Intro To Calc of Feyn Integral

Introduction to the calculation of Feynman Integrals
Peter Marquard
DESY
CAPP 2015
1 / 109
Outline
1 Introduction
2 Mathematical prelude
3 Direct evaluation of Feynman integrals
4 Tensor integrals
5 Integration by parts
6 Mellin Barnes
7 Sector decomposition
8 Differential equations
9 Asymptotic expansions
2 / 109
Introduction
Outline
1 Introduction
4 Tensor integrals
6 Mellin Barnes
3 / 109
Introduction
Introduction
Feynman integrals come in different shapes and colors

one loop ↔ many loops
many legs ↔ no legs
many scales ↔ no scales
for many types of diagrams many results are known
especially one-loop is solved
at three loops and more, massive tadpoles and massless
propagators have been studied in great detail
at two loops, much progress has been made for integrals relevant
for 2 → 2 scattering processes, but every new process requires a
new study of the integrals involved
4 / 109
Introduction
Introduction
many different methods have been invented over the years to

calculate the needed integrals
most methods work well for certain classes but fail for others
the ultimate method/tool is still missing
5 / 109
Mathematical prelude
Outline
1 Introduction
4 Tensor integrals
6 Mellin Barnes
6 / 109
Gamma function
Defining property
zΓ(z) = Γ(z + 1) = z!
Integral representation Z ∞
Γ(z) = t z−1 e−t dt
0
We see immediately from the properties that Γ(−n) is singular for
n = 0, 1, 2, . . ..
The singularities are simple poles
(−1)n 1
Γ(−n + ) = + O(0 )
n!
7 / 109
Series expansion of the Gamma function
In many applications one needs more than the pole of the Γ-function.
This is best done by using the derivative of log Γ(z) and defines the
digamma function Ψ(z)
d log(Γ(z)) Γ0 (z)
Ψ(z) = =
dz Γ(z)
Therefore,
Γ(z − z0 ) = Γ(z0 ) + Γ(z0 )Ψ(z0 )(z − z0 )

1
+ Γ(z0 )Ψ0 (z0 ) + Γ(z0 )Ψ2 (z0 ) (z − z0 )2
2
for regular points z0 .
8 / 109
The Digamma-function Ψ(z)
The digamma function satisfies the relation
1
Ψ(z + 1) = Ψ(z) +
z
For positive integer values the digamma function evaluates to
Ψ(1) = −γE ,
Ψ(2) = 1 − γE
···
n
X 1
Ψ(n + 1) = − γE
k
k=1
with the Euler-Mascheroni constant γE = 0.577216 . . .
9 / 109
The Digamma-function Ψ(z) cont’d
For non-positive integers the digamma functions evaluates again to

simple poles
1
Ψ(−n + ) = − + O(0 )

Γ(z) Ψ(z)
6
10
4
5
2
-2 -1 1 2 3 4 -2 -1 1 2 3 4
-2
-5
-4
- 10
-6
10 / 109
Schwinger Parametrization
From the definition of the Γ function follows immediately the Schwinger

Parametrization
Z ∞
1 1 2 2
= dt t α−1 e−(M −k )t
(−k 2 + M 2 )α Γ(α) 0
by performing the substitution t → t 0 = (M 2 − k 2 )t
11 / 109
Polylogarithms
Function classes appearing in Feynman integral calculations are the

Polylogarithms
∞
X zk
Lin (z) =
kn
k=1
which for z → 1 give the ζ values

∞
X 1
ζn =
kn
k=1
Lin (1) = ζn
π π4
ζ2 = , ζ3 = 1.2020 . . . , ζ4 = , · · · , ζ2n ∝ π 2n
6 90
12 / 109
Polylogarithms
The appear in the calculations from integrals over logarithms

Z
dx log x/(1 − x) = Li2 (1 − x)
Z
dx Li2 (x)/x = Li3 (1 − x)
Z
dx Li2 (1 − x)/x = −2Li3 (x) + Li2 (1 − x) log(x)
+2Li2 (x) log(x) + log(1 − x) log2 (x)
The Polylogarithms are not very systematic and many relations

between them exist.
13 / 109
Harmonic Polylogarithms
A more systematic approach are the harmonic polylogarithms (HPLs)

defined as iterated integrals over the alphabet
1 1 1
f−1 = , f0 = , f1 =
1+x x 1−x
Z x
dx 0 HPL(~n; x 0 )fa (x 0 ) = HPL(a, ~n; x)
0
HPL(1, x) = − log(1 − x), HPL(−1, x) = log(1 + x)

1
HPL(0, x) = log(x), HPL(0. . . . , 0, x) = logn (x)
n!
14 / 109
The Minkowski Metric
gµν = diag(1, −1, −1, −1)
15 / 109
Direct evaluation of Feynman integrals
Outline
1 Introduction
4 Tensor integrals
6 Mellin Barnes
16 / 109
Simple example
Consider simplest diagram possible,

the one-loop tadpole (vacuum diagram)
Z
1
I1 = d4 k
−k + M 2
2
Either introduce an explicit parametrization of the measure or use the

Schwinger parametrization for α = 1
Z Z ∞
2 2
I1 = d k 4
dt e−(M −k )t
0
17 / 109
Simple example
Consider simplest diagram possible,

the one-loop tadpole (vacuum diagram)
Z
1
I1 = d4 k
−k + M 2
2
Either introduce an explicit parametrization of the measure or use the

Schwinger parametrization for α = 1
Z ∞ Z
−M 2 t 2
I1 = dt e d4 k ek t
0
18 / 109
Simple example
perform Wick rotation

k0 → ik0
with the result that
k 2 = k02 − k12 − k22 − k32 → −k02 − k12 − k22 − k32
and we get Z ∞ Z
−M 2 t 2t
I1 = i dt e d4 k e−k
0
19 / 109
Simple example
Z ∞ Z
−M 2 t 2t
I1 = i dt e d4 k e−k
0
doing the Gaussian integral we get
∞ 2t
e−M
Z
I1 = iπ 2 dt
0 t2
This integral does not converge for t → 0 .

⇒ first need a way to give meaning to these kind of integrals.
20 / 109
Dimensional regularization
Most Feynman integrals are not convergent in four space time

dimensions.
Common way out is the use of dimensional regularization, where
the four-dimensional space time is extended to d dimensions.
Divergences of the integrals then become manifest as poles in
d − 4.
d dimensional integrals behave identical to their four-dimensional
counterparts
21 / 109
d-dimensional integration
d-dimensional integrals have to fulfil these axioms

Linearity
Z Z Z
d d
d k (af (k ) + bg(k )) = a d k f (k) + b dd k g(k )
Scaling Z Z
dd k f (s k) = s−d dd k f (k)
Translational invariance
Z Z
dd k f (k + p) = dd k f (k)
22 / 109
d-dimensional integration – properties

Pro: dimensional regularization regularizes both UV and IR
singularities
23 / 109

singularities
Con: dimensional regularization regularizes both UV and IR
singularities in the same way
23 / 109

singularities
Con: dimensional regularization regularizes both UV and IR
singularities in the same way
scaleless integrals vanish
Z
dd k (k 2 )α = 0
integration by parts
Z
∂
dd k f (k) = 0
∂k µ
Interchange of integrations
Z Z Z Z
d p d k f (p, k) = d k dd p f (p, k)
d d d
23 / 109
Gaussian integrals in d dimensions
For many purposes the problem of d-dimensional integrations can be

reduced to one specific integral:
the Gaussian integral in d dimensions
Z
2 d
dd k e−k = π 2
which is the most natural generalization of the integer dimension one.
24 / 109
Gaussian integrals in d dimensions
For many purposes the problem of d-dimensional integrations can be

reduced to one specific integral:
the Gaussian integral in d dimensions
Z π d
d −Ak 2 2
d ke =
A
which is the most natural generalization of the integer dimension one.

The dependence on A follows by rescaling k → √k
A
25 / 109
Simple example – Improved
Z
1
I1 = dd k
+ M2 −k 2
Z ∞ Z
2 2
I1 = i dt e−M t dd ke−k t
0
∞ 2
e−M t
Z
d/2
I1 = iπ dt d/2
0 t
26 / 109
Z
1
I1 = dd k
−k 2 + M2
Z Z
−M 2 t 2t
I1 = i dt e dd k e−k
Z ∞
2t
I1 = iπ d/2
dt t −d/2 e−M
0
27 / 109
Z
1
I1 = dd k
−k 2 + M 2
Z Z
2 2
I1 = i dt e−M t dd k e−k t
Z ∞
2t
I1 = iπ d/2 dt t −d/2 e−M
0
Z ∞
I1 = iπ d/2 (M 2 )d/2−1 dt t −d/2 e−t
0
= iπ d/2 (M 2 )d/2−1 Γ(−d/2 + 1)
28 / 109
Z
1
I1 = dd k
−k 2
+ M2
= iπ d/2 (M 2 )d/2−1 Γ(−d/2 + 1)
(M 2 )d/2−1 overall mass dimension of the integral, could be read off

from the original integral
Γ(−d/2 + 1) contains the real information
singular for d → 4
1
Γ(−d/2 + 1) = − + (γE − 1)

1
+ −6γE2 + 12γE − π 2 − 12
12
29 / 109
1
Γ(−d/2 + 1) = − + (γE − 1)

1
+ −6γE2 + 12γE − π 2 − 12
12
not a very compact result, better to choose a suitable normalization
1
Γ(−d/2 + 1)/Γ(1 + ) = − − 1 −

or
π2

1
Γ(−d/2 + 1)/ exp(−ΓE ) = − − 1 + −1 −
12
30 / 109
Simple example – Extended
Z
1
I1 = dd k
(−k 2
+ M 2 )α
Z Z
1 α−1 −M 2 t 2
I1 = i dt t e dd ke−k t
Γ(α)
π d/2 ∞
Z
2
I1 = i dt t α−d/2−1 e−M t
Γ(α) 0
∞
iπ d/2 2 d/2−α
Z
I1 = (M ) dt t α−d/2−1 e−t
Γ(α) 0
iπ d/2 2 d/2−α
= (M ) Γ(α − d/2)
Γ(α)
Also the integral with α = 2 is divergent

31 / 109
not that simple example: One-loop propagator
Consider the one-loop propagator

Z
1
P1 = d4 k
−k (−(k + q)2 )
2
introduce Feynman parameters
Γ( i ki ) 1 δ( i xi − 1) i xiki −1
P Z Z 1 P Q
1
= Q dx1 · · · dxn P
D1k1 · · · Dnkn ( i xi Di ) i ki
P
i Γ(ki ) 0 0
in their simplest form

Z 1
1 1
= dx
D1 D2 0 [xD1 + (1 − x)D2 ]2
32 / 109
Z
1
P1 = d4 k
−k 2 (−(k + q)2 )
Z Z 1
1
= d4 k dx
0 [−k 2 − 2xk · q − xq 2 ]2
complete the square

Z Z 1
1
P1 = d4 k dx
0 [−(k + xq)2 + x 2 q 2 − xq 2 ]2
shift k = k + (1 − x)q
Z Z 1
4 1
P1 = d k dx
0 [−k 2 + x 2q2 − xq 2 ]2
33 / 109
Z Z 1
1
P1 = d4 k dx
0 [−k 2 + x(x − 1)q 2 ]2
doing the momentum integration gives

Z 1 h id/2−2
d/2
P1 = iπ Γ(2 − d/2) dx x(x − 1)q 2
0
let’s assume q 2 < 0

Z 1
P1 = iπ d/2
Γ(2 − d/2)(−q ) 2 d/2−2
dx [x(1 − x)]d/2−2
0
34 / 109
Z 1
P1 = iπ d/2
Γ(2 − d/2)(−q ) 2 d/2−2
dx [x(1 − x)]d/2−2
0
what is left is special case of the Beta-function

Z
Γ(a)Γ(b)
B(a, b) = dt t a−1 (1 − t)b−1 =
Γ(a + b)
Γ2 (d/2 − 1)
P1 = iπ d/2 Γ(2 − d/2)(−q 2 )d/2−2
Γ(d + 2)
35 / 109
Cuts
branch cuts appear in an Feynman integral when the particles in

the loop can be produced as real particles
in the propagator example before the factor
(−q 2 )d/2−2
could have again be predicted from mass dimension of the the

analytic properties of the diagram
the imaginary parts then corresponds to the total cross section
belonging the the respective cut
36 / 109
Tensor integrals
Outline
1 Introduction
4 Tensor integrals
6 Mellin Barnes
37 / 109
Tensor integrals
Tensor integrals
So far we have only dealt with scalar integrals, i.e. integrals with no
vectors with free indices in the numerator
k µ1 · · · k µn
Z
µ1 ···µn
I = ddk
D1 · · · DN
At one-loop it has long been worked out, how to reduce tensor
integrals to scalar integrals ⇒ Passarino Veltman reduction
38 / 109
Tensor integrals
Passarino Veltman reduction: Example
consider the two-point function with numerator
kµ
Z
ddk 2
k (k + q)2
This has to evaluate to

kµ
Z
ddk = q µ B1 (q 2 )
k 2 (k + q)2
multiply both sides with q µ
k ·q
Z
ddk = q 2 B1 (q 2 )
k 2 (k + q)2
39 / 109
Tensor integrals
k ·q
Z
ddk = q 2 B1 (q 2 )
k 2 (k + q)2
rewrite the scalar product
1
k ·q = ((k + q)2 − k 2 − q 2 )
2
(k + q)2 − k 2 − q 2
Z
1
ddk = q 2 B1 (q 2 )
2 k 2 (k + q)2
40 / 109
Tensor integrals
(k + q)2 − k 2 − q 2
Z
1
ddk = q 2 B1 (q 2 )
2 k 2 (k + q)2
cancel the common factors and solve for B1 (q 2 )
Z Z
2 1 d 1 d 1
B1 (q ) = d k − d k
2q 2 k2 (k + q)2
Z
1 1
− dk 2
2 k (k + q)2
the first two integrals vanish and we have

Z
2 1 1
B1 (q ) = − ddk 2
2 k (k + q)2
41 / 109
Tensor integrals
Projectors
Most of the time it is more convenient not to have open Lorentz

indices from the very start
To avoid this the use of projectors is very convenient
As an example let us consider corrections to the photon
propagator Πµν (q)
from Lorentz covariance we know it can be written in the form
Πµν = (q 2 g µν − q µ q ν )ΠT (q 2 ) + q µ q ν ΠL
to calculate ΠT and ΠL we can use the projectors
PT = (q 2 g µν − q µ q ν )/(q.q)2 /(d − 1), PL = q µ q ν /(q.q)2
42 / 109
Integration by parts
Outline
1 Introduction
4 Tensor integrals
6 Mellin Barnes
43 / 109
state of the art calculations require the calculation of O(103 ) -

O(107 ) Feynman integrals
especially multi-loop or calculations involving expansions require
the calculation of many integrals
individual calculation of all these integrals is not feasible
the number of integrals can be greatly reduced by applying the
so-called integration-by-parts identities
44 / 109
Integration-by-parts identities are based on the property

Z
∂ 1
0 = dd k µ k
∂ki D 1 · · · Dnkn
1
which being the integral of a total derivative evaluates to a surface

term and can be shown to vanish.
45 / 109
To make them more manageable contract with either an external or a

loop momentum
µ µ
∂ {k , qj }
Z
d
0= d k µ k
∂ki D 1 · · · Dnkn
1
which then yields
#loops × (#loops + #(indep ext momenta))
relations
46 / 109
Integration-parts-relations can either be used by

constructing a set of symbolic relations reducing the number of
propagators
LiteRed [Lee]
explicitly applying the relations to a set of integrals and solving the

resulting system of linear equations
Air [Anastasiou, Lazopoulos]
FIRE [Smirnov]
Reduze [v. Manteuffel, (Studerus)]
47 / 109
Simple example
Consider the class of integrals

Z
1
J(n) = dd k
(k 2 − M 2 )n
Applying the only IBP relation leads to
kµ
Z
∂
0 = dd k µ 2
∂k (k − M 2 )n
k2
Z
= dJ(n) − 2n dd k 2
(k − M 2 )n+1
48 / 109
Simple example
Consider the class of integrals

Z
1
J(n) = dd k
(k 2 − M 2 )n
Applying the only IBP relation leads to
kµ
Z
∂
0 = dd k µ 2
∂k (k − M 2 )n
(k 2 − M 2 ) + M 2
Z
= dJ(n) − 2n dd k
(k 2 − M 2 )n+1
= (d − 2n)J(n) − 2nM 2 J(n + 1)
49 / 109
Notation
For the presentation of IBP-identities a notation using lowering and

raising operators is commonly used.
Z
1
J(n1 , . . . , nN ) = dd k n1
D1 · · · DNnN
The operators i + and i − act on J(n1 , . . . , nN ) as
i + J(n1 , . . . , nN ) = J(n1 , . . . , ni + 1, . . . , nN )
i − J(n1 , . . . , nN ) = J(n1 , . . . , ni − 1, . . . , nN )
Note: i + always comes together with an ni and therefore the ni is

sometimes included in the definition of the operator.
50 / 109
Simple example cont’d
Using the new notation the IBP identity can be written as
0 = ((d − 2n1 ) − 2nM 2 1+ )J(n1 )
or simply
(d − 2n1 ) − 2n1 M 2 1+
implying the application to an integral and omitting the left-hand side
51 / 109
Simple example cont’d
In this simple case the IBP identities can easily be solved leading to
d − 2n
J(n + 1) = J(n)
2nM 2
and explicitly to
d −2
J(2) = J(1)
2M 2
d −4 (d − 2)(d − 4)
J(3) = 2
J(2) = J(1)
4M 8M 4
52 / 109
Exercise: one-loop propagator
Derive the IBP relations for the class of one-loop propagator integrals
Z
1
P1 = d d k 2 n
(k ) ((k + q)2 )n2
1
and try to solve them
53 / 109
Exercise: one-loop propagator
The two IBP relations for this class of integrals can be obtained by
evaluating
kµ
Z
∂
0 = ddk µ 2 n
∂k (k ) 1 ((k + q)2 )n2
and
qµ
Z
∂
0= ddk
∂k µ (k 2 )n1 ((k + q)2 )n2
Let us call the integrals
J(n1 , n2 )
54 / 109
Relation I
kµ
Z
∂
0 = ddk
∂k µ (k 2 )n1 ((k + q)2 )n2
k2
Z
= dJ(n1 , n2 ) − 2n1 d d k 2 n +1
(k ) 1 ((k + q)2 )n2
(k + q) · k
Z
−2n2 d d k 2 n
(k ) 1 ((k + q)2 )n2 +1
55 / 109
Relation I
kµ
Z
∂
0 = ddk
∂k µ (k 2 )n1 ((k + q)2 )n2
= dJ(n1 , n2 ) − 2n1 J(n1 , n2 )
(k + q) · k
Z
−2n2 d d k 2 n
(k ) 1 ((k + q)2 )n2 +1
= dJ(n1 , n2 ) − 2n1 J(n1 , n2 )
1
(k 2 + (k + q)2 − q 2 )
Z
−2n2 d d k 2 2 n
(k ) 1 ((k + q)2 )n2 +1
56 / 109
Relation I
kµ
Z
∂
0 = ddk
∂k µ (k 2 )n1 ((k + q)2 )n2
= dJ(n1 , n2 ) − 2n1 J(n1 , n2 )
(k + q) · k
Z
−2n2 d d k 2 n
(k ) 1 ((k + q)2 )n2 +1
= dJ(n1 , n2 ) − 2n1 J(n1 , n2 )
−n2 (J(n1 − 1, n2 + 1) + J(n1 , n2 ) − q 2 J(n1 , n2 + 1))
= (d − 2n1 − n2 )J(n1 , n2 )
−n2 J(n1 − 1, n2 + 1) + n2 q 2 J(n1 , n2 + 1)
57 / 109
Relation I
Taking into account the symmetry
J(n1 , n2 ) = J(n2 , n1 )
and
J(n1 , 0) = J(0, n2 ) = 0
this relation is already enough
n2 q 2 J(n1 , n2 + 1) = −(d − 2n1 − n2 )J(n1 , n2 )

+n2 J(n1 − 1, n2 + 1)
58 / 109
Mellin Barnes
Outline
1 Introduction
4 Tensor integrals
6 Mellin Barnes
59 / 109
Mellin Barnes
Generalized formula and why we need more tools

Consider a general L-loop integral with N internal lines with momenta
qi and masses mi and E external lines with momenta pe
dd k1 · · · dd kL
Z
1
GL =
(iπ d/2 )L 2 )νN
(q12 − m12 )νi · · · (qN2 − mN
The denominators are of the form

!2 !2
X X X
di = αi` k` − Pi − mi2 = αi` k` − βie pe − mi2
` ` e
the numerator in the Feynman parameter representation before the

momentum integration can thus be written in the form
X
N = xi di = kMk − 2kQ + J
i
60 / 109
Mellin Barnes
X
N = xi di = kMk − 2kQ + J
i
X
M``0 = xi αi` αi`0
i
X
Q` = xi αi` Pi
i
X
J = xi (Pi2 − mi2 )
i
before we can do the momentum integration we need to diagonalize

and rescale in k space. This gives rise to the determinant of M.
U(x) = det M
61 / 109
Mellin Barnes
Shifting the loop momenta to complete the square gives rise to the
polynomial
F (x) = −(det M)J + Q M̃Q
with
M̃ = (det M)M −1
and we thus obtain the final Feynman parameter representation
(−1)Nν Γ(Nν − d2 L)
Z Y
dxi xiνi −1 δ(1 −
X
GL = Q xi )
i Γ(νi ) i
U(x)Nν −d(L+1)/2
×
F (x)Nν −dL/2
Note: U(x) = 1 for one-loop integrals
62 / 109
Mellin Barnes
to proceed with the calculation in the next step the Feynman parameter
integral
U(x)Nν −d(L+1)/2
Z Y
X
xi ) ×
F (x)Nν −dL/2
i
has to be performed.
Problem: The polynomials in the Feynman parameters xi (one variable
per line of the integral) become too complicated to be integrated very
fast.
63 / 109
Mellin Barnes
Where do we go from here?
Problem: we have a completely general representation for Feynman

integral properly regularized in d dimensions. It can not (easily)
evaluated any further.
Possible solutions?
expand in = (4 − d)/2
64 / 109
Mellin Barnes

Possible solutions?
most of the time not possible, since Feynman integrals still do not
converge for → 0
64 / 109
Mellin Barnes

Possible solutions?
converge for → 0
integrate numerically
64 / 109
Mellin Barnes

Possible solutions?
converge for → 0
cannot do that in d dimensions
64 / 109
Mellin Barnes

Possible solutions?
converge for → 0
make the objects simpler again
64 / 109
Mellin Barnes

Possible solutions?
converge for → 0
make the objects simpler again
can be done, but for a prize ...
64 / 109
Mellin Barnes
Mellin-Barnes representation
One way relation to achieve this goal is the Mellin-Barnes

representation [Smirnov; Tausk]
i∞
B −λ
Z
1
λ
= dzAz B −z Γ(−z)Γ(λ + z)
(A + B) 2πiΓ(λ) −i∞
the main idea being to chop the long polynomials in smaller pieces in
such a way that they can be integrated over the Feynman parameters
again.
65 / 109
Mellin Barnes

i∞
B −λ
Z
1
λ
(A + B) 2πiΓ(λ) −i∞
again.
Note: If this relation is applied finely enough the Feynman integration
is guaranteed to become doable
65 / 109
Mellin Barnes

i∞
B −λ
Z
1
λ
(A + B) 2πiΓ(λ) −i∞
again.
Note: If this relation is applied finely enough the Feynman integration
is guaranteed to become doable
Note: This might not be the best idea ...
65 / 109
Mellin Barnes
Very important is how the integration contour has to be chosen.

i∞
B −λ
Z
1
λ
(A + B) 2πiΓ(λ) −i∞
There are left poles coming from
Γ(λ + z)
and right poles coming from
Γ(−z)
The integration contour has to separate these poles
66 / 109
Mellin Barnes
Γ(−z)Γ(λ + z)
67 / 109
Mellin Barnes
Mellin-Barnes representations: Singularities
In this approach singularities (in ) arise, when for → 0 left and right
poles coincide such that no valid integration contour can be chosen.
68 / 109
Mellin Barnes
Mellin-Barnes representations: Singularities
In this approach singularities (in ) arise, when for → 0 left and right
poles coincide such that no valid integration contour can be chosen.
This problem can be solved by explicitly taking residues when

necessary.
in the example, instead of the complicated contour shown take a

straight line but explicitly include the residue of the first left pole
68 / 109
Mellin Barnes
Mellin-Barnes representations: How to continue?
Once the optimal MB-representation has been found one can continue
with one or more of the following
further analytical manipulations: e.g. Barnes Lemmas
69 / 109
Mellin Barnes
Mellin-Barnes representations: How to continue?
Once the optimal MB-representation has been found one can continue
with one or more of the following
further analytical manipulations: e.g. Barnes Lemmas
preparation of a series representation and application of
summation techniques
preparation for numerical integration, e.g. MB.m
69 / 109
Mellin Barnes
Barnes Lemmas
First Barnes Lemma

Z i∞
1
Γ(a + z)Γ(b + z)Γ(c − z)Γ(d − z)dz
2πi −i∞
Γ(a + c)Γ(a + d)Γ(b + c)Γ(b + d)
=
Γ(a + b + c + d)
70 / 109
Mellin Barnes
Barnes Lemmas
First Barnes Lemma

Z i∞
1
Γ(a + z)Γ(b + z)Γ(c − z)Γ(d − z)dz
2πi −i∞
Γ(a + c)Γ(a + d)Γ(b + c)Γ(b + d)
=
Γ(a + b + c + d)
Second Barnes Lemma

Z i∞
1 Γ(a + z)Γ(b + z)Γ(c + z)Γ(1 − d − z)Γ(−z)
dz
2πi −i∞ Γ(e + z)
Γ(a)Γ(b)Γ(c)Γ(1 − d + a)Γ(1 − d + b)Γ(1 − d + c)
=
Γ(e − a)Γ(e − b)Γ(e − c)
with e = a + b + c − d + 1
70 / 109
Mellin Barnes
Towards a series representation
If the integrand f (z) of the MB integral

Z i∞
dzf (z)
−i∞
is vanishing fast enough for |z| → ∞ we can close the integration

contour either to the right or to the left and use the residue theorem
and write Z i∞
1 X
dz f (z) = Res(f (z))
2πi −i∞
71 / 109
Mellin Barnes
AMBRE - Automatic Mellin-Barnes Representation
AMBRE written by I. Dubovyk, J. Gluza, K. Kajda, T. Riemann, which

can be obtained from the webpage
http://prac.us.edu.pl/˜gluza/ambre/
is a tool for the automatic construction of a good (best) MB
representation.
72 / 109
Mellin Barnes
MB.m
MB.m Mathematica package by M. Czakon

can be used to extract poles from Mellin Barnes representations
prepare code for numerical integration
and run it
Caveats:
as provided on webpage written to use f77
uses the old Cuba API
needs parts of the CERNLIB
73 / 109
Mellin Barnes
Examples: Hands on
take file from

http://www-zeuthen.desy.de/˜marquard/CAPP.tgz
74 / 109
Sector decomposition
Outline
1 Introduction
4 Tensor integrals
6 Mellin Barnes
75 / 109
Recap
We derived the Feynman parameter representation
(−1)Nν Γ(Nν − d2 L)
Z Y
X
GL = Q xi )
i Γ(νi ) i
U(x)Nν −d(L+1)/2
×
F (x)Nν −dL/2
where U(x) and F (x) are homogeneous functions of xi of degree L

and L + 1, respectively.
For euclidean kinematics F (x) is positive semi definite and we only

have to deal with end-point singularities.
The sector decomposition approach can be used to disentangle
overlapping singularities.
76 / 109
Primary sectors
Let us set νi = 1
1Y
U(x)Nν −d(L+1)/2
Z X
dxi δ(1 − xi )
0 F (x)Nν −dL/2
i
First step is to integrate over the delta function
To do this split the integration into n parts to get the primary sectors G`
Z Z Y XZ Y
n n
d x= d x θ(xi ≥ 0) = dn x θ(x` ≥ xi ≥ 0)
` i6=`
explicit for 2 variables

Z Z Z
dx1 dx2 = dx1 dx2 θ(x1 ≥ x2 ≥ 0) + dx1 dx2 θ(x2 ≥ x1 ≥ 0)
77 / 109
Primary sectors
In each primary sector G` do the variable transformation


 x ` tj j <`
xj = x j =`
 `
x` tj−1 j > `
Then due to the homogeneity x` factors out and
F` (x) → F` (t)x`L+1 , U` (x) → U` (t)x`L
doing the integration over x` to eliminate the δ-function gives for each
primary sector
Z 1 N−1
Y U` (t)Nν −d(L+1)/2
G` = dti
0 F` (t)Nν −dL/2
i=1
78 / 109
Iterated sector decomposition

There are many strategies to do the actual sector decomposition. I
follow here the original version by Binoth and Heinrich
1 Determine a minimal set of parameters, say S = {tα1 ,...,tαr }, such
that U` , respectively F` , vanish if the parameters of S are set to
zero.
2 Decompose the corresponding r -cube into r subsectors.
r
Y r Y
X
θ(1 ≥ tαj ≥ 0) = θ(1 ≥ tαk ≥ tαj ≥ 0)
j=1 k=1 j6=k
3 remap to the unit cube in each subsector

tαk tαj , j 6= k
tαj →
t αj , j =k
tαk factors from U(t) and/or F (t) and we get the form
Z 1 N−1 N −d(L+1)/2
A −B U`k (t) ν
Y Y
G`k = dti tj j j
F (t)Nν −dL/2 79 / 109
Extraction of poles
After the sector decomposition is complete we are left with expressions
of the form Z
A −B
Ij = dtj tj j j I(tj , )
if Aj ≥ 0 we do not get a pole in from the t integration

otherwise we expand I(tj , ) around tj = 0
|Aj ]−1
X (p) tjp
I(tj , ) = Ij (0, ) + R(tj , )
p!
p=0
and obtain for the integral

|Aj ]−1 (p)
X 1 Ij (0, ) Z 1 A −B
Ij = + dtj tj j j R(tj , )
|Aj | + p + 1 − Bj p! 0
p=0
80 / 109
Example: one-loop triangle
Let us look at the one-loop triangle with propagators
{−k 2 , −(k + p1 )2 − M 2 , −(k + p2 )2 − M 2 }, p12 = p22 = 0
we have the U and F polynomials
U(x) = x1 + x2 + x3 , F (x) = sx3 x2 + x2 2 + x1 x2 + 2x3 x2 + x3 2 + x1 x3
81 / 109
To obtain the first primary sector we use the rules
{x1 → x1 , x2 → t1 x1 , x3 → t2 x1 }
integrate over the delta function and obtain the new polynomials
U1 (t) = t1 + t2 + 1, F1 (t) = st2 t1 + t1 2 + 2t2 t1 + t1 + t2 2 + t2
F (t) vanishes for t1 , t2 → 0 ⇒ we get 2 subsectors
82 / 109

We get the 2 subsectors by the transformations t2 → t1 t2 and t1 → t1 t2 ,
respectively.
U1,1 = st1 t2 + t1 t2 2 + 2t1 t2 + t2 + t1 + 1

U1,2 = st2 t1 + t2 t1 2 + 2t2 t1 + t1 + t2 + 1
both are now positive and we can stop here
primary sectors 2 and 3 have no subsectors
all singularities in the Feynman parameter integration are now made

explicit  
Z 1 Y
 t Aj −Bj  f (t)
j
0 j
and f (t) is free of singularities

83 / 109
Non-Euclidean kinematics
If we are not in the Euclidean region (sij < 0) the F -polynomial is

no longer positive semi definite, but changes sign inside the
domain of integration. Leading to (integrable) singularities.
A necessary condition are the Landau equations
xj (qj2 − mj2 ) = 0 ∀j
∂ X
xj (qj2 (k , p) − mj2 ) = 0
∂k µ
j
if there is a solution xi > 0 for the Landau equations, we have the

leading Laudau singularity, which is not integrable
To make these (integration) contour deformation [Borowka, Heinrich] can
be used.
84 / 109
Contour deformation
Reparametrize the integration path

Z N
1Y Z N
1Y

∂zk (x)
dxj Ix = dxj
I(z(x))
0 j=1 0 j=1 ∂xl
a convenient choice is
~z (~x ) = ~x − i~τ (~x )

∂F (~x )
τk = λxk (1 − xk )
∂xk
F expressed in the new variables
2
X ∂F
F (~z (~x )) = F (~x ) − iλ xl (1 − xj ) + O(λ2 )
∂xj
j
85 / 109
Tools
The method of sector decomposition has been implemented in several

tools
FIESTA (http://git.sander.su/fiesta) [Smirnov]
SecDec (https://secdec.hepforge.org/)
[ Borowka, Heinrich, Jones, Kerner, Schlenk, Zirke]
SectorDecomposition (http://wwwthep.physik.uni-
mainz.de/˜stefanw/sector decomposition/)
[Bogner, Weinzierl]
86 / 109
Hands on FIESTA
FIESTA either as complete package from
http://git.sander.su/fiesta
or just the main Mathematica file from
http://www.zeuthen.desy.de/˜marquard/CAPP/
87 / 109
Differential equations
Outline
1 Introduction
4 Tensor integrals
6 Mellin Barnes
88 / 109
Recap of IBP identities
IBP identities are very useful because they allow us to express any
integral Ij as linear combination of master integrals Mi
X
Ij = Cji (d, mi , sij )Mi
i
The set of master integrals is obtain by exploiting all available IBP

identities (and possibly also symmetry relations). The number of
master integral is fixed but there is a freedom to choose the integrals.
The idea is of course to reduce more complicated integrals, i.e. more
lines, more dots, to simpler ones, i.e. fewer lines, fewer dots.
89 / 109
Feynman integrals are functions of masses and kinematical variables
I(mi , sij )
As such one can try to find a differential equation for the integral, e.g.
∂
I(mi , sij ) = f (mi , sij )I(mi , sij ) + R(mi , sij ), {mi , sij }
∂z
and find a solution for it. [Kotikov; Remiddi]
90 / 109
Construction of the differential equation
One way to construct the differential equation is by means of a master

formula !
∂ ∂
DI(mi , sij ) = 2 mi2 + sij I(mi , sij )
∂mi2 ∂sij
where D denotes the mass dimension of the integral. mi and
sij = (pi + pj )2 denote internal masses and kinematical variables,
respectively.
In the simple case of one mass parameter m and one kinematic one
q 2 this turns into

2 2 ∂ 2 ∂
DI(m, q ) = 2 m +q I(m, q 2 )
∂m2 ∂q 2
91 / 109
Construction of the differential equation

2 2 ∂ 2 ∂
DI(m, q ) = 2 m +q I(m, q 2 )
∂m2 ∂q 2
the derivative with respect to m2 can be taken directly leading to
integrals with raised powers of propagators and we remain only with
the differential with respect to q 2 .
Now we can use IBP-relations to rewrite the new integrals in terms of

the original one.
Thus, we obtain a differential equation for the integral
92 / 109
System of differential equation
In general, there is more than one master integral that we want to

calculate. Thus we get a system of coupled differential equations.
∂ X
Ij (z, d) = Cj (z, d)Ij + Djk Ik (z, d)
∂z
k6=j
The integrals in the inhomogeneity are by construction at most as

complicated as the integral we are looking at.
In the best case they are all simpler than the original one.
In the worst case we obtain a system of coupled differential equation
within a sector, i.e. involving integrals where the same lines have
positive powers.
93 / 109
Solving the differential equation
A first-order differential equation can be solved by using the method

”variation of constants”.
Consider the first-order differential equation
∂
f (z) = a(z)f (z) + b(z)
∂z
then Z
A(z)
f̃ (z) = Ce , with A(z) = dz a(z)
is a solution of the homogeneous equation and

Z z
A(z) 0 −A(z 0 ) 0
f (z) = e b(z )e dz + C
z0
a solution of the inhomogeneous equation.
94 / 109
Solving the system of differential equation
In practice, it is best to first perform the expansion in = (4 − d)/2 and

then to solve the differential equations order by order in . Make an
ansatz for the master integrals in the form
m
X
Ij (z) = Ijk (z)k
k=−n
This leads to a system of coupled differential equations for the

coefficients of the Laurent expansion of the master integrals.
The system of differential equations can then be solved in bottom up
approach.
This approach leads naturally to iterated integrals like HPLs.
In case there are several integrals in a sector one can try to decouple
them in be a suitable choice of the master integrals.
95 / 109
Canonical basis
It was recently proposed [Henn] and demonstrated that in many cases a

canonical basis of master integrals can be found in which the
differential equations take the form
∂~
I(x, ) = A(x)~I(x, )
∂x
where A is a n × n matrix. This makes the solution of differential
equations trivial. In addition the alphabet of functions can be read off
from the entries of the matrix.
Furthermore, [Lee] recently proposed an algorithm how to obtain such a
representation.
96 / 109
Asymptotic expansions
Outline
1 Introduction
4 Tensor integrals
6 Mellin Barnes
97 / 109
Motivation
Quite often a problem is to complex at the start to be tackled directly.
In this case it is often possible to expand in some small or large

parameter to simplify the problem.
These expansions sometimes lead to more than a simple power series

and more work than taking a simple Taylor series has to be done.
In general, the procedure than goes by the name of asymptotic

expansion.
There are two procedures to perform an asymptotic expansion:

expansion by regions
expansion by subgraphs
98 / 109
Expansion by regions: The idea
1 Divide the space of the loop momenta into various regions and, in
every region, expand the integrand in a Taylor series with respect
to the parameters that are considered small there.
2 Integrate the integrand, expanded in the appropriate way in every
region, over the whole integration domain of the loop momenta.
3 Set to zero any scaleless integral.
The peculiar thing here is the second step since naively this could lead
to double counting problems.
The problematic thing is how do we find all these regions
99 / 109
Example I: large momentum expansion
k+p
p
m m
k k
Z
F = dd k I
with the integrand I = I1 I2 and the propagators
1 1 1
I1 = n 1 = and I2 =
(k + p)2 (k 2 + 2k · p + p2 )n1 (k 2 − m2 )n2
100 / 109
1 1 1
I1 = n 1 = and I2 =
(k + p)2 (k 2 + 2k · p + p2 )n1 (k 2 − m2 )n2
We want to consider the case where
|p| >> m
What are the regions?
101 / 109
1 1 1
I1 = n 1 = and I2 =
(k + p)2 (k 2 + 2k · p + p2 )n1 (k 2 − m2 )n2
We want to consider the case where
|p| >> m
What are the regions?

We find the following
hard region (h) k ∼ p
soft region (s) k ∼ m
101 / 109
Example I: hard region
In the hard region we have k ∼ p so
m2 k 2 ∼ k · p ∼ p 2
so I1 remains untouched and in I2 we perform an expansion in m2 /k 2

∞
X (h)
X (n2 )j (m2 )j
I2 → T (h) I2 ≡ Tj I2 =
j! (k 2 )n2 +j
j j=0
and we obtain a massless propagator

Z
1
dd k
(k + q)2 (k 2 )n
102 / 109
Example I: soft region
In the soft region we have k ∼ m so
|k 2 | |p2 |, |2k · p| |p2 |
now I2 is untouched and we have to expand I1

∞
X (s)
X (s)
X (n1 )j12 (−k 2 )j1 (−2k · p)j2
I1 → T (s) I1 ≡ Tj I1 ≡ Tj1 ,j2 I1 =
j1 ! j2 ! (p2 )n1 +j12
j j1 ,j2 j1 ,j2 =0
In this region we end with massive tadpoles
k · pn
Z
dd k 2
k − m2
103 / 109
Example II: On-shell integrals
Consider the typical on-shell integral (q 2 = M 2 )

Z
1
dk d dl d 2
(k + 2k · q)k ((k − l)2 − m2 )(l 2 − m2 )
4
we get the regions

k 2 , l 2 ≈ M 2 : expand in m2 → massless on-shell propagator
k 2 ≈ M 2 , l 2 ≈ m2 : one-loop on-shell × massive tadpole
k 2 , l 2 ≈ m2 : expand in k 2 → new class of diagrams
Z
1
d2 k dd l
(k · q)k ((k − l)2 − m2 )(l 2 − m2 )
4
104 / 109
Tools for asymptotic expansions
To help with finding the regions of a Feynman integrals the

Mathematica programs
asy.m [Pak, Smirnov]
asy2.m [Jantzen, Smirnov]
are useful.
Both programs are based on the program Qhull.
105 / 109
Physics
One last example: how everything fits together
Consider the anomalous magnetic moment of the muon g − 2 = aµ
σ µν q ν

µ 2 2
= (−ie)u(p2 ) γ FE (q ) + i FM (q ) u(p1 )
2m
p1 p2
aµ = FM (0)
106 / 109
Physics
One last example: how everything fits together
Consider the anomalous magnetic moment of the muon g − 2 = aµ
σ µν q ν

µ 2 2
= (−ie)u(p2 ) γ FE (q ) + i FM (q ) u(p1 )
2m
p1 p2
aµ = FM (0)
To obtain FM (0) use a suitable projector
106 / 109
Physics
4-loop contribution
I(a) I(b) I(c) I(d) II(a) II(b) II(c)
III IV(a) IV(b) IV(c) IV(d) V
107 / 109
Physics
The calculation
generate the diagrams

map the diagrams to O(100) diagram classes
use (T/P)FORM to apply the projector, expand, and do the algebra
reduce the appearing integrals to O(500) master integrals using
IBPs
calculated the masters using Mellin-Barnes techniques (simpler
ones) or sector decomposition (more complicated ones)
add everything up, do the renormalization and hope for the best.
108 / 109
Physics
Thanks a lot for your attention
109 / 109

Intro To Calc of Feyn Integral

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Intro To Calc of Feyn Integral

Uploaded by

Copyright:

Available Formats

Introduction to the calculation of Feynman Integrals

Feynman integrals come in different shapes and colors

many different methods have been invented over the years to

Series expansion of the Gamma function

Γ(z − z0 ) = Γ(z0 ) + Γ(z0 )Ψ(z0 )(z − z0 )

The Digamma-function Ψ(z)

The digamma function satisfies the relation

with the Euler-Mascheroni constant γE = 0.577216 . . .

The Digamma-function Ψ(z) cont’d

For non-positive integers the digamma functions evaluates again to

From the definition of the Γ function follows immediately the Schwinger

by performing the substitution t → t 0 = (M 2 − k 2 )t

Function classes appearing in Feynman integral calculations are the

which for z → 1 give the ζ values

The appear in the calculations from integrals over logarithms

+2Li2 (x) log(x) + log(1 − x) log2 (x)

The Polylogarithms are not very systematic and many relations

A more systematic approach are the harmonic polylogarithms (HPLs)

HPL(1, x) = − log(1 − x), HPL(−1, x) = log(1 + x)

The Minkowski Metric

gµν = diag(1, −1, −1, −1)

Consider simplest diagram possible,

Either introduce an explicit parametrization of the measure or use the

Consider simplest diagram possible,

Either introduce an explicit parametrization of the measure or use the

perform Wick rotation

k 2 = k02 − k12 − k22 − k32 → −k02 − k12 − k22 − k32

This integral does not converge for t → 0 .

Most Feynman integrals are not convergent in four space time

d-dimensional integrals have to fulfil these axioms

d-dimensional integration – properties

d-dimensional integration – properties

d-dimensional integration – properties

Gaussian integrals in d dimensions

For many purposes the problem of d-dimensional integrations can be

which is the most natural generalization of the integer dimension one.

Gaussian integrals in d dimensions

For many purposes the problem of d-dimensional integrations can be

which is the most natural generalization of the integer dimension one.

Simple example – Improved

Simple example – Improved

Simple example – Improved

Simple example – Improved

(M 2 )d/2−1 overall mass dimension of the integral, could be read off

Simple example – Improved

Simple example – Extended

Also the integral with α = 2 is divergent

not that simple example: One-loop propagator

Consider the one-loop propagator

introduce Feynman parameters

in their simplest form

not that simple example: One-loop propagator

complete the square

not that simple example: One-loop propagator

doing the momentum integration gives

let’s assume q 2 < 0

not that simple example: One-loop propagator

what is left is special case of the Beta-function

branch cuts appear in an Feynman integral when the particles in

could have again be predicted from mass dimension of the the

Passarino Veltman reduction: Example

consider the two-point function with numerator