Feynman Diagrams and Differential Equations: Chemistry Department, Napoli University Via Cinthia, Napoli, Italy
Feynman Diagrams and Differential Equations: Chemistry Department, Napoli University Via Cinthia, Napoli, Italy
MARIO ARGERI
Chemistry Department, Napoli University
Via Cinthia, Napoli, Italy∗
PIERPAOLO MASTROLIA
Institut für Theoretische Physik, Universität Zürich
CH-8057 Zürich, Switzerland†
We review in a pedagogical way the method of differential equations for the evaluation of D-dimensionally
regulated Feynman integrals. After dealing with the general features of the technique, we discuss its appli-
cation in the context of one- and two-loop corrections to the photon propagator in QED, by computing the
Vacuum Polarization tensor exactly in D. Finally, we treat two cases of less trivial differential equations,
respectively associated to a two-loop three-point, and a four-loop two-point integral. These two examples are
the playgrounds for showing more technical aspects about: Laurent expansion of the differential equations
in D (around D=4); the choice of the boundary conditions; and the link among differential and difference
equations for Feynman integrals.
∗ mario.argeri@unina.it
† pierpaolo.mastrolia@physik.unizh.ch
1
2 M. Argeri & P. Mastrolia
1. Preface
That “differentiating is an operation easier than integrating” is a statement that should not sound
too surprising; while, a more pleasant wonder might result when suitable differentiations make us
reduce, if not avoid at all, the number of direct integrations - of course the two operations, being
the inverse of each other, have not to be thought as performed with respect to the same variable!
As paradigmatic example, let us just consider the class of integrals,
Z ∞
2
In (α) = e−αx xn dx .
0
• n = 2s
2 n
∂ ∂ ∂ 2
In = − In−2 = − In−4 = . . . = − I0 ;
∂α ∂α ∂α
• n = 2s + 1
2 (n−1)
∂ ∂ ∂ 2
In = − In−2 = − In−4 = . . . = − I1 .
∂α ∂α ∂α
Therefore the infinite set of integrals In can be computed without any integration, provided the
knowledge of just two basic integrals, namely I0 , and I1 , that in the forthcoming terminology would
be defined as the master integrals of the class In .
The above example was a too lucky one: i) the repeated α-derivative did not entangle integrals
having even and odd indices, therefore I0 and I1 never appear linked by any differential identity; ii)
the value of the master integrals was known, possibly obtained by direct integrations.
In the more general case masters’ are unknown, and their evaluation becomes an open problem. In
the following pages, we will see how the exploitation of integration-by-parts not only yields algebraic
relations among infinite sets of integrals and their masters’, but as well leads to differential equations
satisfied by the master integrals themselves. Solving these differential equations becomes a tool for
computing master integrals, when their direct integration is not viable.
As it happens to (many) Feynman integrals.
Feynman Diagrams & Differential Equations 3
2. Introduction
A perturbative approach to the quantitative description of the scattering of particles in quantum
field theory involves the computation of Feynman diagrams. For a given number of external particles
- the legs of diagram - fixed by the process under study, and a given order in perturbation theory,
the skeletons of diagrams are built up by joining the edges of legs and propagators into vertexes,
forming tree patterns and closed loops.
Beyond the tree level, each Feynman diagram represents an integral which has, in general, a tenso-
rial structure, induced by the tensorial nature of the interacting fields. Therefore, the result of its
evaluation must be a linear combination of the tensors provided by the theory and by the kinematics
of the process under study. The coefficients of this linear combination, usually called form factors,
can be always extracted from each Feynman diagram, before performing any evaluation, by means
of suitably chosen projectors.
These form factors are scalar integrals closely connected to the original Feynman diagram: the nu-
merator of their integrand may contain all the possible scalar products formed by external momenta
and loop variables; whereas its denominator is formed by the denominators of propagators present
in the diagram itself.
Due to the bad convergence of loop integrals in four dimensions, regularization prescriptions
are mandatory. Hereafter the integrals are regularised within the framework of ’t Hooft-Veltman
continuous-dimensional regularisation scheme156 . Accordingly, the dimension D of an extended in-
tegration space is used as a regulator for both infrared (IR) and ultraviolet (UV) divergences, which
finally do appear as poles in (D − 4) when D goes to 4 156.
The aim of a precise calculation is to compute Feynman diagrams for any value of the available kine-
matic invariants. Except in case of simple configurations (e.g. very few legs and/or few scales), quite
generally, approximations have to be taken by limiting the result to a specific kinematics domain,
and, thus, looking for a hierarchy among the scales, to get rid of the ones which anyhow would give
a negligible contribution in that domain.
The puzzling complexity of the Feynman diagrams calculation arises because of two different
sources: either multi-leg or multi-loop processes. In recent years the progress in the evaluation of
higher loop radiative corrections in quantum field theory has received a strong boost, due to the
optimisation and automatising of various techniques (see refs. in 59,79,77 ). In this work we review
one of the most effective computational tools which have been developed in the framework of the
dimensional regularization: the method of differential equations for Feynman integrals.
The method was first proposed by Kotikov1 in the early nineties, while dealing with the evaluation
of 2- and 3-point functions. The basic idea was to consider a given unknown integral as a function
of one of the propagator masses, and to write for it a differential equation in that variable. Thus,
instead of addressing its direct integration, the value of the integral could be found by solving the
differential equation.
The advantages of the novel ideas were soon realized 2,3,4,5,6 , and generalised at a later stage by
Remiddi 7 , who proposed the differentiation with respect to any other available kinematics invariants
formed by the external momenta. That enabled the application of the differential equation method
also to integral with massless propagators (provided the existence of any other non-trivial scale).
4 M. Argeri & P. Mastrolia
Finally, Remiddi and Gehrmann 8,9,10 fully developed the method by showing its effectiveness
through the systematic application to a non-trivial class of two-loop four-point functions, whose
result is still considered as state-of-the-art.
From that moment on, the method of differential equations became to be widely used
in different contexts 11,12,13,14,15,16,17,18,19,20,21,22 . The lists of unprecedented results ob-
tained through its application spans among multi-loop functions from zero to four ex-
ternal legs, 25,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,137,138,139,140,141
and within the most interesting sectors of particle phenomenology, like jets physics
8,9,10,50,51,52,53,54,55,56,57,58 ; QED corrections to lepton form factors 80,81,82 ; Bhabha
Scattering 109,110,111,112,113,114,115,116,117; QCD corrections to lepton form factors and
top-physics 83,84,85,86,87 , forward-backward asymmetry of heavy-quark 95,96,97,98,99 ; Higgs
Physics 103,104,105,106,107,108 ; Electroweak sector 121,122,123,124,125,126 ; Sudakov form fac-
tors 100,101,102 ; semileptonic decay 118,119,120; static parameters and gauge boson properties
88,89,90,91,92,93,94 .
The efforts to achieve analytic solutions of differential equations for Feynman integrals has stimu-
lated new developments on the more mathematical side 127,128 , especially concerning the properties
of transcendental functions 132,133,134,135,136 . In particular, a novel set of functions that generalize
Nielsen polylogarithms, the so called Harmonic Polylogarithms (HPL) 129,130,131, have been found
suitable for casting the result in analytic form - that means without ambiguities due to zeroes hid-
den in functional relations, and supplied with series expansions. While HPL’s can be considered as
iterative integral of rational kernels, recently it has been pointed out that the solution of differential
equations for generic integrals with massive loops demands as well for irrational integration kernels,
yielding elliptic functions 148,102.
The range of applicability of the differential equations technique is broadened by the possibility
of a natural switch toward a semi-numerical approach, since, whenever the analytic integration were
not required or not viable, the differential equation(s), analytically obtained, could be solved with
numerical techniques 23,24 .
Nowadays we are not at the point to have the method for evaluating any Feynman integral,
but certainly we dispose of several tools 60,61,71,62,65,63,64,67,68,72,73,69,70,74,75,76,77 to attack
successfully many problems in perturbation theory, and usually a combination of them is necessary
for the achievement of the final answer. Therefore let us discuss in detail how to build ans solve
differential equations for integral associated to Feynman graphs.
The computational strategy is twofold.
• The second phase consists of the actual evaluation of the MI’s. By using the set of identities
previously obtained, it is also possible to write Differential Equations in the kinematic
invariants which are satisfied by the MI’s themselves. When possible, these equations can
be solved exactly in D dimensions. Alternatively, they can be Laurent-expanded around
suitable values of the dimensional parameter up to the required order, obtaining a system
of chained differential equations for the coefficients of the expansions, which, in the most
general case, are finally integrated by Euler’s variation of constants method.
One of the key advantages of the method is that it yields a clear separation between the merely
algebraic part of the work - which is, not surprisingly, always very heavy in multiloop calculations,
and can be most conveniently processed by a computer algebra program152-, from the actual analytic
issues of the problem, which can then be better investigated without the disturbance of the algebraic
complexity.
The paper is organised as follows. In section 2, it is described how to reduce a generic (combination
of) Feynman integrals to a limited set of MI’s and how to write the system of differential equations
they fulfil. As illustrative applications, respectively in section 3-4, the one-loop and two-loop vacuum
polarization functions in QED are explicitly computed. In the further two sections, we discuss some
less obvious example of differential equations. In section 5, we describe the solution of a system of
three coupled first-order differential equations, to compute three MI’s associated to a class of two-
loop 3-point functions, addressing as well the problem of finding their boundary conditions. While
in section 5, we describe the solution of a fourth-order differential equation involving the MI’s of a
4-loop 2-point diagram, and it will be considered the link between differential and difference equation
for Feynman integrals 65 .
Z N
dD k1 dD k2 dD kℓ Πi=1
sp
Sini
· · · , (1)
(2π)D−2 (2π)D−2 (2π)D−2 Πdj=1 Dj
where
given by
ℓ(ℓ + 1) ℓ 1
Nsp = ℓ(g − 1) + =ℓ g+ − (2)
| {z } | {z2 } 2 2
s.p. external−internal
s.p. internal−internal
One can denote with It,r,s the class of the integrals with: a given set of t denominators; q = Nsp −t
P P
irreducible scalar products; a total of r = i (mi − 1) powers of the t denominators; and s = j nj
powers of the q scalar products. It can be shown that the number of the integrals belonging to the
class It,r,s is
r+t−1 s+q−1
N (It,r,s ) = . (4)
t−1 q−1
In the next sections, we will see that integrals of the type (3), belonging to a given topology,
therefore differing for the values of the indices mi , ni , are not independent. Algebraic relations among
them, can be written in the form of a sum of a finite number of terms set equal to zero, where each
term is given by the product of a polynomial (of finite order and with integer coefficients in the
variable D, masses and Mandelstam invariants) and one of the integrals belonging to It,r,s . They
can be used recursively to express as many as possible integrals of a given class in terms of as
few as possible (suitably chosen) ones. The way to generate those kind of relations goes through
integration-by-parts, Lorentz invariance and symmetry considerations.
Before going on with the discussion let us see explicitly how the general definition we have
introduced do apply in practice.
Feynman Diagrams & Differential Equations 7
k1 − k2
k1 k2 k1
k k
k1 − k
Fig. 1. Feynman diagram with g=2 legs and ℓ = 2 loops
with g = 2 external legs, ℓ = 2 loops and d = 5 internal lines, which gives a number of scalar products amounting to,
„ «
1
Nsp = 2 2 + 1 − =5.
2
The most general set of scalar integrals possibly arising from its computation has the following representation,
n
dD k1 dD k2 Π5i=1 Si i
Z
, (5)
(2π)D−2 (2π)D−2 Π5j=1 Dj
where
k12 + m2 k12
8 8
> D1 = > S1 =
k22 k22
> >
>D
< 2
> = >S
< 2
> =
D3 = (k1 − k2 )2 + m2 S3 = k1 · k2
> D4 = (k1 − k)2 + m2 > S4 = k · k1
>
> >
>
> >
D5 = D1 = k12 + m2 S5 = k · k2
: :
The original diagram contains five internal lines, but two propagators are indeed equal, so there are only four different
propagators. Therefore, the integrals (5) actually belong to the simpler set,
dD k1 dD k2 Π5i=1 Sini
Z
, (6)
(2π)D−2 (2π)D−2 D12 D2 D3 D4
which can be represented by the topology in Fig. 2.
The trivial tensor reduction of the scalar product can be realized according to the following table, where one
chooses to express four (out of five) scalar products in terms of the denominators. In the end, only one of the five
scalar products involving the loop momenta is left over as irreducible, arbitrary chosen to be k2 · k. Therefore, the
integrals in (5), represented by the topology in Fig.2, indeed belongs to the class,
dD k1 dD k2 (k2 · k)n1
Z
I(n1 ; m1 , m2 , m3 , m4 ) . (7)
(2π)D−2 (2π)D−2 D1m1 D2m2 D3m3 D4m4
The trivial tensor reduction might as well lead to the complete cancellation of some denominator. Should it be the
case, the resulting integral can be classified as belonging to the subtopology obtained by pinching the corresponding
internal line.
8 M. Argeri & P. Mastrolia
k1 k1 − k2
k2
k k
k1 − k
Fig. 2. 4-denominators topology
X ∂I(pi ) X ∂I(pi )
I(pi + δpi ) = I(pi ) + = I(pi ) + ωµν pn,ν , (10)
n
∂pn,µ n
∂pn,µ
X ∂ ∂
pn,ν − pn,µ I(pi ) = 0. (11)
n
∂pn,µ ∂pn,ν
Eq. (11) can be contracted with all possible antisymmetric combinations of the external momenta
pi,µ pj,ν , to obtain other identities for the considered integrals.
In case of integral associated to any vertex topologies with two independent external momenta, p1
and p2 , we can build up the identity
p1
∂ ∂ 2 ∂ 2 ∂
(p1 · p2 ) p1,µ − p2,µ + p2 p1,µ − p1 p2,µ p3 = 0 (12)
∂p1,µ ∂p2,µ ∂p2,µ ∂p1,µ p 2
In the case of a richer kinematics, like in the case of integrals associated to box topologies with three
independent external momenta, p1 , p2 and p3 , we ca write down three LI-id’s
p1 p3
X ∂ ∂
(p1,µ p2,µ − p1,ν p2,µ ) pn,ν − pn,µ = 0, (13)
n
∂pn,µ ∂pn,ν p2 p4
p1 p3
X ∂ ∂
(p1,µ p3,µ − p1,ν p3,µ ) pn,ν − pn,µ = 0, (14)
n
∂pn,µ ∂pn,ν p2 p4
p1 p3
X ∂ ∂
(p2,µ p3,µ − p2,ν p3,µ ) pn,ν − pn,µ =0. (15)
n
∂pn,µ ∂pn,ν p2 p4
More identities, originally found by Larin, may arise as well when there is a sub-loop diagram
depending - after its integration - on a specific combination of momenta. By equating the original
integral and the one obtained by projecting onto such a combination of momenta, one gets additional
relations (see 78 for details).
dD k1 dD k2 (k · k2 )n1
ff
∂
Z
D−2
vµ m1 m2 m3 m4 = 0 (i = 1, 2), (16)
(2π) (2π)D−2 ∂ki,µ D1 D2 D3 D4
10 M. Argeri & P. Mastrolia
∂
derivative ∂ki,µ vector vµ corresponding IBP-Id
R
∂ dD k1 dD k2 ∂ k1,µ
k1,µ =0
∂k1,µ
R
(2π)D−2 (2π)D−2 ∂k1,µ
D D
D1 D2 D3 D4
∂ d k1 d k2 ∂ k2,µ
k2,µ =0
∂k1,µ
R
(2π)D−2 (2π)D−2 ∂k1,µ
D D
D1 D2 D3 D4
∂ d k1 d k2 ∂ kµ
kµ =0
∂k1,µ
R
(2π)D−2 (2π)D−2 ∂k1,µ
D1 D2 D3 D4
∂ dD k1 dD k2 ∂ k1,µ
k1,µ =0
∂k2,µ
R
(2π)D−2 (2π)D−2 ∂k2,µ
D D
D1 D2 D3 D4
∂ d k1 d k2 ∂ k2,µ
∂k2,µ k2,µ (2π)D−2 (2π)D−2 ∂k2,µ D1 D2 D3 D4 =0
R
∂ dD k1 dD k2 ∂ kµ
∂k2,µ kµ (2π)D−2 (2π)D−2 ∂k2,µ D1 D2 D3 D4 =0
2k · k1 = D1 − D4 + k 2 , (20)
and substituting Eq. (20) in Eq. (19), we obtain,
dD k1 dD k2 1 dD k1 dD k2 1
Z Z
− 2
+ +
(2π) D−2 (2π) D−2 D1 D2 D3 (2π)D−2 (2π)D−2 D2 D32 D4
dD k1 dD k2 1 dD k1 dD k2 k · k2
Z Z
+k 2 −2 +
(2π)D−2 (2π)D−2 D1 D2 D32 D4 (2π)D−2 (2π)D−2 D1 D2 D32 D4
dD k1 dD k2 k · k2
Z
−2 =0 (21)
(2π) D−2 (2π)D−2 D1 D22 D3 D4
Eq. (21) can be pictorially represented by
− + + k2 +
(k · k2 ) (k · k2 )
−2 −2 =0 (22)
Feynman Diagrams & Differential Equations 11
where a dot on a propagator line means that the propagator is squared; irreducible scalar products in the numerator
are explicitly indicated.
p1
p1 + p2 − k1 − k2
p2 − k2 k1 Q = p1 + p2
k2
p2
Fig. 4. Example of a two-loop three-leg four-denominators topology: p21 = p22 = −m2 , and (p1 + p2 )2 = Q2 = −s
The integrals associated to this topology can have three irreducible scalar products, arbitrary chosen to be (p1 ·k1 ),
(p2 · k1 ) and (k1 · k2 ). In doing so, Eq. (8), for generic values of the indices mi and ni , reads
dD k1 dD k2
ff
∂ p1,µ
Z
D−2
= 0. (24)
(2π) (2π)D−2 ∂k1,µ D1 D2 D3 D4
After taking the derivative with respect to k1 and simplifying the reducible scalar products against the corresponding
propagators, one can write Eq.(24) as follows
(p1 · k1 ) (k1 · k2 )
−2 −2 +
(p2 · k1 )
+2 + +
(D − 2)
− + =0 (25)
2m2
12 M. Argeri & P. Mastrolia
ff
(k1 · k2 ) (p2 · k1 )
−2m2 s + 2s −2 +
s(D − 2)
− (26)
2m2
k1 = p1 + p2 − k1′ − k2 , (28)
that consists in the interchange of the two denominators in the closed electron loop, does not affect the topology of
the integral; nevertheless the explicit form of the integrand can change generating non trivial identities. Taking for
instance n1 = n2 = n3 = 0 and m1 = m2 = m4 = 1 and m3 = 2 in Eq. (27), the substitution (28) gives the following
very simple relation
− = 0. (29)
(k1 · k2 ) (p1 · k1 )
0= + +
(p2 · k1 ) s 1
+ + + (30)
2 2
Feynman Diagrams & Differential Equations 13
Third, we are not able to prove analytically that the MI’s one finds are really the minimal set of MI’s,
i.e. that they are independent from each other; in any case the final number of the MI’s is quite small,
so that reducing the several hundred integrals occurring in a typical calculation to a few of them
is after all a great progress. Studies on the apriori determination of whether a topology could have
or not MI’s, have been performed by Baikov 153 and Smirnov 154 . Although a general algorithm
is still lacking, we think that the analysis of leading and subleading Landau singularities (see 155
for an extensive treatment) could be related to this issue: since a reducible topology without MI’s is
expressible in terms of the MI’s of its subtopologies, it should mean that the leading singularities of
the reducible topology are not independent of the subleading ones, which are leading singularities
for the subtopologies.
As a last remark, let us recall that, at the end of the reduction, there is some freedom for choosing
the basis of MI’s and usually the choice is in general motivated by convenience. For example the
behaviour of the functions in the D-to-4 expansion might determine to select: i) simpler integrands,
in view of a successive analytic computation; ii) more complicated integrands, but with a better
convergence, should their numerical evaluation be of interest.
Z n
dD k1 dD kℓ S1n1 ...Sq q
M (s1 , s2 , · · · , sN ) = ... , (31)
(2π)D−2 (2π)D−2 D1 ...Dt
where {s1 , s2 , · · · , sN } is any set of kinematic invariants of the topology and N is the number of
such invariants.
Let us denote the set {s1 , s2 , · · · , sN } = s and consider the following quantities
∂M (s)
Ojk (s) = pj,µ (j, k = 1, 2, · · · , g − 1), (32)
∂pk,µ
where g − 1 is the number of independent external momenta. By the chain differentiation rule we
have
Feynman Diagrams & Differential Equations 15
N N
X ∂sα ∂M (s) X ∂sα ∂M (s)
Ojk (s) = pj,µ · = pj,µ · . (33)
α=1
∂p k,µ ∂s α α=1
∂p k,µ ∂sα
According to the available number of the kinematic invariants, the r.h.s. of Eq. (32) and the r.h.s.
of Eq. (33) may be equated to form the following system
N
X ∂sα ∂M (s) ∂M (s)
pj,µ · = pj,µ , (34)
α=1
∂p k,µ ∂s α ∂pk,µ
• 3-point case.
p1
∂
P2 p3 =
∂P 2 p
2
p1
∂ ∂ ∂ ∂
= A p1,µ + p2,µ + B p1,µ + p2,µ p3 ,
∂p1,µ ∂p2,µ ∂p2,µ ∂p1,µ p2
(37)
with P = p1 + p2 and A, B rational coefficients.
• 4-point case.
p1 p3
∂ ∂ ∂ ∂
P2 = C p1,µ − p3,µ + Dp2,µ +
∂P 2 p2 p4
∂p1,µ ∂p3,µ ∂p2,µ
p1 p3
∂ ∂ ∂
+ E(p1,µ + p3,µ ) − + , (38)
∂p3,µ ∂p1,µ ∂p2,µ p p4
2
16 M. Argeri & P. Mastrolia
p1 p3
2∂ ∂ ∂ ∂
Q = F p1,µ − p2,µ + Gp2,µ +
∂Q2 p2 p4
∂p1,µ ∂p2,µ ∂p2,µ
p1 p3
∂ ∂ ∂
+ H(p2,µ − p1,µ ) + + , (39)
∂p1,µ ∂p2,µ ∂p3,µ p p4
2
Equation (34) holds for any function M (s). In particular, let us assume that M (s) is a master integral.
We can now substitute the expression of M in the r.h.s. of one of the Eqs.(36-39), according to the
case, and perform the direct differentiation of the integrand. It is clear that we obtain a combination
of several integrals, all belonging to the same topology as M . Therefore, we can use the solutions of
the IBP-id’s, LI-id’s and other identities for that topology and express everything in terms of the
MI’s of the considered topology (and its subtopologies). If there is more than one MI, the procedure
can be repeated for all of them as well. In so doing, one obtains a system of linear differential
equations in s for M and for the other MI’s (if any), expressing their s-derivatives in terms of the
MI’s of the considered topology and of its subtopologies.
The system is formed by a set of first-order differential equations (ODE), one for each MI, say
Mj , whose general structure reads like the following,
∂ X X
Mj (D, s) = Ak (D, s) Mk (D, s) + Bh (D, s) Nh (D, s) (40)
∂sα
k h
where α = 1, · · · , N , is the label of the invariants, and Nk are MI’s of the subtopologies. Note
that the above equations are exact in D, and the coefficients Ak , Bk are rational factors whose
singularities represent the thresholds and the pseudothresholds of the solution.
The system of equations (40) for Mj is not homogeneous, as they may involve MI’s of subtopologies.
It is therefore natural to proceed bottom-up, starting from the equations for the MI’s of the simplest
topologies (i.e. with less denominators), solving those equations and using the results within the
equations for the MI’s of the more complicated topologies with additional propagators, whose non-
homogeneous part can then be considered as known.
n
X
(k)
Mj (D, s) = (D − 4)k Mj (s) + O (D − 4)(n+1) , (41)
k=−n0
where n0 (positive) corresponds to the highest pole, and n is the required order in (D − 4).
When expanding systematically in (D − 4) all the MI’s (including those appearing in the non-
homogeneous part) and all the D-dependent terms of (41), one obtains a system of chained equations
(k)
for the Laurent coefficients Mj of (41). The first equation, corresponding to the highest pole involves
(−n )
only the coefficient Mi 0 as unknown; the next equation, corresponding to the next pole in (D−4),
(−n +1) (−n )
involves the Mi 0 as unknown, but it may involve Mi 0 in the non-homogeneous term (if it
(−n )
appears as multiplied by any power of (D−4)); but such a term, Mi 0 , has to be considered known
once the equation for the highest pole has been solved. For the subsequent equations we have the
(k)
same structure: at a given order k in (D − 4), the equations involve Mj as unknown, and previous
(ℓ)
coefficients Mj (−n0 ≤ ℓ < k) as known non-homogeneous terms.
Let us note that the homogeneous part of all the equations arising from the D → 4 expansion of
(41) is always the same and obviously identical to the homogeneous part of Eq.(40) read at D = 4. It
is, therefore, natural to look for the solutions of the chained non-homogeneous equations by means of
Euler’s method of the variation of the constants, using repeatedly the solutions of the homogeneous
equation as integration kernel, as we will see in the following chapters.
General algorithms for the solution of the homogeneous equations are not available; it turns
out however that in all the considered cases the homogeneous equations at D = 4 have (almost
trivial) solutions, so that Euler’s formula can immediately be written. With suitable changes of
variable, according to the kinematics of the problem, all integrations can further be carried out in
closed analytic form, by exploiting the shuffle algebra induced by integration-by-parts on the nested
integral representation of the solution.
where N is the equation order, and pk (n) are polynomial in n of maximum degree P , whose generic
structure can be parametrised as follows,
P
X i−1
Y
pk (n) = Ak0 + Aki (n + k + j) . (43)
i=1 j=0
where γ is a suitable integration path, and whose effect can be shown to be 65 the translation of
the difference equation (42), into a differential equation for v(t),
P
X d(i)
Φi (t) (−t)i v(t) = 0 , (45)
i=0
dti
with
N
X
Φi (t) = Aki tk . (46)
k=0
The solution v(t) of Eq.(45), still carry unknown integration constants which will be present as well
in U (n), reconstructed by integrating Eq.(44). Finally, the value of the yet undetermined integration
constants can be fixed by imposing boundary conditions in the large-n regime of U (n), easily found
by direct inspection of its original representation as loop integrala .
n
5.0.1. Example: one-loop Tadpole
dD k
Z
D
U (n) = π − 2 = . (47)
(k 2 + 1)n
By means IBP Id’s, one can write the following relationship between the integrals of family (47)
n−1 n
„ «
D
− n−1− + (n − 1) = 0, (48)
2
which, after renaming n → n + 1, can be put in the form (42), and read as a first-order difference equation for the
function U (n), „ «
D
− n− U (n) + n U (n + 1) = 0. (49)
2
a In the rest of the paper, for computational convenience, we adopt two different definitions of the loop measure:
(i) dD k/(2π)D−2 , for integrals treated with the Differential Equations method; (ii) dD k/(π)D/2 , in case of use of
Difference Equations method.
Feynman Diagrams & Differential Equations 19
The boundary condition is determined by the asymptotic behaviour (n → ∞) of the integral (47), which reads
dD k
Z Z
D D 2 D
π− 2 2 n
≈ π− 2 dD ke−nk = n− 2 . (50)
(k + 1)
One way to solve equation (49) is to seek for the solution as a factorial series (see 65,66 for details). Equivalently,
one can convert the difference equation into a differential equation, by means of the Laplace transform of U (n).
Accordingly, with the ansatz Z 1
U (n) = dt tn−1 v(t) , (51)
0
and
D
A00 = , A01 = −1 ,
2
A10 = −1 , A11 = 1 , (52)
equation (49) becomes a differential equation in the form of (45) for the function v(t),
„ «
D
− t(t − 1)v′ (t) + − t v(t) = 0. (53)
2
The solution is,
v(t) = v0 t−D/2 (1 − t)D/2−1 , (54)
where v0 is the yet unknown integration constant. With the above solution, Eq.(51) can be integrated and U (n) reads,
“ ” “ ”
Γ D2
Γ n− D 2
U (n) = v0 . (55)
Γ(n)
By comparing its large-n expansion,
„ „ « „ ««
n→∞ D 1
U (n) = v0 n−D/2 Γ +O (56)
2 n
with the asymptotic limit explicitly computed in Eq.(50), one finally determines the value of the constant,
1
v0 = “ ” . (57)
D
Γ 2
p−k
Fig. 5. 1-loop vacuum polarization diagram.
As an application, we calculate the one-loop correction to the photon propagator: the so called
Vacuum Polarization tensor. The only contributing diagram is shown in Fig. 5, and corresponds to
the expression
Z
α dD p γµ (−ip / − /k) + m]
/ + m)γν [−i(p
i Πµν (k) = (−ie)2 (−1) Tr . (58)
π (2π)D (p + m )[(p − k)2 + m2 ]
2 2
20 M. Argeri & P. Mastrolia
Πµν is a tensor which depends only on the external momentum kµ and the metric tensor δµν and
can be decomposed as a sum of two contributions
1
Πµµ (D, k) = Π(D, k 2 )(k 2 − k 2 D) =⇒ Π(D, k 2 ) = Πµµ (k). (62)
k 2 (1 − D)
e2
From (58), setting α = 4π , we have
Z
dD p Tr(−γµ/pγν/q + m2 γµ γν )
iΠµν (k) = D−2
; (63)
(2π) (p2 + m2 )[(p − k)2 + m2 ]
after some operations within the Dirac algebra, our expression becomes
Tr(−γµ/pγν/q + m2 γµ γν ) =
= −pα qβ Tr(γµ γα γν γβ ) + m2 Tr(γµ γν ) =
= −Tr(I)D pα qβ (δµα δνβ − δµν δαβ + δµβ δνα ) + Tr(I)D m2 δµν
= −Tr(I)D (pµ qν − p · qδµν + pν qµ − m2 δµν ) =
= −Tr(I)D [2pµ pν − pµ kν − pν kµ − (p2 − p · k + m2 )δµν ], (64)
where Tr(I)D is the trace of the unit matrix in D dimensions (this is, in general, different from D).
However, limD→4 Tr(I)D = 4. Substituting (64) in (63) and putting µ = ν we obtain
Πµµ (k) =
Z
dD p 2p2 − 2p · k − D(p2 − p · k + m2 )
= Tr(I)D i =
(2π)D−2 (p2 + m2 )[(p − k)2 + m2 ]
Z
dD p (D − 2)(p · k) − (D − 2)p2 − D m2
= Tr(I)D i . (65)
(2π)D−2 (p2 + m2 )[(p − k)2 + m2 ]
We can define,
D1 = p2 + m2
D2 = (p − k)2 + m2
Feynman Diagrams & Differential Equations 21
Πµµ (k) =
Z
dD p (D − 2)(p · k) − (D − 2)p2 − D m2
= Tr(I)D i =
(2π)D−2 D1 D2
Z
dD p (D − 2)[(D1 − D2 + k 2 )/2] − (D − 2)(D1 − m2 ) − D m2
= Tr(I)D i =
(2π)D−2 D1 D2
Z 2 Z
(2 − D) dD p 1 1 k (D − 2) 2 dD p 1
= Tr(I)D i + + − 2m =
2 (2π)D−2 D1 D2 2 (2π)D−2 D1 D2
2
(2 − D) k (D − 2)
= Tr(I)D i (J10 + J01 ) + − 2m2 J11 , (68)
2 2
where all the integrals can be seen as belonging to the class,
Z
dD p 1
Jm1 m2 = D−2 m1 m2 . (69)
(2π) D1 D2
Finally, the scalar vacuum polarization function can be written as
Π(D, k 2 ) =
Tr(I)D i
= 2 (2 − D)(J10 + J01 ) + [k 2 (D − 2) − 4m2 ]J11 =
2k (1 − D)
Tr(I)D i 2 1 2 2 2
= 2 (2 − D)T (D, m ) + [k (D − 2) − 4m ]J(D, k ) .
k (1 − D) 2
(70)
One may notice the result is expressed in terms of two integrals: the massive Tadpole 7
Z
dD p 1 mD−2
T (D, m2 ) = J10 = J01 = = D−2
= C(D) , (71)
(2π) D1 (D − 2)(D − 4)
and the 1-loop 2-point function
Z
dD p 1
J(D, k 2 ) = J11 = = , (72)
(2π)D−2 D1 D2
where
4−D
C(D) = (4π) 2 Γ(3 − D/2) , (73)
is an overall loop factor whose value in 4 dimension is C(4) = 1. In the following pages, when not
ambiguous, we will give as understood a factor C(D) for every loop.
22 M. Argeri & P. Mastrolia
Z
dD p 1
Jm1 m2 = D−2 m1 m2 . (74)
(2π) D1 D2
By means of IBP id’s, we can reduce every integral of the type (74) to
Z
dD p 1
J11 = J(D, k 2 ) = , (75)
(2π)D−2 D1 D2
which is, actually, the only MI of this topology, and to the tadpole J10 , which is the MI of the only
possible subtopology.
Let us write explicitly some useful IBP id’s for the integral (75). we have
R
dD p ∂ kµ
=0
R
(2π)D−2 ∂pµ D1 D2 , (76)
dD p ∂ pµ
(2π)D−2 ∂pµ D1 D2 =0
where D1 = p2 + m2 and D2 = (p − k)2 + m2 . After the trivial tensor reduction we obtain the
identities
=
, (77)
(D−3) 1
= − (4m 2 +k2 ) + (4m2 +k2 )
with
Z
dD p 1
= D−2
, (78)
(2π) D1 D2
Z
dD p 1
= (79)
(2π)D−2 D12 D2
Z
dD p 1
= , (80)
(2π)D−2 D1 D22
Z
dD p 1 (D − 2)
= D−2 2 =− (81)
(2π) D1 2 m2
Feynman Diagrams & Differential Equations 23
where the last equation has been obtained as IBP identity for the tadpole 7 .
∂J ∂J ∂k 2 ∂J
= 2
= 2kµ 2 . (82)
∂kµ ∂k ∂kµ ∂k
By contracting (82) with the vector kµ we have
∂J ∂J
kµ = 2k 2 2 . (83)
∂kµ ∂k
On the other hand
Z Z
∂J dD p ∂ 1 dD p 2(pµ − kµ )
= = , (84)
∂kµ (2π)D−2 ∂kµ D1 D2 (2π)D−2 D1 D22
so
Z
∂J dD p 2(p · k − k 2 )
kµ = =
∂kµ (2π)D−2 D1 D22 |{z}
2p·k=D1 −D2 +k2
Z Z Z
dD p 1 dD p 1 dD p k2
= − − =
(2π)D−2 D22 (2π)D−2 D1 D2 (2π)D−2 D1 D22
= − − k2 (85)
d 1 1 1
= − − , (86)
dk 2 2k 2 2k 2 2
which is rewritten, thanks to the second identity of the (77) and to (81), as a non-homogeneous
first-order differential equation for J(D, k 2 )
d 1 1 (D − 3)
+ − 2
dk 2 2 k2 (k + 4m2 )
(D − 2) 1 1
=− − 2 . (87)
4m2 k2 (k + 4m2 )
Eq. (87) contains the boundary condition for the solution. In fact, thanks to the analytic properties
of Feynman integrals, we know that J(D, k 2 ) must be a regular function in k 2 = 0, that is
24 M. Argeri & P. Mastrolia
dJ
lim k 2 = 0. (88)
k2 →0 dk 2
By multiplying Eq. (87) by k 2 and taking the limit k 2 → 0, we have
d 1 (D − 3)k 2
lim k2 + lim1− 2 =
k2 →0 dk 2 k →0 22 (k + 4m2 )
(D − 2) k2
= − lim 2
1 − , (89)
k2 →0 4m (k + 4m2 )
2
d 1 1 (D − 3) (D − 2) 1 1
+ − =− − . (91)
dx 2 x (1 + x) 4m2 x (1 + x)
The general solution of Eq. (91) is the sum of solution of the homogeneous equation, say J0 , and
particular solution of the complete equation, say J ∗ . The homogeneous equation is
dJ0 1 1 (D − 3)
=− − J0 (92)
dx 2 x (1 + x)
with solution,
1 D−3
J0 (D, x) = Ax− 2 (1 + x) 2 . (93)
To find a particular solution, J ∗ (D, x) of the complete equation, we use Euler’s variation of
constants method, and write
1 (D−3)
J ∗ (D, x) = x− 2 (1 + x) 2 φ(x), (94)
where φ(x) is an unknown function to be found by imposing J ∗ fulfills Eq.(91). In so doing, we
obtain the following equation for φ(x),
dφ (D − 2) 1 3−D 1 1−D
=− [x− 2 (1 + x) 2 − x 2 (1 + x) 2 ], (95)
dx 4m2
hence
Feynman Diagrams & Differential Equations 25
Z Z
(D − 2) 1 D−3 1 D−1
φ(x) = − x− 2 (1 + x)− 2 dx − x 2 (1 + x)− 2 dx . (96)
4m2
The integrals in Eq.(96) are a representation of hypergeometric functions
R −1 D−3 1
x 2 (1 + x)− 2 dx = 2x 2 2 F1 21 , D−3 2 ; 23 ; −x
R 1 D−1 3 (97)
x 2 (1 + x)− 2 dx = 32 x 2 2 F1 23 , D−1 5
2 ; 2 ; −x ,
hence
J(D, x) = J0 + J ∗ =
1 (D−3) (D − 2) D−3
= Ax− 2 (1 + x) 2 − 2
(1 + x) 2 ×
4m
1 D−3 3 2 3 D−1 5
× 2 2 F1 , ; ; −x − x 2 F1 , ; ; −x . (98)
2 2 2 3 2 2 2
1
By imposing the boundary condition, we see that the constant A = 0, because the term x− 2 (1 +
(D−3)
x) 2 is singular for x → 0, while the MI is regular by inspection.
Finally, the full solution of our Cauchy problem is
(D − 2) D−3
J(D, x) = − 2
(1 + x) 2 ×
2m
1 D−3 3 x 3 D−1 5
× 2 F1 , ; ; −x − 2 F1 , ; ; −x =
2 2 2 3 2 2 2
(D − 2) D−3 D−1 1 3
=− (1 + x) 2 F
2 1 , ; ; −x =
2m2 2 2 2
(D − 2) 4−D 3
=− F
2 1 , 1; ; −x (99)
2m2 2 2
J(D, x) 1 (2 − D) 1
lim = J(D, 0) lim + J ′ (D, 0) = lim + J ′ (D, 0) , (102)
x→0 x x→0 x 2m2 x→0 x
hence
Tr(I)D i D−2 1 ′
Π(D, 0) = J(D, 0) − J (D, 0) . (103)
(1 − D) 2 2
Finally, one has the renormalization counterterm,
(D − 2)
Π(D, 0) = Tr(I)D i (104)
6m2
yielding the renormalised form factor,
Tr(I)D i (2 − D)
ΠR (D, k 2 ) = Π(D, k 2 ) − Π(D, 0) = ×
(1 − D) 4m2
1 1 4−D 3 2
× + (D − 2) − F
2 1 , 1; ; −x + (1 − D) .
x x 2 2 3
(105)
which can be ultimately expanded around D = 4 up to the finite order in terms of HPL’s
5 4 4
ΠR (D, k 2 ) = i − 2 +
9 3 (1 − z) 3 (1 − z)
!
(1 + z) 1 − 4 z + z 2 H(0, z)
+ 3 + O((D − 4)). (106)
3 (1 − z)
with the spacelike variable z defined as
√ √ √ √
−x + 1 − −x −k 2 + 4m2 − −k 2
z= √ √ = √ √ (107)
−x + 1 + −x −k 2 + 4m2 + −k 2
order differential equations in the square of the external momentum. We will see that the system
is equivalent to a pair of second order differential equations, one for each MI. These equations
can be solved exactly in D dimensions and the solution can be written in terms of generalized
Hypergeometric functions.
7.1. Diagrams
The one particle irreducible contributions to the two-loop vacuum polarization are showed in Eq.
(108). They consist of three bare diagrams, plus the corresponding counterterms, namely two fermion
self-energy counterterms and two vertex corrections counterterms. Due to Ward identities of QED,
some of these counterterms cancel each other: the vertex subtractions cancel exactly the wavefunc-
tion renormalization counterterms; so the only effective contribution comes from the fermion mass
counterterm, as shown in Eq. (108). The amplitude reads
α 2
i Πµν (k) =
π
= = + +
iδm
×
+ + (108)
×
iδm
By the systematic application of the reduction algorithm, one can express all the needed integrals
as combination of just five MI’s, depicted in the last column of Tab. 1: one MI with four denominators;
three with three; and one with two.
4
(p1 · p2 )
3
By giving a close look at them, one soon realizes that three of them are indeed product of the
28 M. Argeri & P. Mastrolia
= T 2 (D, m2 ) , (109)
= J 2 (D, k 2 ) , (111)
where T (D, m2 ) and J(D, k 2 ) were given respectively in Eqs. (71, 99).
The two yet unknown MI’s, both belonging to the same topology, are,
Z Z
dD p1 dD p2 1
J1 (k 2 ) = = D−2 D−2
(112)
(2π) (2π) D1 D2 D3
Z Z
2 (p1 · p2 ) dD p1 dD p2 (p1 · p2 )
J2 (k ) = = , (113)
(2π)D−2 (2π)D−2 D1 D2 D3
where D1 = p21 + m2 , D2 = p22 , D3 = (k − p1 − p2 )2 + m2 , p1 and p2 being the two loop momenta and
k being the external momentum. Let’s discuss their evaluation.
2
d3 J1 3 d J1
η 2 (η + 4m2 ) 3 − η (D − 6) η − 14m2 +
dη 2 dη 2
(D − 4)(D − 9) dJ1 (3 − D)(4 − D)
+ η − D(D − 5)m2 + J1 = 0, (116)
2 dη 2
η
which, introducing the dimensionless variable x = 4m2 , becomes
Feynman Diagrams & Differential Equations 29
d3 J1 3(D − 6) 7 d2 J1
x2 (1 + x) 3 − x x− +
dx 2 2 dx2
(D − 4)(D − 9) D(D − 5) dJ1 (3 − D)(4 − D)
+ x− + J1 = 0.
2 4 dx 2
(117)
d3 ψ d2 ψ
x2 (1 + x) 3
+ x[x(a1 + a2 + a3 + 3) + (b1 + b2 + 1)] 2 +
dx dx
dψ
+[x(a1 a2 + a1 a3 + a2 a3 + a1 + a2 + a3 + 1) + b1 b2 ] +
dx
+(a1 a2 a3 )ψ = 0, (118)
where a1 , a2 , a3 , b1 , b2 are parameters, is classified as hypergeometric and its solution space is
spanned by the functions
a1 , a2 , a3
ψ1 (x) = 3 F2 ; −x
b1 , b2
1−b1 1 + a1 − b 1 , 1 + a2 − b 1 , 1 + a3 − b 1
ψ2 (x) = (−x) 3 F2 ; −x (119)
2 − b1 , 1 + b2 − b1
1 + a1 − b 2 , 1 + a2 − b 2 , 1 + a3 − b 2
ψ3 (x) = (−x)1−b2 3 F2 ; −x .
2 − b2 , 1 + b1 − b2
We can easily see that Eq. (117) is hypergeometric if we identify the parameters as follows:
a =1
1
2 =3−D
a
a3 = (4 − D)/2 (120)
b1 = D/2
b2 = (5 − D)/2 .
With the above choice, the three independent solutions of Eq. (117) become
(3 − D), (4 − D)/2, 1
ψ1 (x) = 3 F2 ; −x
D/2, (5 − D)/2
(D−2) (8 − 3D)/2, (3 − D), (4 − D)/2
ψ2 (x) = (−x)− 2 3 F2 ; −x (121)
(4 − D)/2, (7 − 2D)/2
(D−3) (3 − D), 1/2, (D − 1)/2
ψ3 (x) = (−x) 2
3 F2 ; −x
(2D − 3)/2, (D − 1)/2
30 M. Argeri & P. Mastrolia
Finally, the general solution of Eq. (117) can be written as a linear combination of ψ1 (x), ψ2 (x)
e ψ3 (x)
J1 (0) = =
Z Z
dD p1 dD p2 1
= D−2 D−2 2 2 =
(2π) (2π) (p1 + m )p2 [(p1 − p2 )2 + m2 )
2
(D − 2)
=− , (123)
2m2 (D − 3)
where the last line has been obtained by means of IBP-id’s. If we study the behaviour of (123)
in the UV and IR limits, we see that it is regular for 2 < D < 4. This means that B = C = 0.
(D−2) (D−3)
If this was not the case, in fact, terms such as (−x)− 2 and (−x) 2 would give rise, in the
abovementioned range of D, to a divergent behaviour, and this would not be compatible with the
finite result of (123).
Furthermore, (123) allows one to choose once and for all the constant A
A= ,
hence
(D − 2) (3 − D), (4 − D)/2, 1
J1 (x) = − 3 F2 ; −x . (125)
2m2 (D − 3) D/2, (5 − D)/2
Now we can obtain the expression for J2 (x) by simply substituting J1 (x) and dJ
dx in the first
1
equation of (114). To do so, we have to know the expression for the first order derivative of the
Hypergeometric function 3 F2 (x).
Quite in general, a function like
a1 , a2 , a3
3 F2 ; −x
b1 , b2
has the following series representation
X+∞
a1 , a2 , a3 (a1 )n (a2 )n (a3 )n (−x)n
3 F2 ;x = , (126)
b1 , b2 (b1 )n (b2 )n n!
n=0
Feynman Diagrams & Differential Equations 31
d a1 , a2 , a3
3 F2 ;x =
dx b1 , b2
+∞ +∞
d X (a1 )n (a2 )n (a3 )n (−x)n X (a1 )n (a2 )n (a3 )n (−x)n−1
= =− n =
dx n=0 (b1 )n (b2 )n n! n=0
(b1 )n (b2 )n n!
+∞
X (a1 )n (a2 )n (a3 )n (−x)n−1
=− =
n=1
(b1 )n (b2 )n (n − 1)!
+∞
X (a1 )k+1 (a2 )k+1 (a3 )k+1 (−x)k
=− . (127)
(b1 )k+1 (b2 )k+1 k!
k=0
Γ(ξ + k + 1) Γ(ξ + 1 + k)
(ξ)k+1 = =x = ξ(ξ + 1)k ,
Γ(ξ) Γ(ξ + 1)
hence
d a1 , a2 , a3
F
3 2 ; −x =
dx b1 , b2
+∞
(a1 )(a2 )(a3 ) X (a1 + 1)k (a2 + 1)k (a3 + 1)k (−x)k
=− =
(b1 )(b2 ) (b1 + 1)k (b2 + 1)k k!
k=0
(a1 )(a2 )(a3 ) a1 + 1, a2 + 1, a3 + 1
=− F
3 2 ; −x (128)
(b1 )(b2 ) b1 + 1, b2 + 1
Solving (114) with respect to J2 we have
a
J2 (x) = [2(2D − 5)x + (D − 3)]J1 (x) +
3(D − 2)
2m2 dJ1 (x) 1 2
− x(x + 1) − T (D, m2 ). (129)
3(D − 2) dx 3
Substituting (125) in (129) and because of
d (3 − D), (4 − D)/2, 1 2(D − 3)(D − 4) (4 − D), (6 − D)/2, 2
3 F2 ; −x = 3 F2 ; −x , (130)
dx D/2, (5 − D)/2 D(D − 5) (D + 2)/2, (7 − D)/2
we can write down finally the exact expression in D for J2 (x)
32 M. Argeri & P. Mastrolia
( h i
2(2D − 5)x + (D − 3)
(3 − D), (4 − D)/2, 1
J2 (z) = − 3 F2 ; −x +
6(D − 3) D/2, (5 − D)/2
2(D − 4) (4 − D), (6 − D)/2, 2
+ x(x + 1)3 F2 ; −x +
3D(D − 5) (D + 2)/2, (7 − D)/2
)
1
− (131)
3
(p1 · p2 )
Π2L,1R (D, k 2 ) = c1 + c2 + c3
+c4 + c5 , (132)
with
−2 (−2 + D)
c1 = 2 3 2 2 2 2
−16 + 18D − 7D2 + D3 k 4
(−12 + 19D − 8 D + D ) k m (k + 4 m )
2
+4 −14 + 22D − 9D2 + D3 k 2 m2 + 16 (−3 + D) m4 ; (133)
2
12 (−2 + D) 8 − 5 D + D2 k 2 + 4 10 − 7 D + D2 m2
c2 = ; (134)
(−12 + 19 D − 8 D2 + D3 ) k 2 m2 (k 2 + 4 m2 )
−2 (−2 + D)
c3 = −16 + 18 D − 7 D2 + D3 k 4
(−12 + 19 D − 8 D2 + D3 ) k 2 m2 (k 2 + 4 m2 )
2 2 2 4
+4 (−4 + D) (−1 + D) k m − 32 (−3 + D) m ; (135)
1
c4 = − 2 −32 + 30 D − 9 D2 + D3 k 4
(4 − 5 D + D2 ) k 2 (k 2 + 4 m2 )
2 3
2 2 4
−8 −40 + 38 D − 11 D + D k m + 64 m ; (136)
6 (−2 + D)2 8 − 5 D + D2 k 2 + 4 10 − 7 D + D2 m2
c5 = (137)
(−12 + 19 D − 8 D2 + D3 ) k 2 m2 (k 2 + 4 m2 )
and where we used the mass counterterm defined, in terms of the one-loop tadpole, as,
(D − 1)(D − 2)
δm = (138)
2 m (D − 3)
Feynman Diagrams & Differential Equations 33
The completion of the two-loop renormalization procedure requires the subtraction of the value
of Π(2L,1R) at zero momentum, which can be obtained from the above expression,
34 − D (41 + (−12 + D) D)
Π2L,1R (D, 0) = (139)
2 (−5 + D) (−4 + D) (−3 + D) D
The two-loop renormalized expression
Π2R (D, k 2 ) = Π2L,1R (D, k 2 ) − Π2L,1R (D, 0) (140)
agrees with the result in literature 144 .
p1 p 1 − k1 − k2
k1 k2 k
p2 p 2 + k1 + k2
Fig. 6. Two-loop vertex diagram: p21 = p22 = −m2 ; k = p1 + p2 ; k1,2 are loop variables; a curly line stands for
massless propagator; a solid line, for propagator of mass m.
(k1 · p1 )
Φ1 (D, k 2 ) = ; Φ2 (D, k 2 ) = ; Φ3 (D, k 2 ) = . (142)
34 M. Argeri & P. Mastrolia
These MI are found to fulfills the following system of first-order ODE, in the variable k 2 , corre-
sponding to the momentum transfer,
d 22 − 13 D + 2 D2
=−
dk 2 2 k2
(−2 + D) (−7 + 2 D) (−8 + 3 D) (k1 · p1 )
−
(−3 + D) k 2 (k 2 + 4 m2 )
(−4 + D) (−5 + 2 D) m2
+
2 (−3 + D) (k 2 + 4 m2 )
(−4 + D) (−2 + D) (−5 + 2 D)
+
2 (−3 + D) k 2 (k 2 + 4 m2 )
(−4 + D) (−5 + 2 D) (−10 + 3 D) (−8 + 3 D)
− ; (143)
4 (−3 + D) (−7 + 2 D) k 2 (k 2 + 4 m2 )
(k1 · p1 )
d (−3 + D) (−4 + D) k 2 + 4(−3 + D) m2
=
dk 2 8 k2
(20 − 16 D + 3 D2 ) k 2 + (56 − 56 D + 12 D2 ) m2 (k1 · p1 )
+
4 k 2 (k 2 + 4 m2 )
(−4 + D) m2
−
8
10 − 6 D + D2
−
8 k2
−236 + 244 D − 82 D2 + 9 D3
+ ; (144)
16 (−7 + 2 D) k 2
d (−4 + D) (−3 + D) (−2 + D) k 2 + 4 (−3 + D) m2
=−
dk 2 4 k 4 m2
(−4 + D) (−2 + D) (−8 + 3 D) (k1 · p1 )
− 4 2
2 k m
4 − 6 D + D2 k 2 + 4 −2 − 4 D + D2 m2
+
4 k 2 (k 2 + 4 m2 )
(−2 + D) 14 − 8 D + D2 k 2 + 4 8 − 6 D + D2 m2
+
4 k 4 m2 (k 2 + 4 m2 )
(−8 + 3 D)
− 4 2 2 2
−164 + 136 D − 36 D2 + 3 D3 k 2 +
8 (−7 + 2 D) k m (k + 4 m )
+4 −104 + 98 D − 30 D2 + 3 D3 m2 . (145)
After the Laurents expansion, the first-order ODE for Φ3 (D, x) will induce, order-by-order in (D−4),
(k)
a system of chained first-order ODE for the Laurent coeffiecients, Φ3 (x), which at the first two
orders, k = −2, −1, reads,
( )
1 + x2 d (−2)
0= + Φ3 (x) ; (152)
(−1 + x) x (1 + x) dx
(−1)
(2) solve Eq.(153) to find Φ3 (x);
(−1)
(3) solve Eq.(155) to find Φ1 (x);
(−1)
(4) invert the D → 4 expansion Eq.(143), and substitute the expressions of Φ3 (x) and
(−1) (−1)
Φ1 (x) in it, to find Φ2 (x).
The construction of Laurent coefficients for k ≥ 0 goes on by repeating the steps (1–4),
• k = j, j ≥ 0.
(i) (i) (i)
(1) plug the results of the previous orders, Φ1 (x), Φ2 (x) and Φ3 (x) (−2 ≤ i < j), in the
non-homogeneous term of the equations;
(j)
(2) solve the first order equation for Φ3 (x);
(j)
(3) solve the second order equation for Φ1 (x);
(j)
(4) invert Eq.(143) to find Φ2 (x).
Let us see in detail the case k = −2.
(i,−2)
with Φ1 , (i = 1, 2) being integration constants, and K (−2) (x) the non-homogeneous term of
Eq.(154), thus obtaining,
At the end of this three steps, one knows the expressions of the 1/(D −4)2-coefficient of the three MI,
(1,−2) (1,−2) (2,−2)
Φi (D, x), (i = 1, 3) up to the determination of the real integration constants, Φ3 , Φ1 , Φ1 .
(−2) y (1,−2)
Φ3 (y) = − Φ (164)
(−1 + y) (1 + y) 3
(−2) −y(1 + H(0, y)) y (1,−2)
Φ1 (y) = + Φ +
4 (−1 + y) (1 + y) (−1 + y) (1 + y) 1
(2,−2)
−1 + y 2 + 2 y H(0, y) (2,−2) y 1 + 8Φ1
− Φ1 − iπ (165)
(−1 + y) (1 + y) 4 (−1 + y) (1 + y)
Feynman Diagrams & Differential Equations 39
Finally, the three conditions needed to fix the value of the arbitrary constants, order by order in
(k) (k)
D − 4, are: i) the regularity of Φ3 (y) as y → 1; ii) the regularity of Φ1 (y) as y → 1, meaning
the vanishing of the 1/(1 − y)-coefficients in both cases; iii) the realness of the constants, meaning
that real and imaginary part of the 1/(1 − y)-coefficient must vanish separately. From the above
equations, that translates to,
(1,−2) 1
(1,−2) (2,−2) 1
Φ3 = 0; Φ1 ; Φ1 = =− . (168)
4 8
We have all the ingredient to go up in the chain of Laurent coefficients, and repeat the previous
steps for the case k = −1. The iterative structure of the method yields a bottom-up reconstruction
of the three master integrals, Φ1 (D, x), Φ2 (D, x), and Φ3 (D, x), around D = 4.
P P
P− i ki
k2
k4
Fig. 7. Four-loop sunrise diagram: P 2 = −m2 ; k1,2,3,4 are loop variables; a wavy line stands for massless propagator;
a solid line, for propagator of mass M ; a dashed line, for propagator of mass m;
In this case, the heart of the analytic calculation is the study of a homogeneous fourth order
differential equation, whose solutions turned out to be, in a remarkably simple way, either a rational
fraction or repeated quadratures of rational fractions. The required four-loop integral could then be
obtained almost mechanically by repeated quadratures in terms of HPL’s.
Following the reduction algorithm - by now sounding familiar to the reader -, we identify the
MI’s of the problem; write the system of differential equations in x = m/M satisfied by the MI’s;
convert it into a higher order differential equation for a single MI; Laurent-expand it around D = 4;
solve the associated homogeneous equation at D = 4 and then use recursively Euler’s method of the
variation of the constants for obtaining the coefficient of the (D − 4)-expansion in closed analytic
40 M. Argeri & P. Mastrolia
form. The result involves HPL’s of argument x and weight increasing with the order in (D − 4). The
integration constants are fixed at x = 0 . After solving the differential equations for arbitrary x, we
will show how to compute, independently, the numerical value of the solution at x = 1 by using the
Finite Difference method of Laporta 65 , discussed in Sec.5, to show the relation among Differential
and Difference Equations for Feynman integrals.
The system is homogeneous: indeed, quite in general the non homogeneous terms are given by
the MI’s of the subtopologies of the considered graph, obtained by shrinking to a point any of its
propagator lines. When that is done for the parent topology with five propagators of Fig.(7), one
obtains the product of 4 tadpoles; but as the considered graph has two massless propagators, at
least one massless tadpole is always present in the product; as in the D-dimensional regularization
massless tadpoles vanish, the product of the 4 tadpoles is always equal to zero – and therefore the
differential equations are homogeneous.
By inspection, one sees that Φ3 (D, x), Φ5 (D, x) appear in the r.h.s. of Eq.s(171-175) only in the
combination
decouples and can be expressed in terms of the other integrals by means of the trivial 1st order
differential equation
( )
dΨ5 (D, x) 3 (D − 2) 3 (D − 2)
=− − Φ1 (D, x) − 3Φ2 (D, x) − Φ4 (D, x)
dx 2 (1 − x) 2 (1 + x)
( )
5(D − 2) 9 (D − 2) 9 (D − 2)
− + − Ψ3 (D, x) (178)
x 2 (1 − x) 2 (1 + x)
As Ψ5 (D, x) does not enter in the r.h.s. of Eq.s(171-175), the 4 linear equations for Φ1 (D, x),
Φ2 (D, x), Ψ3 (D, x), and Φ4 (D, x) can be written as a fourth order equation for Φ1 (x), which will be
called simply Φ(D, x) from now on, and which is therefore equal to
Z D
C −4 (D) d k1 dD k2 dD k3 dD k4
Φ(D, x) = , (p2 = −x2 ) . (179)
(2π)4(D−2) D1 D2 D3 D4 D5
. (180)
42 M. Argeri & P. Mastrolia
4 ∞
!
X X
αi
Φ(D, x) = x A(i)
n (D)x
2n
, (181)
i=1 n=0
α1 = 0 ,
α2 = (D − 2) ,
α3 = −(D − 2) ,
α4 = (3D − 7) ; (182)
(i) (i)
the A0 (D) are the 4 integration constants, and all the other coefficients An (D) for n > 0 are
determined by the differential equation Eq.(180), once the integration constants are fixed.
A qualitative inspection of the integrals which one tries to evaluate by means of the differential
equation Eq.(179) shows that it is finite (just finite, not analytic!) for x → 0+ and (D − 2) > 0; that
is sufficient to rule out from their expression as solutions of the differential equation the terms with
the behaviour of the third and the fourth exponent (which is negative when D is just above 2).
In the current case, as the equation for Φ(D, x) is homogeneous, the only information one gets out
(3) (4)
is that A0 (D) and A0 (D) are both equal to zero, due to the finiteness for x → 0+ ; by substituting
(3) (4)
the ansatz Eq.(181) in Eq.(180) and dropping A0 (D), A0 (D), one finds for Φ(D, x) Eq.(179) the
x → 0 expansion
(1) 2(2D − 5)(3D − 8) 2
Φ(D, x) = A0 (D) 1 − x + O(x4 )
3D(D − 4)
(2) (D − 3)(D − 4)(3D − 8) 2
+ A0 (D) xD−2 1 + x + O(x4 ) . (183)
2D(2D − 7)
(1) (2)
The expansion depends on the two as yet unspecified integrations constants A0 (D), A0 (D). To fix
them, one has to provide some independent information, such as the value of the required Feynman
integral and of its first derivative at x = 0. Those value can be provided by an explicit conventional
calculation, say in parameter space, which is in any case much easier than a calculation for non-zero
Feynman Diagrams & Differential Equations 43
Fig. 8. Four-loop watermelon diagram: a wavy line stands for massless propagator; a solid line, for propagator of
mass M .
values of the variable x b . That is done explicitly in 147 , and the results are
(1) 3D − 11
A0 = −
8(D − 2)(D − 3)(D − 4)3 (2D − 5)(2D − 7)(3D − 8)(3D − 10)
Γ(1 − (D − 4))Γ(1 − 2(D − 4))Γ2 1 + 21 (D − 4) Γ2 1 − 32 (D − 4)
×
Γ4 1 − 21 (D − 4) Γ(1 − 3(D − 4))
(2) 2 (2D − 7)
A0 = −
3(D − 2)2 (D
− 3)(D − 4)4 (3D − 8)(3D − 10)
Γ 1 + 2 (D − 4) Γ 1 − 32 (D − 4) Γ2 (1 − (D − 4))
1
× , (184)
Γ2 1 − 21 (D − 4) Γ(1 − 2(D − 4))
(1)
where the term A0 is the value of the vacuum graph in Fig.8.
as it is known on general grounds that it develops at most a fourth order pole in (D − 4). By
substituting in Eq.(180) one obtains a system of inhomogeneous, chained equations for the coefficients
Φ(n) (x) of the expansion in (D − 4); the generic equation reads
with
3 d4 22 2 d
3
2 d
2
2 d
D = x (1 − x ) 4 + x (1 + 5x ) 3 − 6x(2 + x ) 2 + 6(2 − 3x ) + 48x , (187)
dx dx dx dx
b In the present case the knowledge of the regularity of the solution at x = 1 does not provide any additional
information.
44 M. Argeri & P. Mastrolia
and
2 d
(n)
K (x) = 24x + 3 + 24x Φ(n−3) (x)
dx
2 d
2
3 d
+ 92x + 16 + 40x − x + 26x Φ(n−2) (x)
dx dx2
2 d
2
3 d 2
3
4 d
+ 116x + 25 − 2x − 13x + 32x − 3x − 9x Φ(n−1) (x)
dx dx2 dx3
, (188)
which shows that the equation at a given order n for Φ(n) (x) involves in the inhomogeneous term
the coefficients Φ(k) (x) (and their derivatives) with k < n (obviously Φ(k) (x) = 0 when k < −4).
Such a structure calls for an algorithm of solution bottom-up, i.e. starting from the lowest value of
n (which is n = −4) and proceeding recursively to the next n + 1 value up to the required order.
The Eq.s(186) have all the same associated homogeneous equation, independent of n,
4 3 2
3 2 d 2 2 d 2 d 2 d
x (1 − x ) 4 + x (1 + 5x ) 3 − 6x(2 + x ) 2 + 6(2 − 3x ) + 48x φ(x) = 0 ; (189)
dx dx dx dx
once the solutions of Eq.(189) are known, all the Eq.s(186) can be solved by the method of the
variation of the constants of Euler.
To our (pleasant) surprise, the solutions of Eq.(189) are almost elementary. By trial and error, a
first solution is found to be
φ1 (x) = x2 . (190)
We then substitute φ(x) = φ1 (x)ξ(x) in Eq.(189), obtaining the following 3rd order equation for the
derivative of ξ(x)
3 2
2 d 2 d 2 d
3 2
x (1 − x ) 3 + 3x (3 − x ) 2 + 6x(1 + 2x ) − 6(5 + 2x ) ξ ′ (x) = 0 ,
2
(191)
dx dx dx
and find that it admits the solution
1
ξ2′ (x) = (1 − x2 + x4 ) . (192)
x3
Substituting ξ ′ (x) = ξ2′ (x)χ(x) in Eq.(191) we obtain the following 2nd order equation for χ′ (x)
d2 d
x2 (1 − x2 )(1 − x2 + x4 ) 2 + 6x5 (2 − x2 ) − 6(2 − 2x4 + x6 ) χ′ (x) = 0 , (193)
dx dx
which admits as solution
1 5 − 2x2 + 5x4
χ′3 (x) = 3
(1 − x2 )4 . (194)
x (1 − x2 + x4 )2
Finally, substituting χ′ (x) = χ′3 (x)τ (x) in Eq.(193), we obtain the equation
d
x(1 − x2 )5 (1 − x2 + x4 )(5 − 2x2 + 5x4 )
dx
−2(1 − x2 )4 (15 − 12x2 + 11x4 + 30x6 − 24x8 + 20x10 ) τ ′ (x) = 0 , (195)
Feynman Diagrams & Differential Equations 45
1
φ2 (x) = − (1 − x4 ) − H(0, x)x2 , (197)
2
in agreement (of course) with the coefficients of the 4th and 3rd x-derivative of φ(x) in Eq.(189).
(n)
where the φi (x) are the solutions of the homogeneous equation given in Eq.s(190,199), the Φi are
the as yet undetermined integration constants, the Wronskian W (x) can be read from Eq.(200),
′′′
the Mi (x) are the minors of the φi (x) in the determinant Eq.(200), and the K (n) (x) are the
(n)
inhomogeneous terms of Eq.(188). The constants Φi are then fixed by comparing the expansion in
46 M. Argeri & P. Mastrolia
1 2
Φ(−4) (x) = − x , (202)
64
1 9 2 1 4 1
Φ(−3) (x) = − + x − x − x2 H(0; x) . (203)
384 256 192 48
(204)
The full results become quickly too lengthy to be reported explicitly here.
4
I5 (D, 1) = [4Γ(1 + ǫ)] Φ(D = 4 − 2ǫ, x = 1) . (206)
By combining identities obtained by integration by parts one finds that I5 (D, n) satisfies the third-
order difference equation
2 2 d2
+(46d + 38D − 144)t + 71D − 284D + 288 dt2 v5 (D, t)
+t (576D − 960)t3 + (−216D2 + 360D)t2
3 2 3 2 d
+ (−18D + 190D − 496D + 384)t + 77D − 533D + 1236D − 960) dt v5 (D, t)
first denominator of Eq.(212) depends on k1 ; expanding for small k1 and performing the angular
integration one gets
Z Z
dΩ(D, k̂1 ) 1 dΩ(D, k̂1 ) 1 k12 − 2k1 · k2 (k12 − 2k1 · k2 )2
= − + + ...
Ω(D) (k2 − k1 )2 + 1 Ω(D) k22 + 1 (k22 + 1)2 (k22 + 1)3
1 1 4 k22
= 2 + k12 − 2 + + ... .
k2 + 1 (k2 + 1)2 D (k22 + 1)3
(215)
The above result gives the expansion of f5 (D, k12 ) at k12 = 0:
(0) (1)
f5 (D, k12 ) = f5 (D) + f5 (D) k12 + O(k14 ) ,
(0)
f5 (D) = I4 (D, 1, p2 ) (216)
(1) 4
2 2 2
f5 (D) = − I4 (D, 2, p ) + I4 (D, 2, p ) − I4 (D, 3, p ) .
D
Note that f5 (D, k12 ) is regular in the origin.
By inspecting the differential equation (209) one finds that the behaviour at t = 1 of the 4
independent solution is ∼ (1 − t)αi , with α1 = D/2 − 1, α2 = D/2, α3 = 0, and α4 = 1; for
comparison with Eq.(214) the behaviours α3 = 0, and α4 = 1 are ruled out and the expansion reads
D (0) (1)
v5 (D, t) = (1 − t) 2 −1 v5 (D) + v5 (D) (1 − t) + O(1 − t)2 ; (217)
d3
2t3 (3t + 1)(t − 1)2 v4 (d, t)
dt3
d2
+t2 (t − 1)(36t2 + (6 − 7D)t − 9D + 18) v4 (D, t)
dt2
d
+t(36t3 − 14Dt2 + (−7D2 + 33D − 36)t + 13D2 − 61D + 72) v4 (D, t)
dt
+(D − 3)(2D − 5)(3D − 8)v4 (D, t) = 0 . (222)
At variance with the previous case, the function f4 (D, k22 ) is not regular for k2 → 0, as at k2 = 0
the value of the external momentum squared (p − k2 )2 becomes the threshold of the 2-loop sunrise
graph associated to I3 (D, p2 ). But it is not difficult to evaluate analytically I3 (D, q 2 ) for generic
off-shell q 2 by using Feynman parameters:
2 D
2 2Γ(5 − D)Γ 3 − D 2 Γ 2 − 1 D D 2
I3 (D, q ) = 2 F1 3 − D, 2 − ; ; −q ,
(D − 4)2 (3 − D)Γ D 2
2 2
2 D
2Γ(5 − D)Γ 3 − D 2 Γ 2 −1 Γ(2D − 5)
a0 (D) = I3 (D, −1) = ,
(4 − D)2 (3 − D)Γ 23 D − 3 Γ(D − 2)
D
b0 (D) = Γ2 − 1 Γ(5 − 2D) .
2
Inserting Eq.(227) into Eq.(224), setting q = p − k2 and performing the angular integration over k̂2
by means of the formula (see Eq.(88) of Ref.65 ) valid for k2 → 0
Z
1 dΩ(D, k̂2 ) 2 −N Γ D 2 Γ 2
N
≈ (k2 ) 2 , k2 → 0 ; (228)
Ω(D) ((p − k2 )2 + 1)N 2Γ(N )Γ 21 (D − N )
50 M. Argeri & P. Mastrolia
Using the variable 1/(k22 +1) = t in Eq.(229) and inserting it in Eq.(226) one gets the initial condition
for v4 (D, t) at the singular point t = 1
a0 (D) D
v4 (D, t) = (1 − t) 2 −1 1 + O(1 − t)
Γ D 2
Γ 25 − D 1
(3D−7)
+ b0 (D)(1 − t) 2 1 + O(1 − t) . (230)
2Γ(5 − 2D)Γ 21 (3D − 5)
By inspecting the equation (222) one gets that the behaviour at t = 0 of v4 (D, t) is
(1) (2) (3) 1
v4 (D, t → 0) ≈ c4 (D) t−D+3 + c4 (D) t−2D+5 + c4 (D) t 2 (−3D+8) , (231)
so that, when D → 4, the integral (208) is surely convergent for n ≥ 5; then we calculate the integral
(208) for n = 5, 6, 7, 8, 9 by integrating the series term by term. By using repeatedly top-down
the recurrence relation (220) starting from n = 9, we finally obtain I5 (D, 6), I5 (D, 5), I5 (D, 4), . . .,
I5 (D, 1). By taking into account the normalization (206) one finds complete agreement with the
value at x = 1 of solution of the differential equation computed in the previous section.
10. Conclusions
The evaluation of multiloop Feynman diagrams in the last years has received a strong boost, thanks
to the ability of turning generic properties of scalar integrals in dimensional regularization into
tools for computing them. Integration-by-parts, Lorentz invariance, and kinematic symmetries have
been exploited to establish infinite sets of relations among integrals sharing (partially) common
integrands. The Laporta algorithm systematizes the by now standard reduction to Master Integrals,
Feynman Diagrams & Differential Equations 51
that is the algebraic procedure for expressing any Feynman integral as a linear combination of few
basic integrals with the simplest integrands.
The completion of the computational task, consisting in the actual evaluation of the Master
Integrals can be as well afforded by employing the same set of identities. In fact, by combining the
differentiation of Master Integrals with respect to their Mandelstam invariants, and the reduction of
the new born integrals, it is possible to derive a system of non-homogeneous first order differential
equations fulfilled by the Master Integrals themselves.
Solving such a system of differential equation amounts to the evaluation of the Master Integrals,
alternatively to their direct parametric integration.
We have reviewed the method of differential equations by its direct application, trying to follow
a didactical path. We discussed the reduction algorithm plus the general derivation of differential
equations for Feynman integrals. Successively, the calculation of Master Integrals in the context of
the evaluation of the one- and two-loop corrections to the photon propagator in QED; whereas,
in the last two sections, we presented two cases of less trivial differential equations, to show more
technical aspects related to the solution of homogeneous equations and to the choice of the boundary
conditions.
In general, solving a system of first order differential equations for more than one Master Integral
is equivalent to solving a higher order equation for a single Master Integral. Despite to the lack of
a theoretical procedure for solving differential equations in the most general case, Euler’s variation
of constants offers a viable procedure. Accordingly, the solution of the non-homogeneous equation
is obtained by quadrature, using as a kernel the Wronskian of the associated homogeneous equation
– whose solution can be preliminarily found by constants’ variation as well.
The main achievement is the integration of the differential equation for exact value of the pa-
rameters (Mandelstam invariants and dimensional-parameter). When that is not possible, one can
Laurent expand the equation, which then becomes a chained system of equations for the Laurent
coefficients of the solution, suitable for a bottom-up solving algorithm, starting from the lowest
Laurent coefficient.
As a natural feature of Euler’s variation of constants, the solution manifests an analytic integral
representation, generic of transcendental functions: a flexible nested structure of multiple integra-
tions (or equivalently, iterative summations) which benefits of the shuffle algebras induced by the
integration-by-parts. Within this framework, the actual efforts required by the computation are
the finding out of the homogeneous solutions, and the definition of new occurring functions, or-
dered according to their increasing trascendentality – as required by the iterative fulfilment of the
non-homogeneous equation.
Boundary conditions are found by imposing the regularity or the finiteness of the solution at the
pseudo-thresholds of the Master Integrals. This qualitative information is sufficient for the quan-
titative determination of the arbitrary integration constants. At the diagrammatic level, boundary
conditions usually correspond to integrals related to simpler diagrams.
The use of Differential Equation in the external invariants is a very powerful tool for computing
Feynman (master) integrals. Dimensional regularization was fundamental for the derivation of the
differential equations we discussed in this review.
In principle differential identities for integral functions can be derived whenever it is allowed
52 M. Argeri & P. Mastrolia
by the algebra of the integral representation under use - as induced by integration-by-parts. And
their use is not limited to the perturbative description of Feynman diagrams. Therefore we like to
conclude by remarking that the use of Differential Equations for integrals’ evaluation, is not just a
technique, but a point of view from which any integral is seen under a new perspective, where there
appear, explicitly exposed, its analytic structure, its singularities which finally determine its value.
11. Acknowledgment
We wish to thank Roberto Bonciani, Thomas Gehrmann, and Ettore Remiddi, for their careful
reading and comments on the manuscript, but especially for their invaluable collaboration during
these years. The research of PM was supported by Marie-Curie-EIF under the contract MEIF-CT-
2006-024178. MA wishes to anknowledge the Institute for Theoretical Physics of Zürich for the
kind hospitality while part of this work was performed, with the support of the abovementioned
Marie-Curie-EIF founds.
References
1. A. V. Kotikov, Phys. Lett. B 254 (1991) 158.
2. E. E. Boos and A. I. Davydychev, Theor. Math. Phys. 89 (1991) 1052 [Teor. Mat. Fiz. 89 (1991) 56].
3. A. V. Kotikov, ITF-91-26E
4. A. V. Kotikov, Phys. Lett. B 259 (1991) 314.
5. F. A. Lunev, Phys. Rev. D 50 (1994) 6589 [arXiv:hep-th/9407174].
6. R. Scharf, PhD theis, Wurzburg 1994.
7. E. Remiddi, Nuovo Cim. A 110 (1997) 1435 [arXiv:hep-th/9711188].
8. T. Gehrmann and E. Remiddi, Nucl. Phys. B 580, 485 (2000) [arXiv:hep-ph/9912329].
9. T. Gehrmann and E. Remiddi, Nucl. Phys. B 601, 248 (2001) [arXiv:hep-ph/0008287].
10. T. Gehrmann and E. Remiddi, Nucl. Phys. B 601, 287 (2001) [arXiv:hep-ph/0101124].
11. C. Ford and D. R. T. Jones, Phys. Lett. B 274 (1992) 409 [Erratum-ibid. B 285 (1992) 399].
12. M. Caffo, H. Czyz, S. Laporta and E. Remiddi, Acta Phys. Polon. B 29, 2627 (1998)
[arXiv:hep-th/9807119].
13. M. Caffo, H. Czyz, S. Laporta and E. Remiddi, Nuovo Cim. A 111, 365 (1998) [arXiv:hep-th/9805118].
14. R. Bonciani, Acta Phys. Polon. B 30 (1999) 3463.
15. A. V. Kotikov, arXiv:hep-ph/0102178.
16. M. Czachor and H. Czyz, Acta Phys. Polon. B 32, 3823 (2001) [arXiv:hep-ph/0110351].
17. M. Caffo, H. Czyz and E. Remiddi, Nucl. Instrum. Meth. A 502, 613 (2003) [arXiv:hep-ph/0211171].
18. M. Caffo, H. Czyz and E. Remiddi, Nucl. Phys. Proc. Suppl. 116, 422 (2003) [arXiv:hep-ph/0211178].
19. P. Mastrolia and E. Remiddi, Nucl. Phys. Proc. Suppl. 116, 412 (2003) [arXiv:hep-ph/0211210].
20. U. Aglietti, arXiv:hep-ph/0408014.
21. R. Bonciani, Acta Phys. Polon. B 35, 2587 (2004) [arXiv:hep-ph/0410210].
22. V. A. Smirnov, “Feynman integral calculus,” Berlin, Germany: Springer (2006) 283 p
23. M. Caffo, Acta Phys. Polon. B 34, 5357 (2003) [arXiv:hep-ph/0311052].
24. M. Caffo, H. Czyz, A. Grzelinska and E. Remiddi, Nucl. Phys. B 681, 230 (2004)
[arXiv:hep-ph/0312189].
25. J. A. Gracey, Phys. Lett. B 277 (1992) 469.
26. O. V. Tarasov, Phys. Rev. D 54 (1996) 6479 [arXiv:hep-th/9606018];
Nucl. Phys. Proc. Suppl. 89 (2000) 237 [arXiv:hep-ph/0102271].
27. O. V. Tarasov, Nucl. Phys. B 502 (1997) 455 [arXiv:hep-ph/9703319].
28. J. Fleischer, A. V. Kotikov and O. L. Veretin, Phys. Lett. B 417 (1998) 163 [arXiv:hep-ph/9707492].
Feynman Diagrams & Differential Equations 53
58. L.W. Garland, T. Gehrmann, E.W.N. Glover, A. Koukoutsakis and E. Remiddi, Nucl. Phys. B 627
(2002) 107 [hep-ph/0112081] and 642 (2002) 227 [hep-ph/0206067]. T. Gehrmann and E. Remiddi,
Nucl. Phys. B 640 (2002) 379 [hep-ph/0207020].
91. C. Anastasiou, L. J. Dixon, K. Melnikov and F. Petriello, Phys. Rev. D 69 (2004) 094008
[arXiv:hep-ph/0312266].
92. K. G. Chetyrkin, J. H. Kuhn, P. Mastrolia and C. Sturm, Eur. Phys. J. C 40 (2005) 361
[arXiv:hep-ph/0412055].
93. Y. Schroder and A. Vuorinen, JHEP 0506, 051 (2005) [arXiv:hep-ph/0503209].
94. M. Faisst, P. Maierhoefer and C. Sturm, arXiv:hep-ph/0611244.
95. M. Neubert, Phys. Rept. 245 (1994) 259 [arXiv:hep-ph/9306320].
96. A. Juste et al., arXiv:hep-ph/0601112.
97. W. Bernreuther, R. Bonciani, T. Gehrmann, R. Heinesch, T. Leineweber, P. Mastrolia and E. Remiddi,
PoS HEP2005, 229 (2006) [arXiv:hep-ph/0601207].
98. W. Bernreuther, R. Bonciani, T. Gehrmann, R. Heinesch, T. Leineweber, P. Mastrolia and E. Remiddi,
arXiv:hep-ph/0604031.
99. R. Bonciani, arXiv:hep-ph/0607037.
100. U. Aglietti and R. Bonciani, Nucl. Phys. B 668, 3 (2003) [arXiv:hep-ph/0304028].
101. U. Aglietti and R. Bonciani, Nucl. Phys. B 698, 277 (2004) [arXiv:hep-ph/0401193].
102. U. Aglietti, R. Bonciani, L. Grassi and E. Remiddi, arXiv:0705.2616 [hep-ph].
103. P. N. Maher, L. Durand and K. Riesselmann, Phys. Rev. D 48 (1993) 1061 [Erratum-ibid. D 52 (1995)
553] [arXiv:hep-ph/9303233].
104. U. Aglietti, R. Bonciani, G. Degrassi and A. Vicini, Phys. Lett. B 595, 432 (2004)
[arXiv:hep-ph/0404071].
105. U. Aglietti, R. Bonciani, G. Degrassi and A. Vicini, Phys. Lett. B 600, 57 (2004)
[arXiv:hep-ph/0407162].
106. W. Bernreuther, R. Bonciani, T. Gehrmann, R. Heinesch, P. Mastrolia and E. Remiddi, Phys. Rev. D
72, 096002 (2005) [arXiv:hep-ph/0508254].
107. C. Anastasiou, S. Beerli, S. Bucherer, A. Daleo and Z. Kunszt, arXiv:hep-ph/0611236.
108. U. Aglietti, R. Bonciani, G. Degrassi and A. Vicini, arXiv:hep-ph/0611266.
109. V. A. Smirnov, Phys. Lett. B 524, 129 (2002) [arXiv:hep-ph/0111160].
110. R. Bonciani, A. Ferroglia, P. Mastrolia, E. Remiddi and J. J. van der Bij, Nucl. Phys. B 681, 261
(2004) [Erratum-ibid. B 702, 364 (2004)] [arXiv:hep-ph/0310333].
111. R. Bonciani, A. Ferroglia, P. Mastrolia, E. Remiddi and J. J. van der Bij, Nucl. Phys. B 701, 121
(2004) [arXiv:hep-ph/0405275].
112. G. Heinrich and V. A. Smirnov, Phys. Lett. B 598, 55 (2004) [arXiv:hep-ph/0406053].
113. R. Bonciani, A. Ferroglia, P. Mastrolia, E. Remiddi and J. J. van der Bij, Nucl. Phys. B 716, 280
(2005) [arXiv:hep-ph/0411321].
114. M. Czakon, J. Gluza and T. Riemann, Phys. Rev. D 71, 073009 (2005) [arXiv:hep-ph/0412164].
115. R. Bonciani and A. Ferroglia, Phys. Rev. D 72, 056004 (2005) [arXiv:hep-ph/0507047].
116. R. Bonciani and A. Ferroglia, Nucl. Phys. Proc. Suppl. 157, 11 (2006) [arXiv:hep-ph/0601246].
117. M. Czakon, J. Gluza, K. Kajda and T. Riemann, Nucl. Phys. Proc. Suppl. 157, 16 (2006)
[arXiv:hep-ph/0602102].
118. D. Seidel, Phys. Rev. D 70, 094038 (2004) [arXiv:hep-ph/0403185].
119. K. Melnikov and A. Mitov, Phys. Lett. B 620, 69 (2005) [arXiv:hep-ph/0505097].
120. H. M. Asatrian, T. Ewerth, A. Ferroglia, P. Gambino and C. Greub, Nucl. Phys. B 762 (2007) 212
[arXiv:hep-ph/0607316].
121. F. De Fazio, arXiv:hep-ph/0010007.
122. F. Jegerlehner, M. Y. Kalmykov and O. Veretin, Nucl. Phys. B 658 (2003) 49 [arXiv:hep-ph/0212319].
123. M. Awramik, M. Czakon, A. Freitas and G. Weiglein, Phys. Rev. Lett. 93, 201805 (2004)
[arXiv:hep-ph/0407317].
124. M. Awramik, M. Czakon, A. Freitas and G. Weiglein, Nucl. Phys. Proc. Suppl. 135, 119 (2004)
[arXiv:hep-ph/0408207].
125. A. Freitas, M. Awramik and M. Czakon, arXiv:hep-ph/0507159.
56 M. Argeri & P. Mastrolia