult for the general Feynman graph; only some loworder or specialtype graphs can be analyzed in details. In 1951, Eden [1] first discussed the threshold behavior of a Feynman integral as an analytic function of an energy variable. In 1958, the existence of anomalous thresholds, which could not be expected from the unitarity of the Smatrix, was discovered by Karplus, Sommerfield, and Wichmann [2], by Nambu [3], and by Oehme [4], independently. In 1959, systematic investigations of the singularities of the general Feynman integral were made by Landau [5], by Nakanishi [6], and by Bjorken [7], independently. Since then, many authors have investigated detailed analytic properties for various Feynman graphs. Especially, people of the Cambridge school [8] extensively developed a general but heuristic theory on the analyticity of the Feynman function. Cutkosky [9] proposed an interesting rule for giving the discontinuity around a singularity of the Feynman function. In 1964, Fotiadi, Froissart, Lascoux, and Pham [10] introduced the homological method, which is based on rigorous mathematics and is powerful in investigating the Riemannsheet structure of the Feynman function. In this chapter, we develop the theory of the singularities of the general Feynman function on the basis of the Feynmanparametric formula. In Section 11, the Feynman function is defined, but we do not discuss the 90
SINGULARITIES OF THE FEYNMAN INTEGRAL
91
homological method in detail. In Section 12, we derive the Landau equations and discuss some general properties of the Landau singularities. Section 13
is devoted to the investigation of real singularities such as normal and anomalous thresholds. Physicalregion singularities and infrared divergence are discussed in Section 14. In Section 15, we investigate the region swept by
the singularities as internal masses are varied. Finally, the secondtype singularities are discussed in Section 16. We do not claim mathematical rigor throughout this chapter.
§11
FEYNMAN FUNCTIONS
It is the purpose of this chapter to investigate how the Feynman integral behaves when external .momenta Pb (b E g) vary. Throughout this chapter, n(> 2) stands for the number of the external vertices of a connected Feynman
graph G. The collection of n  1 (independent) external momenta is symbolically denoted by p. The renormalized Feynman integral fG(p) with a > +0 is a distribution over external momenta p. According to a wellknown result in the axiomatic
field theory [11], the extended transition amplitude is a certain boundary value of an analytic function of external momenta. Because of Theorem 78, therefore, the Feynman integral should also be a boundary value of an analytic function. This analytic function is called a Feynman function and denoted by fG(p). It is an analytic function of n  1 complex 4momenta.
Given a Feynman integral, it is a boundary value of an analytic function
of n  1 4momenta as mentioned above. One should note, however, that it may not necessarily be a boundary value of an analytic function if some constraints are imposed on external momenta, for example, if they are fixed on the mass shells.
Since fG(p) is Lorentzinvariant, it is sometimes more convenient to regard it as a function of invariant variables. According to a theorem due to Hall and Wightman [12], the Feynman function fG(p) is an analytic function of independent scalar products of external momenta. We regard it as a function of independent ones of invariant squares Sh = (by ph)2
(heY Ipb)2
(111)
for subsets h of g, because they are physically more convenient quantities than scalar products. When nonzerospin particles are taken into account, the Feynman function of course can be introduced in exactly the same way as above, but it is not Lorentzinvariant but covariant. It can be written as a linear combination of Lorentzinvariant functions, each of which is a function of invariant squares (111). The analytic properties of the Feynman function is determined to a large extent by the denominator function V of the Feyn
manparametric formula, and the modifications due to the presence of spins bring only minor changes. Hence, we are mainly concerned with the theory of spinless particles alone. 92
93
SINGULARITIES OF THE FEYNMAN INTEGRAL
Let a e g and
DV = {pb (beg  a) V(p, a) + 0 for all a E d }
(112)
where V(p, a) is given by (745) or (746) together with E Pb = 0 and a e d beg
means a1 z 0 for all I e IGI and Y. at = 1. Then the integral Reg fG(p, A) with
1= 1
`
fG(P, a') =
F(20  2ls)
11 r(2,)
f
d '10C
J
i
(113)
1
a)]'2F"
UZ(a) [V(P,
4
for p complex defines a holomorphic function of p in each connected part of D. When it is analytically continued, it defines the Feynman functionjG(p),
which is, in general, not singlevalued. The Riemann sheet on which the boundary value yields the Feynman integral is called the physical sheet. The Riemannsheet structure of the Feynman function can be investigated
by the homnological method [13, 14], which is an elegant application of modern mathematics. In the homological method, one usually starts from the momentumspace Feynman integral (626) without PE Pb), where the indices 2, (1 a IGI) are chosen to be positive integers. The Feynman function can be defined from it by dropping iel(1 a IGI) andregardingit as a multiple contour integral. This definition is, however, somewhat ambiguous because contours are not compact. Therefore, one has to compactify the momentum
integrations before applying the homological theory. The difficulty remarked in (2) of Subsection 62 is inherited by the ambiguity of the method
of compactification. The homological approach based on the Feynmanparametric integral [15] is free from this difficulty. In this chapter, we are mainly concerned with the locations of the singu
larities and the behaviors in their neighborhoods, which are independent of the problem of ultraviolet divergence. Hence, for simplicity of descriptions, we hereafter use the unrenormalized Feynman function fG(p) instead of the renormalized onejG(p), where f0(p) is formally defined to be fG(p, 1). More precisely, 1 fG (P) =
(N2,u+r 1)! A" r (r  1)!
J
0
f d"la
dao
J
4
a0)N2R1
X
(114)
U2[a0A + (1  a0) V]N2/s+r where A > 0 and r is a positive integer such that N  2,u + r  1 > 0. When rigor is required, we should of course go back to fG(p).
§ 12 LANDAU SINGULARITIES 121
Landau Equations
In this section, we investigate the locations of the singularities of the Feyn
man function fG(p). Since p is involved only in V, they are determined almost completely by V. The singularities of fG(p) purely caused by V are called Landau singularities. We confine ourselves to Landau singularities until Section 16, in which some exceptional singularities are discussed. Since the integrand of (114) is an analytic function of a, (l e IGI) except for 6 (1  a,1, we first carry out one of Feynmanparametric integrationsfi J
to eliminate it. Then the integrand is an analytic function of p and N  1 (complex) Feynman parameters together with ao, and the integration region is compact. As discussed in Appendix A2, the singularities of fG(p) can appear when some of Feynman parameters (including ao) lie at the end points and a pinch occurs for the remaining Feynman parameters. The endpoint condition implies that ao = Off and a, = 0 for all l E I where I c IG1. Then a pinch can occur if V at a(I) = 0 has a double zero under the restriction E 1. Because of Remark 92, V at a(I) = 0 equals VR, where 1EIGlI
R = G/I. By using Lagrange multipler2, we obtain VR = 0 and aVR/aa, = 2 for all I E SRI. Since VR is homogeneous, one finds
VR = Y a, aVR/aa, = 2
(121)
LEIRI
whence aVR/&x1 = 0 for all 1 E SRI. Thus:
121 [6] A necessary condition for a Landau singularity of .fG(P) A(P) is
a, = 0 for all l E I aVR/aa, = 0 for all 1 E IRI
(122)
where I is a proper subset of IGI and R = G(I.
We call the singularity for R = G (i.e., I = 0) the leading singularity or the principal singularity, and those for R G nonleading singularities. Since a nonleading singularity of G is the leading singularity of a Feynman t It is more elegant than doing this to work in a projective space [15]. ft A pinch with respect to ao cannot occur because the denominator function is linear in ao. Furthermore, it is nonvanishing for as = 1. 94
95
SINGULARITIES OF THE FEYNMAN INTEGRAL
graph R, we may discuss the leading singularity alone without loss of generality. Hence, until the end of Subsection 123, we confine ourselves to the leading singularity. Then (122) becomes
aV/aat = 0
for all 1 E IG1
(123)
In order to investigate (123), it is most convenient to make use of the circuit representation (726) of V. First, according to (619), internal momenta qI are related to p as
Z [b: 1]q1 + pb = 0 (b + a)
(124)
IEIG]
together with E pb = 0. As remarked in Subsection 62, we can arbitrarily b
choose qI for all 1 belonging to a chord set T* without contradicting (124). Hence we may require qI to satisfy E [CJ : 1] algl = 0
(j = 1,
..., Iu)
IEIGI
(125)
where circuits C1, ..., C,, belong to the fset of circuits corresponding to T* in the sense of Theorem 222, and the Feynman parameters aI should be understood to assume those specific values corresponding to the pinch. The
compatibility between (124) and (125) is assured by Theorem 311 as long as U + 0. Because of Theorem 219, we can extend (125) to
Y [C: 1] a,q, = 0 for all C e C
(126)
IEIGI
Then from (726), (123) reduces to
mi ql =0 for all lEIGI
(127)
We call (124), (125), and (127) altogether the Landau equations. We may take (747) and (126) instead of (124) and (125), respectively, in order to exhibit the duality character of the Landau equations. THEOREM 122 [5] The location of the (leading) singularity of ff(p) is determined by the Landau equations with E aI = 1. IEIGI
REMARK 123 [5]
The Landau equations can be derived also from
(74) together with (76) and (79). The double zero of 0 occurs when
ao/ak, = 0 for all I E T* a0laal = 0 for all"l E IGI
(128)
(129)
96
GRAPH THEORY AND FEYNMAN INTEGRALS
On account of (618) and (624), (128) and (129) coincide with (125) and (127), respectively, if we replace k, by q1. REMARK 124 [16]
If we regard G as an electrical network, al as an
Ohm resistance, and q, as an electric current [17], then (124) and (126) correspond to Kirchhoff's first law and second law, respectively. Furthermore,
Q
almi
 V
(1210)
IEIGI
corresponds to the power dissipated in G.
According to Theorem 311, if U + 0 then the simultaneous equations (124) and (125) can be solved uniquely with respect to q1. Indeed, the unique solution is given by ql = YI for all 1 e iGi
(1211)
because of (920) together with (126) [18]. From (925), therefore, we have
ql = U_1
(1212)
bega
We call any solution of the Landau equations such that U = 0 but fl a1+0, IEIGI
a peculiar solution. In order for a peculiar solution to exist, (1212) has to be an indefinite form 0/0. Hence if pb (b + a) are indepedent we have WI°lb) = 0 for all 1 e IGI and all beg  a
(1213)
In this case, ql can be linearly independent of external momenta, that is, the condition for the singularity generally admits extra freedoms for q1. The pinch in such a case becomes a nonsimple pinch in the momentumspace Feynman integral [19].
Example According to (927), Y,, ..., Y5 of the selfenergy graph shown in Fig. 121 are as follows: UY1 =
a5a2 + a2a3)P
UY2 =
a5a1 + ala4)P
UY3 =
a5aZ + ala4)p
UY4 =
a5a1 + a2a3)P
UYS = (alas  ala4)P
SINGULARITIES OF THE FEYNMAN INTEGRAL
97
Fig. 121 A Feynman graph yielding a peculiar solution to the Landau equations.
with U = (1X1 + a2) (a3 + aa) + (a3 + aa) a5 + 0'5(at + a2)
The condition U = 0 for f ai + 0 is compatible with the pinch condition if ` alai + a3a5 + a5a1 = a2aq + aga5 + a5a2 =  alaq =  a2a3
In what follows, we consider the ordinary case U + 0 only. Let d be the number of linearly independent external momenta Pb when they are fixed to general values, that is,
d = min (n  1, 4)
(1214)
because the momentum space is of four dimensions.
Since each internal momentum q1 is a linear combination of external momenta, the numbers of independent Landau equations (124), (125), and (127) are d(M  1), du, and N, respectively. Hence the total number
is (d + 1) N. On the other hand, the numbers of unknown quantities q, and a1 (with Y at = 1) are dN and N  1, respectively. Hence the total
number is (d + 1) N  1. The deficiency of one unknown quantity leads to one relation for external momenta. In the Landau equations, however, the Feynman parameters are involved only in (125). Accordingly, d(N  ,u) 7
Nakaniebi
98
GRAPH THEORY AND FEYNMAN INTEGRALS
+ N equations (124) and (127) involve dN unknown quantities q,. Therefore unless
(1215) N d, 1 there have to exist two or more relations for external momenta. As re
marked in Appendix A2, however, the singularity of an analytic function cannot be such a lowdimensional manifold; it should be a submanifold of a nonleading singularity. Thus we conclude that if (1215) is not satisfied, the leading singularity is absent. The same condition has to be satisfied in every subFeynmangraph of G, because every subsystem of the Landau equations also should not impose excessive relations on momenta.
THEOREM 125 A necessary condition for the presence of the leading singularity is that the inequality
N(H) < d(H),u(H) + 1
(1216)
holds for every subFeynmangraph H of G.
For example, no leading singularity exists for singleloop (i.e., ,u = 1) Feynman graphs with N > max (n, 5). It may be expected that if the condition of Theorem 125 is satisfied then the leading singularity will be present on some Riemann sheets except for some possible exceptional values of m, (1 E IGI). If N = du + 1 then the internal momenta q, are determined without solving (125), which determines a, ft JGD) alone. For example, this is the case for singleloop Feynman graphs with 2 < N = n < 5. 122
DualGraph Analysis
It is in general very difficult to solve the Landau equations analytically. It is therefore helpful to analyze them in a geometrical way.
As defined in Section 4, the Feynman graph G is called planar if G' is planar, where G' is obtained from G by regarding all the free ends of the external lines of G as a new common vertex "infinity". In this case, G has a dual graph G. By definition, G has ,u + n vertices.t We identify each line 1 of G with an "arrow" representing the related complex 4momentum (q, if 7 corresponds to an internal line 1 of G, or p, if it corresponds to an external line incident with b), that is, 7 is regarded as a complex ddimensional geometrical vector. Then the Landau equations have the following geometrical meanings. T Without loss of generality, we suppose that every external vertex is an end point of only one external line.
SINGULARITIES OF THE FEYNMAN INTEGRAL
99
The conservation law (124) expresses that the momenta incident with each vertex in G form a closed polygon in the complex ddimensional space embedding 6. Certain ,u equations of (126) imply that all vectors incident with each vertex a of G are linearly dependent if a corresponds to a circuit in G. Finally, (127) assigns a "length" m, in the Minkowskimetric sense to each vector q,. A more concrete analysis of using dual graphs is presented in Subsection 132.
As remarked in Section 4, dual graphs are not necessarily unique. In that case, according to Theorem 410 and Definition 48, a dual graph 6
Fig. 122 A dual graph to Fig. 121.
consists of two connected subgraphs I? and 11* such that there are only two vertices a and b' belonging to both H and H*, provided that G' is nonseparable. It is possible, therefore, to rotate either ft or fY* in the geometrical sense without violating the Landau equations. This freedom is closely related to the existence of a peculiar solution, because then internal momenta may not lie in the space spanned by external momenta. For example, a dual graph of Fig. 121 is Fig. 122; the central part can be rotated. 123
Behavior near the Singularity
In this subsection, we investigate the behavior of fG(p) in a neighborhood of a singularity. Let A V be a variation of V when p is varied by dp from the position of the singularity while the Feynman parameters are held fixed. Furthermore, we denote the value of at at the position of singularity by 7*
100
GRAPH THEORY AND FEYNMAN INTEGRALS
a, and the difference a,  a, by ,,; of course E ?i, = 0. When A V and q, 1EIGI
(1 E IGI) are infinitesimal, we can write
V= J V+ I E
(1217)
1.1'EIGI
with
a,,. ° a2V/aa,aa,.la=a
(1218)
in a neighborhood of the singularity, where a = a means a, = a, for all 1 e IGI. Since the Landau equations are invariant under the scale change of Feynman parameters, the second term of (1217) vanishes when ,j, = 26E, (2 + 0), whence (1219)
det (a,,. 11, I' c IGI) = 0
This variation, however, is not admissible because then E 77, = 2 + 0. If i
there is no special reason, therefore, the rank of the matrix (a,,) should be N  1, the dimension of the admissible variation. By diagonalizing the matrix (a,,.), the second term of (1217) can be written as N1
z
blCi , J=1
(H bf + 0)
(1220)
J
Then the integral (114) may be approximated by 1
N1
1
f daoao 1 JJ 0
f dCj I (d V + Z Y bbd j +
i=1 `1
aoll)N+zr&
(1221)
f
near the singularity (apart from a multiplicative constant). On carrying out the integrations successively by taking account of the contribution from a
neighborhood of S1 = ... = bNi = ao = 0 only, we finally find that fG(p) behaves like (d V)k log (d V)
for k = 0, 1, 2, ...
(d V)k
otherwise
(1222)
where
k(N+2,ur)+J(N 1)+r=1(4,uN 1)
(1223)
Since N 1,
that is, the strongest possible singularity is a simple pole, and multiple poles are absent except for some special values of internal masses. For k ? 1, the singularity is a branch point.
SINGULARITIES OF THE FEYNMAN INTEGRAL
101
The above discussion has been made for the leading singularity. If we extend it to the case of a nonleading singularity, some care has to be taken.
We have to take account of the N(I) integrations over at (1 e I) and the multiplicity 2p(I) of the zero of UZ [see (91)]. Since V varies with a, (1 E I) approximately linearly, (1223) should be replaced by
k = N + 21u + [N(R)  1] + N(I)  21C(I) =  [4,u(R)  N(R)  11
(1224)
For the case in which nonzerospin particles are present, (1224) presents a lower bound on k.
REAL SINGULARITIES
§ 13
131
General Consideration on Real Singularities
It is extremely difficult to discuss the detailed analytic properties of the general Feynman function. Since the quantity of physical interest is not the Feynman function but the Feynman integral, we investigate the singularities
of fo(p) with e  +0 in this section. The singularities of fG(p) are those of fG(p) which are characterized by the reality of V. More precisely, we consider the Landau singularities of fG(p) specified by Theorem 121 under additional conditions
a, > 0
for all 1 E SRI
(131)
Im s,, = 0 for all h c g
(132)
where s,, is defined by (111). If they he on the physical sheet, then they are
called real singularities. The singularities of fG(p) are nothing but real singularities of ff(p). Let DR be the region defined by (132), and Do = Dv n DR, where Dv is defined by (112). We may also characterize Do as the set of points such
that V > 0 for all a e A because V is continuous in d and V > 0 in some subregion of d (see Remark 92). Evidently, no real singularities exist in Do. We note that Do is not necessarily nonempty. Real singularities are much simpler than the general Landau singularities. For example, any peculiar solution to the Landau equations cannot correspond to a real singularity because (131) implies UR > 0. THEOREM 131 [20] If all of the conditions (122), (131), and (132) are satisfied, then a real singularity is always present (except for some possible exceptional values of internal masses).
Proof As in (1217) and (1220), we may write N(R)  1
VR =
2
y bJC2
A,1
J=1
(133)
near the point which satisfies (122). Since all Feynman parameters and invariant squares are real, zi VR, b j, and Cj(j = 1, ..., N(R)  1) are all real. Except for the case JJ bj = 0, as a > +0 the solutions to equation
j
VR  ie = 0 approach to the real axis of Cj from both sides. Thus a pinch 102
SINGULARITIES OF THE FEYNMAN INTEGRAL
103
occurs at every stage of integrations. Therefore a real singularity is present. In the exceptional case 11 b f = 0, the singularity will also be present except r
for particular values of internal masses because the singularity manifold is a closed set of points. q.e.d. Hereafter in this section, we regard fG(p) as a distribution of only one particular invariant square s, and we denote it by fe(s). Here we choose as s one of invariant squares such that aV/as < 0 in a neighborhood of the real
singularity under consideration. [This may be always possible as seen from the modified cutset representation (746) of V (see Section 17 for more detail).] If aV/as < 0 holds for all x Ed, then fG(s) can be regarded as a boundary value of an analytic function fG(s) from the upper half plane of s. This function is holomorphic in Im s + 0. Now, we consider a real singularity specified by (122), (131), and (132).
We denote the solution of (122) by s = s and a = a (i.e., x, = ai for all 1 e IRI). We write
VR(s, a) = A(a)s + B(a)
(134)
Then by assumption A(a) is negative in a neighborhood of a = a. Putting we have
C(a)  A(x)9 + B(a)
(135)
VR(S, LX) = A(a) (s  s) + C(a)
(136)
with ac/aaiI,,=a =
0
(1 E IRI)
(137)
The following theorem assures that VR has no maximal zero point, though it may have maximal nonzero points. THEOREM 132 [16]
As a function of a e d, V has no maximal zero
point.
Proof Let s = s and a, = at (1 E IGI) be the solution of (123), and define m = ai  al with Y qt = 0 as before. If a = a is a maximal point s
of V(s, a), then we always have
V(s, a + 7]) < 0 for 1771 °
(138)
127,1 infinitesimal. On substituting the Landau equations (126)
and (127) in the circuit representation (726) of V, we find V(s, a + z!) _ Ma)]1 Y_ Uc(a) ( Y_ [C: 1],7:4,)2 + 0(11713) CEC
IEjGj
(139)
104
GRAPH THEORY AND FEYNMAN INTEGRALS
where q = q is the solution of the Landau equations. Let 171 = no(alk  ai)
(1310)
(1 E IGI)
Then we obtain V(s, a + '1) =
2 8U U ax k a=a
qk + 0(I,II3)
(1311)
with the aid of (126) and (733). Since qk = Mk > 0 because of (127), (1311) is positive in contradiction with (138). q.e.d.
From this theorem, we see that a = a is a minimal or a saddle point of
VR(s, a). If x = a is a minimal point of VR(s, a), then we call s = s a threshold. If s = s is a threshold, we have
C(a) > 0 near a = a
(1312)
together with A(a) < 0. Therefore, in a neighborhood of s = s, we see from (136) that when s < s VR(s, a) >_ 0
for any anear a
(1313)
but when s > s VR (s, a) < 0 for some a near a = a
(1314)
Thus, according to (766), the contribution to Im fG(s) from a neighborhood
of a = a is zero for s < s but in general nonzero for s > s, that is, a new contribution to the imaginary part begins to occur at s = s. Physically, this fact can be interpreted as the appearance of a new competing process if we are in the physical region [see (765)]. As shown in Subsection 123, singularities of fG(p) are poles and branch points; it has no essential singularities. If we differentiate fG(p) with respect to a relevant variable s sufficiently many times, then it becomes infinite at the position of the singularity. Therefore we can characterize nonsingular
points by the property that all derivatives with respect to s are finite. In this way, we can define the singularities of Re fG(s) and Im fG(s).
THEOREM 133 [16] A necessary and sufficient condition for Re fG(s) being singular at s = s is that s = s is a threshold.
Proof If a = a is a saddle point of VR(s, a), then in (133) there exist j and j' such that bj > 0 and b f. < 0, provided that fl bi + 0. Then VR i
is linear with respect to new variables u = i + Cf, and v = i  ;,, and the Jacobian is a constant. According to (A16) of Appendix A, the integrand
105
SINGULARITIES OF THE FEYNMAN INTEGRAL
of Re fG(s) is expressed in terms of Hadamard's finite part. Since neither
u = 0 nor v = 0 is an end point of the integration region, therefore, the integrations over u and v are well defined. The same is true also for any finiteorder derivative of Re fG(s). Thus Re fG(S) is not singular at s = s.
For the case of a threshold, all of bi (j = 1, ..., N(R)  1) are positive, provided that fl bi + 0. Then with a new variable r = Y bJCj, the intet i gration range is r >_ 0, whence (d/ds)k Re fG(s) becomes divergent if k is sufficiently large, because we no longer have Hadamard's finite part. Thus Re fG(s) is singular at s = s. We may extend the above reasoning to the exceptional case fl bi = 0 ` by considering higher order terms. q.e.d. THEOREM 134 At s = s, Im fG(s) is always singular.
Proof From the discussion of the threshold behavior, it is clear that ImfG(s) is singular at s = s if it is a threshold. For the nonthreshold case, Theorems 131 and 133 imply that Im fG(s) has to be singular at s = s. q.e.d.
Example
Consider an integral 1
1
f(s) = f da1 f dal (x1a2  s  ie)2 (e > + 0) 1
1
It has a saddle point at s = 0. By direct calculation, we find
Ref(s) =
2
s
log
s+l =4(1 s1 ``
+++... SZ
s4
3
5
Imf(s) = 2'rs1[O(s  1)  0(s + 1)] Thus at s = 0 Re f(s) is nonsingular but Im f(s) is singular.
132 Normal and anomalous thresholds In this subsection, we investigate the properties of thresholds in more detail.
DEFINITION 135 A set of complex Minkowski vectors {p1, ..., pk) is called euclidean if
k
2
(Y_ cipi) > 1=1
for any real numbers c1, ..., ck.
0
(1315)
106
GRAPH THEORY AND FEYNMAN INTEGRALS
THEOREM 136 If Minkowski vectors pl, ..., pk are all real but nonzero and if they form a euclidean set, then they are timelike or lightlike (i.e., P12 z 0) and parallel to each other.
Proof The first statement directly follows from (1315) by setting ej = 6 j, .
We consider two arbitrary vectors, say p1 and p2. Since they are real and not spacelike, we have pi O) + 0 and p2) + 0. Hence we can find a linear combination c1p1 + c2p2(c1 + 0, c2 + 0) such that c2p2 = 0. Then (1315) and the reality require clp1 + cep2 = 0, that is, pl and p2 are parallel. q.e.d.
We denote the set of external vertices of the reduced graph R by gR, the external momenta outgoing from each external vertex b e gR by pb, and the number of linearly independent pb by d(R). From Theorem 136, we see that [PR I b e gR} can be euclidean in the physical region (i.e., Pb b real) only when d(R) = 1. THEOREM 137 [18] Let pb = pb (b a gR) be the solution of the Landau equations and s = s be the corresponding one. If {pb I b e g) is euclidean, s = s is a threshold.
Proof Since all internal momenta qi in R are linear combinations of pb, they form a euclidean set. Then (139) tells us that VR(s, a) > 0 in a neighborhood of a = a. q.e.d. The inverse proposition of Theorem 137 is true only for p = 1. Thresholds which do not correspond to euclidean sets can easily be found for u >_ 2. If R is a planar Feyriman graph and if {pb I b E gR} is euclidean, then it is very convenient to employ the dualgraph analysis mentioned in Subsection
122. We can now embed the dual graph R in a real d(R)dimensional euclidean space. We can give the Landau equations more precise geometrical
meanings: (124) means that the momenta incident with each vertex of R form a closed euclidean polygon (not planar in general); (127) implies that the euclidean length of the vector qi is equal to m, > 0; (126) leads to the property that each vertex a of R corresponding to a circuit in R lies in the subspace which is determined by its adjacent vertices, and moreover (131) means that a lies inside the polyhedron whose vertices are its adjacent onest. t For example, for N(R) = 3 the Feynman parameters can be regarded as the triangular coordinates in the subspace.
SINGULARITIES OF THE FEYNMAN INTEGRAL
107
Example [21] The dual graph of the "triangle graph" shown in Fig. 131 is Fig. 132. A dual graph of the "square graph" (u = 1, N = n = 4) is a pentahedron.
Fig. 131
The triangle graph.
Pi
Fig. 132 A dual graph to Fig. 131.
The simplest case is d(R) = 1. In this case, from i7 = m1 > 0 and qt being parallel to p', we have (pR)2 > 0, whence s = s is a threshold because
of Theorem 137. We call a threshold of d(R) = 1 a normal threshold, and that of d(R) ? 2 an anomalous threshold. For a normal threshold, we know N(R) 0 for all l E JRI, the existence of loop lines is excluded by (127). Hence the only possible graph R consists of N(R) parallel lines alone which link the two external vertices of R, that is, JRI is a cutset Sin G. An example of R is shown in Fig. 133. Let as and bs be the two vertices of R and define [S: Z] by [as : 1] of R (see
108
GRAPH THEORY AND FEYNMAN INTEGRALS
Definition 212). We denote the external momentum outgoing from bs in R by pR. Then the Landau equations read pR
Y [S:1]gl IES
[S: 1] aigt = [S: 1'] aI gI'
(1, 1' E S)
gi=mi (iES)
OCRI
(1316)
Fig. 133 A reduced graph corresponding to a normal threshold.
It is straightforward to solve (1316) tinder the condition (131). We find
=m1/ y inkI (IES) kES
gI = IS: 1] (mI/ Y_ ink) pR
(1 E S)
k6S
(pR)2 = (E inI)2
(1317)
ICS
together with
A(a) = _
nzI
(
IES
1
1
Y nzk
i such that V(i', a) > 0 for all a c d in contradiction with (1326). Therefore s = i presents a minimal zero point of V, that is, it is a threshold. Furthermore, we observe that i z so on account of (1324), that i < s because 8V/8s < 0 at the value of a giving the normal threshold (see (1318)), and that i + s because of Theorem 139 and (1325). Thus there is a threshold s = s such that so < i < s, and it cannot be a normal threshold. q.e.d.
Let G be the triangle graph shown in Fig. 134. It is straightforward to show that V(s, a) >_ 0 for s  (p. + Pb)' < (ml  m2)2 Example
Fig. 134 The triangle graph.
SINGULARITIES OF THE FEYNMAN INTEGRAL
111
if pQ < (m2 + m3)2 and pb 0
(1327)
A similar consideration also applies to the square graph ifs is a square of a sum of two external momenta associated with adjacent vertices. As described in Appendix B, (1327) is indeed a necessary and sufficient condition for the existence of the anomalous threshold in s in the triangle and square graphs [2, 21]. Furthermore, we note that (1327) is equivalent to the condition under which the dispersion relation for forward scattering (see next chapter) cannot be proven in the axiomatic field theory, but we omit the detail [16].
§ 14
PHYSICALREGION SINGULARITIES
This section is only of physical interest. We use some physical terms without explaining them explicitly. 141
Multiple Scattering
Physicalregion singularities are real singularities for which all external momenta are real [of course this condition is stronger than (132)]. Normal thresholds are physicalregion singularities, but anomalous thresholds are not in general. As seen later, if all external momenta satisfy the "stability condition" (i.e., pQ is less than the lowest normal threshold MQ in the channel
(a I g  a)), then there exist no physicalregion singularities other than normal thresholds. Hence we should have po > Mo for at least one vertex a. If no unstable particles are involved, that is possible on the mass shell only when two or more external lines are incident with a. Thus, in this section, it is convenient to distinguish various external lines incident with the same vertex from each other. As in Section 12, we consider the leading singularity alone. The leading physicalregion singularity appears if and only if the Feynman graph G is consistently interpreted as a real multiple scattering. The simplest nontrivial example is shown in Fig, 141.
Fig. 141 A Feynman graph corresponding to a physicalregion singularity. The 0th component of the momentum associated with each line is positive along its orientation.
THEOREM 141 [20]
The leading physicalregion singularity exists if and
only if G can be consistently interpreted as the trajectories of classical, 112
SINGULARITIES OF THE FEYNMAN INTEGRAL
113
relativistic particles in the position space when we identify the Feynman parameters a, with r,/m, > 0 (1 e IGI), where r, stands for the proper time interval of the particle corresponding to 1.
Proof Given the Landau equations, we can choose the orientation of each line 1 in such a way that q,(°) z 0. For each vertex b, we associate a 4vector xb with it. We can put a1g1 = 
[b: 1] xb
(141)
beIGI
because then the circuit equations (125) are identically satisfied on account of (28). On substituting z,/1n1 for a, in (141), we find that xb can be inter
preted as the spacetime position of the "event" b. The other Landau equations (124) and (127) are the 4momentum conservation and the massshell conditions, respectively. Thus the Feynman graph is regarded as trajectories of classical, relativistic particles. Furthermore, in this case, the singularity is always present because of Theorem 131. q.e.d.
It is important to note that the time ordering is introduced between any two adjacent vertices because r, > 0. Thus the set (xb I b e v(G)} has a semiordering. Hence there are at least one first vertex and at least one last one. For such a vertex a, we have to havepa > MQ to satisfy the 4momentum conservation law, as mentioned at the beginning of this subsection. The above theorem can be made more precise if we consider the singularity
directly in the position space [22]. First, we rewrite the Landau equations in terms of the relative positions xb  xb  xa (b e v(G)  a) and the inverse Feynman parameters fl, (1 e IGI). According to (141), (124) becomes c#a
W1,,X  Pb = 0
(b rr a)
(142)
where ivbc
[b : 1] [c : 1] f3,
(143)
The massshell equations (127) are rewritten as
,
(Y [b : 1] xb)2 = m, b*a
(1 e IGI)
(144)
Since (125) becomes trivial under (141), (142) and (144) are equivalent to
the Landau equations. If we solve (142) and substitute xb in (144), then according to Theorem 95 the lefthand side of (144) becomes YY in the vertex representation. 8
Nakanishi
114
GRAPH THEORY AND FEYNMAN INTEGRALS
After the above preliminaries, we consider the Feynman integral in the f,/2 in (86) and position space. By changing the parameters f, into omitting the primes, (87) yields 00
ff(P) = const bfl (f d4xb)
(f d8,) exp [iO(P, x, fl)  e]
(145)
0
where s 
e1,RR9
1 and
O(P, X, N) _ E Pbxb  2 b4.a
iNl(: [b : l] ((/ i
1111
bTa
Q
ml /F'l
(146)
Since the differentiations with respect to Pb simply yield a monomial of xb in the integrand of (145), any derivative of fG(p) will not in general diverge because of the rapid oscillation of the integrand. Hence a real singularity for p real can occur only when the phase factor is stationary, that is, aO/axb = 0 (b $ a)
(147)
aO/afl, = 0
(148)
(1 e I GI)
apart from the cases in which some of j, become infinite. It is straightforward to see that (147) and (148) are equivalent to (142) and (144), respectively. Since 0 has the dimension of action, it should be replaced by 0/Il if we do
not employ natural units. If we take the classical limit h * 0, the only surviving contribution comes out from the stationary point, which satisfies (147) and (148). Thus the classical limit of the Feynman integral precisely yields the condition for the (leading) physicalregion singularity, and xb is the spacetime position relative to xa by definition. Therefore, Theorem 41 holds without making the ad hoc assumption a, = T,/m,. 142
Discontinuity Formula
Since most of the singularities of fG(p) are branch points, it is important to
know the discontinuities around them. The best way of obtaining the discontinuity formula of fG(p) is the homological method [13, 14], but we do not describe it here. The discontinuity around a real singularity is purely imaginary; 2iImf0(s)
is a superposition of discontinuities. In order to extract a discontinuity around a single real singularity, we have to compute the nonanalytic change of Im f0(s) at that point s = s, which is a difference between Im fG(s) for
SINGULARITIES OF THE FEYNMAN INTEGRAL
115
s > s and the analytic continuation to s > s of Im fG(s) for s < S. The nonanalytic change is easy to be found only when Im fG(s) = 0 for s < s. As remarked in Subsection 73, the Feynman integral satisfies the unitarity condition by itself. The discontinuity around a normal threshold corresponding to an intermediate state S can be seen to be given by fesl(p) in the physical region [see (765)]. Therefore the expression for ffs)(p) can be regarded as a discontinuity formula. The Cutkosky rule [9] is a formal generalization of it to other singularities. The original prescription was as follows: The discontinuity around a Landau singularity corresponding to a reduced graph R should be given by an integral which is obtained from the Feynman integral by replacing the propagator (615) by 27rSp(ml  k2) for every 1 E IRI, where Sn is called a "proper" afunction. When external momenta p are complex, the meaning of this rule is ambiguous because the contours of k1 are not specified. For physicalregion singularities, however, the above rule is well defined if we set ks) = 6(k,'0)) 6(mi  ki)
6 (mi 
(149)
The validity of the Cutkosky rule for the physicalregion singularity can be verified by the homological method [23, 24].
According to Pham [23], the Cutkosky rule for the physicalregion singularity can be regarded as a more fundamental rule than the unitarity of the Smatrix. The Smatrix satisfies a kind of macrocausality condition, that is, if several elementary processes occur in very large spacelike separations, the Smatrix for the total process is a product of the Smatrices for those subprocesses. This property is called the cluster property. Since the physicalregion singularity can be understood in terms of multiple scattering, it may be natural to call the Cutkosky rule a "cluster property for very large timelike separations". 143
Infrared Divergence
The infrared divergence, which is well known in quantum electrodynamics, can be regarded as a pathological case of the physicalregion singularity [25]. Because of the 4momentum conservation (124) and the massshell equations (127), the reduced graph R corresponding to a physicalregion singularity cannot involve any vertex of Fdegree 3 if no unstable particles are involved and if no massless particles are considered. Only in this subsection, however, we admit the presence of massless particles. Then vertices of Fdegree 3 are 8*
116
GRAPH THEORY AND FEYNMAN INTEGRALS
of only one type, that is, R can involve a vertex b if two lines 1 and 1' incident with b has the same mass m and the remaining one line k has zero mass. The Landau equations can be satisfied only if [b: 1] qi + [b: 1'] q1, = 0, qk = 0, and q, = m2. We call such a vertex an infravertex. It is convenient to use the terminology of infravertex also for any vertex of Fdegree d + 3 if two lines incident with it have the same mass in (m may equal zero) and if all the other lines incident with it have zero mass. The infrared singularity is defined to be the physicalregion singularity for. a reduced graph in which no pair of noninfravertices is linked by nonzeromass line alone. For practical applications, we always consider infrared singularities for reduced graphs, but for simplicity we here consider only the leading infrared singularity as in Section 12. As seen from the definition of the infrared singularity, the set of all lines
which have nonzero mass is a disjoint union of paths and circuits which pass through no common vertices except for noninfravertices. Those paths and circuits are called infrapaths and infracircuits, respectively. More precisely, an infrapath is a path consisting of nonzeromass lines either between two external vertices or between one external vertex and one noninfravertex. By definition, the masses on any infrapath or infracircuit are equal to each other.
I
Fig. 142 A Feynman graph corresponding to infrared singularity. Dotted lines indicate masslessparticle lines, and solid ones are of nonzeromass particles. The vertex a is a noninfravertex.
Now, we consider the Landau equations. As noted above, so as to satisfy (124) and (127), the momenta qi are equal to each other for all I belonging to any infrapath P or infracircuit C if the orientation of 1 is chosen so as to coincide with that of P or C, and qk = 0 for all lines k of zero mass
SINGULARITIES OF THE FEYNMAN INTEGRAL
117
(ink = 0). Next we consider the circuit equations (126). If Cis an infracircuit,
(126) implies Y_ a, = 0 and therefore a, = 0 for all 1 e C under the con1CC
dition a, > 0. Likewise, if C is an arbitrary circuit, which intersects m infrapaths P1i ..., P,,,, then (126) implies m
i ( E a,) (±qPj) = 0
(1410)
j=1 l 1ECnPj
where qPj is the momentum common to all lines of Pj, and the double sign depends only on j. If we assume that qp, , ..., qPm are linearly independent because they are equal to momenta of distinct external particles,] then (1410) implies Y_ a, = 0, that is, a, = 0 for all 1 e P n C, where P is any infraZEPnC
path. Thus we have a, = 0 for all (noncut) lines 1 of nonzero mass. The value
a, = 0 should not be identified with the endpoint value, but it should be regarded as a limiting case of a, > 0. This feature is a pathological aspect of the infrared singularity. It is important to note that (126) cannot determine ak if k is a line of zero mass because qk = 0. If G involves.v lines of zero mass, v  1 Feyn man
parameters are undetermined (the minus one is due to X a, = 1). ThereleIGI
fore, when we calculate the behavior near the singularity, v  1 coefficients
b j in (1220) should vanish. Hence, the contribution from the leading infrared singularity$ behaves like (1222) with
k=J(4ju N1)z(v1) =J[4,u(Nv) 2v]
(1411)
where .J V should be proportional to a fictitious mass squared of the massless
particles. If k < 0 then the Feynman integral is divergent at the infrared singularity. This divergence is known as infrared divergence..§ As seen from
(1411), the occurrence of infrared divergence can easily be judged by counting the order of zero in the denominators of (623) [26]. t Here we need not consider C which intersects more than two infrapaths incident with a common noninfravertex. $ Here we neglect the contributions from nonleading infrared singularities. § The infrared divergence is usually the divergence in the case in which all infrapaths are incident with one noninfravertex, because only in this case the singularity becomes independent of all invariant squares of momenta other than external masses squared. The divergence in the case in which we have no noninfravertex and two infrapaths is known as the Coulombforwardscattering infinity.
EXISTENCE REGION OF SINGULARITIES
§ 15
Landau singularities of a Feynman function fG(p) can be present on the surfaces defined by the Landau equations. When internal masses m, (1 e IGI ) vary in a certain region, say, in mt >_ 0
(151)
(1 E IGI)
those surfaces sweep some region in the complex p space. A Feynman function with arbitrary internal masses satisfying (151) is holomorphic outside that region (apart from the possible secondtype singularities described in the next section). To find the existence region of singularities is closely related to the problem of discussing the analyticity in the axiomatic field theory.
For the moment, we regard the Feynman function as a function of only one invariant s by holding all others fixed. THEOREM 151
Let s = s and a = a be a (complex) solution of the
Landau equations for the leading Landau singularity. When mt varies, the variation of s is given by (i,
as
(152)
amt
av/asla=a
where V is considered as a function of s, a, (1 e IGI), and mt (1 E IGI).
Proof By differentiating V(s, a, m2) by mt , we have
av amt
a,
av as + as amt +
aV &k kEI I aak
(153)
amt
Since the variations of s and a, (1 e IGI) occur in such a way that the Landau equations are still satisfied, we have aV
amt =
0,
aV = 0
aa,
(1 E IGI)
(154)
On substituting (154) in (153), we find as
amt
(155) (0/as) V(s, a., m2)
a result which is identical with (152) because V(s, a, m2) is linear with respect to s. q.e.d. 118
SINGULARITIES OF THE FEYNMAN INTEGRAL
119
An important consequence of Theorem 151 is that al1 8s/8m2 is independent of l [27].t THEOREM 152 When internal masses vary in (151), the boundary of the singularity region is part of the union of curves r, over all I c JGJ. Here I'I is defined by s = s and the equations
mi = 0 8s/8m2I
=
r,C
for all l c 1
(156)
for all 1 e 1*
(157)
where I* _ JGJ  I, r, is real, and C is complex but independent of 1. We omit the proof of the above theorem; here we merely note that (157) can be obtained by rewriting the wellknown condition for the envelope of a family of curves into that in the complex plane [28]. For 1 e I, the quantity r, _ C18s/0mi is of course complex, but the signs of lm r, are common for all 1 c L This is because the change of s for arbitrary variations of internal masses is written as
ds = > (8s/8m2) dmi IEJGJ
r, dmi + Y_ r, dmi) 1EI*
(158)
1E1
with Im r, = 0 for all 1 c 1* and dmI > 0 for all 1 c 1, and Im (ds/C) should have a definite sign since the singularity region lies in one side of its boundary. From Theorems 151 and 152, we obtain the following theorem. THEOREM 153 [27] On the boundary of the singularity region, the Feynman parameters a, are proportional to rt. Therefore, Im at = 0 for
I e I* and Im a, > 0 for l e I if we neglect a common factor. This theorem simplifies the analysis of the singularity region considerably.
It is still prohibitively difficult, however, to analyze the general Feynman graph in detail. Fortunately, the singularity region seems not to vary with Feynman graphs very much. It seems natural to conjecture [29]$ that all
Landau singularities of the general Feynman graph having n external t The original proof of this fact was more complicated than the one presented here.
t There is also a conjecture [30] that the npoint Wightman function, which is a positionspace analytic function, has a singularity region whose boundary would be characterized by that of the complete Feynman graph having n external vertices. It seems, however, that this conjecture is groundless for n z 4 [31, 32].
120
GRAPH THEORY AND FEYNMAN INTEGRALS
vertices would lie in the singularity region defined by the internalmass superposition of the Landau singularities of the complete Feynman graph having n external vertices (v(G) = g). Hence it is important to analyze the complete Feynman graph. For this purpose, it is convenient to employ the modified form, (142) and (144), of the Landau equations. In the matrix notation, (142) is rewritten as (see Subsection 81) y1,°xa  p° = 0 (159) Hence Pa(Pa)t = `/Taxa(xa)t .//fa
(1510)
In a complete graph G, since every pair of distinct vertices uniquely correspond to a line, it is convenient to denote every line 1 by the pair of its end vertices b and c, and correspondingly we rewrite Nt and mi as F'bc(= F'cb) and mb,(= mCb), respectively. Then (144) becomes
(xb 
(b, c e g  a)
xc)2 = jnbclF'bc
(b c g  a)
xb = mab(fab
(1511)
with M2bb = 0. Hence 2
xxc b b
2 m2c
2
F'ab
flat
M
2
r6c
)
(b, c e g  a)
(1512)
Let _#a be the square matrix of order n  1 whose elements are defined by the righthand side of (1512); then x°(xa)t = _#a
(1513)
On substituting (1513) in (1510), we obtain (1514) Pa(pa)l =1Y°.fa a In (1514), n(n  1)/2 scalar product of external momenta are expressed in terms of n(n  1)/2  1 auxiliary variables Pbc/ E together with b'$cl n(n  1)/2 mass parameters.
Now, we apply the theorems given above to our complete Feynman graph. There are several nontrivial choices of I, but the most interesting case is the case in which I is a star. Hence we put I = S[a]. Then Theorems 152 and 153 (remember that fl, is inversely proportional to ai) lead us to mab = 0, Im flbc = 0
Im flab < 0
for b e g  a
for b, c e g a
(1515)
SINGULARITIES OF THE FEYNMAN INTEGRAL
121
where we have redefined Nbc by dividing it by a single complex factor. On account of (1515), we have y r a = ( > [b : 1] [c : 1] z I b, c e g  a IeIGI
= (E f3dbbbc  flbc I b, c e g  a
(1516)
dEA
jfa
 J(mbc/F'bc I b, c e g  a)
(1517)
From (1516) and (1515)
IM f4'° = (Im Pal  bbc I b, c c g  a)
(1518)
Thus the matrix W11° is symmetric, and its imaginary part is diagonal and negative definite; _#a is real and symmetric and its diagonal elements are zero. Therefore, p°(pa), is expressed in the Jost canonical form, usually called
DANAD, except for the overall sign [30]. In the axiomatic field theory, DANAD appears as the boundary of the singularity region of a positionspace analytic function obtained without using the microcausality condition (612). Apart from the trivial case n = 2, the implication of (612) was fully utilized only when n = 3 [33]. The main part of the boundary surface found in this case exactly coincides with
the boundary of the singularity region for the n = 3 complete Feynman graph obtained above.
§ 16 SECONDTYPE SINGULARITIES 161
Pure SecondType Singularities
The integrand of a Feynmanparametric integral has two denominators U
and V. If a e A, U can vanish only at the end points. If one considers complex singularities of a Feynman function on unphysical sheets, however,
U can vanish for complex nonzero values of Feynman parameters. The singularity caused by a complex zero point of U is called a secondtype singularity.t A characteristic feature of a secondtype singularity is that its location is (at least partially) independent of internal masses. If it is completely independent of them, the singularity is called a pure secondtype singularity; otherwise it is called a mixed secondtype singularity. The former is discussed in this subsection. It is known that the existence of secondtype singularities depends on the dimensionv of the momentum space. By extending (114) to the vdimensional case (see (772)), we have 1
r
IG(P) = const
Ur12 [and o
J
_J
0
ar1
f dN1 a
dao d
1
a0)N_}vA1
+ (11  a V N
rµ+r
(161)
where N  Ivy + r > 0 and r > 0. It is convenient to transform the integration variables by (1  ao) ai = oci (1 e JGI). After dropping the primes of the new variables, we obtain ar1
r
fc(P) = const
dNa
(162)
1vu+r
4,
where dJ
dNa
OJ
d 0 If
dai I S (1  ao OJ
J
ai)
(163)
i
and
V'  a0A + V (A > 0)
(164)
The use of the parameter ao is suitable for the investigation of the secondtype singularity. t The existence of the secondtype singularities was found by Cutkosky [9]. Their systematic investigation was made first by Fairlie, Landshoff, Nuttall, and Polkinghorne [34]. 122
123
SINGULARITIES OF THE FEYNMAN INTEGRAL
Since r can be arbitrarily large, we may choose it in such a way that
mNv(p+1)+r
(165)
is a nonnegative integer. Then (162) is rewritten as
fc(p) = const
ar1 U"`
dN a
°
(UV')"
J Al
"
(166)
Since the numerator of its integrand is a polynomial, we have only to investigate the denominator. Hence, as in Section 12, the leading singularity is given by
(a/aa°) UV' = 0
(a/aa) UV' = 0 Since a U/aa° = 0 and 8 V'/0a° = A
(1 e JGI)
(167) (168)
0, (167) reduces to
U=0
(169)
Thus it characterizes the secondtype singularity. On the contrary, if a° = 0 instead of (167), (168) reduces to the equations which determine a Landau singularity. In order to solve (168) with (169) for a° + 0, it is convenient to employ (722), that is, UV = UL + (qT*)` (1610) qT* where 11, l' e T*)
°u T =
(1611)
and L is defined by (710). By definition, °11T.°ICT;
°11T; °ILT. = U 
(1612)
where d' denotes the unit matrix of order p. Suppose that the internal momenta q1 (l E JGJ) are expressed as linear combinations of external momenta Pb (b + a). Then we can write qT.
(1613)
where _V is a certain u x (n  1) matrix depending on Feynman parameters. If there exists a lightlike vector which is orthogonal to Pb for all b c g, we denote it by e; if not so, let = 0. Then we make an ansatz qT* = °I1T.q + ee
(1614)
124
GRAPH THEORY AND FEYNMAN INTEGRALS
where q is a column vector consisting of u momenta orthogonal to q, and e is a column vector consisting of u arbitrary numbers. Since a/JT is singular in the present case, we can find a column vector r consisting of ,u numbers
such that dhT,r = 0. [If we assume rank (Q4T,) = It  1, r is unique apart from a multiplicative factor.] Then q exists only if rtgT = erte
(1615)
From (1613) and (1615), we see that a certain linear combination of external momenta is equal to a lightlike vector which is orthogonal to them. By multiplying this equation bypb (b e g  a) and eliminating the coefficients of the scalar products pp,, we have the Gramdeterminant equation
det [p°(p°)t] = 0
(1616)
which is the characteristic equation of the pure secondtype singularity [34]. Now, we substitute (1614) in (1610) and differentiate it with respect to al. By using (1612) and (169), we have
UV = L
a
au
adi
+ gto&T
act,
am,
am,
T,q
(1617)
a
(1618)
Since (1612) implies GICT
am,
am,
am,
(1617) is rewritten as a
as
UV
au
as
(L + gt,&Tq)
(1619)
Thus (167) and (168) reduce to (169), (1614), and
a0A + L + gtO&Tq = 0
(1620)
The last equation (1620) determines a,, that is, it yields no restriction on a, and q, (1 e IGI). Thus we need not solve q from (1614). Since the above method is based on the ansatz (1614), it is instructive to analyze (168) more directly. By means of the modified cutset representation (746), we have a
am,
UV =au  E akmk  Y am,
(hlgh)
k
am,
beh
Pb) 2
(1621)
with the aid of (169). On setting
c,  a0A + I akmk k
(1622)
SINGULARITIES OF THE FEYNMAN INTEGRAL
125
we find that (168) becomes 8W(hlgh)
(b>J Pb) 2 = S as
8ai
(h1Y h)
(l E IGI)
(1623)
This equation shows the reason why the pure secondtype singularity is independent of internal masses. We consider a singleloop graph having no internal vertex (i.e., ,u = 1 and N = n). We number the internal lines cyclically. Then n
U = Y_ aJ
(1624)
J=1
W(hl9h) = Ws = aJak for S = {j, k} e S(h I g  h)
On writing
Y_ Pb
(1625)
= PJk = Pkj, (169) and (1623) become
beh n
Y, a; = 0
(1626)
J=1
aJP.2ik = E
II
J*k
1
1
...
1
2
0 2
1
P21
1 2
Pie
Pin
0
Pen
Pn2
0
Pn,
1
(1627)
and al , ..., an , we find
respectively. By eliminating 10
(k = 1, ... , n)
2
=0
(1628)
a result which is equivalent to (1616) because of a wellknown identity between determinants, called the Caylay transformation. When (1628) is satisfied, the ratio of ao , al , ..., a is uniquely determined. Hence the behavior of ff(p) near the singularity is given by (1222) with
k = (Iv + m) + m + IN=  1(v  N)
(1629)
according to (166). It is known, however, that the secondtype singularity is present only if v > n(= N) in the singleloop graph [34]. Hence the behavior near the singularity is given by
(d
V)(vn)I2
for v > n
(1630)
In the general graph, the pure secondtype singularity can be present only
for n < v + 1. This is because if n > v + 1 then the external momenta p
126
GRAPH THEORY AND FEYNMAN INTEGRALS
are always linearly dependent so that they identically satisfy (1616), but it is impossible that fG(p) is singular for all values of p.
We note that the behavior near a secondtype singularity becomes stronger if G involves particles with nonzero spins, because Y, and X,,. contain U in their denominators. Thus the secondtype singularity depends not only on the dimension of the space but also on the spins of particles. Example 1 The N = 2 selfenergy graph (,u = 1, n = 2). Let s = p2; (1626) and (1627) with (1622) lead us to
a2=al, ao = Alal (m2  ml) The case nil = m2 is exceptional because then ao = 0 (a Landau singularity coincides with the secondtype singularity); the behavior near the singularity
is always likes * independently of v. For ml + m2, (1630) tells us that fG(p) behaves like s(°2)12. Indeed, Im fG(P) = const S(v2)12 ([s
 (ml  m2)2] [s  (ml + m2 )2])(V3)/2
x 6(s  (ml + m2)2)
whence ff(p) is singular at s = 0 on an unphysical sheet, provided that
vz3.
Example 2 v = 4, u = 1, n = N. The only secondtype singularities are those of the above selfenergy graph and of the triangle graph (Fig. 131) because of the condition N < 4. In the latter, fG(p) behaves like [9] [P1
2  (PjP2)2]
Example 3 u = 2, N = 3, n = 2 (Fig. 133). Since (1623) is satisfied by = 0, the ratio al : a2 : a3 is not uniquely determined as in the case of the infrared singularity. Correspondingly, N in (1629) should be replaced
s=
by N  1 = 2. Thus fG(p) behaves like (4
V)(12)'2.
Indeed, for v = 3 we
have
Im f0(p) = const sl(sl  ml  m2  m3) 6 (s  (ml + m2 + m3)2) 162
Mixed SecondType Singularities
The mixed secondtype singularities appearwhen,u > 2. The simplest example
of them is provided by the socalled icecreamcone graph (u = 2, N = 4, n = 3) [34, 35].
SINGULARITIES OF THE FEYNMAN INTEGRAL
127
The ansatz (1614) in Subsection 161 has been made under the implicit assumption that all Feynman parameters are of the same order of magnitude.
If some of them tend to zero as U + 0, then we can find another type of solutions to (168), which are nothing but mixed secondtype singularities. Let H be a subgraph such that JHJ is a union of independent circuits ,u(H) > 0). Suppose that (1631) S ° Y_ Ia,I IcIHI
is infinitesimal and that U  &(17)}1 so that U/&(H) > 0 as 6 + 0. Then, since U, +a 0 as 6 + 0, where R = G/I HI , we find UH ,., &M(H)+1
(1632)
because of (91), that is, UH/S(H) + 0 as 6 > 0. Now, we consider (168). By using the vertex representation (745) of V, (168) is written as 8U Um, 8m, +  +(b.c)=g F a
aw(blc) amt
pbpc = 0 (1 a IGI)
(1633)
where E is defined by (1622). For 1 e JHI, the leading order (. &Z(H)1) becomes independent of 1. Indeed, we obtain URA +
(b,c)cg
WR,clpbpc
=0
(1634)
with the aid of (91) and (96). Therefore, in order to obtain N(H)  1 independent equations, we should consider the next order (~ S"(H)). Because U6"(1)1, those equations are independent of in, (l e JHI). On the other hand, for I e SRI, the leading order is P(H)+' because of (1632) together with
(91) and (96). Hence those N(R) equations of (1633) are dependent on in, (i e IRS) in the leading order. Thus the mixed secondtype singularity depends only on the internal masses of R. References (1) R. J. Eden, Proc. Roy. Soc. A 210, 388 (1952). (2) R. Karplus, C. M. Sommerfield, and E. H. Wichmann, Phys. Rev. 111, 1187 (1958). (3) Y. Nambu, Nuovo Cim. 9, 610 (1958). (4) R. Oehme, Phys. Rev. 111, 1430 (1958). (5) L. D. Landau, Nucl. Phys. 13, 181 (1959). (6) N. Nakanishi, Progr. Theoret. Phys. 22, 128 (1959).
128
GRAPH THEORY AND FEYNMAN INTEGRALS
(7) J. D. Bjorken, unpublished (thesis).
(8) R. J. Eden, P. V. Landshoff, D. I. Olive, and J. C. Polkinghorne, The Analytic SMatrix (Cambridge University Press, London, 1966). (9) R. E. Cutkosky, J. Math. Phys. 1, 429 (1960). (10) D. Fotiadi, M. Froissart, J. Lascoux, and F. Pham, Topology 4, 159 (1965). (11) H. Araki, Progr. Theoret. Phys. Suppl. 18, 83 (1961). (12) D. Hall and A. S. Wightman, Mat. Fys. Medd. Dan. Vid. Selsk. 31, No. 5 (1957). (13) R. C. Hwa and V. L. Teplitz, Homology and Feynman Integrals (W. A. Benjamin Inc., N. Y., 1966). (14) F. Pham, Introduction a L'etude Topologique de Singularites de Landau (GauthierVillars, Paris, 1967). (15) J. B. Boyling, Nuovo Cim. 53A, 351 (1968).
(16) N. Nakanishi, Progr. Theoret. Phys. Suppl. 18, 1 (1961); Errata, Progr. Theoret. Phys. 26, 806 (1961). (17) J. Mathews, Phys. Rev. 113, 381. (1959). (18) N. Nakanishi, Progr. Theoret. Phys. 23, 284 (1960). (19) J. M. Westwater, Helv. Phys. Acta 40, 596 (1967). (20) S. Coleman and R. E. Norton, Nuovo Cim. 38, 438 (1965). (21) R. Karplus, C. M. Sommerfield, and E. H. Wichmann, Phys. Rev. 114, 376 (1959). (22) N. Nakanishi, Progr. Theoret. Phys. 39, 768 (1968). (23) F. Pham, Annales de l'Institute Henri Poincare A 6, 89 (1967). (24) J. B. Boyling, Nuovo Cim. 44, 379 (1966). (25) T. Kinoshita, J. Math. Phys. 3, 650 (1962). (26) N. Nakanishi, Progr. Tlzeoret. Phys. 19, 159 (1958). (27) B. Andersson, Nucl. Phys. 74, 601 (1965). (28) G. Kallen, Nucl. Phys. 25, 568 (1961). (29) N. Nakanishi, J. Math. Phys. 3, 1139 (1962). (30) A. C. T. Wu, Phys. Rev. 135, 222 (1964). (31) M. Minami and N. Nakanishi, Progr. Theoret. Phys. 40, 167 (1968). (32) N. Nakanishi, Progr. Theoret. Phys. 42, 966 (1969). (33) G. Kallen and A. Wightman, Mat. Fys. skr. Dan. Vid. Selsk. 1, No. 6 (1958).
(34) D. B. Fairlie, P. V. Landshoff, J. Nuttall, and J. C. Polkinghorne, J. Math. Phys. 3 594 (1962).
(35) I. T. Drummond, Nuovo Cim. 29, 720 (1963).
CHAPTER 4
PerturbationTheoretical Integral Representations The extended transition amplitude is a sum of the Feynman integrals which are generated by a particular Lagrangian. It is therefore important to find a domain in which all Feynman functions corresponding to a certain set of Feynman graphs are holomorphic. This approach is complementary to the one developed in Chapter 3. Without analyzing the explicit dependence on the Feynman parameters, we can obtain a certain amount of knowledge about such a domain as mentioned above from the Feynmanparametric formula. For this purpose, it is convenient to introduce the perturbationtheoretical integral representation (PTIR), which is obtained from the Feynmanparametric integral by changing integration variables in such a way that the number of new variables is equal to that of independent invariant squares of momenta. The most essential problem is to find support properties of the weight distributions of PTIR. If the Feynman function is regarded as a function of only one invariant square, PTIR reduces to a (single) spectral representation, which is called a dispersion relation when applied to the scattering amplitude. Aside from this case, PTIR (a very special case of it) was used first by Wick [1] and Cutkosky [2] in 1954 in order to solve the BetheSalpeter equation for bound states. At the end of 1954, Nambu [3] proposed an integral representation for the scattering amplitude on the basis of Feynman integrals. In 1955, he [4] further proposed some multiple spectral representations, which unfortunately turned out to be wrong. In 1958, Mandelstam [5] proposed a
double spectral representation for the scattering amplitude, called the Mandelstam representation. On the other hand, an integral representation for the vertex functiont, which was a special case of PTIR, was derived, but t This representation had, already in 1957, been applied to the BetheSalpeter equation as a natural generalization of the WickCuskosky one [6]. 9
NakanisLi
129
130
GRAPH THEORY AND FEYNMAN INTEGRALS
unfortunately incorrectly, on the basis of the axiomatic field theory independently in 1959 by Fainberg [7], by Deser, Gilbert, and Sudarshan [8], and later by Ida [9]. The systematic proposal and detailed study of PTIR were made by Nakanishi [1016] in 19611964. He also investigated support
properties for various cases, and in particular proved the validity of the abovementioned integral representation (under the stability condition) for any Feynman integral [10]. The PTIR for the scattering amplitude is quite akin to the Mandelstam representation in its appearance. Some important improvements of support properties were made by Boyling [17]. In this chapter, we develop the theory of PTIR. We derive PTIR from the
Feynmanparametric formula in Section 17. Mathematical properties of PTIR are presented in Section 18. Methods of finding support properties are discussed in Sections 1921; Section 19 is an application of the transport problem (see Section 5) and Section 21 offers an interesting graphtheoretical problem. In Section 22, we summarize the results on the support properties of PTIR.
§ 17 171
DEFINITION OF PTIR
General Remarks
As is well known in hadron physics, the dispersion relation is one of the most powerful tools for the analysis of strong interaction because it involves no approximation. Here a dispersion relation is a spectral representation of a scattering amplitude in a single invariant square (more precisely, it consists of two terms each of which has a spectral representation). Since the scattering amplitude is a function of two variables, however, the dispersion relation becomes unsuitable for investigating detailed dynamical problems. We shall need analyticity in two or more variables. In this direction, the Mandelstam representation] is the most interesting hypothesis for the twovariable
analyticity of the scattering amplitude. Unfortunately, however, this representation is not yet proven even in the case of no anomalous threshold, and it is hopeless to extend it to the case of the production amplitudes. It may be worthwhile, therefore, to investigate a family of integral representations which are valid for various transition amplitudes and extended ones. As described in Subsection 61, an extended transition amplitude is ex
pressed in terms of an infinite sum of the Feynman integrals which are generated by a particular Lagrangian. Therefore, its analyticity domain can be obtained by investigating the analyticity domains of all those Feynman integrals, aside from the convergence problem of the perturbation series (see Section 24). For this purpose, we introduce the perturbationtheoretical integral representation (PTIR) in the next subsection. We consider certain (infinite) sets of Feynman graphs G having n external lines. Those external lines are incident with n external vertices which may not necessarily be distinct. For simplicity of notation, however, we use n different names for the external vertices throughout this chapter. [Therefore, the meaning of n is somewhat different from that in Chapter 3.] The set of all external vertices of G is denoted by g as before. The set, Y',,, of (connected) Feynman graphs is characterized by the kinds
of particles associated with n external lines and the admissible types of vertices, where the type of a vertex a is a list of the particles of the lines (including external lines) incident with a. Of course, as required physically,
particles of the same kind should have the same mass. A more detailed specification of Y' (or Y';,) is stated in Subsection 201. t Its explicit form is presented in Subsection 182. 9*
131
132
GRAPH THEORY AND FEYNMAN INTEGRALS
The above characterization of the set W. is based on the Lagrangian field theory because the interaction Lagrangian determines the admissible types of vertices. In the axiomatic field theory, however, one cannot specify the type of each vertex because the Lagrangian is not considered there, and therefore one uses the spectral conditions as its substitute. Accordingly, in order to see the connection with the axiomatic field theory, it is interesting to consider another set, O,,, of (connected) Feynman graphs specified by the spectral conditions. Here the spectral conditions are defined in terms of Feynman graphs as follows : There exists a nonnegative constant M,,( M9_,,) for each nontrivial division (h I g  h) of g such that for any Feynman graph G C On we have
Y_ M, ? Mh for any SeSG(h I g  h)
(171)
1ES
that is, the total mass of any intermediate state in the channel (h I g  h) has a lower bound Mh. In particular, if h = {a} then we write M, as M,,; if h {a, b} then Mh = Mab, etc. t It is noteworthy that the set Y' also satisfies the spectral conditions if we define
Mh  inf
min
E m,
(172)
GEW SESG(bbI9h) IES
and that (172) can easily be computed if the admissible types of vertices are given. Thus, in this sense, the conditions imposed on 0n is weaker than those on ,,. The distinction between On and iP becomes very important if we have, for example, the baryonnumber conservation law, which implies the existence of baryon paths between pairs of baryon external lines (see Subsection 201 for details). There is no way of expressing such a property in
the spectral conditions. Owing to such circumstances, the supports of the weight distributions of PTIR for !a become more restrictive than those for On, provided that they satisfy the same spectral conditions. For a transition amplitude (on the mass shell), we usually require pa for any a e g to satisfy the stability condition
pa 5 Ma
(173)
Throughout this chapter, we neglect spins, that is, all particles are regarded as spinless. t We use this notation for any quantity having a suffix h.
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
172
133
Derivation of PTIR
In this subsection, we derive PTIR from the Feynmanparametric formula. We note that it contains external momenta only in the denominator function Vas scalar products or invariant squares, and that V is a linear function of them. Since we often consider the quantities on the mass shell, it is convenient to employ invariant squares as variables. Hence the modified cutset representation is suitable for our purpose. The number of independent invariants, that of scalar products, and that of invariant squares vary with n as follows:
n
independent variables
invariant squares
scalar products
2
1.
1
1
3
3
3
3
6
6
4n10
n(n1)/2
4
n?5
7
2"11
Since the Feynman function should be expressed as a function of independent invariants, we have to eliminate some of invariant squares for n ? 4. For n = 4, 5, all relations between squares are linear, and therefore it is not difficult to eliminate the redundant ones. For n > 6, however, since the external momenta are no longer linearly independent, the Gram determinants for all sets of five momenta vanish identically, and they present nonlinear relations for squares. If one eliminates redundant squares by
using them, then the linearity of V in squares is lost. Thus it is not convenient to employ squares for n z 6 if our momentum space is of four dimensions. Hereafter, we do not consider the complication due to the dimension of the Minkowski space, that is, we neglect the non linear relations.
We denote the set of all independent squares by s symbolically. All functions of external momenta are regarded as functions of s. Hence we write (113) as L%
f
fG(S, A) = const
dNa
1
`
(174)
U2(a) [V(s' a)]1020 A
The modified cutset representation (746) of V reads
V = E semi  Y_ C,,s,,
(175)
134
GRAPH THEORY AND FEYNMAN INTEGRALS
where
r
W(Mgh)/U
Sh =
(176) (177)
(_ pb) 2 beh
and the summation Y_ goes over all nontrivial divisions (h I g  h) of g. Of course h >= 0 for a e A. We multiply (174) by oo
1 = H
dzbl
a(zb 
dx 6(x 
I
alnzi/s)
(178)
Jo
0
where (179) Is
and interchange the order of the integrations. Then we obtain 1
fG(s, 2) = fj
dzh
6(1 
f
Z')
h
dx
,J
Is
0
Pc(z, x
(x 
0
ZpSh) l°
Zre
(1710)
h
with
6(1  Y zr) (PA., x; 7.) = const h
irMAJ1
f dNla
J
U2 AO2r`
A
r$(x 
x fT 6(Z"
armi /)
h
(1711)
Now, we perform the regularization defined by (1031) and integrate (1710)
by parts with respect to x N  2,u  1 times. [If N  2,u < 0 then we need subtraction procedure, which is described in Subsection 183. In this section, we neglect this complication.] Then we have
f
fc(s) _ h
dzr,
6 (1  y
J
zf .J
0
0
d;
X
(1712) `frSh b
where cpG(z, x) _
(8/8x)Nzk1 Reg (PG(z,
x; 2)/(N  21u  1)!
(1713)
1=1
It is noteworthy that in (1712) the dependence on G of the integral is contained only in the weight distribution OG(z, x). Therefore we obtain an integral representation associated with the set rh or P if we replace 9G(z, x)
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
135
by a weight distribution T (z, x), which is a formal linear combination of g9G(z, x) over all G E 0a or Pa. When tpc(z, is replaced by 97a(z, x), the righthand side of (1712) is denoted by a(s), that is, 1
A(S) _ H
dz,,
8(1
zb)
f dx
h
S
0
(1714)
994Z,
x  Y_ ZhSh
0
h
By definition, fa(s) is dependent only on fi or iV' but not on the individual Feynman graph. In this way, we can express some common properties of the Feynman functions associated with rha or If,.
Since the invariant squares sb are independent for n = 2 and n = 3, the representation (1714) is meaningful in those cases. Hence for n = 2 and 77 = 3, we call (1714) the perturbationtheoretical integral representation (PTIR) of fa(s). More precisely, for the selfenergy, writing x = s', we have'
r
f2(s) =
ds'
,J
0
(1715)
Ss
a (single) spectral representation which was obtained in the axiomatic field theory [18, 19, 20, 21]. For the vertex function, we have 1
/a(s) =
1
f dza
f
1
d,,,
oo
f dz,8(1  Za  Zb  Zc)
(P3(Zn, Zb, Zc, x)
x
r dx
(1716)
x  Zasa  ZbSb  zcSc
with g = {a, b, c}.
Now, we consider the extended scattering amplitude (n = 4). Let g = {a, b, c, d}. Then there are seven invariant squares sa, sb, s., sd, sab Scd, Sac = Sbd, Sad = sbc Since an identity Sa + Sb + Sc + Sd = Sab + Sac + Sad
(1717)
holds, we have to eliminate one of these seven squares. Any point a E A in the Feynmanparameter space belongs to one of the three regions Uad = {a E d I Cad
pS Cab,
pad
< Cac}
dab = {a e d I Cab G Sac, Cab < Sad}
dac = {a E d I
Sac
C Cad, Sac < Cab}
(1718)
136
GRAPH THEORY AND FEYNMAN INTEGRALS
For x e dad, because of (1717), V can be written as d
V=
alm;
 E (v + Sad) S, + (Cab  Sad) p
/pp
p
L=a
7g
Sab
p
l
+ (bac  Sad) Sac]
(1719)
In the square bracket of (1719), the coefficients of invariant squares are positive semidefinite. In this form, we can introduce the variables za, ..., Zac and x in the same way as in the derivation of (1712). We can do the same thing for the other two regions dab and day. In this way, instead of (1714) we obtain the PTIR for the extended scattering amplitude:
A(S) =h*ad n I dzh

S
f dx x 
h*ad
(1
(01
9ad(za a ... , Zac I x)
Zh)
0
zhsh
h*ad
+ two other terms obtained by cyclic permutations of b, c, d (1720) where h + ad means that h runs over a, b, c, d, ab, ac.
The PTIR for the scattering amplitude is obtained by putting sa, sb, s,, sd on the mass shell s, = m, (v = a, b, c, d) and then by setting y = (x  E Zvmv)l(Zab + Zac) \
v
(1721)
Z = Zabl (Zab + Zac)
in the first term and likewise in the other terms. It is customary to write Sab = S, sac
t, sad
u. Then from (1720) the scattering amplitude
f4(s, t) is represented as ao
f4(s, t) =
Jdzfd/___ oo
0
+
f dz f
Pta(Z, Y)
dy Y
i
est(Z, Y)
zs(1 z)t
zt(1z)u
00
Pas(Z' Y)
zu(1 z)s
(1722)
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
137
where Qom, etc. are defined in terms of g),,, etc. If we derive (1722) directly
from (174), the correspondence between integration variables and the Feynman parameters are as follows: d
Y^
[
OG:mi
 1 =a(Cl +
Cad)mV /E
Z (Cab Sad)l with
Cab + Sac  2Sad
(1723)
(1724)
for the first term, and likewise for the other two terms.
We note that the representation (1722) is useless if we cannot find finite quantities y,,(z), etc. such that 2sr(z, y) = 0 for y < Ys:(z), etc.
(1725)
because if not the Feynman integral cannot be obtained as a boundary value of f4(s, t). We see later that (1725) can be proven in most cases of practical interest, but we should remember that it is not always true even under the stability conditions. A counterexample is provided by a complete Feynman graph Fig. 41(b) with the common external mass m = m being larger than J6 times the common internal mass mi [11, 22]. Finally, we consider the case n > 5. In order to find the PTIR for n z 5, we have to solve the following problems. a) To find sets of n(n  1)/2 linearly independent squares among 2111  1 ones.
b) To find the region of the Feynmanparameter space d in which the coefficients in  V of the linearly independent squares of each set are positive semidefinite.
c) To find some combinations of the sets such that the corresponding regions found in (b) cover the whole 4 without overlapping except for their boundaries. This program was worked out for fs(s) on the mass shell [13]. In this case, we have to find five linearly independent squares among ten. The number of such sets is found to be 162 instead of (s) = 252. By the investigation of (b) and (c), we find that it is impossible to construct PTIR in such a way that it is symmetric with respect to all external vertices. Accordingly, at least one external vertex has to be dealt with asymmetrically. If one external
GRAPH THEORY AND FEYNMAN INTEGRALS
138
vertex is dealt with asymmetrically, then for any value of n we can always construct PTIR for the production amplitude in the following way [14, 23].
Let a be a particular external vertex. For h c g  a,
Sh  (> Pb)2 beh
beh
pb + 2{b,c}ch E Pip,
(1726)
We consider a set of (n  1) (n  2)/2 squares sbc = (Pb + bc)2 for {b, c} c: g  a. Then we have (1727)
2PbPc = Sbc  Sb  Sc
Hence
Sh = (nh  2) y sb + beh
Y_
{b,c}ch
(1728)
Sbc
where nh denotes the number of the external vertices belonging to h. For h = g  a, since Sh = so, from (1728) we obtain an identity Y_
Sbc = (n  3)
Y_
(1729)
Sb + Sa
bega
{b,c}cgh
If we eliminate one of squares Sbc by means of (1729), then we obtain
(n  1) (n  2)/2 sets of n(n  3)/2 linearly independent squares other than s (v e g). [Note n(n  1)/2  n = n(n  3)/2.] The combination of those sets satisfies the requirements stated in (b) and (c), because (1728) leads us to V = Y_ !X1111! 
Y_
bega
1
1J,Sb 
Y
{b,c}egh
where % is a certain linear combination of Ch (h
bcSbc
g  a), and
y C,, 0 (hcg a)
Sbc
(1730)
(1731)
and because the region of 4 described in (b) is characterized by the minimum one of Ebc. In this way, we obtain {'
I('
f (S) =
b'c' # be
(
,/
dZb ,c,)
6(1 
b' cl*
Z. bc
be
oo
f
J ao
dy eebc({zb'c' I b'c' + be}, y)
7
(1732)
Y, Zb'c'Sb,c' b'c'*bc
where any of be = {b, c} and b'c'  {b', c'} is contained in g  a. For n = 4, (1732) reduces to (1722). For n > 5, we have n different representations according to the choices of a.
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
173
139
PositionSpace PTIR
As in Subsection 172, PTIR is usually considered for the momentumspace
extended transition amplitude. If we regard the positionspace transition amplitude as a function of invariant distances squared
sbc  (xb  xc)2 for {b, c} c g
(1733)
then from (833) together with (844), we can derive the positionspace PTIR
t
f
be
J 0
dzb
b(1 
r dx
zbc) be
J 0
(1734) Y  Y_ Zbcsbc be
It is noteworthy that (1734) consists of a single term for any value of n. Since there are no spectral conditions in the position space, we have no restriction on the support of rpn. Hence we do not investigate (1734) any more.
§18 MATHEMATICAL PROPERTIES OF PTIR 181
Analyticity
We consider the analyticity of a function defined by oo
f
f dm1z
F(sl, ..., Sm) =
J
J
d
zm, y)
dy
 ao
(181)
M
Y zisj
y
j=1
where
LI  {z1i ... , Z. I z j z 0 (j = 1, ... , m),
z j = 1}
j
(182)
only in this section and
y) = 0 unless y >
(183)
z jc j j=1
with some constants cj. By linear transformations of s1, ..., s,, and y, however, (183) can be reduced to yi(z1, ..., zm,
y) = 0 unless y z 0
(184)
Hence we consider (181) with (184) without loss of generality. Then it is straightforward to see that F(s1, ..., s,,) is holomorphic in D, where D
{sl, ..., S. I y  Z zjsj * 0 for all zed, j=1
lllll
0)
(185)
1111
Let D+
{s1,...,sml Imsj > 0 (j= 1,...,m)}
D_
{s1i ..., sm l Im sj < 0
E
{s1,...,smllmsj=O,Resj 0, say, or there exist s, and s2 i say, such that 0 < 9)1 < r and 99, + v < ip2 < 2ir, where z;sj z 0 for z1 = 1, z2 = ... = zm = 0. 99;  arg s;. For the first case, ZJSJ _> 0 for zl =  c IM s2 , z2 = c IM sl , z3 = ... i = z m = 0 with c > 0. Thus (sl, ..., sm) D. q.e.d. Combining the above two theorems, we find that D is the envelope of
For the second case,
holomorphy of D+ u D_ u E. The fact that D is a natural domain of holomorphy can be proven more easily by constructing a function which is holomorphic in D but not so in an arbitrary point outside D [14]. 182
Uniqueness
THEOREM 183 [14]f
The integral representation (181) with (184) is
unique, that is, if F(sl , ..., sm) = 0 then 6(1  z;) yq(z1, ..., Zm, y) vanishes J as a distribution.
Proof We take the difference between the boundary values of F from D+ and from D_, and introduce new (real) variables tl = s1, ti = s;  si
(j = 2, ..., m). The assumption F =_ 0 then yields m
J A
m
dzic (z2 i ..., z,,,, tl + Y zit;) = 0 J=2 j=2
(1811)
where
zi =(z21,Z.I z;>0 (j=2,...,m), Y z; oo.
Roughly speaking, subtraction procedure is necessary for the functions which are nondecreasing in all directions (in D). Given a function F which is holomorphic in D and has certain asymptotic
behavior, we can find the weight distribution 99 in terms of F [16]. The extension of this problem to the case (1722) is solved partially [15]. 10
Nakanishi
146
GRAPH THEORY AND FEYNMAN INTEGRALS
The PTIR is useful mainly for the quantities off the mass shell, e.g., for the analysis of the BetheSalpeter equation. For the amplitude on the mass shell, (1722) is useful only as a theoretical tool, e.g., in the proof of the dispersion relations. This is because PTIR does not exhibit the analyticity in a complex neighborhood of the physical region. Therefore it is difficult to combine (1722) with the unitarity of the Smatrix, in contrast with the Mandelstam representation.
§ 19 SUPPORT PROPERTIES IN THE GENERALMASS CASE
The most important problem in the theory of PTIR is to find the supports of the weight distributions. This and the succeeding three sections are devoted
to the investigation of support properties of PTIR. 191
General Method
Given a Feynman graph G, the corresponding Feynman function fG(s) can be rewritten in the form of PTIR, and therefore the support of each weight distribution, say, 97G(z, x), of it is well defined. The problem which we consider in this section is to investigate (191)
Supp [0n] ° U supp wG Geo
It is very difficult in most cases to find (191) exactly, and we have to content ourselves by finding an upper bound on it. An upper bound on supp 4T9G can be found in the following way. First,
we find a point p = p in the complex 4(n  1)dimensional momentum space such that V(p, a) >_ 0 for all
aed
(192)
Second, as done in Subsection 172, we rewrite V as n(n 1)/2
ainh  , ?7j(a) sJ
V(p, a)
(193)
where 77J(a) z 0 for the relevant values of a and s1, ..., sn(n_1)/2 are independent squares of p. Let cJ be the value of sJ at p = p. Then (192) implies a:mi
(194)
71JCJ
t
J
Since 9G(z, x) has contributions only from
(j
1, ..., n(n  1)/2), where
Y _
o in / and zJ = qj/
E ?7j, (194) yields that J
99,(z,;,) = 0
unless
x>
zJCJ
(195)
J
Thus it is important to find a point p = p such that (192) holds for all G e J,,. For this purpose, the following theorem is useful. 10*
147
148
GRAPH THEORY AND FEYNMAN INTEGRALS
THEOREM 191 [26] If the set of external momenta p is euclidean (see Definition 135 and also Theorem 136), and if we can find internal mo
menta qt (1 c JGJ) such that
(a) they are linear combinations of p and satisfy the conservation law (124) at each vertex, and (b) mi Z qi for all I e JGJ, then (192) holds.
Proof Since {qi I 1 c JGI} is euclidean, the circuit representation (726) of V implies V(p, a) z ai(mI  qi) 0 (196) q.e.d. In particular, if all momenta Pb are timelike and parallel, each qt can be regarded as a onedimensional quantity. In this case, therefore, if we regard m1 as a capacity c(l), Pb as a demand d(b), and qt as a flow f(l), the problem of finding {q, 1 1e IGI} is exactly reduced to the transport problem discussed in Section 5. Accordingly, Theorem 53 is translated as follows. THEOREM 192 When all external momenta Pb are timelike and parallel, a necessary and sufficient condition for the existence of qt (l c JGJ) satisfying the conditions of Theorem 191 is (bY_ pb)2
(1Y m,)2
for any h e g and any S e S(h I g h)
(197)
If we combine Theorem 192 with the spectral conditions (171), we obtain the following theorem. THEOREM 193 For p tirnelike and parallel, (192) holds for all G e 0a
if
(E pb\2 < Mh for any h c g
(198)
beh
For n = 2, let a and b be the two external vertices. From Theorems 191 and 192 together with (745), we have
alms > (W(alb)/U) min i
SES(alb)
mJ)2
(199)
lcS
In particular, if all internal masses are equal to each other, (199) reduces to where
W(alb)/U < 11r2
for
r = min N(S) SeS(alb)
a e/,
(1910) (1911)
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
If we apply Theorem 193 to the case n = 2, we obtain tp2(s') = 0 unless s' ? Mo
149
(1912)
Since pG(s') + 0 for s' > Ma in the N = 2 selfenergy graph G, Supp [l2] is given by s' Z M.2. This result is well known in the axiomatic field theory [19, 20, 21]. 192
Vertex function
In this subsection, we discuss the support properties of Cp3(za, Zb, ZC, x) involved in (1716) by applying Theorem 193 to 03 Since pa, Pb, p. are parallel, we can write p,, = 1,,p
(v = a, b, c)
(1913)
where p2 = 1 and 2a, 2b, Ac are constants satisfying
Aa + Ab + A = 0
(1914)
The conditions (198) are written as I2,,I < M,,
(v = a, b, c)
(1915)
With those values of A,, Theorem 193 implies 9p3(za, Zb, zc, x) = 0 unless
(1916)
x
Thus our task is to maximize E zv2v under the restrictions (1914) and (1915). v
For simplicity, we assume that Ma, Mb, Mc satisfy the triangular inequalities
IMaMbI5Mc 0 for v = a, b, c; v
Y
the exceptional case JJ z,, = 0 implies d
of Y
z,,A2 = 0. Therefore, the maximum Y
has to occur at one of the three extreme points of the region
V
defined by (1914) and (1915). Thus 993 = 0 unless
x>max zaM2 a +zbMb + [ zaMa + zb(Ma
 MJ2 +
2
Mb) ,
zcMc
za(Mb  MX +zbMb + zzMa]
(1918)
150
GRAPH THEORY AND FEYNMAN INTEGRALS
where "max" should be taken for each (za, zb, zj fixed; hereafter we always use max and min in this sense if they appear in the expressions for support properties. Unfortunately, (1918) is not exactly Supp [rh3], but (1918) is a good upper bound on it. For s, = mC, if it satisfies the stability condition in, < MM, setting
Z=
Z,,
y=
Za+Zb
x  zCmC Za+Zb
(1919)
in (1716), we obtain a twovariable representation W
1
f dz Jdy
L(Sa, sb, mc) =
0
where
49ab(Z, y)
y  ZSa  (1  Z) Sb
(1920)
0
1
99ab(z,
y) = f dy y g3(yz, y(1  Z),1  y, yy + (1  y) m2)
(1921)
0
and it vanishes unless
y _> max [zMQ + (1  z) (Ma  MM)2, z(Mb  Mc)2 + (1  z) Mb 2]
(1922)
In (1922), under the stability condition, Mc can be replaced by its smallest
possible value in, if me > IM.  MbI. Then, in contrast with the threevariable case, (1922) is realized in the triangle graph (see Fig. 134). In this example, (1920) is directly obtained from the Feyumanparametric formula by a transformation of variables
y= .
Zw2
011 +012 Since al + 012
(1923)
(al +0'2)013
+013=1, [C(z)]2 013
where
aim j  a1a2mC
e(z)
2
+
n13
+ z(1  z) mC
(1924)
1  013
Z) m1 + zm2  z(1  z) mC >_ 0
(1925)
Hence
inf y = [((z) + m3]2 + z(1  z) inc a;
(1926)
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
151
If in, = MM = m1 + m2 then
infy=(1 z)mi+zm2+m2+2m3Im1(1 z)
m2zI
a3
max [z(m2 + M3 )2 + (1  z) (m1  in3)2, Z(m2  rn3 )2 + (1  Z) (M1 + m3)2]
(1927)
Since Ma = 7n2 + 7n3, Mb = m1 + m3, and MC = m1 + m2, (1927) coincides with (1922). It is interesting to rewrite (1920) together with (1922) and Mc = me into
another form. Let P = pc and p
(Pb  pa)!2, so that
S. = (+P + p)2, Sb = (iP  p)2,
mC2 = P2
(1928)
By a transformation of variables
+ (1  2) mC
z = (1 + C)/2, y =
(1929)
we have 1
PA P) = f
oo
,J
1 where
f
dC
fl)
d1
(1930)
1  (P +  'P)2
J
0
j9) = 0 unless
9 > Max {[Ma  M(l  C) mJ2,
[Mb  M(1 + C) ma]2}
(1931)
The integral representation (1930) together with (1931) is the one proposed (but not correctly proven) in the axiomatic field theory [7, 8, 9]. Thus it is
true in every order of perturbation theory [10]. If m. does not satisfy the stability condition, however, this representation is no longer true even in the triangle graph (see Appendix B2). 193
Extended Scattering Amplitude
In this subsection, we discuss the support properties of Pad involved in (1720); those of 92ab and 99aa are similar.
As before, we set p,, = Avp (v = a, b, c, d), where
Aa + 2b + Aa + td = 0
(1932)
and p2 = 1. Then (198) reads
IA,I<M.
(v=a,b,c,d)
IAa + 2vl < May
(v = b, c, d)
(1933)
GRAPH THEORY AND FEYNMAN INTEGRALS
152
Theorem 193 then implies that Pad(za, ..., Zac, x) = 0 unless d
x
aa
Y_ Zv"v + Zab(Aa + 4)2 + Zac(Aa + 2c)2
(1934)
y=a
Since the extremum condition for the righthand side of (1934) under (1932) yields a particular relation for Za, ..., zac, it is probably unimportant.
Hence we consider the extreme points of the region defined by (1932) and (1933) as before. For example, if we put
A. = Ma, 2b =
Mb,
A. + 2 = Mac
(1935)
(1934) reduces to
X_
zaMa + ZbMb + Zc(Mac  Ma)2 + zd(Mb  Mac)2
+ zab(Ma  Mb)2 + Zac. C (1936) However, the point (1935) does not necessarily satisfy all of the inequalities (1933), that is, (1936) is true under the additional assumptions
IMacM.I <Me, IMb  MacI <Md Ma  Mbj Mab, Mc  Mb Mad = Mbc
(1937)
Ma + Mb < Mac + Mad
(1938)
and
As shown below (see Theorem 198), the triangular inequalities (1937) are always satisfied for anyreasonable set 0a, but (1938) is a very stringent assumption on 04 as is seen by the following example. Example
Consider the set of all Feymman graphs G having four external
lines of mass m such that all vertices of G have Fdegree 3 and the lines incident with each vertex have masses m, m, 1, and suppose m > 1. [This is a model corresponding to the nucleonnucleon scattering.] Then we have
Ma = Mb = m + 1 and Mac = Mad = 2, and therefore (1938) is not satisfied. It should be noticed, however, that both Mac = 2 and Mad = 2 are never realized in a single Feynman graph. As is suggested by the above example, it is better to prove the inequality corresponding to (1938) in each Feynman graph and later consider the
spectral conditions of J,,. Since originally the spectral conditions are determined implicitly from a certain interaction Lagrangian, it is quite reasonable to suppose that 0a satisfies the following condition. ASSUMPTION 194
Let g = {a1, ...,
For any G1, ..., Ge0a, we
require G e 0a, where G is any connected Feynman graph having n external
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
153
..., a such that it can be constructed out of G1, ..., G. through repeated use of the following operation: to replace two external lines associated with aj belonging to two graphs by an internal line j' incident vertices al ,
with the relevant two vertices (in short, to "join" two external lines of the same kind). gh
Fig. 191 A Feynman graph composed of G1, G2, G3.
For example, if G1, G,, G3 e J,, then the above assumption implies that the Feynman graphs shown in Fig. 191 and in Fig. 192 belong to 0n
Given a Feynman graph G, we introduce a complete Feynman graph, G, in the generalized sense (that is, we admit parallel lines, loop lines, and internal vertices) in such a way that

G10=G with Io{1lm,=0} and define
t We need not specify its mass. t Any cutset in d corresponds to a cut in G.
(1939) (1940)
154
GRAPH THEORY AND FEYNMAN INTEGRALS
Fig. 192Another Feynman graph composed of G1, G2, G3, where
h1uhzuh3=g.
Furthermore, we introduce the following notation: m(I)
(1941)
mi ie7
M,,(G) = MB_,,(G)
min
m(S)
(1942)
SESG,(hI9h)
where Mh(G) = oo if SG(h I g  h) = 0, and Mh(G) is similar. For any S e SG(h I g  h), we can always find S e SG(h I g  h) such that S c S u to . Since m(S) = m(S) because m(I0) = 0, we have Mh(G) > 0
Mh(G)
(1943)
The spectral conditions (171) imply
M,, = inf M,,(G)
(1944)
Geo[
THEOREM 195 [11]
M,, = inf Mh(G)
(1945)
Gem
Proof Because of (1943), it is sufficient to show that for any G e 0. and for any h c g there exists G' e J such that Mh(G) > M,,(G')
(1946)
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
155
Such a Feynman graph G' is given by Fig. 191 with Gl = GZ = G3 = G. Indeed, for any S E SG(h I g  h), by regarding S as a cutset in G2, we can find a cutset S' in G' such that S' c S [hence m(S')  m(S)] because of Theorem 21(d) (S to because G is connected), and moreover S' ESG.(h I g  h) because S' is a cutset in G'/(IGII + IG31) [see Theorem 22(d)]. q.e.d. THEOREM 196
For a complete graph G, if S e Sj;(h I g  h) and S'
E Sff(h' I g  h') for any h, h' c g, then we have
S(D S'ESG[h(D h'Ig(hph')]
(1947)
where the symbol p denotes the symmetric difference (see Subsection 11).
S u S'), which is Proof We consider a reduced graph R also a complete graph and has only four vertices; they may naturally be
called h n h', h n (g  h'), (g  h) n h', and (g  h) n (g  h'). Since R is essentially the graph shown in Fig. 41(b), it is easy to see that S (D S' is a cutset belonging to SR(h n h',
(g  h) n (g  h') I h n (g  h'),
(g  h) n h'). Hence in G, S (D S' is a cutset because of Theorem 22(d) and belongs to SG[h (D h' I g  (h (D h')] because the division [h (D h' I g
 (h p h')] is nothing but the division ([h n (g  h')] u [(g  h) n h'] [h n h'] u [(g  h) n (g  h')]). q.e.d. THEOREM 197
I
For any h, h' c g,
Mh(G) + Mh' (G)
Mheh' (G) > I WO  Mh' Al
(1948)
Proof Mh(G) + Mh. (G) =
min
[m(S) + m(S')]
SESU(hjgh)
S' ESZi(h' I g  h')
z min m(S (D S') min
m(S)
SESGth h'Ig(h®h')]
(1949)
with the aid of Theorem 196. The latter inequality of (1948) follows from the former one by substituting h E h' for h or h'. q.e.d.
The corresponding formula for Mh(G) cannot be proven because for S, S' e S, S (D S' is not necessarily a cutset in G.
GRAPH THEORY AND FEYNMAN INTEGRALS
156
If h n h' = 0 for h, h' c g, then
THEOREM 198
Mh + Mh. >
IM,,  Mh'I
(1950)
Proof According to (1945), for any e > 0 we have 015 G2, G3, e such that IMh  Mh(G0I < e, M.  Mh'((72)I < e (1951)
IMhuh'  Mh,,h'(G3)I < e
We construct G e C from Gl, G2 , G3 as in Fig. 192t with hl = h, h2 = h',
h3 =ghuh'.Then 6e1,and I m"  M,(G)I < e, I m",  Mh'(G)I < I Mh,,h'
(1952)
 Mhi,h'(G)I < e
Since e can be arbitrarily small, Theorem 197 implies (1950) because
hQh' =huh'for hnh' = 0. q.e.d. In particular, for any a, b e g we obtain the triangular inequalities
M. + Mb > Mab > I Ma  MbI
(1953)
No proof of (1953) seems to exist (but it was used tacitly [27]) in the axiomatic field theory. We also note that if h n h' + 0 then M,, + Mh. >_ Mh®h' does not hold in general.
Now, we return to the case n = 4 (g
{a, b, c, d}).
THEOREM 199 [11] d
Y_ M,(G) < Mab(G) + Mac(G) + Mad(G)
(1954)
,,=a
Proof' We may assume without loss of generality that a, b, c, d are distinct because otherwise the righthand side is infinite. We have only to show that for any S,, e SG(h I g  h) (h = ab, ac, ad) there exist S, e SG(v I g  v) (v = a, b, c, d) such that d
YN(S, n I) < N(Sab n I) + N(Sac n I) + N(Sad n I)
,=a
(1955)
with I  {l} for any 1 e IGI. We define 9, (v = a, b, c, d) by the stars S[v] in a reduce graph R  GI(IGI  Sab u Sac u Sad). Then because of Theorem t Use of this graph was suggested to the author by Prof. H. Araki.
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
157
22(d), S,, e SG(v I g  v). Let N1(1) and N2(1) be the lefthand and the right
hand side of (1955), respectively. It is evident that N1(I) _ 1 then d
l e U Sy v= a
C Sab U Sac U Sad
(1956)
whence N2(1) > 1. Finally, if N1(I) = 2 then 1 e S. n Sb, say, that is,
1 e S[a] n S[b] in R. That means that a and b are adjacent in R, and therefore l e Sac and 1 e Sad, that is, N2(1) > 2. q.e.d. Combining (1954) with (1948), we obtain six inequalities
Ma(G) + Mb(G) < Mac(G) + Mad(G), etc.
(1957)
which are always valid under Assumption 194. Thus we no longer need ad hoc assumptions such as (1938). Hence instead of (1936) we obtain that G,ad = 0 unless Za[Ma(G)]2 + Zb[Mb(G)]2 + zc[Mac(G)  Ma(G)]2
+ zd[Mb(G)  Mac(G)]2 + Zab[M.(G)  Mb(0]2 + zac[Mac(G)]2,
(1958)
etc.
Hence Theorem 195 yields that qua = 0 unless x > max [2a + Zb + Zac, Zc + id + !.c,
Za + Za + Zac) Zb +'C + A e. , (1959)
Za + Zc + tab, Zb + id + Zab, Za + Zd + Zab, Zb + Z2 + Zab]
with z,, = zMM (h = a, b, c, d, ab, ac). For example, from (1959) we have
I d
x?
y_a
max (ZabM ry,
(1960)
§20 MAJORIZATION METHOD 201
Models
As explained in Subsection 61, hadrons are only particles which strongly interact each other. If other interactions are neglected, several hadrons are regarded as stable. Those stable hadrons have particular masses and specific quantum mumbers. We construct some models by abstracting the following qualitative features from their properties. First of all, there exist particles having the smallest mass, which are called
pious. Their mass is taken to be unity. We sometimes require pious to satisfy the parity conservation law; its definition is given in Subsection 203. We call G an equalmass Feynman graph if all particles involved in G are pions and if the parity conservation law is not taken into account.
In general, we also have various mesons and baryons (pious are a kind of mesons). Each baryon line has an intrinsic orientation, that is, for any line associated with a baryon we cannot arbitrarily assign outgoing and
ingoing to its two end points. If a Feynman graph G involves baryon lines, we require G to satisfy the baryonnumber conservation lain, that is, for any vertex of G, an outgoing baryon line and an ingoing one (without distinguishing external lines from internal ones) can be incident with it only in a pair (or pairs). This rule immediately implies that the number r of the outgoing baryon external lines of G is equal to that of the ingoing ones, whence 2r 1. As emphasized in Subsection 171, the existence of the baryonnumber conservation law has an important effect on the supports of weight distributions. The following example explains how Y'3 is different from 03 when they satisfy the same spectral conditions. 158
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
159
Example We consider the condition for the absence of an anomalous threshold of the "nucleon vertex function" in the triangle graph shown in Fig. 134; the particles of the external lines incident with a and b are nucleons.
For simplicity, we put ml = m2 and sq = sb S (in, + m3)2. From (1327), the condition for no anomalous threshold in s, reads sa < m2 + m3
(201)
If the baryonnumber conservation law is taken into account, all the minima
of in, + m3, rn2 + m3, and ml + m2 are realized in the case in which the line 3 is of a nucleon and the lines 1 and 2 are of pions. Then the condition (201) reduces to sQ < m2 + 1; the nucleon's mass shell sq = m2 is allowed
for any value of in. Since the above identification of particles implies Mq = Mb = in + 1 and MM = 2, if we impose only the spectral conditions with those values of M, Mb, MM on the internal masses, we can choose
ml = m2 = m3 = (n? + 1)/2. Then (201) becomes so < (m + 1)2/2; the mass shell is allowed only for in < 1/2 + 1. Thus if in is much larger than unity as observed experimentally (in  6.7), the baryonnumber conservation law is very important. Now, we have to define the set P;; precisely. If we strictly adhere to the Lagrangian field theory, we shall be bothered by inessential complications. For example, a particular interaction Lagrangian L;,,, admits only particular types of vertices, but there is no inevitability of choosing Lip, . To be general as well as possible, we will not specify the types of vertices more than the baryonnumber conservation law. The set Y',', consists of all strongly connected Feynman graphs G satisfying the following conditions: DEFINITION 201
(a) in, z 1 for anyleIGj. (b) The baryonnumber conservation law is satisfied. (c) G has 2r baryon external lines and n  2r meson ones.
It is important to note that if G e I' then G/I e !P for any I c IGJ. According to Definition 201, Yf, does not contain the Feynman graphs which have cutlines. We have excluded cutlines because since they induce
only poles, it is easy and convenient to take them into account later on. [We do not explicitly describe them any more, however.] For any G' satisfying (a), (b), (c) of Definition 201, Y'.' does contain G'/I' instead of G', where I' is the set of all cutlines of G'.
160
GRAPH THEORY AND FEYNMAN INTEGRALS
The problem of discussing
Supp [V12 = U supp 9'G
(202)
can be reduced to finding p = p at which (192) holds, as in Subsection 191. It is no longer good, however, to consider the case in which all of Pb (b e g) are parallel, because then no characteristics of P can be taken into account. If we still wish to apply Theorem 191 to our problem, we have to solve a nonlinear problem, which is beyond the scope of the theory of linear programming. We quote here the following theorem, which can be proven by means of the path theorem (Theorem 54). We call the method based on the path theorem the path method. THEOREM 202 [10]
If G is a strongly connected equalmass Feynman
graph with n = 4, we can find qi (1 e J GI) satisfying the conditions of Theorem 191 for the euclidean external momenta defined by
Pa= klk2, Pb= 'l+Ice PO = k1  k2,
(203)
pd = k1 + k2
where k21 = k22 = 1 and k1k2 = 0.
This theorem is particularly useful for the case in which all four external particles are of the same kind (e.g., T4). In the other cases, it is better to employ majorization methods, where a majorization method is a method of reducing the inequalities (192) for given complicated graphs to those for
very simple ones. Two majorization methods are known: one is called euclidean [28, 25] and the other is noneuclidean [29, 30] according as {Pb b e g} is restricted to a euclidean set or not. Generally speaking, the latter method, which we describe in the next subsection in detail, is very powerful for the transition amplitudes but not for the quantities off the mass shell. We omit the euclidean majorization method and quote its results only when it yields better results than those which can be obtained by the noneuclidean majorization method. I
202
Majorization Rules
In this subsection, we present some basic majorization rules of the noneuclidean majorization method. Since we consider the V functions corresponding to various Feynman graphs, we write the V of G as VG(p, a) and
thezi ofGasdG.
161
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
If
DEFINITION 203
VG(p, a) z 0 for all a e dG
(204)
follows from the assumption that VGf(p, a) >_ 0
for all a e AG,
(205)
for j = 1, ..., k, then G is said to be majorized by the Feynman graphs G1, ..., Gk, and the majorization is denoted by G > {Gl , ...,
(206)
Gk}
For k = 1, we omit the curly bracket in (206). The following theorem is a direct consequence of Definition 203. THEOREM 204
If there exist some j (1 < j< k), some a e dGj, and some
c > 0 such that
cVG(P, a) Z VGJ(P, a)
(207)
where VG(p, a) with p fixed assumes the minimum value at a = a in do {a I at >_ 0 for all at, Y at = 2 > 01, then G is majorized by G,, ..., G. IEIGI
THEOREM 205
1
Let (QG)t be the increment (with p and al, for all 1' + 1
fixed) of
QG
akmk
 VG
(208)
keIGI
when at is increased by St (not infinitesimal). Then /QG)I
l
=
0161
(209)
et + bt
where Qt =
I,t2
U
(2010)
8 U/bat
and Yt is given in Subsection 92.
Proof According to the circuit representation (726) of V, we have QG = k
akgk 
(UC/U) C
(
[C: k] akgk)2
(2011)
k
We can choose qk (k e JGI) in such a way that all the circuit equations (126) are satisfied identically, that is, we set qk = Yk (k e IGI) [see (1211)]. Using the fact that U and UU are linear in at, from (2011) we find (QC,)I 11
Nakanishi
= SIYt 
Uc + St 8Uc/8a, [C: 1]2 b2I'i U + St 8U/8at
(2012)
162
GRAPH THEORY AND FEYNMAN INTEGRALS
By means of (733) and aUc/8a, = 0 for I e C, (2012) is reduced to U 6,Y; U + S, aU/8a,
(QG
(2013)
q.e.d.
THEOREM 206
If 1 is not a cutline, (2014)
QG L  QG = e!
If If 1 is a cutline, of course the lefthand side of (2014) equals
2
Proof We note from (745) that (2015)
urn QG = QGi
because the linearity in a, implies W(b1C)
(2016)
a U/8a,
a,
and if 1 is not a cutline then aU/8a, and 0W(b1 `)/8a, are the U and the W(blc) in G  1, respectively, according to Theorem 319. Therefore, (209) with 61 > oc yields (2014). q.e.d.
We rewrite (2014) as
VG(P, a) = VG_I(P, a) + a,mi
O,Yi
(2017)
If VG(p, a) assumes the minimum at a = a in dG, we should have either
a, = 0 or 2 2 M1  Y,la=a=
a VG arc, [a=s
= 21 VG(P, a)
(2018)
according to (931) and the method of Lagrange multiplier. We therefore have either VG(P, a) = VG/t(P, a)
(a' = 0)
(2019)
because of Theorem 91, or, from (2017) and (2018),
c,VG(P, a) = VG_l(p, a) + (a, + 1) mi where p, . Q,la=B and c,
1 + 2_1 Q, > 0 (with 2 = 1 + a,).
(2020)
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
163
After the above preparation, we consider majorization rules. When G is majorized by G1, ..., Gk, we usually require that
(a)ifGeP then Gje jnforj=1,...,k,and (b) N(G) > N(Gj) for j = 1, ..., k. The majorization is called admissible if (a) is satisfied and normal if (b) is satisfied. If each of G1, ..., G, is further majorized by some among G'j, ..., Gm, then evidently G is majorized by G'1, ..., G,., that is, majorizations can be combined. This combined majorization is admissible and normal if all the constituent majorizations are admissible and normal, respectively. Thus by combining many admissible, normal majorizations, we can reduce a given Feynman graph G into a set of much simpler graphs.
Majorization rules are general ways of majorizing the Feynman graphs which have certain characteristics. Majorization rules should usually give us admissible, normal majorizations, but we sometimes use nonadmissible or abnormal ones. A nonadmissible majorization has to be followed by other nonadmissible ones in such a way that the combined majorization is admissible. Abnormal majorizations are used only in the initial and final stages of majorizations to simplify the descriptions. The following auxiliary majorization rules give us abnormal majorizations (except for G Let G be an arbitrary Feynman graph. (Al) G > G, where G is G with some internal masses reduced in value. THEOREM 207
(A2) G > G' and G' + G, where G' is obtained from G by replacing some parallel lines in G by a single line whose mass equals the sum of theirs. (A3) Gfl > G for any 1 e IGI.
Rules (Al) and (A2) are not necessarily admissible, but it is easy to find in what cases they are admissible.
Proof (Al). Evident. (A2). We use the inverseFeynmanparametric representation. If 1 and l' are parallel lines, Q depends on ,9, and fl, only through r't + Nt' = fl,.. Furthermore, 2
mI
inf C
+H 2 l
(mt + mt )2 (2021)
Thus inf VG = inf VG.. (A3). This is evident from VGIt = VGIa:=o. q.e.d.
Now, we state main majorization rules, which give us normal majorizations. 11*
164
GRAPH THEORY AND FEYNMAN INTEGRALS
THEOREM 208 [29, 30]
to
Let G be an arbitrary Feynman graph belonging
,.
(M1) G > G11 if there exists an internal vertex a such that S[a] = {l, l'} and m, > m1.. (M2) G + {G,, G2} if G is separable at a cutvertex a into Gl = G/IG21 and G2 = G/IG11. (M3) G + {G/l, G  1} for any 1 e JGI, but this majorization is admissible only when G  I is strongly connected and 1 is not a baryon line. (M4) G  {G/l, (G  1)/1'} if there exists an external vertex a such that
S[a] = {l, I'} and if s,, < rn + mi , (see Fig. 20 1). This majori zation is admissible only when (G  1)/i' is strongly connected and 1 is not a baryon line.
Fig. 201
The situation in Rule (M4).
Fig. 202 The situation in Rule (M5).
(M5) G > {G/k, (G  k)1{1, l'}} if there exist two adjacent external
vertices a and b such that S[a] = {l, k} and S[b] = {l', k} and if sa < mi + mk and Sb < tni. + mk . This majorization is admissible only when (G  k)/{l, l'} is strongly connected and k is not a baryon line. Proof: (M1) We observe that QG depends on ai and cx,. only through ai + x,,. (M2). This follows from V. = VG1 + VGZ. (M3). This is a direct consequence of (2019) and (2020) according to Theorem 204. (M4). Since (2019) gives us G/I, we have only to consider (2020). Since 1' is a cutline in G  1, we have VG1 = V(Gq/,' + oc,,(mi.  sa)
(2022)
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
165
Therefore, from (2020) together with s, < m + mi , we find c1VG(PI a)
V(cq/1'(P, a) + (041 + Q1 
mi
(2023)
Since, in G, U depends on al and a1. only through ml + ml., we see
U Z (a1 + al.) aUlaa,
(2024)
01?a1+a1,
(2025)
Hence (2010) implies
Thus the last term of (2023) is positive. (M5). We have only to consider the case ak + 0 as before. Since
VGk = V(Gk)/{l.l'} + a1(m1  s,) + ar(mi  Sb)
(2026)
(2010) implies CkVG(P, LC)  V(Gk)1{1.1'}
(ak + Qk) Mk + LC1(771  Sa) + IXi.(mi
 Sb)
d' (1n2 + ink  Sa) + IX1.(m1 + Mk  Sb) + 2IXkmk2 (2027)
0
with the aid of k o>_ a1 + a1. + a
(2028)
q.e. d. 203
Majorization Procedure
According to Theorem 208, Majorization Rules (M4) and (M5) do not allow external masses to become larger than certain values. Hence we should consider the transition amplitudes rather than the extended ones. For simplicity of description, we assume that any external line is of either a
pion or a nucleon, but we may replace it by a meson or a baryon if its mass squared is not greater than 2 (for a meson) or m2 + 1 (for a baryon). We first consider TO,, for which all external lines are of pious. We assume sa S 2 for any a e g. By means of Rule (Al), we may put !n1 = 1 for all 1 e IG1, that is, we have only to consider the equalmass Feynman graphs. Let .Q.1 be the set of all equalmass Feynman graphs which cannot be majorized by others of by means of Rules (M1)(M5). Any graph G e .Q;, is nonseparable because of Rule (M2) and has no internal vertex of degree 2 because of Rule (M1). We call a vertex a such
166
GRAPH THEORY AND FEYNMAN INTEGRALS
that N(S[a]) >_ 3 a node. [Note that we do not count the number of external lines incident with a.] We also call a path between two distinct nodes which passes through no other nodes an n path. Since G has no internal vertex of degree 2, on each npath there can lie external vertices only. If G has an npath P on which less than three external vertices lie, we can majorize G by means of one of Rules (M3), (M4), and (M5) if this majorization is admissible. It is not admissible only when there is another npath P' such that S  {l, l'} is a cutset for any 1 e P and any 1' a P'. Let H and H'
be two connected subgraphs of G  S. Since the end points of P and P' are nodes, H contains at least two nodes because if P and P' had a common node then it would be a cutvertex, and moreover H includes a circuit which passes through at least two nodes, because they are not cutvertices. This circuit is composed of at least two npaths, on any of which there lie at least three external vertices because otherwise G is majorized in the admissible way by means of one of Rules (M3), (M4), and (M5). Therefore H contains at least six external vertices. The same is true also for H'. Hence n >_ 12.
Thus, for n < 11, on every npath in G there lie at least three external vertices. Hence for 71 < 8 G can have no node, that is, it is a singleloop
graph. For 9 < n < 11, some twonode graphs are possible. Let Q. be the subset of Q consisting of the Feynman graphs having n distinct external
vertices which are also distinct from nodes. Then any graph of Q' .  S2 is majorized by a graph of Q. by repeated use of Rule (A3). Thus we obtain the following theorem. THEOREM 209 [29]t
If s, < 2 for any a e g, then any graph of
is
majorized by the graphs of 52,,, where Q for n < 8 consists of singleloop Feynman graphs having n distinct external vertices alone, and Q9 and 5210 ,
k o0
Fig. 203 (a) & (b) The nonsingleloop graphs of 0, and 010
t The above explicit proof was not presented in the original paper.
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
167
in addition to those graphs, contain the graphs shown in Fig. 203 (a) and (b), respectively.
It can be shown that for n ? 11 the domain in which VG > 0 for all G e W,° is empty even for sa = 1 for all a e g [29]. Next, we consider V. with n/2 z r z 1. By means of Rule (Al), all baryon
and meson internal lines can be replaced by nucleon and pion internal lines, respectively.
In the present case, the majorizations given by Rules (M3), (M4), and (M5) are not admissible when a nucleon line is deleted, because then the baryonnumber conservation law is violated. It turns out to be convenient, however, to use such nonadmissible majorizations sometimes. Furthermore, the majorization rules presented in Theorem 208 are not sufficient for the present purpose, that is, one has to add some other majorization rules. Then, according to Boyling [30], one can explicitly carry out the majorization of W, for n < 6 under the restrictions that sa S 2 for all pion external lines, sb < m2 + 1 for all nucleon ones, and s1, 5 M,, for all other invariant squares. Any graph of P is majorized by the graphs of .Q', where .Q4 and Q5 consist of singleloop graphs alone but S76 contains several topologically
different graphs. For example, Q' and Q4 consist of the graphs shown in Fig. 204 and in Fig. 205, respectively.
\11;1_ Fig. 204 The graphs of 01. Double lines stand for nucleon lines.
Id
d
C
b
Fig. 205 The graphs of Q. Double 1in,s stand for nucleon lines.
168
GRAPH THEORY AND FEYNMAN INTEGRALS
For r >_ 2, some care has to be taken. For example, consider VJ4, the nucleonnucleon scattering. According to the above majorization, every graph G of T4 is majorized by a pair of the graphs shown in Fig. 205. Actually, it turns out, however, that G is majorized by either of those two graphs alone. This latter majorization is called the classified majorization, and correspondingly the former is called the majorization.
Although both majorizations yield the same Symanzik region Do [see (1829)], the classified majorization leads us to better support properties for PTIR. For example, for (1722) we find that y) vanishes unless y > 4 if the unclassified majorization is employed, and that it vanishes unless
y z min [4m2z + 4(1  z), 4z + 4mz(1  z)]
(2029)
if the classified one is employed. Finally, we briefly mention the parity conservation latia% Physically, this
law implies that the strong interactions are invariant under the space reflection. For the present purpose, we can state it as follows: If pion lines (including external lines) only are incident with a vertex of G or of any reduced graph of G, then their number (i.e., the Fdegree of that vertex) has to be even; if there are mesons other than pious, the above law should be applied after replacing each. of them by two or three pious appropriately in the sense of Rules (Al) and (A2), where we assume that the masses of the particles other than pions are not smaller than 2. We note that the circuits formed by baryon lines alone (called baryon loops) can consistently be replaced by the one formed by pion lines in the sense of Rules (Al) and (A2). In this way, we have graphs such that any vertex which does not lie on any baryon path has an even Fdegree. We denote by Pn* the subset of consisting of all such graphs.
It is very difficult to discuss; in general because Rule (M3) usually violates the parity conservation law. The path method is more convenient to discuss Pn (n even).
§21 W FUNCTION INEQUALITIES
In this section, we consider the case n = 4 in more detail. As a result of the majorization procedure presented in Subsection 203, in order to show V > 0 for all graphs of Y'4, we have only to investigate the square graph, namely, the N = 4 singleloop graph, provided that
s, <my +1
(v= a, b, c, d)
(211)
where a, b, c, d are the four external vertices and in,, = 1 or m according as the corresponding external line is of a pion or of a nucleon. Since, as stated in Subsection 132, no anomalous thresholds are present in the square graphs under (211), we have V > 0 for any graph G of Y'4 if s, t, u are not greater than the normal thresholds M. b, M , Mme, respectively. The Symanzik region (212)
is nonempty for s, = inv + 1 (v = a, b, c, d) if d
My+Mc+Ma< Z m,2+4
(213)
v=a
This inequality holds for any of YJ4 (r = 0, 1, 2). Hence for s = in, + 1
(v=a,b,c,d),s=Mb,and t=Mb,we have V=

almI
=a
(C,, + Caa) (my + 1) p
(cab  bad) Mab  // (sac  Cad) Mac
0
(214)
namely, Lr1/772 I

d pp
(Sv + Cad) Mv v=a d
(bat,  cad) Mab + (Cac  Sad) M2 + Z (Cv + sad)
(215)
v=a
In dad [see (1718)], we obtain from (215) a support property of Qst(z, y) of (1722),
y>zMb+(1
Z) M.2
(216)
with the aid of (1723). In (216), however, the last term of the righthand side of (215) is wasted. In order to utilize it fully, we need some inequalitiest t Those inequalities were proposed by Nakanishi [12] and proven by Boyling [31 ]. 169
170
GRAPH THEORY AND FEYNMAN INTEGRALS
among C functions. In this and later sections, we write Wh instead of W(hI g")
for simplicity. Since C,, = Wn f U, we should find inequalities among W functions or W functions (see Subsection 32). THEOREM 211 [31]
The W functions satisfy a quadratic identity
(F'. + Wad) (Wd + Wad)  (Wab  Wad) (Wac  Wad) = UWaad)
(217)
where Waad) denotes the Wad function in GO,), the graph obtained from G by identifying a and d.
Proof From Theorems 316 and 318 together with (345), we have [/fa]d,d = W(ald)=Wa +Wd+ Wab + Wac W(albc)
= Wa + Wad
[//ta]b,d = W(albd)
Wa + Wac
W(alcd)
= Wa + Wab
[/ra]b,c =
[ 1` 'a]C,d =
(218)
It is straightforward to rewrite (218) as Wa + Wad = [//ta]b.C Wd + Wad = [/Va]d,d + [/Ka]b,c  [/Va]b.d  [/Va]c.d
Wab  Wad = [/yb]C.d  [ Wac  Wad = [ /a]b,d

^a]b.c
(219)
[/ya]b,c
Hence K = (Wa + Wad) (Wa + Wad)  (Wab  Wad) (Wac  Wad) [.J/ta]b,c [/ya]d,d
[.Yy'a]b,d [/y'a]c,d
(2110)
Because of Jacobi's theorem for determinants, (2110) becomes[ K = (det /I/'a) [,Yya]bd,ed
=
U[*,ad]b,c = UW;d )
(2111)
because W(ad) = 0. q.e.d.
Since the righthand side of (217) is nonnegative, we obtain an inequality (Wa + Wad) (Wd + Wad) ? (Wan  Wad) (Wac  Wad)
for all a, Z 0 t /mad denotes the matrix //'a in G(ad).
(2112)
171
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
in the Feynmanparametric form. Hence (Wa + Wa + 2Wad)2 _> 4(Wab  Wad) (Wac  Wad)
(2113)
In dad, therefore, we find and likewise
p p C. + pp Sd + 2Sad
2[(Sab  Cad) Sac  Cad)]"
(2114)
Cb + Cc + 2Cad
2(Gab  Sad) (Sac  Ca0]1
(2115)
7p
(pp
Thus (215) together with (1723) yields that ost(z, y) vanishes unless instead of (216). REMARK 212
211 reduces to
1'
> zM b + (1  z) M.2. + 4[z(1
 z)]l
(2116)
If we identify b and c so that JVGb = Wac = 0, Theorem WaWd + WdWdd + WadW. = UWa1cad)
(2117)
Since we now have a Feynman graph with n = 3, we relabel the external vertices as a' = a, b' = d, and c' = b = c. Then (2117) is rewritten as W,'Wb. + Wb.Wc, + Wc,Wa. = UWa
(2118)
where Wa denotes the 3tree product sum with respect to a', b', c'. [A 3tree with respect to a', b', c' is a tree in the graph obtained by identifying
a', b', c'.] It is interesting to note that (2118) presents an exact relation between the numbers of trees, 2trees, and 3trees in an arbitrary graph.
If we confine ourselves to the application to the support properties of PTIR, Theorem 211 is sufficient. The remainder of this section is devoted to the problems of mathematical interest only. First, the inequality (2112) can be made more precise. THEOREM 213 [12, 31] W.W, ? Wab Wac,
WbWe ? Wab Wac
WaWb  Wac Wad,
WeWd 3 WacW.d
W. We > Wad Wab,
Wb Wd
(2119)
Wad Wab
for all at>_0
Proof Of course we have only to prove the first inequality of (2119). In the inverseFeynmanparametric form, we find WaWd  WabWac = (Wa + Wad) (Wd + Wad)  (Wab  Wad) (Wac  Wad)
 Wad(Wa + Wd + Wab + Wac)
= UW(ad) 
WadU(ad)
(2120)
172
GRAPH THEORY AND FEYNMAN INTEGRALS
from (217) and the first formula of (218) together with Cl(ad) =
(2121)
Tf7(aId)
where U(°') denotes the U function in G(°d). For S e S(ad I bc), let H and H' be the two connected subgraphs of G  S,
where H has the two external vertices a and d. According to Definition 314, we can write
U. U.
Y
Wad =
(2122)
SES(adlbc)
(d)
Wad =
Ug Ua, (ad)
(2123)
SsS(a= djbc)
where S(a = d I be) = SG(ad)(ad I be) S(ad I be) (a cutset in G(ad) is not necessarily a cutset in G). Therefore, the righthand side of (2120) is nonnegative if UUAd) > U(ad)U,, (2124)
for any S eS(ad I be). Thus we have only to prove that U d)/UF,
U(ad)/U
(2125)
117
for any subgraph H of G. We may assume H = G  I without loss of generality. Since U(ad)  W(°1d), the quantity (U(ad)/U)p2 is nothing but
agni  V of a selfenergy graph having two external vertices a
Q
and d. Hence Theorem 206 with p2 = 1 implies UG(ad)l

UG_1
U(ad)
U
2
= _1I
U
(2126)
if I is not a cutline; the lefthand side of (2126) equals 0 if 1 is a cutline. Thus (2120) is nonnegative. q.e.d. Example 1
For the square graph shown in Fig. 211(a), we have
Wa = aia2,
Wb = a21X3,
Wab = aia3,
Wac = a2a4,
WW = ala4,
Wd = a3a4
Wad = 0
whence R RWd = aia2Cr3a4 = WahWac 2
W°Wb = ala2a3 ; 0 = WacWad
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
173
Example 2 For Fig. 211(b), Wa = cx1a2(a4 +'l'S + a6) Wd = csa6(x1 + a2 + c3 + a4) + oc3oc4as Wab = aja3(a4 + as + cc6) + a1a4a6 Wac = cc2a4a5
whence
WaWd > Wab Wac
Fig. 211 (a) & (b)
Examples for Conjecture 214.
In the proof of Theorem 213, we have used an algebraic inequality at the last step (2126), but the above examples suggest that any term contained
in WabWac is contained also in W,Wd. Thus we are led to the following conjecture. CONJECTURE 214 [12]t When we write Wa Wd = r
ar
77
r
1EIGI
(2127)
Wab Wac = Y b, JJ cc' J.
1EIGI
where r is a sequence of r1(l e IGI), r1 = 0 or 1 or 2 with
Y rl = 2(lu + 1)
(2128)
IEIGI
t Fairlie [32] proposed a proof of Conjecture 214, but unfortunately it was incomplete.
174
GRAPH THEORY AND FEYNMAN INTEGRALS
and a, and b, are nonnegative integers, we shall always have a, >_ b, for any r
(2129)
We note that a, and b, are not restricted to 0 and 1; for example, in Example 2, a, = 2 for r = (1, 1, 1, 1, 1, 1). Evidently, Conjecture 214 is more precise than Theorem 213. It is an interesting graphtheoretical problem to prove Conjecture 214. By means of mathematical induction with respect to N, Conjecture 214 is reduced to the proposition that a, >_ b, for r = (1, 1, ..., 1)
(2130)
only for all graphs such that N = 2(,u + 1). This is because if there exists 1
such that r, = 0 then the proof of (2129) reduces to that of the similar inequality in G11, and if there exists 1 such that r, = 2 then it reduces to that for G  1 by applying (8f 8x1)2 to (2127). Thus one has only to count the number of the ways of dividing IGI into two 2trees of certain types. The inequality (2130) was explicitly checked in this way for ,u = 1, 2, 3 [121.t Thus Conjecture 214 is true at least for N :!g 9.
t Unfortunately, Table I of Ref. [12] contains some errors. First row, column g : 3 instead of 2; second row, column z : 3 instead of 1; third row, column k : 0 instead of 1.
§22 SUPPORT PROPERTIES IN PARTICULAR MODELS
In this section, we summarize the best known results on the support properties of PTIR for Y'.' and Y'n. 221
Vertex Function
Since in this case we cannot put all external momenta on the mass shells, the euclidean majorization method is better than the noneuclidean one. Since is obtained as a special case in = 1 of Y'3, we consider only the latter, namely, the case of the nucleon vertex function. The euclidean majorization method [33] yields the result that any graph of Y'3 is majorized by the triangle graph shown in Fig. 221 in the euclidean
region of p. From it, one can show [34] that g73(za, zb, zz, x) of (1716) vanishes unless 7G
(Zgzb + ZbZc + ZcZa) (Za 1 + zb 1 + m2Zc 1)
for x
IZa1
(m + 1)2 (Za + Zb)
 zb1I :!g mzc1 < Za1 + Zb1
for inzc 1 > zq 1 + Zb
1
(221)
(m + 1)2 Z. + (m  1)2 zb + 4zc for zb 1 > z,1 + mz 1
x>_(m1)2za+(m+1)2zb+4zc for za1>zb1+mzcl
a Fig. 221
b
The graph of Q. Double lines stand for nucleon lines.
The bounding curve of the first case of (221) is realized in the triangle graph (Fig. 221). Support properties in the other cases, which arise from the contributions from outside the euclidean region is nothing but (1918)
with Ma=Mb=m+ landMM=2. 175
176
222
GRAPH THEORY AND FEYNMAN INTEGRALS
Scattering Amplitude
We consider the support properties of (1722). (a) Equalmassparticle scattering ( 4). From (2116) we have QS=(z, y) = e:u(z, Y) = Cus(Z, Y) = 0
y >_ 4 + 4[z(l  z)]l
unless
(222)
This inequality gives us the exact shape of the support because (222) is realized in the square graph. (b) Nucleonnucleon scattering (P4). As above, the exact shapes of the supports are found to be e,,(z, y) = 0 unless y > 4m2z + 4(1  z) + 4[z(l  z)]' e,, ,(z, y) = 0
unless y
min[4z + 4m2(l  z), 4m2z + 4(1  z)] + 4[z(1  z)]
eeu,(z, y) = 0
unless y
4z + 4m2(1  z) + 4[z(l  z)]i
(223)
As remarked in Subsection 203, the unclassified majorization for e,. yields
only y >_ 4 + 4[z(1  z)]*. We here present two proofs of the support of e w in (223).
Proof No. 1 [17] Since we should consider the case s,, = mz + 1 (v = a, b, c, d), we have only to show that VG 0 for all G e T4 for either
{s = 0,t =4,u=4m2) or{s= 0,1 =4m2,u=4}. We choose the former set of values if two baryon paths Pl and P2 are such that Pl e P(ac) and Pee P(bd); the latter set is chosen if Pl e P(ad) and P2 e P(bc). Because of symmetry, it is sufficient to consider the former case alone. Since s,, = (s + t + u)/4, we can write MIMI _ V = 4(Q(I)s + Q(z)t Qcs1u) Q= (224) I
On substituting s = 0, t = 4, and u = 4m2 in (224) and using the baryonnumber conservation law
El alms >
I
ai +
IePI VP2
a,(m2  1)
(225)
we obtain
V>
ryaIQ(2)QI3J
l
I
E aIQ13))(m21) IePIUPZ
(226)
177
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
The quantity in the square bracket is just the V function for the equalmass Feynman graph, whence it is nonnegative. To prove Y_
Oct > Q(3)
(227)
IEP1uPz
we note that (3)
QG(PlUP2)* _
IEPIUP2
at
(228)
where (P1 u P2)* _ JGJ  PI U P2, provided that Pl and P2 pass through no common vertex [see Rule (A3)]. Since the point {Sa = sb = SC = Sa = 1,
s = t = 0, u = 4} belongs to the euclidean region, Theorem 206 implies (3)
QG(PiuPz)
Q(3). q.e.d.
Proof No. 2 [11] We employ the path method. Let Pl e P(ac) and P2 e P(bd) be two baryon paths as before. Because of Theorem 202 and the existence of Pl and P2, we can find q1(1 a JGD) satisfying the conditions of Theorem 191 if
P. = )nkl  k2, Pb = mkl + k2 (229)
Pc = ink,  k2, Pa = mkl + k2 where ki = k2 = 1 and k1k2 = 0. Since (229) implies that
sv = py = tn2 + 1
(v = a, b, e, d)
s =(Pa+Pb)2=0, t=(Pa+pJ2=4, u=(Pa+Pa)2=4m2 (2210)
we obtain V >_ 0 for (2210). q.e.d.
(c) Pionpion scattering ('4). If one takes account of the parity conservation law, the exact shape of the supports is found to be ns:(z, Y) = &(Z' y) = Qas(z, y) = 0 unless y >_ min [16z + 4(1  z), 4z + 16(1  z)] + 16[z(1  z)]!(2211) instead of (222).
Proof [12] It is convenient to employ the path method. By means of Rules (A 1) and (A2), we have only to consider strongly connected, equalmass Feynman graphs G such that every vertex is of even Fdegree. We assume that all external vertices are distinct; otherwise G is obtained as a 12
Nakanishi
178
GRAPH THEORY AND FEYNMAN INTEGRALS
reduced graph of another. Then, since three or more internal lines are incident with every external vertex, there exist six disjoint paths between external vertices. Topologically, there are only three possible situations A, B, C of those paths shown in Fig. 222. Since G is connected, however, Situation C has to be of the form D, whence it is a special case of Situation A. Since Situation A is not symmetric with respect to the external vertices, there
A
C
B
D
Fig. 222 Possible structures of the pionpion scattering graphs. Each
line indicates a path, and paths may cross at some vertices. Possible unimportant vertices and internal lines are omitted.
are six ways of identifying a, b, c, d with them. For example, for the support property of Qst, we should investigate the three cases A1, A2, B shown in Fig. 223. We can find q1 (1 e IGI) satisfying the conditions of Theorem 191 is shown in Fig. 223 if
Po =  2k1  k2, Pb =  2k1 + k2 Pc = 2k1  k2,
Pd = 2k1 + k2
(2212)
179
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
A,
B
A2
Fig. 223 The way of applying Theorem 191 to the pionpion scattering.
for Al and B and if
Pa = 2k1  k2, Pb = 2k1  k2 Pc _  2k1 + k2,
(2213)
Pa = 2k1 + k2
for A2, where ki = k2 = 1 and k1k2 = 0. Since (2212) and (2213) imply {sv = 5 (v = a, b, c, d), s = 16, t = 4} for A 1 and B and {s = 5 (v = a, b, c, d), s = 4, t = 16} for A2, respectively, we obtain (2211). q.e.d. (d) Pionnucleon scattering (14 and W4). The noneuclidean majorization method yields the following results [17]. For W4 (b and d are of pions), one obtains that Cst(z, y) = 0
unless
y
y) = 0 unless y y) = 0 unless y
(m + 1)2 z + 4(1  z) + ;,.,[z(1  z)]l
4z + (m + 1)2 (1  z) + x[z(1  z)]! (2214) (m + 1)2 + 2[z(1  z)]
with x = 2 = 4. For VJ4, one obtains (2214) with x = 8 and 2 = 10 under some plausible assumptions for certain loworder graphs. Furthermore, if one employs also the euclidean majorization method, one can improve (2214) except for neighborhoods of z = 0, Z, 1. Contrary to the above three cases, one does not yet succeed in proving the exact shapes of the supports. (e) Mandelstam bounding curves. and Y'4, we have obtained the exact shapes of the supports. In For tl'4 the Mandelstam representation (1830), the boundaries of the supports of s') are called the the double spectral functions 6s,(s', t'), 6tu(t', u'), and 12*
180
GRAPH THEORY AND FEYNMAN INTEGRALS
Mandelstam bounding curves. Their equations were conjecturedt on the basis of the unitarity of the Smatrix in the elastic region [36]. We show that the Mandelstam bounding curves can be derived for the equalmass, pion
pion, and nucleonnucleon scatterings if we assume the validity of the Mandelstam representation. Suppose osr(z, y) = 0 unless y >_ c,z + c,(1  z) + c[z(1  z)]l (2215)
As shown in Subsection 182, Ost(z, y) is uniquely related with d,,(s', t') through (1831). To be consistent with (2215), we should have that$
d "(s" t') = 0 unless z(s'  cs) + (1  z) (t'  ct) >_ c[z(1  z)]l (2216)
for any 0 < z < 1. Since the envelope of a family of those straight lines in the (s', t') plane is the right branch of a hyperbola
(s'  cs) (t'  c=) = c2/4
(2217)
we see d,,(s', t') = 0 unless the point (s', t') lies in the upper region bounded by it. For the abovementioned three cases, (2217) yields the conjectured Mandelstam bounding curves precisely. 223
Proof of Dispersion Relations
As an immediate consequence of PTIR, we can prove dispersion relations for t fixed. Suppose that the following support properties are proven already: p,,(z, y) = 0
unless y > csz + c,(1  z)
et (z, Y) = 0 unless y
y) = 0 unless y
c'z + c"(1  z)
(2218) c
z)
Since t is fixed to a real value, we substitute in (1722)
s' = z1[Y  t(1  z)] in e. u' = (1  z)1 (y  tz) in «< s' = (1  2z)1 [y  (ca  t)z] for z <  in Qus u' _ (2z  I )  ' [ y  ( c ,  t) (1  z)] for z > J in
(2219) errs
f Recently, Martin [35] has proven the Mandelstam bounding curve for the pionpion scattering except for 8 < s < 32 in the axiomatic field theory. t Rigorously speaking, we need reasoning similar to the proof of Theorem 183.
PERTURBATIONTHEORETICAL INTEGRAL REPRESENTATIONS
181
d
with co =
mv. Then the dispersion relation v=a
a(s) + fa (s) = J ds'
s's
CS
r
du'
J
(u')
u'u
(2220)
Ct,
follows if
 ct :!_  t _< cs + c + c  CO
(2221)
where a(s) and j9(u') are certain distributions. It is customary to denote the maximum possible value of t by 44max, that is,
dmax = (c3 + c + c  co)/4
(2222)
Then the values of dmax obtained from the results in Subsection 222 are as follows :
j2. = 2 dmax
for equalmass particle sacttering,
= 2 for nucleonnucleon scattering,
dmax = 8 for pionpion scattering, dm ax = m + 42
for pionnucleon scattering.
Finally, we note that PTIR can also be used for the proof of the partialwave dispersion relation [37, 38]. 224
OneParticle Production Amplitude
As mentioned in Subsection 203, t9' reduces to Qs, which consists of the singleloop graphs alone. They do not have the leading singularity on the mass shell. Relevant singularities are those of the reduced graphs R of nR = 2 and nR = 3. Let a be the external vertex which is dealt with asymmetrically. Then we can show [17] that the weight distributions of (1732) with n = 5 vanish unless
(a) 9'50 :y7/3, 7/3 if a nucleon line is incident with a, y >= 2 + /3 if a pion line is incident with a, (c) iW : y > 4 if a pion line is incident with a, (d)Y'S:y>>=4. (b) tJ'S : y
182
GRAPH THEORY AND FEYNMAN INTEGRALS
References (1) G. C. Wick, Phys. Rev. 96, 1124 (1954). (2) R. E. Cutkosky, Phys. Rev. 96, 1135 (1954). (3) Y. Nambu, Phys. Rev. 98, 803 (1955). (4) Y. Nambu, Phys. Rev. 100, 394 (1955). (5) S. Mandelstam, Phys. Rev. 112, 1344 (1958). (6) G. Wanders, Hely. Phys. Acta 30, 417 (1957).
(7) V. Ya. Fainberg, J. E. T. P. 36, 1503 (1959), [translation: Sov. Phys. J. E. T. P. 9, 1066 (1959)].
(8) S. Deser, W. Gilbert, and E. C. G. Sudarshan, Phys. Rev. 115, 731 (1959). (9) M. Ida, Progr. Theoret. Phys. 23, 1151 (1960).
(10) N. Nakanishi, Progr. Theoret. Phys. Suppl. 18, 1 (1961); Errata, Progr. Theoret. Phys. 26, 806 (1961). (11) N. Nakanishi, Progr. Theoret. Phys. 26, 337 (1961); Errata, ibid. 28, 406 (1962). (12) N. Nakanishi, Progr. Theoret. Phys. 26, 927 (1961). (13) N. Nakanishi, J. Math. Phys. 3, 1139 (1962). (14) N. Nakanishi, Phys. Rev. 127, 1380 (1962). (15) N. Nakanishi, J. Math. Phys. 4, 1385 (1963). (16) N. Nakanishi, J. Math. Phys. 5, 1458 (1964). (17) J. B. Boyling, Ann. Phys. 28, 435 (1964). (18) H. Umezawa and S. Karnefuchi, Progr. Theoret. Phys. 6, 543 (1951). (19) G. Kallen, Hely. Phys. Acta 25, 417 (1952). (20) H. Lelunann, Nuovo Cim. 11, 342 (1954). (21) M. GellMann and F. E. Low, Phys. Rev. 95, 1300 (1954). (22) N. Nakanishi, J. Math. Phys. 4, 1539 (1963). (23) A. Z. Patashinskii, J. E. T. P. 43, 1371 (1962), [translation: Sov. Phys. J. E. T. P. 16, 973 (1963)]. (24) J. Bros and V. Glaser, L'enveloppe d'holomorphie de l'union de deux polycercles (1961). (25) K. Symanzik, Progr. Theoret. Phys. 20, 690 (1958). (26) N. Nakanishi, Progr. Theoret. Phys. 21, 135 (1959). (27) H. J. Bremermann, R. Oehme, and J . G. Taylor, Phys. Rev. 109, 2178 (1958). (28) Y. Nambu, Nuovo Cim. 9, 610 (1958). (29) T. T. Wu, Phys. Rev. 123, 678 (1961). (30) J. B. Boyling, Ann. Phys. 25, 249 (1963). (31) J. B. Boyling, J. Math. Phys. 6, 1469 (1965). (32) D. B. Fairlie, Proc. Camb. Phil. Soc. 59, 157 (1963).
(33) A. A. Logunov, I. T. Todorov, and N. A. Chernikov, J. E. T. P. 42, 1285 (1962), [translation: Sov. Phys. J. E. T. P. 15, 891 (1962)]. (34) Liu ICh'en and I. T. Todorov, Doklady 148, 806 (1963), [translation: Sov. Phys. Dokl. 8, 157 (1963)]. (35) A. Martin, Nuovo Cim. 44, 1219 (1966). (36) S. Mandelstam, Phys. Rev. 115, 1752 (1959). (37) K. Yamamoto, Progr. Theoret. Phys. 26, 1014 (1961). (38) N. Nakanishi, Phys. Rev. 126, 1225 (1962).
CHAPTER 5
Miscellaneous Topics In this chapter, we discuss some topics concerning Feynman graphs which are not related to the analyticity of the Feynman integral. First, we count the number of the Feynman graphs having Mvertices which appear under a given interaction Lagrangian. Here all vertices are regarded as distinguishable. It is extremely difficult to count the number of topologically distinct graphs, and we are not concerned with this problem. Second, by making use of the above result, we prove the divergence of the unrenormalized perturbation series for the case of spinless particles alone.
The third topic is a graphtheoretical problem concerning Feynman graphs. We investigate how the topological structure of Feynman graphs is restricted if certain types of cutsets are absent.
Finally, we briefly discuss the asymptotic behavior of the Feynman integral when one of the invariant squares is very large.
183
§ 23
NUMBER OF FEYNMAN GRAPHS
Given an interaction Lagrangian density 2i i(x), the Smatrix is given by S = S/ with (613). Any Smatrix element or (extended) transition amplitude is expressed in a perturbation series. Its Mth order term is a sum of the Feynman integrals corresponding to the Feynman graphs having M vertices, each of which is admissible under 2i.,(x) in the sense explained in Subsection 61. The disconnected Feynman graphs involving vacuum polari
zation parts are removed by the factor , but, for the moment, we include them. Let G be a Feynman graph having M vertices. As mentioned briefly in Subsection 73, we can assign the M positions xl , ..., xM to the M vertices generally in M! different ways; this number M! just cancels the denominator
of the coefficient in (613). For certain graphs, which are invariant under some permutations of vertices, however, such a cancellation is not complete so that some fractional factor remains. To avoid this complication, in this section, we regard all vertices as distinguishable, that is, two topologically identical Feynman graphs with the same particle assignment are regarded as distinct if the names of the corresponding vertices are not identical. Furthermore, for the moment, we assume that 2i ,(x) is linear with respect to each distinct field operator. In an exceptional case, the Smatrix element may admit the presence of external particles passing without interaction, that is, we may have some "pairs of external lines" which are incident with no vertex. The contribution from such a case is trivially factored out, and the Smatrix element reduces to the one having a smaller number of external lines. We exclude this trivial
case in accord with our definition of the Feynman graph stated in Subsection 14.
All fermions and some bosons are distinct from their antiparticles. For such particles,fi the corresponding field operators ?p(x) are different from their conjugate fields yi(x), which are essentially the hermitian conjugates of ip(x). The other bosons are not distinguishable from their own antiparticles, that is, the corresponding fields are hermitian. To indicate this, those boson fields are denoted by q((x) [or qpj(x)] in this section.
by writing field We may specify the essential structure of operators only symbolically. We call Yin,  yry,T the Yukawa coupling, t For brevity, we call them fermions only in this section. 184
185
MISCELLANEOUS TOPICS
Lpint
(i1V2 + V)2 i) 9q with VI + v12 the modified Yukawa coupling, and
Y;,,,
11 qqj the multilinear boson coupling. We count the number of
k
j=1 Feynman graphs by "drawing" lines between M given distinct vertices. We note that we can independently draw the lines of differenttype particles to construct Feynman graphs [1]. We first consider the y'v, partof the Yukawa coupling. All lines corresponding to fermions have an intrinsic orientation in such a way that one outgoing and one ingoing lines are incident with every vertex. Let r be the number of outgoing fermion lines, which is equal to that of ingoing ones. The number
of the ways of choosing r outgoing external vertices is I M). From the r remaining M  r vertices, M  r fermion internal lines are outgoing; their ingoing end points have to be selected from M vertices (then the ingoing external vertices are uniquely determined). The number of the ways of doing this is
(MM
r) (M  r) !. Thus the number of the ways of drawing fermion
lines for the fp part is given by NF(M,
r) =
 r)! _ (r!)2(M!)2 \rM/M\M (M (M  r) r)
(231)
Next, we consider the VIP2 + yi2'Wl part. Let r, and ri (i = 1, 2) be the numbers of the outgoing and ingoing external lines, respectively, of the ?P,
particle. Of course rI + r2 = PI + r2. Each vertex is classified into either of two classes 1 and 2 according to the name of the outgoing particle incident with it. Let M, (i = 1, 2) be the number of Class i; then we have M = MI + M2,
MI = rI + (M2  rI), and M2 = r2 + (MI  r2). Hence M + ri  Pi has to be even and
M, = (M + r,  rJ/2
(i = 1, 2)
The number of the ways of dividing M vertices into two classes is ( Cla ss 1, rI vertices are assigned to the outgoing
I
(232)
). In
external lines, and the v!I
internal lines outgoing from the remaining MI  rI = M2  rI vertices choose their ingoing end points from M2 vertices of Class 2. Thus the number
of the ways of drawing the, lines is
`MI) (
l rI
M2
M2  rI
)
(M2  r1)!
(233)
186
GRAPH THEORY AND FEYNMAN INTEGRALS
We have a similar formula for the 7p2 lines. Thus, the number for the )112 + 1P2'P1 part is given by NF(M, r1 , r2, rl, r2)
M! [(M + rl  r1)/2] ! [(M + r2  r2)l2] ! r1! rl! r2! r2! [(M  1.1  rl)/2]! [(M  r2  r2)/2]!
(234)
The number of the ways of drawing the boson (T) lines is counted as. follows. Let n; be the number of the external lines of the qqJ particles. Since we assume that only one qqJ line is incident with every vertex, the number of the ways of drawing the 99j lines is equal to that of classifying the M  ni vertices into pairs, that is, it is given by
NB(M, n;) = (nj) (M  n;  1) ! !
(235)
where M  nJ has to be even, and (2m  1)!! _ JJ (2k  1). k=1
On combining the above results, we can find the number of Feynman graphs (having M vertices and n external lines) generated by the abovementioned interaction Lagrangians. We have NN(M, r) NB(M, n  2r) for the Yukawa coupling,
(236)
N;(M, r1, r2, r'1, r2) NB(M, n  r1  r2  r1  r2) for the modified Yukawa coupling, k
11 NB (M, nJ] with
J=1
(237)
k
E nJ = n
'J=1
for the multilinear boson coupling.
(238) k
In (238), all qqJ fields are distinct. If they are all identical, fl TJ should J= 1
be replaced by elk!. The number of the graphs in the qI coupling case cannot be obtained from (238), because the graphs which are mutually transmuted by permuting 991, ..., 9)k (totally or partially) are not distinguish
able in the 0 coupling case. Furthermore, in this case, there are certain k
special graphs which cannot be generated by the 11 qqJ coupling.fi Therefore, J=1 j' For example, consider selfenergy graphs with M = 3, N = 5, and k = 4.
187
MISCELLANEOUS TOPICS
we have to content ourselves with obtaining upper and lower bounds on the number of the graphs. f Let Nk(M, n) be the number of the Feynman graphs having M vertices and n external lines in the 99k coupling case. Of course, kM  n = 2N has to be even. To find an upper bound on Nk(M, n), for a moment we suppose
that all internal lines are distinguishable. There are ZM(M + 1) ways of choosing two end points of any internal line and M ways of choosing one end point of any external line. Hence if we neglect the fact that each vertex is of degree k, the number of the ways of drawing all lines is [+M(M + 1)]NMn.
Since the internal lines should not be distinguished, we have Nk(M, n) < M2N+n/N!
(239)
with N = (kM  n)/2. By means of Stirling's formula, the righthand side of (239) asymptotically behaves like MkMI2
(2310)
apart from a factor of order (const)M. Next, we consider a lower bound. In this case, it is preferable to exclude
the Feynman graphs which include vacuum polarization parts. We here consider only the Feynman graphs which have a Hamilton circuit [2], where a Hamilton circuit is a circuit which passes through all vertices. Those graphs
are of course strongly connected, but strongly connected graphs do not k2 necessarily have a Hamilton circuit. We first consider the q9o JJ 99j coupling. J=1
Let Nk(M, n) be the number of the Feynman graphs which have M vertices, n external lines, and a Hamilton circuit composed by the quo lines alone. Then we evidently have k2
NN(M, n) _ z(M  1)!
Z
[J NB(M, ni)
("I.....nmx) !=1
(2311)
k2
where the summation goes over all partitions such that Y_ n f = n. When J=1
we identify all the fields quo, ..., q'k2, we have to take care of the fact that several graphs become indistinguishable. Consider an arbitrary graph G
in the q coupling case. The number of Hamilton circuits in G cannot exceed k(k  1)M2 because every vertex is of degree k or less. Furthert We also note that the Feynman integral involves some extra factors, that is, each of r parallel lines of identical particles induces a factor 1/r!. # This point is not taken into account in Ref. [2].
188
GRAPH THEORY AND FEYNMAN INTEGRALS
more, given a Hamilton circuit C of G, the number of the ways of assigning (pl, ..., 9?k, to all the lines of IGI  C and the external lines of G at random is k2 (k  2)N M+". Therefore, the number of the graphs in the quo fl 4pj coupling j=1
case to which G corresponds cannot exceed k(k  1)M2 (k  2)NM+,, Thus the number Rk(M, n) of the strongly connected Feynman graphs in the q coupling case is bounded by 1Vk(M, n) > Nk(M, n)IkN+111
(2312)
The righthand side of (2312) asymptotically behaves like (2310). Therefore,
the number of the topologically different, strongly connected Feynman graphs generated by the qqk coupling asymptotically behaves like Nk(M, n)IM! '" M(k2)Ml2
(2313)
apart from a factor of order (const)M. We can also find a lower bound of the same order on the number of the strongly connected Feynman graphs which involve no selfenergy parts.
§24 DIVERGENCE OF THE PERTURBATION SERIES
As stated in Section 23, the number of the Feynman graphs having M vertices (in nontrivial theories) increases asymptotically like MM12 or more.
Hence unless the contributions from them cancel with each other very strongly, the perturbation series will diverge for any nonzero value of the coupling constant. In fact, such cancellations do not occur in the theory involving bosons only. In the theory involving fermions, however, it may be possible that such cancellations happen because of the Pauli principle for
fermions. j Indeed, according to Yennie and Gartenhaus [3]$, the unrenormalized perturbation series for the Yukawacoupling theory (namely, .Ti,,,  ippgq where V is a fermion field and 99 is a boson one) is convergent for all values of the coupling constant if the spacetime volume is finite and
if high momenta are cut off, and therefore the transition amplitude is an entire function of the coupling constant. The reason for this result can be explained as follows. Under the above assumptions, the number of the possible quantummechanical states of fermions becomes a finite constant, and therefore the contribution from the fermion modes increases like at most
(const)M. Accordingly, we have only to count the number of the ways of drawing boson lines. Thus the Mth order term of the perturbation series asymptotically behaves like Mml', and the series is absolutely bounded by an exponential function of the coupling constant squared. Thirring and others [5, 6, 7, 8] showed that the ordinary unrenormalized
perturbation series for the theory involving spinless particles alone is divergent for any nonzero value of the coupling constant. We here present a proof of this divergence based on the Feynmanparametric integral [9, 2]. In order to avoid ultraviolet divergence, we have either to cut off high momenta or to employ a twodimensional model. We here employ the first alternative. By using the Feynman cutoff (1029) with AI = 1, the cutoff Feynman integral reads A,2 Oct &2 f dN1 a fG(p, A) = (2M  3)! H I (241) UZ(V I ,I 2E)2M2
min
d
where all mI (1 e I GI) in V are replaced by xI, respectively. If p belongs to the domain Do = {p I V(p, a) > 0 for all a e Al (242) t The Pauli principle forbids two identical fermions to be in the same state. # For related work, see Simon's paper [4]. 189
190
GRAPH THEORY AND FEYNMAN INTEGRALS
then ass * +0 (241) is absolutely convergent, and ff(p, A) is positive definite. Let Vmax be the maximum value of V with x1 = AI (1 C IGI) for p fixed. Then
(Ai  nit)
ff(p, A) >_ (2M  3) !
i
Vm2M
ax
(243)
A
Since Theorem 93 implies
U
AM where A is a certain positive quantity independent of M. For simplicity of description, suppose that Yj., contains only one coupling constant g. Then, according to (2313), the perturbation series for the extended transition amplitude is of the form (247) iM=1 E gM x(M)
where X(M) > 0 for kM  n(>_ 0) even, and x(M) asymptotically increases as M112)M/2 (or more). Thus the perturbation series is divergent for any nonzero value of g.
We remark that the above proof also applies to the socalled parityviolating Yukawa coupling [10]]'
Y;.(x) = g V(x) (1 + iy(") V(x) 4i(x) with
(s) = (0)
(248)
(1) (2) (3)
y(s)y(v> + y(v>y(5) = 0
(v = 0, 1, 2, 3)
(249)
(y(5))2 = 1 because the numerator of the fermion Feyntnan propagator reduces to a constant when sandwiched by 1 + iy(s): (1 + iy(s)) (mi + yki) (1 + iy(s)) = 2nit(1 + iy(5)) (2410) Since the convergence proof of Yennie and Gartenhaus applies also to (248), we see that the assumption of the finite spacetime volume is more essential for the convergence than the Pauli principle is. t Though (248) is not hermitian, this point is irrelevant to our reasoning.
MISCELLANEOUS TOPICS
191
Finally, we briefly mention the renormalized perturbation series. Here we are concerned with mass and couplingconstant renormalizations rather than
the removal of ultraviolet divergence itself. Unrenormalized masses and coupling constants are expressed in power series of renormalized masses and coupling constants. The amounts of the changes from the former to the latter are called renormalization constants. The renormalized perturbation series, which is expressed in terms of renormalized masses and coupling constants,
is formally equal to the unrenormalized perturbation series, which is expressed in terms of unrenormalized masses and coupling constants. Thus the mass renormalization implies certain rearrangement of the perturbation series, and the couplingconstant renormalization implies the change of expansion parameters. As stated in Subsection 102, all Feynman integrals become free from ultraviolet divergence without cutoff in the renormalizable field theories.
For definiteness, we consider the case Pint ^ gq with g t 0 real. Since x(M) = 0 for kM  n odd, all terms of unrenormalized perturbation series have the same sign unless g < 0 and k is even. In that case, therefore, at least
one of renormalization constants should be divergent if we hope that the renormalized perturbation series would be convergent. This divergence is independent of ultraviolet divergence because we are considering Feynmancutoff quantities. This result may also be understood in the following way. It can be shown [1 I] that if all renormalization constants are assumed to be finite, the ggik theory has no ground state (zeroenergy state) except for the case in which g < 0 and k is even (only in this case the expectation value of the Hamiltonian is positive definite). As seen from (1715), however, the perturbation series for the selfenergy has no imaginary part for s < 0. This fact physically implies that the theory has a ground state if the renormalized perturbation series is meaningful. Thus, to be consistent, except in the case g < 0 and k even, either at least one of renormalization constants is divergent or the renormalized perturbation series is divergent. It is extremely difficult to show convergence or divergence of the completely renormalized series, but it is easy to discuss the series for which mass renormalization alone is carried out. We have only to exclude all Feynman graphs involving selfenergy parts and to replace the propagator (615) by
i mi  k  iei + 1
Q(s')
s'  ki  isi
(2411)
192
GRAPH THEORY AND FEYNMAN INTEGRALS
where mi denotes the renormalized mass and we know from the axiomatic field theory that ci > mi and e(s') z 0. By taking the first term of (2411) as a lower bound, we can still make the same proof of the divergence of the series. For k = 3, all Feynman integrals become finite after mass renormalization, whence the massrenormalized perturbation series can be considered without cutoff, and it is divergent.
§ 25
CUTSET PROPERTIES OF FEYNMAN GRAPHS
The Feynman integrals corresponding to certain Feynman graphs G do not contribute to some or any weight distributions (except for the end points of the z variables) of PTIR for the transition amplitude (see Section 17) and to some or any double spectral functions of the Mandelstam representation for the scattering amplitude (see Subsection 182). The presence or absence of the contributions is determined by some topological properties of G, namely, by the existence or nonexistence of cutsets of certain types. Thus we are led to a graphtheoretical problem. Let g = {a, b, c, ...} be the set of all external vertices of G as before. They are not necessarily distinct because they are supposed to be in onetoone
correspondence to the external lines of G as in Chapter 4. A cutset S e S(h I g  h) is called of mass type if either h or g  h consists of only one
external vertex, and it is called of energy type if both h and g  h contain at least two external vertices. Of course, there exist cutsets of energy type only when n z 4. The structure of a connected Feynman graph G is much restricted if some sets S(h I g  h) of energytype cutsets are empty. We call this property the cutset property of Feynman graphs. In the following, we consider the case n = 4 only. The sets of energytype cutsets are S(ab I cd), S(ac I bd), and S(ad I bc). The number of nonempty ones among them is denoted by a ; a = 0, 1, 2, 3. We can prove the following theorems, which appear to be rather selfevident, but their proofs are not necessarily trivial. Ti3EoIEM 251 empty.
If a > 2 then all four sets of masstype cutsets are non
THEOREM 252 A necessary and sufficient condition for a = 0 is the existence of a vertex e such that all paths j between any pair of external vertices pass through it. This vertex e is a cutvertex if all or three external vertices are distinct. THEOREM 253
A necessary and sufficient condition for a < 1, say
S(ac I bd) = S(ad I bc) = 0 for definiteness, is the existence of at least one vertex e such that any path between a or b and c or d passes through it. This vertex e is a cutvertex if a, b, c, d are all distinct. f If P E P(ab) and if a and b are not distinct, then we understand that P = 0 passes through a = b. The same applies also to Theorem 253. 13
Naksniahi
193
194
GRAPH THEORY AND FEYNMAN INTEGRALS
We outline the proofs of the above theorems. We can prove Theorem 251 by noticing that if S e S(ab I cd) and S' a S(ac I bd) then there exists a cutset
S e S(v I g  v) for any v = a, b, c, d such that S e S u S'. For Theorems 252 and 253, only the proofs of necessity are nontrivial. The graph G under consideration will in general consist of several maximal nonseparable subgraphs. Any two of them have at most one common vertex, which is a
cutvertex. We can successively contract nonseparable subgraphs which
have only one cutvertex and at most one external noncutvertex of G without destroying any cutset of energy type. Let R be the reduced graph of G which cannot further be contracted in this way. For a = 0, we can show that R is nonseparable and that three external vertices are not distinct in R. Theorem 252 follows from this result. It is sufficient to prove Theorem 253 in the case in which a, b, c, d are all distinct in R and Slab I cd) + 0. Then, R has at least one cutvertex e such that any paths between a or b and c or d passes through e. The existence of such a cutvertex follows from Menger's theorem [12], which states: In a nonseparable graph, for any two distinct vertices a' and c', there exist two disjoint paths between
a' and c' which pass through no common vertices except for a' and c'. If R is nonseparable, we construct a graph R' from R by adding two vertices a' and c' and four lines which link a' and a, a' and b, c' and c, and c' and d.
Then by applying Menger's theorem to R', we see that R has a cutset belonging to S(ac I bd) or S(ad I bc). This contradicts a < 1. For a 0. Thus it can vanish only when some of at (1 e IGI) vanish.
Let Pa, and Pbd be the path between a and c and that between b and d included in C, respectively. Any path between a vertex on Pac and that on Pb,, is called a t path (see Fig. 261). The set of all tpaths is denoted by Ps,
and its subset consisting of all the tpaths of the minimum "length" is denoted by P=, that is, Pt = {P I P e Pt, N(P) < N(P) for any P e Pt}
(264)
c t
L
b
S
Fig. 261 An example of a tpath. Notched lines form a tpath.
THEOREM 261
In a planar Feynman graph G having four external
vertices a, b, d, c as specified above, a necessary and sufficient condition for
Cac = 0 but U r 0 is that al = 0 for all 1 e P, where P is a tpath.
Proof If a(P) = 0 (i.e., a, = 0 for all 1 e P) then V reduces to VG/P because of (92). Since SG,P(ac I bd) = 0, we see Wac I acP>o = 0, but of course U I a(P)=o * 0. Conversely, Ca, = 0 implies Wac = 0, whence a!(1) = 0 for some I c IGI. Since U r 0, I includes no circuit. If there exists S e SG(ac I bd) such that r W =
W(hi9b)
198
GRAPH THEORY AND FEYNMAN INTEGRALS
I n S = 0, then the contribution, Ws, from S to W,,, is nonvanishing by
definition. Hence for any S e SG(ac I bd) we have I n S * 0, that is, SG/,(ac I bd) = 0 on account of Theorem 22 (d). Since of course SG,I(ad I be = 0, G/I has a < 1. According to Theorem 253, therefore, in GJI there is a vertex e such that any path between a and c and any path between b and d pass through it. Thus I has to include a tpath in G. q.e.d.
We decompose the integration region of (263) into 2N subregions by dividing the positive real axis of each ai into two intervals [0, 77] and (q, oo), where 77 > 0. Each subregion is characterized by the set, I, of all lines I such that 0 < a, < 77. Our task is to find the set I corresponding to the subregion which yields the leading asymptotic behavior of (263) as It!  oo.
Let 6 = Y, al. If I contains no circuit, Theorem 261 implies that as fez
d + 0, S,c behaves like dk, where k is the maximum number of disjoint
tpaths included in I. If I includes a circuit, let I' be a union of disjoint tpaths of the maximum number k included in I such that I' n P,, = I' n Pbd = 0 (this condition is independent of determining Ic). Then I' includes no circuit because every vertex is of Fdegree 3. Now, as d + 0, C,, behaves like
at best dkI because U = 0(d). On the other hand, as d' * 0, where y ai, C., behaves like d'k. Thus it is sufficient to consider only the
d'
IEI'
case in which I is a union of disjoint tpaths: k
I= U P;, P,ci' J=1
P;nP;=0 for i+j
(265)
and I includes no circuit. Then the asymptotic contribution from the subregion corresponding to I is, apart from a possible logarithmic factor, 77
('
(5'
' dd
(dkt + C)'" +1
.v t_m/k,
(C + 0)
(266)
0
where m  N(I). According to (266), the power of III is equal to the average of the length N(PI) (j = 1, ..., k). Hence in order to minimize it, we should have Pj e Pt for all j.
199
MISCELLANEOUS TOPICS
The power of log t then can be calculated as follows. Let S; _ E ai LEP.j
(j = 1, ..., k), and r = N(PJ) (independent of j). By induction, we can show
that J =1
8i 1 d8J 1L
18J ) t+ \ J=1
Cl
t
(log t)1 1
(267)
Thus we obtain the following theorem. THEOREM 262 [14]
The Feynman integral fG(s, t) with s fixed asympto
tically behaves like
const t (log t)kI
(268)
as Itl + oo, where r = N(P) (P e P), and k denotes the maximum number of disjoint tpaths of P, (their union can include no circuit). Though Theorem 262 is a general rule, the constant coefficient in (268) is sometimes divergent. This situation can happen when G involves a selfenergy part or a vertex part of certain special types, called the "singular configurations" [14]. We omit details. For nonplanar Feynman graphs, the coefficient of t in V can vanish for none of o being zero. Hence the investigation of the asymptotic behavior for nonplanar graphs is more involved [13, 16]. References
(1) R. J. Riddell, Jr., Phys. Rev. 91, 1243 (1953). (2) A. Jaffe, Commun. Math. Phys. 1, 127 (1965). (3) D. R. Yennie, and S. Gartenhaus, Nuovo Cim. 9, 59 (1958). (4) B. Simon, Nuovo Cim. 59 A, 199 (1969). (5) C. A. Hurst, Proc. Roy. Soc. A 214, 44 (1952). (6) W. E. Thirring, Hely. Phys. Acta 26, 33 (1953). (7) M. A. Peterman, Hely. Phys. Acta 26, 291 (1953). (8) R. Utiyama and T. Imamura, Progr. Theoret. Phys. 9, 431 (1953). (9) N. Nakanishi, Progr. Theoret. Phys. 17, 401 (1957). (10) K. Yokoyama, Progr. Theoret. Phys. 39, 830 (1968). (11) G. Baym, Phys. Rev. 117, 886 (1960). (12) C. Berge, Theoree des Graphes et Ses Applications, 2nd ed. (Dunod, Paris, 1963), p. 192. (13) G. Tiktopoulos, Phys. Rev. 131, 2373 (1963). (14) G. Tiktopoulos, Phys. Rev. 131, 480 (1963). (15) M. Martinis, J. Math. Phys. 6, 136 (1965). (16) C. S. Lam, Nuovo Cim. 62 A, 97 (1969).
Appendix A Some Analytical Concepts In this appendix, we explain without rigor some mathematical concepts related to distributions and analytic functions. Readers who are familiar with those subjects need not read this appendix. A1
Distributions
The distribution is a generalized concept of the function. Dirac's Sfunction, its derivatives, and Cauchy's principal value are examples of distributions. According to Schwartz [I], a distribution is a linear, continuous functional
q.[f] over test functions f. Test functions are functions such that they are infinitely many times differentiable and such that the asymptotic behaviors of them and their derivatives are specified. The continuity of q.[f] means that if fk and any derivative of fk uniformly tend to f and the corresponding derivative of f, respectively, then gq[fk] tends to qq[f].
Practically, a distribution is a generalized function such that a product of it and any test function is integrable through integrations by parts. Let f(x) be any test function of n real variables x1, ..., x". If 91[.f] = f 91(x) f(x) d"x
(A1)
is well defined, where the integration range in (A1) is the ndimensional euclidean space R" (or its subregion), then qq(x) can be regarded as a distribu
tion practically. For example, the kth derivative of 6function is defined by
5 a(k)(X) f(x) dx = ( l)kf (k)(0)
(A2)
The support of a function f(x) is the complement of the maximum open set in which f(x) = 0. The support of a distribution qq(x) is the complement of the maximum open set D such that (A1) vanishes for any f(x) whose support is included in D. We denote the support of q(x) by supp 99. [Distinguish it from sup, which means supremum.] 200
SOME ANALYTICAL CONCEPTS
201
Let {0 , 02, ...} bean open covering of R°, that is, let O j (for any j) be an open set and U O; = R". Then an arbitrary distribution q((x) can be decomposed into q,(x) _
Tj(x),
supp 'Tj c 0,
(A3)
J
Thus we can speak about qq(x) at a point x by considering its neighborhood, through the "value" of qq(x) at x is mathematically meaningless.
As for the asymptotic behavior, in most cases, test functions are restricted to those which, together with all their derivatives, tend to zero as Ixi + oo faster than I xI _" for any v > 0. In this case, the corresponding distribution is called tempered. A tempered distribution has its fourier transform [2]. Any distribution which has a compact (i.e. bounded and closed) support can be represented as a certain (finiteorder) derivative of a continuous function [3]. [This property can be extended to a tempered distribution.] Thus the distribution is a natural extension of the continuous function. Let a distribution cp(x) be a certain derivative of a continuous function ?p(x). We call a point x = a a singularity of qq(x) if and only if ?p(x) is not infinitely many times differentiable at x = a. Though the choice of V(x) is not unique, this definition is not ambiguous. Dirac's 6function 6(x) and Cauchy's principal value P. 1/x have a singularity only at x = 0. A product of two distributions ggi(x) and gg2(x) can be defined as a dis
tribution if both of them do not have a singularity at the same point. Though q l(x) q2(x) can be defined also in some other cases (e.g. Ixi 6(x) = 0), products of three distributions are not in general associative if two of them have a singularity at the same point. For example,
0 = [6(x) x] P 1/x r 6(x) [x P. 1/x] = 6(x)
(A4)
Let f(x, y) be a real analytic function of a variable x containing parameters yl, ..., yk. For general values of y, all real roots x = al(y), ..., ak(y) of the equation f(x, y) = 0 will be simple, that is, they will not coincide
with any root of f'(x, y) = 0, where f'(x, y) _ (8/8x) f(x, y). Since the support of 6[f(x, y)] consists of points x = MAY), ..., ak(y), we have k
6Lf(x, y)] = Y_ I.f'(a;(y), y)h
8[x  a.i(y)l
(A5)
with the aid of S(ax) = Ia11 6(x) for a * 0, real
(A6)
202
GRAPH THEORY AND FEYNMAN INTEGRALS
After integrating S[f(x, y)] (multiplied by a test function of x) over x, therefore, we have a function F(y) defined except for the values which yield
nonsimple real roots of f(x, y) = 0. For those values, F(y) will in general be divergent. Let qq(z) be a distribution of one dimension whose singularities
are discrete. As done in the above for S(z), we can define gq[f(x, y)]. For example, we have P.
1
=
i
(P
1
 P.
(A7)
1
xy x+y/ x2  y2 2y Next, we consider a finite part of a singular function. If f(x) is a test function which vanishes asymptotically faster than [x1_' for any v > 0, the integral
I(A) = f x"O(x) f(x) dx
(A8)
exists for Re A > 1, where 0(x) = 1 for x > 0 and 0(x) = 0 for x < 0. Hadamard's finite part [4] is a generalization of this integral to all values of A not equal to negative integers. It is defined by
f Pf xz0(x) f(x) dx  Jim ,J
+
f xlf(x) dx
J
k
j=o
f j!
(J)(O)
sz+j+l
A+ j + 1
e
(A9)
where Re A + k + 2 > 0. Then it coincides with the analytic continuation of I(A), and therefore it satisfies (d/dx) Pf x20(x) = A Pf xA10(x)
(A10)
for A + 1, 2, .... To extend Hadamard's finite part to negative integer values of A, we have to consider some modifications. We can define Pf x' for 1 = 1, 2, ... by adding Pf x 'O(x) and Pf x 10(x) formally, because then the troublemaking term so/0 in (A9) cancels out. It satisfies
(d/dx) Pf x' = (1) Pf x`1
(A11)
We note that Pf xt is a generalization of Cauchy's principal value because
Pf x 1 = P. (1/x)
(A12)
Another way of avoiding difficulty at A = 1 is to consider
v + 0, 1, 2, ... = 5°'°(x) for v = m = 0, 1, 2, ...
Y,,(x) = [r(v)]1 Pf x""1 0(x)
for
(A13)
203
SOME ANALYTICAL CONCEPTS
Then as seen from the definition (A9), the integral
f Y,(x) p(x) dx
(A14)
becomes an entire function of v. Distributions of the following type are also important: lim 1 /(x ± is),
(A15)
e+0
where the limit should be taken after the integration over x is carried out. The distributions (A15) are related to Hadamard's finite part as follows:
lim 1/(x ± is)` = Pf xt { i7L[(l 
1)!]_1
6('1)(x)
e+0
for I = 1, 2, ...
Im Jim 1/(x ± is), e+0
(A16)
Y_x+1 (x)
(A17)
If a test function f(x) is an analytic function, then the integral
f dx f(x)/(x ± is)'
(A18)
can be regarded as a contour integral (see the next subsection) along real axis but avoiding the origin by using a small semicircle. In general, distributions can be defined also as boundary values of analytic functions [5].
A2 Analytic Functions A function f(z) of a complex variable z is holomorphic at z = a if the derivative df/dz at z = a exists independently of the direction of the increment dz. If f(z) is holomorphic at every point in a domain D, f(z) is said to be holomorphic in D. Let .P be a (Jordan) curve entirely lying in a domain D. Then one can define a contour integral
f f(C) dd
r
(A19)
in a way quite analogous to the definition of the curvilinear integral. If D is simplyconnected, that is, if every closed curve lying in D can be continuously shrunk into a point of D, then (A19) is dependent only on the
two end points of l but not on the shape of r, that is, (A19) vanishes if
204
GRAPH THEORY AND FEYNMAN INTEGRALS
T is closed (Cauchy's theorem). An important theorem for a holomorphic function is Cauchy's formula, which reads
f(.,)(z) = m 2zi
r
.f(') ni}1 (c  Z)
dC
(m = 0, 1, 2, ...)
(A20)
= z counterclockwise. From (A20) together with the expansion of (C  z)1 in powers of z  a for a e D, f(z) is expanded into a Taylor series where .V is any closed curve in D surrounding once the point
M
.f(z) _
f 1
(B5)
+1
ai,i+1
respectively. For la,,i+,1 S 1 (i = 1, ..., n), we set a'i,i+1 = cos 0i,i+1
(0 < Bt,t+1 < z)
(B6)
The location of the leading singularity is determined by det (vtj i, j = 1, ..., n) = 0
(B7)
I
When we solve (B7) with respect to a particular invariant square siJ, we have two solutions. If they are real, the larger one gives the location of the leading anomalous threshold when it is present. For n > 3, the Feynman integral corresponding to G is fG(s) = (n  3) !
J
do1 a(UV  is)e}2
(B8)
A
For n = 2, the selfenergy Feynman integral is ultravioletdivergent. Its explicit formula is
IG(s) = 14
Nakanishi
R+
v+R4
2s log
v  R
miin2 2s
in1
log m22
+ const
(B9)
GRAPH THEORY AND FEYNMAN INTEGRALS
210
with s = s12, v
v12 i and R = 2(mi, m?, s), where
x2 + y2 + z2  2(xy + yz + zx)
.Z(x, y, z)
(B10)
The constant in (B9) is infinite, and should be made finite by renormalization. The imaginary part is free from divergence: 1 Im fG(s) = (R'/s) 6(s  (In1 + ln2)2)
(B11)
B2 Vertex Function
For n = 3, it is convenient to write si j as sk and vi j as v, where (i, j, k) is a permutation of (1, 2, 3). We define the following functions:
0 = det(vij I i,j = 1,2,3) 3
2 Z mk Sk + k=1
mfmk(St Y15j_ (si + s2)2, namely, the condition for the s3 physical region, in which the Cutkosky rule for the discontinuity formula (see Subsection 142) is valid; we can also show that L3 > 0 there.
If
ml + m2 < si + s2
(B21)
we have to take account of the A < 0 case of (B20) (the analytic continuation of theA > 0 case), which exhibits the correct normal threshold behavior at s3 = (ml + m2)2 if and only if L3 > 0 at s3 = (m1 + rn2)2; in this case, tan' takes its principal value. 14*
212
GRAPH THEORY AND FEYNMAN INTEGRALS
If L3 < 0 at s3 = (m1 + m2)2, we have an anomalous threshold s3 = s3 in (m1  m2)2 < s3 < (m1 + m2)2, that is, it appears if and only if 623 + 631 < 0,
i.e.,
023 + 031 > Z
(B22)
Its location s3 is the larger root of 0 = 0, i.e., 012 + 023 + 031 =27t
(B23)
From (B22) together with 10231 S 1 and 16311 < 1, it is easy to show that
(B21) is always satisfied. Since L3 > 0 at s3 = (SI + s2)2 and L3 < 0 at s3 = (m1 + m2)2, L3 vanishes once between those two points. Hence tan' in (B20) passes through oz/2 and it becomes 7z at s3 = (in1 + m2)2. The imaginary part of fG(s) for s3 < (m1 + m2)2 due to the anomalous threshold is 7c 1[Im.fG(S)]anomalous = 27t(A)+0(53  s3) 0[(m1 + m2)2  S3]
(B24)
so that the total {imaginary part IM MS) _ [Tm fG(S)]uormal + [Im fGIS)]anomalous
(B25)
is still continuous at s3 = (m1 + m2)2. We note that ImfG(s) > 0 for s3 > s3.
If 53 < (m1  m2)2, then the vertex function has a double spectral representation [4, 5, 6, 7,]: fG(S) =
dS1
612(51 S2 S3)
CIS2
(B26)
(s1  SO (si  S2) 0
0
where 612(S1, S2, S3) _
"0(') 0(51  (m2 + M3 )2) 0(52  (m3 + tn1)2)
A
for S3 < 0
(B27)t
and there is an additional term due to the secondtype singularity for s3 > 0.
For (m1  m2)2 < s3 < (in1 + ln2)2, we have complex singularities in (s1 , S2) space [8, 3].
Nambu's triple spectral representation is valid only when m1 = m2 M3 = 0. The weight distribution then is given by [8]
7L1(A)} 0(A)
0(sl) 0(52) 0(53)
t One may omit either of the last two 0 functions, owing to 0(0).
(B28)
213
FEYNMAN INTEGRALS FOR SINGLELOOP GRAPHS
The twovariable PTIR [cf. (1920)] is as follows [8]:
fG(s) = J dz 5 dy

o
'T 12(Z, Y)
yZS1 (1 z)s2
(B29)
where ns3
991 .(Z' Y) _
PD1
B1N  ( + In3)2) B(q) +fln+m3DAB(n)
(B30)
2(iD4
with =y Z(1 Z)S3
(1 z)ini+zm2z(1 z)s3 D=
(B31)
27, m3)
If s3 (ml + m2)2 then 0, whence the first term alone survives in (B30). If s3 > (MI + m2)2 then the support of g912 is not bounded below in y. B3
Scattering Amplitude
Throughout this subsection, we assume ja i.t+11 < 1 (i = 1, 2, 3, 4), and write S13  s and s24 = t. The conditions for the physical region in which the scattering in the s channel can take place are S
77
Max [(S12 + 523)2, (S34 + 541)2]
1cos 61 < 1
(B32)
where 0 stands for the scattering angle in the centerofmass system; more explicitly
cos 0 = Lj(A2A4)!
(B33)
where LS  s2  s(s12 + s23
+ S34 + S41  2t)  (s12  S23) (S34  S41)
A2 = 2(512, 523, S)
A4 = x'(534, 541, S)
The explicit formula for fG(s, t) is obtainable but complicated [9].
(B34)
214
GRAPH THEORY AND FEYNMAN INTEGRALS
The conditions for the presence of anomalous thresholds are as follows [10]:
(a) Leading 27L < 012 + 023 + 034 + 041 < 2m + 2 min (012, 023, 034, 041)
(B35)
(b) Nonleading (s channel)
012 + 023 > 7t or 034 + 041 > 7s (B36)
(t channel) 023 + 034 > z
or 041 + 012 > z
The imaginary part due to the normal threshold in the s channel is [1 1]]'
7L 1[IiLi.fG(S, t)]normnl =
log 024 + (R13fl+ 6(S  (m1 + m3)2)
1
W}
024  (R13
(B37)
where
Vf  det(vij I i,j = 1,2,3,4) V11 V12 V13
024
_ I ayf/at
V31 V32 V33 V41 V42 V43
R13 = 2(ml, m3 S) _
a2/at2
Because of Jacobi's theorem, we have an identity o24 R13yf = 022044

(B38)
(B39)
where O is the (i, i) cofactor of W, namely, the 0 of G/i, and /,i > 0 (i = 2, 4) for s > (m1 + m3)2.
The imaginary part due to an anomalous threshold can be obtained by deforming the branch cut of the normal threshold according to the movement of the anomalous threshold [12, 13]. The Mandelstam representation W
f.(s, t) =
f 0
f
ds'
dt'
6."(S, t
(s'  s) (t'  t)
0
t For 'P S 0, we should analytically continue the righthand side as in (B20).
(B40)
FEYNMAN INTEGRALS FOR SINGLELOOP GRAPHS
215
is possible if and only if [14] 012 + 023 +034+041 S 27r
(B41)
If there is no anomalous threshold, the double spectral function is given by [11, 15]
aat(s, t) = 2W 40(x) 0(s  (ml + m3)2) 0(t  (m2 + m4)2) B4
(B42)t
Production Amplitudes
For n = 5, the leading singularity is a simple pole [16]. The explicit formula for fG can be written as [17, 18]$
fG/1

fG15
1
V11
...
V15
1
V21
...
V25
1
V51
...
V55
fG
0
(B43)
where fGl j is a Feynman integral for n = 4. The analyticity of fG is unfavorable to multiple spectral representations [19, 20]. For n 6, there is no leading singularity [21, 22]. If we replace any row
of a matrix (v;; I i, j = 0, 1, ..., n) with voo as 0 and vol  vio  1 (i then the resultant matrix is of = 1, ..., n) by (fG,  fGl1, ..., rank 6 [18, 23, 24]. References
(1) G. KAll6n, and A. Wightman, Mat. Fys. Skr. Dan. Vid. Selsk, 1, No. 6 (1958). (2) K. Yamamoto, Progr. Theoret. Phys. 25, 361 (1961). (3) P. V. Landshoff, and S. B. Treiman, Plays. Rev. 127, 649 (1962). (4) C. Fronsdal and R. E. Norton, J. Math. Phys. 5, 100 (1964). (5) K. Yamamoto, Progr. Theoret. Phys. 25, 720 (1961). (6) C. Fronsdal, K. T. Mahanthappa, and R. E. Norton, Phys. Rev. 127, 1847 (1962). (7) Yu. A. Simonov, J. E. T. P. 43, 2263 (1962), [translation: Sov. Phys. J. E. T. P. 16, 1599 (1963).
(8) N. Nakanishi, Progr. Theoret. Phys. 24, 1275 (1960). (9) A. C. T. Wu, Mat. Fys. Medd. Dan. Vid. Selsk. 33, No. 3 (1961).
t One may omit either of the last two 0 functions, owing to 0(W). In Ref. [17], a numerical factor is erroneous.
216
GRAPH THEORY AND FEYNMAN INTEGRALS
(10) R. Karplus, C. M. Sommerfield, and E. H. Wichmann, Phys. Rev. 114, 376 (1959). (11) S. Mandelsatm, Phys. Rev. 115, 1741 (1959). (12) S. Mandelstam, Phys. Rev. Letters 4, 84 (1960). (13) J. Otokozawa, Progr. Theoret. Phys. 25, 277 (1961). (14) C. Fronsdal, R. E. Norton, and K. T. Mahanthappa, J. Math. Phys. 4, 859 (1963). (15) T. W. B. Kibble, Phys. Rev. 117, 1159 (1960). (16) R. E. Cutkosky, J. Math. Phys. 1, 429 (1960). (17) F. R. Halpern, Phys. Rev. Letters 10, 310 (1963). (18) D. B. Melrose, Nuovo Cim. 40, 181 (1965). (19) L. F. Cook, Jr. and J. Tarski, J. Math. Phys. 3, 1 (1962). (20) O. I. Zav'yalov and V. P. Pavlov, J. E. T. P. 44, 1500 (1963), [translation: Sov. Phys. J. E. T. P. 17, 1009 (1963)]. (21) L. M. Brown, Nuovo Cim. 22, 178 (1961). (22) V. E. Asribekov, J. E. T. P. 43, 1826 (1962), [translation: Sov. Phys. J. E. T. P. 16, 1289 (1963)]. (23) G. K&116n, and J. S. Toll, J. Math. Phys. 6, 299 (1965). (24) B. Petersson, J. Math. Phys. 6, 1955 (1965).
Author Index (f indicates a footnote, and * stands for a page of References.)
Andersson 128* Appelquist 87, 89* Araki 128*, 156f
Gale 40, 43* Gartenhaus 199, 190, 199* GeIIMann 182* Gilbert 130, 182*
Asribekov 216*
Glaser
Baym 199* Berge 43*, 199* Bjorken
182*
Hall 92, 128* Halpern 216*
89*, 90, 128*
Bochner 207*
Harary 43* Hartogs 205
Bogoliubov 85, 89* Boyling 128*, 130,'167,169f, 182*
Hepp 85, 89* Hurst 199* Hwa 128*, 207*
Bremmermann 182* Bros 182*
Brown 216* Chernikov 182* Chisholm 44, 88*
ICh'en 182* Ida 130, 182*
Chow 43*
Imamura 199*
Coleman 128*
Cook 216*
Jaffe 199*
Cutkosky 90, 122f, 128*, 129, 182*, 216*
Jost 88*
Deser 130, 182* Drell 89*
Kahane 89*
Drummond 128* Dyson 81, 89*
Kamefuchi 182* Karplus 89*, 90, 127*, 128*, 216*
Eden 90, 127*, 128*, 207*
Kibble 216*
Fainberg 130,182* Fairlie 122f, 128*, 173f, 182* Feynman 1, 43*, 44, 88*
Kleitman 43*
Ki;116n
128*, 182*, 215*, 216*
Kinoshita 45, 88*, 128* Kirchhoff 1
Kroll 89* Kuratowski 32, 43*
Ford 38, 43* Fotiadi 90, 128* Froissart 90, 128*
Lam 87, 89*, 199* Landau 90, 127*
Fronsdal 215*,216* Fulkerson 38, 43*
Landshoff 122f, 128*, 207*, 215* 217
218
AUTHOR INDEX
Lascoux 90,128* Lebrun 89*
Lehmann 182* Logunov
182*
Low 182*
Schwartz 207* Schweber 88* Seshu 43* Shimamoto 45, 88* Simon 189f, 1.99*
Simonov 215* Mahanthappa 215*, 216* Mandelstam 129, 182*, 216* Martin (A.) 180f, 182* Martin (W. T.) 207* Martinis 199* Mathews 128* Melrose 216*
Sommerfield 90, 127*, 128*, 216* Speer 30f, 43*, 85, 89* Sudarshan 130, 182* Symanzik 44, 88*, 182*
Minami 128*
Teplitz 128*, 207* Thirring 189, 199*
Nakanishi 43f, 43*, 44, 45, 88*, 89*, 90, 127*, 128*, 130, 169f, 182*, 199*,
Tiktopoulos
215*
Nambu 44, 88*, 90, 127*, 129, 182* Norton 128*, 215*, 216* Nuttall 122f, 128*
Oehme 90, 127*, 182* Olive
Tarski
216*
Taylor
182*
Todorov
195f, 199*
182*
Toll 216*
Treiman 215* Turnbull 43* Umezawa 88*, 182* Utiyama 199*
128*, 207*
Ore 43*
Vladimirov
207*
Otokozawa 216*
Parasiuk 85, 89* Patashinskii 182*
Pavlov 216* Peterman 199* Petersson 216*
Pham 90,115,128"' Polkinghorne 122f, 128*, 207*
Reed 43* Riddell Jr. 199* Riesz 85, 89* Sato
207*
Wanders 182* Westwater 128* Whitney 34, 43* Wichmann 90, 127*, 128*, 216* Wick 1,29,182* Wightman 92, 128*, 141 f, 215* Wu (A.C.T.) 128*, 215* Wu (T.T.) 89*, 1.82* Yamamoto 182*, 215* Yennie 189,190,199* Yokoyama 199*
Zav'yalov 216*
Subject Index (f indicates a footnote.)
Absorptive part 65
Additive function 4 Adjacent 5 Admissible (majorization) Analytic completion 205
163
Analytic continuation 204 Analytic function 203207 Anomalous threshold 107 Antiparticle 46 apath 39 Articulation point 6 Asymptotic behavior 1.96199 Axiomatic field theory 50
Baryon 46 Baryon loop 168 Baryonnumber conservation law Baryon path 158
Basic momenta 54 BetheSalpeter equation 146, 196 BinetCauchy theorem 26
Boson 46 Branch point 204
158
Cluster property 115 Commutation relation 48 Complement 3 Complete Feynman graph 9 Complete graph 7 Connected 5, 6 Connected component 5 Connected graph 6 Connectivity 16 Contour integral 203 Contract (a line) 6 Coupling constant 51 Couplingconstant renormalization Crossing symmetry 145 Cut 8 Cutkosky rule 115 Cutline 6 Cutset
8
Cutset matrix 21 Cutset number 15 Cutset property 193 Cutset representation 60, 73, 78 Cutvertex 6 Cyclically connected 6
Capacity 37
Cauchy's formula 204 Cauchy's theorem 204 Channel 9 Chord set 8
Chordset product sum 25 Circuit 8 Circuit matrix 21
Circuit number 14 Circuit representation 60, 73, 78 Classified majorization 168
DANAD 121 Degree (of a vertex) Delete (a line) 6
5
Demand 37 Dirac equation 47 Dirac particle 47
Directed graph 4f Discontinuity 204 Discontinuity formula 114115 Disjoint 3 219
5
220
SUBJECT INDEX
Disjoint union 3 Dispersion relation 131, 180181 Distribution 200203 Doperator 77 Double spectral function 144 Dual graph 3336, 9899, 106107 Duality 20
Finite graph 5 Finite part 85, 202 Frst Bietti number 16
Dynamics 49
Fundamental set (fset) of circuits 17 Fundamental set (fset) of cutsets 18
Dyson's powercounting theorem 8184 Electromagnetic (interaction) Element 3 Empty set 3 End point 4, 7 Endpoint singularity 206 Energy type (cutset) 193 Entire (function) 205
46
Envelope of holomorphy 205 Equalmass Feynman graph 158 Equivalence class 4 Equivalence relation 4 Essential singularity 205 Euclidean (set of vectors) 105
Euclidean majorization method 160 Extended production amplitude 52 Extended scattering amplitude 52 Extended transition amplitude 51 External line 9 External momentum 52 External vertex 9 Fdegree 9 Fernlion 46 Feynman cutoff 85 Feynman function 9293 Feynman graph 89 Feynman identity 44, 56 Feynman integral 5255 Feynman parameter 24, 44 Feynmanparametric function 7580 Feynmanparametric integral (or formula) 5667
Feynman propagator 52 Field (operator) 47 Final state 50
Flow 37 Forest 8 Free (particle) 47 Free Lagrangian
49
Generalized Feynman identity
56
Graph 4 Graph theory 143 Hadamard's finite part 202
Hadron 46 Hamilton circuit 187 Heisenberg operator 49 Highenergy asymptotic behavior 196199 Holomorphic 203, 205 Homological method 93
Hyperon 46 Identify (vertices) 6 Image 3 Incidence matrix 21 Incidence number 14 Incidence relation 4 Incident 4 Independent (circuits) 15 Independent (cutsets) 18 Infracircuit 116 Infrapath
116
Infrared divergence
54, 115117
Infrared singularity
116
Infravertex
116
Ingoing (end point) 4 Initial state 50
Integral representation 206 Interaction Lagrangian 49 Interaction representation 50 Intermediate state 9 Internal line 9 Internal mass 53 Internal momentum 53
SUBJECT INDEX
Internal vertex 9 Intersection 3 Inverse Feynman parameter 24 InverseFeynmanparametric integral
221
Modified cutset representation 62 Modified Yukawa coupling 185 Multilinear boson coupling 185 Multiple pole 205
6871
Isolated point 5 Isolated singularity 204
Natural boundary 205 Natural domain of holomorphy 205
Jost canonical form 121
Natural unit 46f Node 166
Kirchhoff's first and second laws 96
KleinGordon equation 47 Kuratowski graphs 32
Landau equations 95 Landau singularity 94101 Laurent series 204 Leading singularity 94 Leadingterm summation 196
Lepton 46
Line 4 Linear graph 4 Loop 8 Loop line 5 Lorentz group 47 Lorentz rotation 47 Lorentz transformation 47 Lowest normal threshold 109 Majorize (a graph) 161
Noneuclidean majorization method 160 Noneading singularity 94 Nonseparable (graph) 6 Normal (majorization) 163 Normal threshold 107 npath 166 ntree
8
ntree product stun 73
Nucleon 46 Nullity 16 Off the mass shell
51
Oka's theorem 205 On the mass shell 51
Oneparticle production graph 9 Onetoone correspondence Onto 3 Open covering 201 Oriented graph 4 Outgoing (end point) 4
3
Majorization method 160 Majorization rule 160165 Mandelstam bounding curve 180 Mandelstam representation 144 Mapping 3 Mass renormalization 85 Mass type (cutset) 193
Parallel (lines) 5 Parity conservation law 168 Pass through (a vertex) 5 Path 7 Path method 160
Maximal 4
Pauli principle 189f Peculiar solution 96 Perturbation series 51, 189192 Perturbationtheoretical integral representation 129182 Perturbation theory 50
Menger's theorem
194
Meromorphic 205
Meson 46 Microcausality condition 49
Minimal 4 Minimum cut theorem
38
Mixed secondtype singularity 122, 126127
Path theorem 43
Photon 46 Physicalregion singularity 11.2117 Physical sheet 93
222
SUBJECT INDEX
Pinch 206
Spectral representation 207
Pion
Spin 46
158
Planar graph 7, 3236 Poincar6 group 47 Pole 205 Position space 46 Positionspace Feynmanparametric integral 7174 Principal singularity 94 Production amplitude 52 Proper (subset) 3 Pseudotree 8
Square graph
107
Stability condition 132 Star
5
Strong (interaction) 46 Strongly connected 6
SubFeynmangraph 9 Subgraph 5 Subset 3 Subtraction procedure 145
Support 200
PTIR 129182
Pure secondtype singularity
47
Spinless
122126
Quantum field theory 46 Quantummechanical state 47
Rank (of a graph) 18 Real singularity 102111 Reduced graph 7 Reggepole theory 196f Relative degree 5 Renormalizable field theory 85 Renormalization 8488 Renormalization constant 191 Renormalized perturbation series 191 Restriction 4 Riemann sheet 204
Scattering amplitude 52 Secondquantized state 48 Secondtype singularity 122127 Selfenergy 52 Selfenergy graph 9 Selfenergy part 9 Set 3
Simple pinch 206
Symanzik region 144 Symmetric difference
3
Tadpole graph 9 Tadpole part 9 Taylor series 204
Tempered 201 Test function 200 Theorem of unicity 204
3tree product sum 171 Threshold 104, 105111 Topological formula 2431, 5964 tpath 197 Transition amplitude 51 Translation 47 Transport problem 3743 Tree 8
Tree product sum 25 Triangle graph 107, 110 Triple spectral representation 212 2isomorphic 35 Twoparticle scattering graph 9 2tree 8
2tree product sum 28 Type (of a vertex)
131
Simple pole 205
Singleloop graph 208215 Singularity (of a distribution) 201 Singularity (of an analytic function) 204 Skeleton 194 ,S matrix 46 Spectral condition 50, 132
Ultraviolet divergence 54, 8188 Unclassified majorization 168
Union
3
Unoriented graph 4 Vacuum 48
SUBJECT INDEX
Vacuum polarization graph 9 Vacuum polarization part 9 Vertex 4 Vertex function 52 Vertex graph 9 Vertex part 9 Vertex representation 62, 74, 78
223
Wave function 47 Weak (interaction) 46 Weight distribution 206 Wfunction inequality 169174
Yukawa coupling
184