NE W J E RSE Y • L ONDON • SI NGAP ORE • BE I J I NG • SHANGHAI • HONG KONG • TAI P E I • CHE NNAI
World Scientifc
Matilde Marcolli
California Institute of Technology, USA
British Library CataloguinginPublication Data
A catalogue record for this book is available from the British Library.
For photocopying of material in this volume, please pay a copying fee through the Copyright
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to
photocopy is not required from the publisher.
ISBN13 9789814271202
ISBN10 9814271209
ISBN13 9789814304481 (pbk)
ISBN10 9814304484 (pbk)
All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means,
electronic or mechanical, including photocopying, recording or any information storage and retrieval
system now known or to be invented, without written permission from the Publisher.
Copyright © 2010 by World Scientific Publishing Co. Pte. Ltd.
Published by
World Scientific Publishing Co. Pte. Ltd.
5 Toh Tuck Link, Singapore 596224
USA office: 27 Warren Street, Suite 401402, Hackensack, NJ 07601
UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
Printed in Singapore.
FEYNMAN MOTIVES
Zvezda v Galaktike i voron na suku
ivut, ne muqimy xekspirovsko i dilemmo i.
No my — my pozdni i stil starewe i Vselenno i.
Ona skvoz nas poet svo tosku.
Yuri Manin – “(not yet quite) The Late Style”
This page intentionally left blank This page intentionally left blank
Preface
This book originates from the notes of a course on “Geometry and Arith
metic of Quantum Fields”, which I taught at Caltech in the fall of 2008.
Having just moved to Caltech and having my ﬁrst chance to oﬀer a class
there, I decided on a topic that would fall in between mathematics and
theoretical physics. Though it inevitably feels somewhat strange to be
teaching Feynman diagrams at Caltech, I hope that having made the main
focus of the lectures the yet largely unexplored relation between quantum
ﬁeld theory and Grothendieck’s theory of motives in algebraic geometry
may provide a suﬃciently diﬀerent viewpoint on the quantum ﬁeld the
oretic notions to make the resulting combination of topics appealing to
mathematicians and physicists alike.
I am not an expert in the theory of motives and this fact is clearly
reﬂected in the way this text is organized. Interested readers will have to
look elsewhere for a more informative introduction to the subject (a few
references are provided in the text). Also I do not try in any way to give
an exhaustive viewpoint of the current status of research on the connection
between quantum ﬁeld theory and motives. Many extremely interesting
results are at this point available in the literature. I try, whenever possible,
to provide an extensive list of references for the interested reader, and
occasionally to summarize some of the available results, but in general I
prefer to keep the text as close as possible to the very informal style of the
lectures, possibly at the cost of neglecting material that should certainly
be included in a more extensive monograph. In particular, the choice of
material covered here focuses mostly on those aspects of the subject where
I have been actively engaged in recent research work and therefore reﬂects
closely my own bias and personal viewpoint.
In particular, I will try to illustrate the fact that there are two possible
vii
viii Feynman Motives
complementary approaches to understanding the relation between Feynman
integrals and motives, which one may refer to as a “bottomup” and a “top
down” approach. The bottomup approach looks at individual Feynman
integrals for given Feynman graphs and, using the parametric representa
tion in terms of Schwinger and Feynman parameters, identiﬁes directly the
Feynman integral (modulo the important issue of divergences) with an inte
gral of an algebraic diﬀerential form on a cycle in an algebraic variety. One
then tries to understand the motivic nature of the piece of the relative coho
mology of the algebraic variety involved in the computation of the period,
trying to identify speciﬁc conditions under which it will be a realization of
a very special kind of motive, a mixed Tate motive. The other approach,
the topdown one, is based on the formal properties that the category of
mixed Tate motives satisﬁes, which are suﬃciently rigid to identify it (via
the Tannakian formalism) with a category of representations of an aﬃne
group scheme. One then approaches the question of the relation to Feyn
man integrals by showing that the data of Feynman integrals for all graphs
and arbitrary scalar ﬁeld theories also ﬁt together to form a category with
the same properties. This second approach was the focus of my joint work
with Connes on renormalization and the Riemann–Hilbert correspondence
and is already presented in much greater detail in our book “Noncommu
tative geometry, quantum ﬁelds, and motives”. However, for the sake of
completeness, I have also included a brief and less detailed summary of this
aspect of the theory in the present text, referring the readers to the more
complete treatment for further information. The bottomup approach was
largely developed in the work of Bloch–Esnault–Kreimer, though in this
book I will mostly relate aspects of this approach that come from my joint
work with Aluﬃ. I will also try to illustrate the points of contact between
these two diﬀerent approaches and where possible new developments may
arise that might eventually unify the somewhat fragmentary picture we
have at the moment. This book only partially reﬂects the state of the art
on this fastmoving subject at the speciﬁc time when these lectures were
delivered, but I hope that the timeliness of circulating these ideas in the
community of mathematicians and physicists will somehow make up for
lack of both rigor and of completeness.
Matilde Marcolli
Acknowledgments
A large part of the material covered in this book is based on recent joint
work with Paolo Aluﬃ and on previous joint work with Alain Connes, whom
I thank for their essential contributions to the development of those ideas
and results. I also thank Abhijnan Rej for conversations and collaboration
on some of the topics in this book, and for providing detailed comments on
an earlier draft of the manuscript. I thank the students of my “Geometry
and Arithmetic of Quantum Fields” course, especially Domenic Denicola
and Michael Smith, for their valuable comments and questions during the
class. Many thanks are due to Paolo Aluﬃ for his very detailed and useful
feedback on the ﬁnal version of the manuscript. I also thank Arthur Green
spoon for a ﬁnal proofreading of the text. This book was partly written
during stays of the author at MSRI and at the MPI. I thank both institutes
for the hospitality and for support. This work was partially supported by
NSF grants DMS0651925 and DMS0901221.
ix
This page intentionally left blank This page intentionally left blank
Contents
Preface vii
Acknowledgments ix
1. Perturbative quantum ﬁeld theory and Feynman diagrams 1
1.1 A calculus exercise in Feynman integrals . . . . . . . . . . 1
1.2 From Lagrangian to eﬀective action . . . . . . . . . . . . . 6
1.3 Feynman rules . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4 Simplifying graphs: vacuum bubbles, connected graphs . . 12
1.5 Oneparticleirreducible graphs . . . . . . . . . . . . . . . 14
1.6 The problem of renormalization . . . . . . . . . . . . . . . 18
1.7 Gamma functions, Schwinger and Feynman parameters . 20
1.8 Dimensional Regularization and Minimal Subtraction . . 21
2. Motives and periods 25
2.1 The idea of motives . . . . . . . . . . . . . . . . . . . . . 25
2.2 Pure motives . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3 Mixed motives and triangulated categories . . . . . . . . . 34
2.4 Motivic sheaves . . . . . . . . . . . . . . . . . . . . . . . . 37
2.5 The Grothendieck ring of motives . . . . . . . . . . . . . . 38
2.6 Tate motives . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.7 The algebra of periods . . . . . . . . . . . . . . . . . . . . 44
2.8 Mixed Tate motives and the logarithmic extensions . . . . 45
2.9 Categories and Galois groups . . . . . . . . . . . . . . . . 49
2.10 Motivic Galois groups . . . . . . . . . . . . . . . . . . . . 50
xi
xii Feynman Motives
3. Feynman integrals and algebraic varieties 53
3.1 The parametric Feynman integrals . . . . . . . . . . . . . 54
3.2 The graph hypersurfaces . . . . . . . . . . . . . . . . . . . 60
3.3 Landau varieties . . . . . . . . . . . . . . . . . . . . . . . 65
3.4 Integrals in aﬃne and projective spaces . . . . . . . . . . 67
3.5 Nonisolated singularities . . . . . . . . . . . . . . . . . . 71
3.6 Cremona transformation and dual graphs . . . . . . . . . 72
3.7 Classes in the Grothendieck ring . . . . . . . . . . . . . . 76
3.8 Motivic Feynman rules . . . . . . . . . . . . . . . . . . . . 78
3.9 Characteristic classes and Feynman rules . . . . . . . . . 81
3.10 Deletioncontraction relation . . . . . . . . . . . . . . . . 84
3.11 Feynman integrals and periods . . . . . . . . . . . . . . . 93
3.12 The mixed Tate mystery . . . . . . . . . . . . . . . . . . . 94
3.13 From graph hypersurfaces to determinant hypersurfaces . 97
3.14 Handling divergences . . . . . . . . . . . . . . . . . . . . . 112
3.15 Motivic zeta functions and motivic Feynman rules . . . . 115
4. Feynman integrals and Gelfand–Leray forms 119
4.1 Oscillatory integrals . . . . . . . . . . . . . . . . . . . . . 119
4.2 Leray regularization of Feynman integrals . . . . . . . . . 121
5. Connes–Kreimer theory in a nutshell 127
5.1 The Bogolyubov recursion . . . . . . . . . . . . . . . . . . 128
5.1.1 Step 1: Preparation . . . . . . . . . . . . . . . . . 128
5.1.2 Step 2: Counterterms . . . . . . . . . . . . . . . . 128
5.1.3 Step 3: Renormalized values . . . . . . . . . . . . 129
5.2 Hopf algebras and aﬃne group schemes . . . . . . . . . . 130
5.3 The Connes–Kreimer Hopf algebra . . . . . . . . . . . . . 133
5.4 Birkhoﬀ factorization . . . . . . . . . . . . . . . . . . . . . 135
5.5 Factorization and RotaBaxter algebras . . . . . . . . . . 137
5.6 Motivic Feynman rules and RotaBaxter structure . . . . 139
6. The Riemann–Hilbert correspondence 143
6.1 From divergences to iterated integrals . . . . . . . . . . . 143
6.2 From iterated integrals to diﬀerential systems . . . . . . . 145
6.3 Flat equisingular connections and vector bundles . . . . . 146
6.4 The “cosmic Galois group” . . . . . . . . . . . . . . . . . 147
Contents xiii
7. The geometry of DimReg 151
7.1 The motivic geometry of DimReg . . . . . . . . . . . . . . 151
7.2 The noncommutative geometry of DimReg . . . . . . . . . 155
8. Renormalization, singularities, and Hodge structures 167
8.1 Projective Radon transform . . . . . . . . . . . . . . . . . 167
8.2 The polar ﬁltration and the Milnor ﬁber . . . . . . . . . . 170
8.3 DimReg and mixed Hodge structures . . . . . . . . . . . . 173
8.4 Regular and irregular singular connections . . . . . . . . . 176
9. Beyond scalar theories 185
9.1 Supermanifolds . . . . . . . . . . . . . . . . . . . . . . . . 185
9.2 Parametric Feynman integrals and supermanifolds . . . . 190
9.3 Graph supermanifolds . . . . . . . . . . . . . . . . . . . . 199
9.4 Noncommutative ﬁeld theories . . . . . . . . . . . . . . . 201
Bibliography 207
Index 215
Chapter 1
Perturbative quantum ﬁeld theory
and Feynman diagrams
This ﬁrst chapter gives a very brief and informal introduction to pertur
bative quantum ﬁeld theory. We start with a simple ﬁnite dimensional
example, which is aimed at illustrating the main ideas in a context fa
miliar to most readers, where the more serious diﬃculties of the inﬁnite
dimensional setting are not present, but one can already see the logic be
hind the perturbative expansion as well as the roles of Feynman graphs.
We then brieﬂy describe the inﬁnite dimensional setting, in the case of a
scalar ﬁeld theory, and the corresponding Feynman rules. We also include a
brief discussion of dimensional regularization (DimReg) of divergent Feyn
man integrals, a subject to which we return later in the book from a more
geometric perspective. We include here also a brief discussion of some com
binatorial properties of Feynman graphs, which we need in later chapters.
We state the renormalization problem, which we discuss in more detail in
a later chapter, in the setting of the Connes–Kreimer approach. There are
many very good books on quantum ﬁeld theory. The subject is very rich
and can be approached from many diﬀerent perspectives. For the material
we are going to cover in this text, the reader may wish to consult [Bjorken
and Drell (1964)], [Bjorken and Drell (1965)], [Itzykson and Zuber (2006)],
[LeBellac (1991)], [Nakanishi (1971)]. A lively and very readable introduc
tion, by which the beginning part of this chapter is inspired, can be found
in the recent book [Zee (2003)].
1.1 A calculus exercise in Feynman integrals
To understand the role of Feynman graphs in perturbative quantum ﬁeld
theory, it is useful to ﬁrst see how graphs arise in the more familiar setting
of ﬁnite dimensional integrals, as a convenient way of parameterizing the
1
2 Feynman Motives
terms in the integration by parts of polynomials with respect to a Gaussian
measure. It all starts with the simplest Gaussian integral
_
R
e
−
1
2
ax
2
dx =
_
2π
a
_
1/2
, (1.1)
for a > 0, which follows from the usual polar coordinates calculation
_
∞
−∞
e
−
1
2
ax
2
dx
_
∞
−∞
e
−
1
2
ay
2
dy = 2π
_
∞
0
e
−
1
2
ar
2
r dr =
2π
a
_
∞
0
e
−u
du.
Similarly, the Gaussian integral with source term is given by
_
R
e
−
1
2
ax
2
+Jx
dx =
_
2π
a
_
1/2
e
J
2
2a
. (1.2)
This also follows easily from (1.1), by completing the square
−
ax
2
2
+Jx = −
a
2
(x
2
−
2Jx
a
) = −
a
2
(x −
J
a
)
2
+
J
2
2a
and then changing coordinates in the integral to y = x −
J
a
. In this one
dimensional setting a ﬁrst example of computation of an expectation value
can be given in the form
¸x
2n
) :=
_
R
x
2n
e
−
1
2
ax
2
dx
_
R
e
−
1
2
ax
2
dx
=
(2n − 1)!!
a
n
, (1.3)
where (2n−1)!! = (2n−1) (2n−3) 5 3 1. One obtains (1.3) inductively
from (1.1) by repeatedly applying the operator −2
d
da
to (1.1). It is worth
pointing out that the factor (2n−1)!! has a combinatorial meaning, namely
it counts all the diﬀerent ways of connecting in pairs the 2n linear terms x
in the monomial x
2n
= x x x in the integral (1.3). In physics one refers
to such pairings as Wick contractions. As we discuss below, the analog of
the Gaussian integrals in the inﬁnite dimensional setting of quantum ﬁeld
theory will be the free ﬁeld case, where only the quadratic terms are present
in the Lagrangian. The onedimensional analog of Lagrangians that include
interaction terms will be integrals of the form
Z(J) =
_
R
e
−
1
2
ax
2
+P(x)+Jx
dx, (1.4)
where P(x) is a polynomial in x of degree deg P ≥ 3. The main idea in such
cases, which we’ll see applied similarly to the inﬁnite dimensional case, is to
treat the additional term P(x) as a perturbation of the original Gaussian
integral and expand it out in Taylor series, reducing the problem in this
Perturbative quantum ﬁeld theory and Feynman diagrams 3
way to a series of terms, each given by the integral of a polynomial under
a Gaussian measure. Namely, one writes
Z(J) =
_
R
_
∞
n=0
P(x)
n
n!
_
e
−
1
2
ax
2
+Jx
dx. (1.5)
The perturbative expansion of the integral (1.4) is deﬁned to be the series
∞
n=0
1
n!
_
R
P(x)
n
e
−
1
2
ax
2
+Jx
dx. (1.6)
Notice then that, for a monomial x
k
, the integral above satisﬁes
_
R
x
k
e
−
1
2
ax
2
+Jx
dx =
_
d
dJ
_
k
_
R
e
−
1
2
ax
2
+Jx
dx. (1.7)
Using (1.2), this gives
_
R
x
k
e
−
1
2
ax
2
+Jx
dx =
_
2π
a
_
1/2
_
d
dJ
_
k
e
J
2
2a
.
Thus, in the case where the polynomial P(x) consists of a single term
P(x) =
λ
k!
x
k
,
one can rewrite each term in the perturbative expansion using (1.7), so that
one obtains
∞
n=0
1
n!
_
R
_
λ
k!
x
k
_
n
e
−
1
2
ax
2
+Jx
dx =
∞
n=0
1
n!
_
λ
k!
_
d
dJ
_
k
_
n
_
R
e
−
1
2
ax
2
+Jx
dx.
Thus, the perturbative expansion can be written in the form
Z(J) =
_
2π
a
_
1/2
exp
_
λ
k!
_
d
dJ
_
k
_
exp
_
J
2
2a
_
. (1.8)
Two examples of this kind that will reappear frequently in the inﬁnite
dimensional version are the cubic case with P(x) =
g
6
x
3
and the quartic
case with P(x) =
λ
4!
x
4
.
To see then how the combinatorics of graphs can be used as a convenient
device to label the terms of diﬀerent order in λ and J in the perturbative
series of Z(J), ﬁrst observe that the term of order λ
α
and J
β
in Z(J) is
produced by the combination of the term of order α in the Taylor expansion
4 Feynman Motives
of the exponential exp(
λ
k!
(
d
dJ
)
k
) and the term of order β + kα in J in the
Taylor expansion of the other exponential exp(
J
2
2a
) in (1.8). All the resulting
terms will be of a similar form, consisting of a combinatorial factor given by
a ratio of two products of factorials, a power of J, a power of λ and a power
of 2a in the denominator. The graphs are introduced as a visual way to
keep track of the power counting in these terms, which are associated to the
vertices and the internal and external edges of the graph. The combinatorial
factor can then also be described in terms of symmetries of the graphs.
Here, as in general in perturbative quantum ﬁeld theory, one thinks of
graphs as being constructed out of a set of vertices and a set of half edges.
Each half edge has an end that is connected to a vertex and another end
that may pair to another half edge or remain unpaired. An internal edge
of the graph consists of a pair of half edges, hence it is an edge in the
usual graph theoretic sense, connecting two vertices. An external edge is
an unpaired half edge attached to a vertex of the graph. The graphs we
consider will not necessarily be connected. It is sometimes convenient to
adopt the convention that a graph (or connected component of a graph)
consisting of a single line should be thought of as consisting of an internal
and two external edges.
The way one assigns graphs to monomials of the form
λ
α
J
β
a
κ
is by the
following rules.
• To each factor of λ one associates a vertex of valence equal to the
degree of the given monomial P(x) =
λ
k!
x
k
. This means a total of
α vertices, each with k half edges attached.
• To each factor J one associates an external edge.
• The power of a
−1
is then determined by the resulting number of
internal edges obtained by pairing all the half edges to form a graph.
Notice that the procedure described here produces not one but a ﬁnite
collection of graphs associated to a given monomial
λ
α
J
β
a
κ
, depending on all
the diﬀerent possible pairings between the half edges. This collection of
graphs can in turn be subdivided into isomorphism types, each occurring
with a given multiplicity, which corresponds to the number of diﬀerent
pairings that produce equivalent graphs. These combinatorial factors are
the symmetry factors of graphs. To see more precisely how these factors
can be computed, we can introduce the analog, in this 1dimensional toy
model, of the Green functions in quantum ﬁeld theory. The function Z(J)
Perturbative quantum ﬁeld theory and Feynman diagrams 5
of (1.4) can be thought of as a generating function for the Green functions
Z(J) =
∞
N=0
J
N
N!
_
R
x
N
e
−
1
2
ax
2
+P(x)
dx = Z
∞
N=0
J
N
G
N
, (1.9)
where Z =
_
R
e
−
1
2
ax
2
+P(x)
dx and the Green functions are
G
N
=
_
R
x
N
N!
e
−
1
2
ax
2
+P(x)
dx
_
R
e
−
1
2
ax
2
+P(x)
dx
. (1.10)
Upon expanding out the interaction term exp(P(x)), with P(x) =
λ
k!
x
k
and
formally exchanging the sum with the integral, one obtains
G
N
=
∞
n=0
_
x
N
N!
(λx
k
)
n
(k!)
n
n!
e
−
1
2
ax
2
dx
∞
n=0
_
(λx
k
)
n
(k!)
n
n!
e
−
1
2
ax
2
dx
. (1.11)
Using (1.9), we then see that one way of computing the coeﬃcient of a term
in
λ
α
J
β
a
κ
in the asymptotic expansion of Z(J) is to count all the pairings
(the Wick contractions) that occur in the integration
_
R
x
N
x
kn
e
−
1
2
ax
2
dx. (1.12)
As we have seen in (1.3), these are (N +kn−1)!! for even N +kn, with the
odd ones vanishing by symmetry. Taking into account the other coeﬃcients
that appear in (1.9) and (1.11), one obtains the factor
(N +kn − 1)!!
N! n! (k!)
n
.
The meaning of this factor in terms of symmetries of graphs can be ex
plained, by identifying (N + kn − 1)!! with the number of all the possible
pairings of half edges, from which one factors out N! permutations of the
external edges, k! permutations of the half edges attached to each valence
k vertex and n! permutations of the n valence k vertices along with their
star of half edges, leaving all the diﬀerent pairings of half edges. These then
correspond to the sum over all the possible topologically distinct graphs ob
tained via these pairings, each divided by its own symmetry factor. Thus,
in terms of graphs, the terms of the asymptotic series in the numerator of
(1.10) become of the form
Γ∈graphs
λ
#V (Γ)
J
#E
ext
(Γ)
#Aut(Γ) a
#E
int
(Γ)
. (1.13)
Notice also how, when computing the terms of the asymptotic series using
either the Taylor series of the exponentials of (1.8) or by ﬁrst using the
6 Feynman Motives
expansion in Green functions and then the terms (1.12), one is implicitly
using the combinatorial identity
(2n − 1)!! =
(2n)!
2
n
n!
.
Passing from the 1dimensional case to a ﬁnite dimensional case in many
variables is notationally more complicated but conceptually very similar.
One replaces an integral of the form
_
R
N
e
−
1
2
x
t
Ax+Jx
dx
1
dx
N
=
(2π)
N/2
(det A)
1/2
e
1
2
JA
−1
J
t
, (1.14)
where the positive real number a > 0 of the 1dimensional case is now
replaced by an N N real matrix A with A
t
= A and positive eigenvalues.
The real number J is here an Nvector, with Jx the inner product. The
form of (1.14) is obtained by diagonalizing the matrix and reducing it back
to the 1dimensional case. One can again compute the asymptotic series
for the integral
_
R
N
e
−
1
2
x
t
Ax+P(x)+Jx
dx
1
dx
N
,
where the interaction term here will be a polynomial in the coordinates
x
i
of x, such as P(x) =
λ
4!
(
N
i=1
x
4
i
). One can use the same method of
labeling the terms in the asymptotic series by graphs, where now instead
of attaching a factor a
−1
to the internal edges one ﬁnds factors (A
−1
)
ij
for
edges corresponding to a Wick contraction pairing an x
i
and an x
j
.
The conceptually more diﬃcult step is to adapt this computational pro
cedure for ﬁnite dimensional integral to a recipe that is used to make sense
of “analogous” computations of functional integrals in quantum ﬁeld theory.
1.2 From Lagrangian to eﬀective action
In the case of a scalar ﬁeld theory, one replaces the expression
1
2
x
2
+P(x)
of the onedimensional toy model we saw in the previous section with a
nonlinear functional, the Lagrangian density, deﬁned on a conﬁguration
space of classical ﬁelds. Here we give only a very brief account of the basics
of perturbative quantum ﬁeld theory. A more detailed presentation, aimed
at giving a selfcontained introduction to mathematicians, can be found in
the book [Connes and Marcolli (2008)].
Perturbative quantum ﬁeld theory and Feynman diagrams 7
In the scalar case the classical ﬁelds are (smooth) functions on a space
time manifold, say φ ∈ (
∞
(R
D
, R), and the Lagrangian density is given by
an expression of the form
L(φ) =
1
2
(∂φ)
2
−
m
2
2
φ
2
− T(φ), (1.15)
where (∂φ)
2
= g
µν
∂
µ
φ∂
ν
φ for g
µν
the Lorentzian metric of signature
(1, −1, −1, . . . , −1) on R
D
and a summation over repeated indices under
stood. The interaction term T(φ) in the Lagrangian is a polynomial in
the ﬁeld φ of degree deg T ≥ 3. Thus, when one talks about a scalar ﬁeld
theory one means the choice of the data of the Lagrangian density and the
spacetime dimension D. We can assume for simplicity that T(φ) =
λ
k!
φ
k
.
We will give explicit examples using the special case of the φ
3
theory in
dimension D = 6: while this is not a physically signiﬁcant example because
of the unstable equilibrium point of the potential at φ = 0, it is both suﬃ
ciently simple and suﬃciently generic with respect to the renormalization
properties (e.g. nonsuperrenormalizable), unlike the more physical φ
4
in
dimension D = 4 .
To the Lagrangian density one associates a classical action functional
S
L
(φ) =
_
R
D
L(φ)d
D
x. (1.16)
The subscript L here stands for the Lorentzian signature of the metric and
we’ll drop it when we pass to the Euclidean version. This classical action
is written as the sum of two terms S
L
(φ) = S
free,L
(φ) + S
int,L
(φ), where
the free ﬁeld part is
S
free,L
(φ) =
_
R
D
_
1
2
(∂φ)
2
−
m
2
2
φ
2
_
d
D
x
and the interaction part is given by
S
int,L
(φ) = −
_
R
D
P(φ)d
D
x.
The probability amplitude associated to the classical action is the expression
e
i
S
L
(φ)
, (1.17)
where = h/2π is Planck’s constant. In the following we follow the conven
tion of taking units where = 1 so that we do not have to write explicitly
the powers of in the terms of the expansions. An observable of a scalar
ﬁeld theory is a functional on the conﬁguration space of the classical ﬁelds,
8 Feynman Motives
which we write as O(φ). The expectation value of an observable is deﬁned
to be the functional integral
¸O(φ)) =
_
O(φ)e
iS
L
(φ)
T[φ]
_
e
iS
L
(φ)
T[φ]
, (1.18)
where the integration is supposed to take place on the conﬁguration space
of all classical ﬁelds. In particular, one has the Npoint Green functions,
deﬁned here as
G
N,L
(x
1
, . . . , x
N
) =
_
φ(x
1
) φ(x
N
) e
iS
L
(φ)
T[φ]
_
e
iS
L
(φ)
T[φ]
, (1.19)
for which the generating function is given again by a functional integral
with source term
_
e
iS
L
(φ)+J,φ
T[φ], (1.20)
where J is a linear functional (a distribution) on the space of classical
ﬁelds and ¸J, φ) = J(φ) is the pairing of the space of ﬁelds and its dual. If
J = J(x) is itself a smooth function then ¸J, φ) =
_
R
D
J(x)φ(x)d
D
x.
Although the notation of (1.18) and (1.19) is suggestive of what the
computation of expectation values should be, there are in fact formidable
obstacles in trying to make rigorous sense of the functional integral involved.
Despite the successes of constructive quantum ﬁeld theory in several im
portant cases, in general the integral is illdeﬁned mathematically. This is,
in itself, not an obstacle to doing quantum ﬁeld theory, as long as one re
gards the expression (1.18) as a shorthand for a corresponding asymptotic
expansion, obtained by analogy to the ﬁnite dimensional case we have seen
previously.
A closer similarity between (1.20) and (1.4) appears when one passes
to Euclidean signature by a Wick rotation to imaginary time t → it. This
has the eﬀect of switching the signature of the metric to (1, 1, . . . , 1), after
factoring out a minus sign, which turns the probability amplitude into the
Euclidean version
e
iS
L
(φ)
→ e
−S(φ)
, (1.21)
with the Euclidean action
S(φ) =
_
R
D
_
1
2
(∂φ)
2
+
m
2
2
φ
2
+ T(φ)
_
d
D
x. (1.22)
Thus, in the Euclidean version we are computing functional integrals of the
form
G
N
(x
1
, . . . , x
N
) =
_
φ(x
1
) φ(x
N
) e
−S(φ)
T[φ]
_
e
−S(φ)
T[φ]
, (1.23)
Perturbative quantum ﬁeld theory and Feynman diagrams 9
for which the generating function resembles (1.4) in the form
Z[J] =
_
e
−
R
D
1
2
(∂φ)
2
+
m
2
2
φ
2
+P(φ)+J(x)φ(x)
d
D
x
T[φ], (1.24)
satisfying
Z[J]
Z[0]
=
∞
N=0
1
N!
_
J(x
1
) J(x
N
) G
N
(x
1
, . . . , x
N
) d
D
x
1
d
D
x
N
, (1.25)
for
Z[0] =
_
e
−
R
D
1
2
(∂φ)
2
+
m
2
2
φ
2
+P(φ)
d
D
x
T[φ]. (1.26)
In order to make sense of this functional integral, one uses an analog of
the asymptotic expansion (1.6), where one expands out the exponential of
the interaction term S
int
(φ) =
_
R
D
T(x) d
D
x of the Euclidean action and
one follows the same formal rules about integration by parts as in the ﬁnite
dimensional case to label the terms of the expansion by graphs. What is
needed in order to write the contribution of a given graph to the asymptotic
series is to specify the rules that associate the analogs of the powers of λ,
J and a
−1
to the vertices, external and internal edges of the graph. These
are provided by the Feynman rules of the theory.
1.3 Feynman rules
By analogy to what we saw in the 1dimensional model, where one writes
the Green functions (1.11) in terms of integrals of the form (1.12), and
the latter in terms of sums over graphs as in (1.13), one also writes the
Green functions (1.23) in terms of an asymptotic series whose terms are
parameterized by graphs,
(
N
(p
1
, . . . , p
N
) =
Γ
V (Γ, p
1
, . . . , p
N
)
#Aut(Γ)
, (1.27)
where (
N
(p
1
, . . . , p
N
) is the Green function in momentum space, i.e. the
Fourier transform
(
N
(p
1
, . . . , p
N
) =
_
G
N
(x
1
, . . . , x
N
)e
i(p
1
x
1
+···+p
N
x
N
)
d
D
p
1
(2π)
D
d
D
p
N
(2π)
D
.
(1.28)
The reason for writing the contributions of Feynman integrals in momen
tum space is that in physics one does not only think of the Feynman graphs
as computational devices that do the bookkeeping of terms in integration
10 Feynman Motives
by parts of polynomials under a Gaussian measure, but one can think of a
diagram as representing a (part of) a physical process, where certain parti
cles with assigned momenta (external edges) interact (vertices) by creation
and annihilation of virtual particles (internal edges). The momenta ﬂowing
through the graph then represent the physical process. In fact, it is clear
from this point of view that what has physical meaning is not so much
an individual graph but the collection of all graphs with given external
edges and assigned external momenta, and among them the subset of all
those with a ﬁxed number of loops. The latter speciﬁes the order in the
perturbative expansion one is looking at. The terms V (Γ, p
1
, . . . , p
N
) are
constructed according to the Feynman rules as follows.
• Each internal edge e ∈ E
int
(Γ) contributes a momentum variable
k
e
∈ R
D
so that
V (Γ, p
1
, . . . , p
N
) =
_
1
Γ
(p
1
, . . . , p
N
, k
1
, . . . , k
n
)
d
D
k
1
(2π)
D
d
D
k
n
(2π)
D
,
(1.29)
for n = #E
int
(Γ). The term 1
Γ
(p
1
, . . . , p
N
, k
1
, . . . , k
n
) is con
structed according to the following procedure.
• Each vertex v ∈ V (Γ) contributes a factor of λ
v
(2π)
D
, where λ
v
is
the coupling constant of the monomial in the Lagrangian of order
equal to the valence of v and a conservation law for all the momenta
that ﬂow through that vertex,
δ
v
(k) := δ(
s(e)=v
k
e
−
t(e)=v
k
e
), (1.30)
written after chosing an orientation of the edges of the graph. In
the case of vertices with both internal and external edges (1.30) is
equivalently written in the form
δ
v
(k, p) := δ(
n
i=1
v,i
k
i
+
N
j=1
v,j
p
j
), (1.31)
where the incidence matrix of the graph Γ is the #V (Γ)#E(Γ)
matrix with
v,e
=
_
¸
¸
_
¸
¸
_
+1 for v = t(e)
−1 for v = s(e)
0 for v / ∈ ∂(e).
(1.32)
Perturbative quantum ﬁeld theory and Feynman diagrams 11
• Each internal edge contributes an inverse propagator, that is, a
term of the form q
−1
e
, where q
e
is a quadratic form, which in the
case of a scalar ﬁeld in Euclidean signature is given by
q
e
(k
e
) = k
2
e
+m
2
. (1.33)
• Each external edge e ∈ E
ext
(Γ) contributes a propagator q
e
(p
e
)
−1
,
with q
e
(p
e
) = p
2
e
+m
2
. The external momenta are assigned so that
they satisfy the conservation law
e
p
e
= 0, when summed over
the oriented external edges.
• The integrand 1
Γ
(p
1
, . . . , p
N
, k
1
, . . . , k
n
) is then a product
v∈V (Γ)
λ
v
(2π)
D
δ
v
(k
e
i
, p
e
j
)
e
i
∈E
int
(Γ)
q
e
i
(k
e
i
)
−1
e
j
∈E
ext
(Γ)
q
e
j
(p
e
j
)
−1
,
(1.34)
with linear relations among the momentum variables k
e
i
and p
e
j
imposed by the conservation laws δ
v
(k
e
i
, p
e
j
) at the vertices of the
graph.
We can then write the Feynman integral associated to a Feynman graph
Γ of the given theory in the form
V (Γ, p
1
, . . . , p
N
) = ε(p
1
, . . . , p
N
) U(Γ, p
1
, . . . , p
N
), (1.35)
where the factor ε(p) is the product of the inverse propagators of the ex
ternal edges
ε(p
1
, . . . , p
N
) =
e∈E
ext
(Γ)
q
e
(p
e
)
−1
, (1.36)
while the factor U(Γ, p) is given by
U(Γ, p
1
, . . . , p
N
) = C
_
δ(
n
i=1
v,i
k
i
+
N
j=1
v,j
p
j
)
q
1
(k
1
) q
n
(k
n
)
d
D
k
1
(2π)
D
d
D
k
n
(2π)
D
,
(1.37)
with C =
v∈V (Γ)
λ
v
(2π)
D
.
We work here for convenience always with Euclidean signature in the
Feynman integrals, so that our propagators contain quadratic terms of the
form (1.33). The study of parametric Feynman integral we discuss in greater
detail in Chapter 3 can be done also for the Lorentzian case, as shown for
instance in [Bjorken and Drell (1965)].
12 Feynman Motives
1.4 Simplifying graphs: vacuum bubbles, connected graphs
There are some useful simpliﬁcations that can be done in the combinatorics
of graphs that appear in the formal series (1.27).
The basic property that makes these simpliﬁcations possible is the mul
tiplicative form (1.34) of the Feynman integrand 1
Γ
(p
1
, . . . , p
N
, k
1
, . . . , k
n
).
This implies the following property.
Lemma 1.4.1. The Feynman integral V (Γ, p
1
, . . . , p
N
) is multiplicative on
connected components of the graph Γ.
Proof. This follows immediately from the form (1.34) and (1.35) with
(1.36), (1.37) of the Feynman integral. In fact, if the graph Γ has diﬀer
ent connected components, no linear relations arise between momentum
variables of the edges of diﬀerent components (as these have no common
vertices) and the corresponding integrals split as a product.
Moreover, one also has the following multiplicative form of the symme
try factors of graphs.
Lemma 1.4.2. For a graph Γ that is a union of connected components Γ
j
with multiplicities n
j
(i.e. there are n
j
connected components of Γ all iso
morphic to the same graph Γ
j
), the symmetry factor splits multiplicatively
on components according to the formula
#Aut(Γ) =
j
(n
j
)!
j
#Aut(Γ
j
)
n
j
. (1.38)
Proof. The factorials come from the symmetries of the graph Γ that
permute topologically equivalent components. All symmetries of Γ are
obtained by composing this type of symmetries with symmetries of each
component.
One then has a ﬁrst useful observation on the combinatorics of the
graphs that appear in the asymptotic expansion of the Green functions. A
graph with no external edges is commonly referred to as a vacuum bubble.
Lemma 1.4.3. The graphs of (1.27) do not contain any vacuum bubbles.
Proof. As we have seen in the ﬁnite dimensional toy model, these cor
respond to the terms with J = 0 in the asymptotic series. Thus, when
one writes the expansion (1.25) into Green functions, and then the expan
sion (1.27) of the latter into Feynman integrals of graphs, the expansion
Perturbative quantum ﬁeld theory and Feynman diagrams 13
of the functional integral Z[J] would count the contribution of all graphs
including components that are vacuum bubbles as in the case of (1.8) in
the ﬁnite dimensional case. The expansion of Z[0] on the other hand only
has contributions from the vacuum bubble graphs, and the multiplicative
properties of Lemma 1.4.1 and Lemma 1.4.2 then imply that the expansion
for Z[J]/Z[0] only has contributions from graphs which do not contain any
connected components that are vacuum bubbles.
One can then pass from multiconnected to connected graphs by rewrit
ing the functional integral Z[J] in an equivalent form in terms of
W[J] = log
_
Z[J]
Z[0]
_
. (1.39)
One can again write a formal asymptotic series for W[J] as
W[J] =
∞
N=0
1
N!
_
J(x
1
) J(x
N
) G
N,c
(x
1
, . . . , x
N
) d
D
x
1
d
D
x
N
,
(1.40)
where now the Green functions G
N,c
(x
1
, . . . , x
N
) will also have an expan
sion on graphs of the form (1.27), where, however, only a smaller class of
graphs will be involved.
Lemma 1.4.4. The connected Green functions G
N,c
(x
1
, . . . , x
N
) of (1.40)
have an expansion
(
N,c
(p
1
, . . . , p
N
) =
Γ connected
V (Γ, p
1
, . . . , p
N
)
#Aut(Γ)
, (1.41)
where (
N,c
(p
1
, . . . , p
N
) is the Fourier transform
(
N,c
(p
1
, . . . , p
N
) =
_
G
N,c
(x
1
, . . . , x
N
)e
i(p
1
x
1
+···+p
N
x
N
)
d
D
p
1
(2π)
D
d
D
p
N
(2π)
D
(1.42)
and the V (Γ, p
1
, . . . , p
N
) in (1.41) are computed as in (1.35).
Proof. We only sketch brieﬂy why the result holds. More detailed ex
positions can be found in standard Quantum Field Theory textbooks (for
example in [LeBellac (1991)]). Suppose that Γ is a disjoint union of con
nected components Γ = ∪
j
Γ
j
, with multiplicities n
j
and with N
j
external
edges, so that
j
n
j
N
j
= N. Then by Lemma 1.4.1 and 1.4.2 we get
V (Γ, p
1
, . . . , p
N
)
#Aut(Γ)
=
j
V (Γ
j
, p
1
, . . . , p
N
j
)
n
j
(n
j
)!#Aut(Γ
j
)
n
j
.
14 Feynman Motives
Thus, we can write
Z[J]
Z[0]
=
N
1
N!
_
J(x
1
) J(x
N
) G
N
(x
1
, . . . , x
N
)
i
d
D
x
i
=
N
j
n
j
N
j
=N
j
1
n
j
__
J(x
1
) J(x
N
j
)G
N
j
,c
(x
1
, . . . , x
N
j
)
_
n
j
= exp
_
N
1
N!
_
J(x
1
) J(x
N
) G
N,c
(x
1
, . . . , x
N
)
_
.
1.5 Oneparticleirreducible graphs
Further simpliﬁcations of the combinatorics of graphs can be obtained by
passing to the 1PI eﬀective action, or higher loop versions like the 2PI eﬀec
tive action, etc. We discuss here brieﬂy only the 1PI eﬀective action, though
we will later need to return to discussing higher connectivity conditions on
Feynman graphs. We ﬁrst recall the following notions of connectivity of
graphs.
Deﬁnition 1.5.1. The notion of kconnectivity of graphs is given as fol
lows:
• A graph is kedgeconnected if it cannot be disconnected by removal
of any set of k or fewer edges.
• A graph is 2vertexconnected if it has no looping edges, it has
at least 3 vertices, and it cannot be disconnected by removal of a
single vertex, where vertex removal is deﬁned as below.
• For k ≥ 3, a graph is kvertexconnected if it has no looping edges
and no multiple edges, it has at least k + 1 vertices, and it cannot
be disconnected by removal of any set of k − 1 vertices.
Here what one means by removing a vertex from a graph is the following.
Given a graph Γ and a vertex v ∈ V (Γ), the graph Γ v is the graph with
vertex set V (Γ) ¦v¦ and edge set E(Γ) ¦e : v ∈ ∂(e)¦, i.e. the graph
obtained by removing from Γ the star of the vertex v. Thus, 1vertex
connected and 1edgeconnected simply mean connected, while for k ≥ 2
the condition of being kvertexconnected is stronger than that of being
kedgeconnected. The terminology more commonly in use in the physics
literature is the following.
Perturbative quantum ﬁeld theory and Feynman diagrams 15
Deﬁnition 1.5.2. For k ≥ 2 a (k + 1)edgeconnected graph is also called
kparticleirreducible (kPI). For k = 1, a 2edgeconnected graph that is
not a tree is called oneparticleirreducible (1PI) graph. These cannot be
disconnected by removal of a single (internal) edge.
Notice that trees are all considered not to be 1PI, even though a tree
consisting of just n edges attached to a single valence n vertex cannot be
disconnected by removal of a single edge (such edges are not internal though
in this case).
Lemma 1.5.3. Any connected graph can be obtained from a tree, after
replacing the vertices by 1PI graphs with the number of external edges equal
to the valence of the vertex.
Proof. If the connected graph Γ is 1PI the tree consists of a single vertex
with the number of edges attached equal to the number of external edges of
the graph. Suppose the graph is not 1PI. Find an edge that disconnects the
graph. Look at each component and again repeat the operations ﬁnding
edges that disconnect them further until one is left with a collection of 1PI
graphs, Γ
1
, . . . , Γ
n
. It then suﬃces to show that the graph obtained from Γ
by shrinking each Γ
i
to a vertex is a tree. Since each of the internal edges
that remain in this graph was an edge whose removal increased the number
of connected components, this must still be true in the graph obtained after
collapsing all the Γ
i
. A graph that is disconnected by the removal of any
one internal edge is a tree with at least one internal edge.
This suggests that there should be a way to further simplify the combi
natorics of graphs in the asymptotic expansion of the functional integrals,
by counting separately the contributions of trees and that of 1PI graphs,
and getting back from these the contributions of all connected graphs.
This is done by passing to the 1PI eﬀective action, which is deﬁned as
the Legendre transform of the functional W[J], namely
S
eﬀ
(φ) = (¸φ, J) −W[J]) [
J=J(φ)
, (1.43)
evaluated at a stationary J, that is, a solution of the variational equation
δ
δJ
(¸φ, J) −W[J]) = 0.
The asymptotic expansion of the eﬀective action collects the contri
butions of all the 1PI graphs, so that the semiclassical calculations (i.e.
involving only graphs that are trees) done with the eﬀective action recover
16 Feynman Motives
the full quantum corrections to the classical action given by all the con
nected graphs that appear in the expansion of W[J]. We do not prove here
how one gets the asymptotic expansion for the eﬀective action and we refer
the interested reader to [Connes and Marcolli (2008)] and [LeBellac (1991)].
To explain more precisely the diﬀerence between edge and vertex con
nectivity in Deﬁnition 1.5.1 above, and their relation to the 1PI condition,
we recall the following observations from [Aluﬃ and Marcolli (2009a)].
Deﬁnition 1.5.4. A graph Γ
is a splitting of Γ at a vertex v ∈ V (Γ) if it
is obtained by partitioning the set E ⊂ E(Γ) of edges adjacent to v into
two disjoint nonempty subsets, E = E
1
∪ E
2
and inserting a new edge e
to whose end vertices v
1
and v
2
the edges in the two sets E
1
and E
2
are
respectively attached, as in the example in the ﬁgure.
We have the following relation between 2vertexconnectivity and 2
edgeconnectivity (1PI).
Lemma 1.5.5. Let Γ be a graph with at least 3 vertices and no looping
edges.
(1) If Γ is 2vertexconnected then it is also 2edgeconnected (1PI).
(2) Γ is 2vertexconnected if and only if all the graphs Γ
obtained as split
tings of Γ at any v ∈ V (Γ) are 2edgeconnected (1PI).
Proof. (1) We have to show that, for a graph Γ with at least 3 vertices and
no looping edges, 2vertexconnectivity implies 2edgeconnectivity. As
sume that Γ is not 1PI. Then there exists an edge e such that Γe has two
connected components Γ
1
and Γ
2
. Since Γ has no looping edges, e has two
distinct endpoints v
1
and v
2
, which belong to the two diﬀerent components
after the edge removal. Since Γ has at least 3 vertices, at least one of the
two components contains at least two vertices. Assume then that there
exists v ,= v
1
∈ V (Γ
1
). Then, after the removal of the vertex v
1
from Γ,
the vertices v and v
2
belong to diﬀerent connected components, so that Γ
is not 2vertexconnected.
(2) We need to show that 2vertexconnectivity is equivalent to all split
tings Γ
being 1PI. Suppose ﬁrst that Γ is not 2vertexconnected. Since
Perturbative quantum ﬁeld theory and Feynman diagrams 17
Γ has at least 3 vertices and no looping edges, the failure of 2vertex
connectivity means that there exists a vertex v whose removal disconnects
the graph. Let V ⊂ V (Γ) be the set of vertices other than v that are end
points of the edges adjacent to v. This set is a union V = V
1
∪ V
2
where
the vertices in the two subsets V
i
are contained in at least two diﬀerent
connected components of Γ v. Then the splitting Γ
of Γ at v obtained
by inserting an edge e such that the endpoints v
1
and v
2
are connected by
edges, respectively, to the vertices in V
1
and V
2
is not 1PI.
Conversely, assume that there exists a splitting Γ
of Γ at a vertex v
that is not 1PI. There exists an edge e of Γ
whose removal disconnects the
graph. If e already belonged to Γ, then Γ would not be 1PI (and hence not
2vertex connected, by (1)), as removal of e would disconnect it. So e must
be the edge added in the splitting of Γ at the vertex v.
Let v
1
and v
2
be the endpoints of e. None of the other edges adjacent
to v
1
or v
2
is a looping edge, by hypothesis; therefore there exist at least
another vertex v
1
,= v
2
adjacent to v
1
, and a vertex v
2
,= v
1
adjacent to v
2
.
Since Γ
e is disconnected, v
1
and v
2
are in distinct connected components
of Γ
e. Since v
1
and v
2
are in Γ v, and Γ v is contained in Γ
e, it
follows that removing v from Γ would also disconnect the graph. Thus Γ is
not 2vertexconnected.
Lemma 1.5.6. Let Γ be a graph with at least 4 vertices, with no looping
edges and no multiple edges. Then 3vertexconnectivity implies 3edge
connectivity.
Proof. We argue by contradiction. Assume that Γ is 3vertexconnected
but not 2PI. We know it is 1PI because of the previous lemma. Thus, there
exist two edges e
1
and e
2
such that the removal of both edges is needed
to disconnect the graph. Since we are assuming that Γ has no multiple or
looping edges, the two edges have at most one endpoint in common.
Suppose ﬁrst that they have a common endpoint v. Let v
1
and v
2
denote the remaining two endpoints, v
i
∈ ∂e
i
, v
1
,= v
2
. If the vertices v
1
and v
2
belong to diﬀerent connected components after removing e
1
and e
2
,
then the removal of the vertex v disconnects the graph, so that Γ is not 3
vertexconnected (in fact not even 2vertexconnected). If v
1
and v
2
belong
to the same connected component, then v must be in a diﬀerent component.
Since the graph has at least 4 vertices and no multiple or looping edges,
there exists at least one other edge attached to either v
1
, v
2
, or v, with
the other endpoint w / ∈ ¦v, v
1
, v
2
¦. If w is adjacent to v, then removing
v and v
1
leaves v
2
and w in diﬀerent connected components. Similarly,
18 Feynman Motives
if w is adjacent to (say) v
1
, then the removal of the two vertices v
1
and
v
2
leave v and w in two diﬀerent connected components. Hence Γ is not
3vertexconnected.
Next, suppose that e
1
and e
2
have no endpoint in common. Let v
1
and
w
1
be the endpoints of e
1
and v
2
and w
2
be the endpoints of e
2
. At least
one pair ¦v
i
, w
i
¦ belongs to two separate components after the removal of
the two edges, though not all four points can belong to diﬀerent connected
components, else the graph would not be 1PI. Suppose then that v
1
and w
1
are in diﬀerent components. It also cannot happen that v
2
and w
2
belong
to the same component, else the removal of e
1
alone would disconnect the
graph. We can then assume that, say, v
2
belongs to the same component
as v
1
while w
2
belongs to a diﬀerent component (which may or may not be
the same as that of w
1
). Then the removal of v
1
and w
2
leaves v
2
and w
1
in two diﬀerent components so that the graph is not 3vertexconnected.
Conditions of 3connectivity (3vertexconnected or 3edgeconnected)
arise in a more subtle manner in the theory of Feynman integrals, in the
analysis of Laundau singularities (see for instance [Sato, Miwa, Jimbo, Os
hima (1976)]). In particular, the 2PI eﬀective action is often considered
in quantum ﬁeld theory in relation to nonequilibrium phenomena, see e.g.
[Rammer (2007)], ¸10.5.1.
In the following we restrict our attention to Feynman integrals of graphs
that are at least 1PI.
1.6 The problem of renormalization
So far we have treated the integrals U(Γ, p
1
, . . . , p
N
) of (1.37) as purely
formal expressions. However, if one tries to assign to such integrals a nu
merical value, one soon realizes that most of them are in fact divergent.
This was historically one of the main problems in the development of per
turbative quantum ﬁeld theory, namely the renormalization problem: how
to extract in a consistent and physically signiﬁcant manner ﬁnite values
from the divergent integrals (1.37) that appear in the asymptotic expan
sion of the functional integrals of quantum ﬁeld theory.
The problem of renormalization consists of three main aspects:
• Regularization
• Subtraction
• Renormalization
Perturbative quantum ﬁeld theory and Feynman diagrams 19
Regularization consists of a procedure that replaces the divergent integrals
(1.37) by functions of some regularization parameters, in such a way that
the resulting function has a pole or a divergence for particular values or
limits of the additional parameters that correspond to the original diver
gent integral, but has ﬁnite values for other values of the regularization
parameters. Subtraction then consists of removing the divergent part of
the regularized integrals by a uniform procedure (such as removing the po
lar part of a Laurent series in the main example we use below). This is
not all there is yet. Renormalization means being able to perform the sub
traction procedure by modifying the parameters in the Lagrangian (which
become themselves functions of the regularization parameter).
To understand more clearly the last point, it is important to stress the
fact that the parameters that appear in the Lagrangian, such as masses and
coupling constants of the interaction terms, are not physical observables,
nor are they the same as the actual masses and physical quantities that one
can measure in experiments. In fact, the parameters in the Lagrangian can
be modiﬁed without aﬀecting the physics one observes. This is what makes
it possible, in a renormalizable theory, to correct for divergent graphs by
readjusting the parameters in the Lagrangian. Thus, if one introduces a
regularization of the divergent Feynman integrals in terms of a complex
parameter z (as we discuss below) or in terms of a cutoﬀ Λ, then the
Lagrangian can be modiﬁed by changing the coeﬃcients to
L(φ) =
_
1 +δZ
2
(∂φ)
2
+
m
2
+δm
2
2
φ
2
+
λ +δλ
k!
φ
k
_
, (1.44)
where the functions δZ, δm
2
and δλ depend on the regularization param
eter. They consist, in fact, of a formal series of contributions coming from
all the divergent graphs of the theory. Notice that a theory is still renor
malizable if, in addition to the modiﬁcations of the coeﬃcients of the terms
initially present in the Lagrangian, to compensate for divergent graphs one
needs to also add a ﬁnite number of other terms, i.e. other monomials
δλ
i
(k
i
)!
φ
k
i
, that were not initially present. In fact, one can just think of this
as being the eﬀect of having chosen an arbitrary value = 0 for the coef
ﬁcients of these terms in the initial Lagrangian. However, a theory is no
longer renormalizable if an inﬁnite number of additional terms is needed to
compensate for the divergences.
20 Feynman Motives
1.7 Gamma functions, Schwinger and Feynman parameters
We digress momentarily to recall some useful formulae involving Gamma
functions, which are extensively used in Feynman integral computations
and are the basis of both the dimensional regularization procedure we de
scribe below and the parametric representation of Feynman integrals that
we discuss later and which is the basis for the relation between Feynman
integrals and periods of algebraic varieties.
First recall that the Gamma function is deﬁned by the integral
Γ(t + 1) =
_
∞
0
s
t
e
−s
ds. (1.45)
It satisﬁes Γ(t+1) = t Γ(t), hence it extends the factorial, namely Γ(n+1) =
n! for nonnegative integers. The function Γ(t) deﬁned in this way extends
to a meromorphic function with poles at all the nonpositive integers.
As we are going to see in more detail below, a typical way of dealing
with divergences in Feynman integrals is to ﬁrst identify them with poles of
some meromorphic function and typically a product of Gamma functions.
The ﬁrst useful operation on Feynman integrals is the introduction of
Schwinger parameters. These are based on the very simple identity
1
q
=
_
∞
0
e
−sq
ds. (1.46)
This allows the reformulation of integrals where quadratic forms q
i
in the
momentum variables appear in the denominator in terms of Gaussian inte
grals where quadratic forms appear in the exponent. In terms of Schwinger
parameters, one writes the denominator of a Feynman integral of the form
(1.37) as
1
q
1
q
n
=
_
R
n
+
e
−(s
1
q
1
+···+s
n
q
n
)
ds
1
ds
n
. (1.47)
This is a special case of the more general useful identity
1
q
k
1
1
q
k
n
n
=
1
Γ(k
1
) Γ(k
n
)
_
R
n
+
e
−(s
1
q
1
+···+s
n
q
n
)
s
k
1
−1
1
s
k
n
−1
n
ds
1
ds
n
.
(1.48)
Another related way to reformulate the expression (1.37) of Feynman
integrals is based on the Feynman parameters. Here the basic example,
analogous to (1.46) is the Feynman trick
1
ab
=
_
1
0
1
(ta + (1 −t)b)
2
dt. (1.49)
Perturbative quantum ﬁeld theory and Feynman diagrams 21
The more general expression analogous to (1.47) is obtained by considering
the general formula (1.48) and performing a change of variables s
i
= St
i
with S = s
1
+ +s
n
so as to obtain
1
q
k
1
1
q
k
n
n
=
Γ(k
1
+ +k
n
)
Γ(k
1
) Γ(k
n
)
_
[0,1]
n
t
k
1
−1
1
t
k
n
−1
n
δ(1 −
i
t
i
)
(t
1
q
1
+ +t
n
q
n
)
n
dt
1
dt
n
.
(1.50)
In the particular case of the denominators of (1.37) this gives
1
q
1
q
n
= (n − 1)!
_
[0,1]
n
δ(1 −
i
t
i
)
(t
1
q
1
+ +t
n
q
n
)
n
dt
1
dt
n
. (1.51)
Thus, the integration is performed on the ndimensional topological simplex
σ
n
= ¦t = (t
1
, . . . , t
n
) ∈ R
n
+
[
i
t
i
= 1¦. (1.52)
1.8 Dimensional Regularization and Minimal Subtraction
Dimensional regularization (DimReg) is based on a formal procedure aimed
at making sense of integrals in “complexiﬁed dimension” D−z, with z ∈ C
∗
,
instead of integral dimension D ∈ N. It would seem at ﬁrst that one needs
to develop a notion of geometry in “complexiﬁed dimension” along with
measure spaces and a suitable theory of integration, in order to deﬁne
dimensional regularization property. In fact, much less is needed. Due to
the very special form of the Feynman integrals U(Γ, p
1
, . . . , p
N
), it suﬃces
to have a good deﬁnition for the Gaussian integrals in D dimensions,
_
e
−λt
2
d
D
t = π
D/2
λ
−D/2
, (1.53)
in the case where D is no longer a positive integer but a complex number. In
fact, one can then reformulate the Feynman integrals in terms of Gaussian
integrals, using the method of Schwinger parameters described brieﬂy in
¸1.7.
Clearly, since the right hand side of (1.53) continues to make sense for
D ∈ C, one can use the right hand side of (1.53) as the deﬁnition of the
left hand side and set
_
e
−λt
2
d
z
t := π
z/2
λ
−z/2
, ∀z ∈ C. (1.54)
22 Feynman Motives
One then obtains well deﬁned Feynman integrals in complexiﬁed dimen
sion D −z, which one writes formally as
U
z
µ
(Γ(p
1
, . . . , p
N
)) =
_
1
Γ
(p
1
, . . . , p
N
, k
1
, . . . , k
) µ
z
d
D−z
k
1
d
D−z
k
.
(1.55)
The variable µ has the physical units of a mass and appears in these inte
grals for dimensional reasons. It will play an important role later on, as it
sets the dependence on the energy scale of the renormalized values of the
Feynman integrals, hence the renormalization group ﬂow.
It is not an easy result to show that the dimensionally regularized in
tegrals give meromorphic functions in the variable z, with a Laurent series
expansion at z = 0. See a detailed discussion of this point in Chapter 1
of [Connes and Marcolli (2008)]. We will not enter into details here and
talk loosely about (1.55) as a meromorphic function of z depending on the
additional parameter µ.
The method of minimal subtraction consists of removing the divergent
part of the Laurent series expansion at z = 0 of a meromorphic function
obtained from the regularization of the Feynman integrals. One denotes by
T the operator of projection of a Laurent series onto its polar part. With
this notation, if U(Γ) = U(Γ)(z) is shorthand for the meromorphic function
obtained from the dimensional regularization of the Feynman integral asso
ciated to the graph Γ (suppressing the explicit dependence on the external
momenta), then (1 − T)U(Γ) would be its minimal subtraction. This is a
convergent power series in the variable z ∈ C
∗
, which then has a ﬁnite value
at z = 0.
As we see below (see also [Connes and Marcolli (2008)] for a more de
tailed explanation), taking (1 −T)U(Γ)[
z=0
does not suﬃce as a renormal
ization method, due to the role of subdivergences, i.e. smaller subgraphs
γ of the Feynman graph Γ which themselves already contribute divergent
Feynman integrals U(γ). The correct procedure that uses dimensional reg
ularization and minimal subtraction to extract ﬁnite values from divergent
Feynman integrals is given by the Bogolyubov recursion described below in
¸5.1.
We report here from [Connes and Marcolli (2008)], [Collins (1986)] the
simplest example of dimensional regularization, for the selfenergy graph of
the theory T with L
(φ) = φ
3
and D = 6.
Perturbative quantum ﬁeld theory and Feynman diagrams 23
k
p + k
p p
This corresponds to the divergent Feynman integral (neglecting the con
stant multiplicative factor containing the coupling constant λ and powers
of 2π) is given by
_
1
k
2
+m
2
1
(p +k)
2
+m
2
d
D
k
for the theory with Lagrangian L(φ) =
1
2
(∂φ)
2
+
m
2
2
φ
2
+
λ
6
φ
3
. This follows
directly from the application of the Feynman rules: the variables assigned
to the two internal edges are −k and p+k, with p the external momentum,
due to the eﬀect of the delta functions imposing the conservation law at the
two vertices. It is easy to see that this integral is divergent, for instance in
dimension D = 4 or D = 6, where it is usually interesting to consider this
scalar ﬁeld theory.
In this case, the method of Schwinger parameters described in ¸1.7 con
sists of replacing the integral above by
1
k
2
+m
2
1
(p +k)
2
+m
2
=
_
s>0, t>0
e
−s(k
2
+m
2
)−t((p+k)
2
+m
2
)
ds dt.
One can then diagonalize the quadratic form in the exponential to get
−Q(k) = −λ((k +xp)
2
+ ((x −x
2
)p
2
+m
2
))
with s = (1 − x)λ and t = xλ. One obtains in this way a Gaussian in
q = k +xp and using the prescription (1.54) one then gets
_
1
0
_
∞
0
e
−(λ(x−x
2
)p
2
+λm
2
)
_
e
−λq
2
d
D
q λdλ dx
= π
D/2
_
1
0
_
∞
0
e
−(λ(x−x
2
)p
2
+λm
2
)
λ
−D/2
λdλdx
= π
D/2
Γ(2 −D/2)
_
1
0
((x −x
2
)p
2
+ m
2
)
D/2−2
dx,
which makes sense for D ∈ C
∗
and shows the presence of a pole at D = 6.
It seems then that one could simply cure these divergences by removing
the polar part of the Laurent series obtained by dimensional regulariza
tion. Slightly more complicated examples with nested divergent graphs
24 Feynman Motives
(see [Connes and Marcolli (2008)] [Collins (1986)]) show why a renormal
ization procedure is indeed needed at this point. In fact, the main point is
that one would like to cancel the divergence by correcting the coeﬃcients
of the Lagrangian by functions of z, the counterterms, that is, by redoing
the Feynman integral computation for the modiﬁed Lagrangian
L(φ) =
1
2
(∂φ)
2
(1 −δZ) +
_
m
2
−δm
2
2
_
φ
2
−
g +δg
6
φ
3
.
While one can check that this works for the example given here above, it
does not seem to work any longer for more complicated examples where
the graph Γ contains divergent subgraphs. In such cases, one needs to
account for the way the divergences of the subgraphs have already been
renormalized, by correcting the Laurent series U(Γ) by a linear combination
of other Laurent series associated to the subgraphs. This will be explained
in more detail in ¸5.1 below.
We conclude this introductory section by recalling more formally the
main distinction we already mentioned between renormalizable and non
renormalizable theories.
Deﬁnition 1.8.1. A quantum ﬁeld theory with Lagrangian L(φ) =
1
2
(∂φ)
2
+
m
2
2
φ
2
+ T(φ), with polynomial interaction terms T(φ) deﬁnes
a renormalizable quantum ﬁeld theory if all the divergences arising from
the corresponding Feynman integrals U(Γ) can be corrected by repeatedly
altering the bare constants of the existing terms in the Lagrangian or by
adding a ﬁnite number of new polynomial terms to L(φ).
In the case of a nonrenormalizable theory, one can still follow the same
renormalization procedure. However, since inﬁnitely many new terms will
have to be added to the Lagrangian to correct for the divergences of increas
ingly complicated graphs, the theory will end up depending on inﬁnitely
many parameters.
Chapter 2
Motives and periods
After reviewing the main tools from quantum ﬁeld theory that we need,
we give here a very brief and sketchy account of the other subject whose
relation to quantum ﬁeld theory will then be the main topic of the rest of
the book, the theory of motives of algebraic varieties. This will be a simple
introduction, were we will inevitably avoid the deeper and more diﬃcult
aspects of the theory, and focus only on the minimal amount of information
that is necessary in order to explain how the connection to quantum ﬁeld
theory arises. We ﬁrst sketch the construction and main properties of pure
motives. As will become clearer in the following chapter, these do not suﬃce
for the quantum ﬁeld theoretic applications, for which the more complicated
theory of mixed motives is needed. We are even sketchier when it comes to
describing the world of mixed motives, but we give some concrete examples
of mixed Tate motives, which will be the type of motives that, at least
conjecturally, mainly arise in quantum ﬁeld theory. We describe brieﬂy
the role of periods, which will be the main point of contact with Feynman
integral computations, with special attention to the case of multiple zeta
values. We also show how one can estimate the complexity of a motive
through an associated invariant, a universal Euler characteristic given by
the class in the Grothendieck ring of varieties. Finally, we recall the role of
Galois symmetries in categories of motives.
2.1 The idea of motives
Grothendieck introduced the idea of motives as a universal cohomology the
ory for algebraic varieties. There are diﬀerent notions of cohomology that
can be deﬁned for algebraic varieties: Betti cohomology, algebraic de Rham
cohomology, ´etale cohomology. There are relations between them, like the
25
26 Feynman Motives
period isomorphism between de Rham and Betti cohomologies. The idea
of a universal cohomology that lies behind all of these diﬀerent realizations
asserts that properties which occur in all the reasonable cohomologies and
relations between them should come from properties that are motivic, i.e.
that already exist at the level of this universal cohomology. Setting up a
theory of motives is no easy task. The ﬁeld is in fact still a very active and
fast developing area of algebraic and arithmetic geometry. For a long time
the only available written reference giving a detailed account of the theory
of motives, along with some interesting algebrogeometric applications, was
[Manin (1968)], which is still an excellent short introduction to the subject.
A more recent reference that gives a broad survey of the current state of
the ﬁeld is [Andr´e (2004)]. An extensive series of papers on motives is
available in the proceedings of the 1991 Seattle conference, collected in the
two volumes [Jannsen, Kleiman, Serre (1994)]. For our purposes we will
try to follow the shortest path through this intricate theory that will allow
us to draw the connection to quantum ﬁeld theory and discuss the main
problems currently being investigated. Thus, this will be by no account an
accurate introduction to motives and we refer the readers interested in a
more serious and detailed account of the current state of research in this
area to the recent book [Andr´e (2004)].
The theory of motives comes in two ﬂavors, roughly corresponding to
the distinction between working with smooth projective varieties or with
everything else, meaning varieties that may be singular, or “noncompact”.
In the ﬁrst case, where one only considers motives of smooth projective
varieties, one obtains the theory of pure motives, which is built by ﬁrst
abstracting the main properties that a “good” cohomology theory of such
algebraic varieties should satisfy (Weil cohomologies) and then constructing
a category whose objects are smooth projective varieties together with the
data of a projector and a Tate twist, and where the morphisms are no longer
just given by morphisms of varieties but by more general correspondences
given by algebraic cycles in the product. Under a suitable choice of an
equivalence relation on the set of algebraic cycles, one can obtain in this
way an abelian category of pure motives and in fact a Tannakian category,
with a corresponding Galois group of symmetries, the motivic Galois group.
This part of the theory of motives is at this point on solid ground, although
problems like explicitly identifying the motivic Galois groups of various
signiﬁcant subcategories of the category of pure motives, or identifying
when a correspondence is realized by an algebraic cycle, remain extremely
challenging. While the general theory of pure motives is a beautiful and
Motives and periods 27
well established part of arithmetic algebraic geometry, the main problem of
establishing Grothendieck’s standard conjectures remains at this time open.
The situation is far more involved when it comes to mixed motives,
namely motives associated to varieties that are no longer necessarily smooth
and projective. The ﬁrst issue in this case is that of identifying the right
set of properties that a good cohomology theory is expected to have and
that should then be also associated to a good category of motives. As it
happens in topology, also in the context of algebraic varieties passing to the
“noncompact” case profoundly alters the main properties of the cohomol
ogy, with Poincar´e duality being replaced by the pairing of cohomologies
and cohomologies with compact support, and with a prominent role played
by Mayer–Vietoris type long exact sequences. The theory of mixed mo
tives, in the form which is presently available after the work of Voevodsky,
is based on triangulated categories, where these types of cohomology ex
act sequences manifest themselves in the form of distinguished triangles.
Only in very special cases, and for very particular subcategories of the
triangulated category of mixed motives, is it possible to extract from the
triangulated structure a heart giving rise to an abelian category. This is the
case for the subcategory of mixed motives that is most directly of interest
to quantum ﬁeld theory, namely the category of mixed Tate motives over a
number ﬁeld. As we are going to see in the next chapter, the reason why
it is mixed motives instead of the simpler pure motives that arise in the
quantum ﬁeld theoretic context lies in the fact that certain algebraic vari
eties naturally associated to Feynman graphs are typically singular, so that
they cannot be described, at the motivic level, using only pure motives.
When one is studying the motivic nature of certain classes of varieties,
as will be the case here with certain hypersurfaces associated to Feynman
graphs, it is often too compicated to work directly in the triangulated cate
gory of mixed motives, but one can sometime get useful information about
the motives by looking at their classes in the Grothendieck ring. This con
sists of isomorphism classes of varieties up to equivalence relations induced
by very rough “cutandpaste” operations on algebraic varieties, with the
same inclusionexclusion property of an Euler characteristic.
28 Feynman Motives
2.2 Pure motives
The main properties of a good cohomology theory of algebraic varieties are
summarized by the concept of Weil cohomology. We assume we are working
with smooth projective varieties over a ﬁeld K, which for our purposes we
can assume is Q or a ﬁnite algebraic extension (a number ﬁeld). This
means that, at each archimedean place of the number ﬁeld, such a variety
X deﬁnes a smooth complex manifold X(C), which is embedded in some
suﬃciently large projective space P
N
. The axioms of a Weil cohomology
are the following.
• H
·
is a contravariant functor from the category 1
K
of smooth projective
varieties to the category GrVect
K
of graded ﬁnite dimensional Kvector
spaces.
• H
i
(X) = 0 outside of the range i ∈ ¦0, . . . , 2 dim(X)¦.
• Poincar´e duality holds: H
2 dim(X)
(X) = K and there is a non
degenerate bilinear pairing
¸, ) : H
i
(X) H
2 dim(X)−i
(X)
−→K. (2.1)
• The K¨ unneth formula holds:
H
n
(X
1
K
X
2
) =
i+j=n
H
i
(X
1
) ⊗
K
H
j
(X
2
), (2.2)
with the isomorphism induced by the projection maps π
i
: X
1
K
X
2
→
X
i
, for i = 1, 2.
• Let Z
i
(X) denote the abelian group of algebraic cycles in X (smooth
connected subvarieties) of codimension i. Then there exists a cycle map
γ
i
X
: Z
i
(X) → H
2i
(X), (2.3)
which is functorial with respect to pullbacks and pushforwards and
satisfying
γ
n
X×Y
(Z W) = γ
i
X
(Z) ⊗γ
j
Y
(W),
with i + j = n, and normalized by γ
i
pt
: Z → K being the standard
inclusion.
• The weak Lefschetz isomorphisms hold: if H denotes a smooth hy
perplane section of X, with ι : H → X the inclusion, then ι
∗
:
H
i
(X) → H
i
(H) is an isomorphism for i ≤ dim(X) − 2 and injec
tive for i = dim(X) −1.
Motives and periods 29
• The hard Lefschetz isomorphisms hold: let L : H
i
(X) → H
i+2
(X)
be given by L : x → x ∪ γ
1
X
(H). Then L
dim(X)−i
: H
i
(X) →
H
2 dim(X)−i
(X) is an isomorphism for all i ≤ dim(X).
Algebraic de Rham cohomology H
∗
dR
(X, K) and Betti (singular) cohomol
ogy H
∗
B
(X) = H
∗
sing
(X(C), Q) are both examples of Weil cohomologies.
It is in general extremely diﬃcult to determine whether a given class in a
Weil cohomology H
∗
(X) lies in the image of the cycle map γ
∗
X
. The Hodge
conjecture, for example, is closely related to this question. The fact that
in general algebraic cycles are very complicated objects and one does not
know much about how to characterize them homologically constitutes the
main diﬃculty in the theory of pure motives, as the construction of the
category of pure motives that we now recall will show more clearly.
Grothendieck’s idea that gave birth to the whole theory of motives was
that there should be a cohomology theory for (smooth projective) algebraic
varieties that should map via realization functors to any cohomology satis
fying the axioms recalled above. The existence of isomorphisms between the
diﬀerent known Weil cohomologies, such as the period isomorphism between
Betti and de Rham cohomology, motivated this search for an underlying
universal theory.
The approach to constructing such a universal cohomology theory is
to remain suﬃciently close to the geometry of the varieties themselves,
instead of passing to associated vector spaces or modules over a ring, as
cohomologies typically do. One modiﬁes instead the maps between varieties
so as to obtain a category that still contains the usual category of smooth
projective varieties. It has additional objects and morphisms that make it
into a “linearization” of the category of varieties, that is, a category that
is good enough to behave as a category of modules over a ring.
The precise notion for a category that behaves like a category of modules
over a ring, namely a category that is good enough to do homological
algebra, is that of an abelian category, see [Gelfand and Manin (1994)].
Deﬁnition 2.2.1. An abelian category ( is a category satisfying the fol
lowing properties.
• For all X, Y ∈ Obj(() the set of morphisms Hom
C
(X, Y ) is an abelian
group (i.e. ( is preadditive). The composition of morphisms is com
patible with the abelian group structure of the Homsets.
• There is a zeroobject 0 ∈ Obj(() with Hom
C
(0, 0) = 0, the trivial
abelian group.
30 Feynman Motives
• There are ﬁnite products and coproducts in (.
• All morphisms have kernels and cokernels in the category and every
morphism f : X → Y has a canonical decomposition
K
k
→ X
i
→ I
j
→ Y
c
→ K
, (2.4)
with j ◦ i = f, and with K = Ker(f) ∈ Obj((), K
= Coker(f) ∈
Obj((), and I = Coker(k) = Ker(c) ∈ Obj(().
The property that an abelian category “behaves like a category of mod
ules over a ring” is made precise by the Freyd–Mitchell theorem, which
shows that indeed a small abelian category can be faithfully embedded as
a subcategory of a category of modules over a ring.
To obtain a category of pure motives one then aims at constructing
an abelian category of smooth projective varieties and correspondences,
which admits realization functors to vector spaces determined by the various
possible Weil cohomologies.
A morphism f : X → Y of varieties can be regarded as a correspondence
by considering its graph
((f) = ¦(x, y) ∈ X Y [ y = f(x)¦ ⊂ X Y. (2.5)
This is a closed connected subvariety of the product X Y of codimension
codim(((f)) = dim(Y ). (2.6)
Since one is aiming at constructing a universal cohomology theory, which
should be a contravariant functor, we regard the graph ((f) of a morphism
f : X → Y as a correspondence from Y to X,
((f) ∈ Corr
dim(Y )
(Y, X), (2.7)
where Corr
i
(Y, X) is the abelian group of correspondences from Y to X
given by algebraic cycles in Z
i
(X Y ). Thus, the generalization of mor
phisms f : X → Y is given by correspondences of the form Z =
j
n
j
Z
j
,
that are ﬁnite linear combinations with n
j
∈ Z of irreducible smooth alge
braic subvarieties Z
j
⊂ X Y . The usual composition of morphisms
X
1
f
1
→ X
2
f
2
→ X
3
is replaced by a composition of correspondences given by the ﬁber product
Z
◦ Z = Z
X
2
Z, (2.8)
for Z ⊂ X
1
X
2
and Z
⊂ X
2
X
3
. This means that, to obtain the
composition Z
◦ Z one does the following operations:
Motives and periods 31
• Take the preimages π
−1
32
(Z
) and π
−1
21
(Z) in X
1
X
2
X
3
, with π
ij
:
X
1
X
2
X
3
→ X
i
X
j
the projections.
• Intersect π
−1
32
(Z
) ∩ π
−1
21
(Z) in X
1
X
2
X
3
.
• Push forward the result along the remaining projection π
31
: X
1
X
2
X
3
→ X
1
X
3
,
Z
◦ Z = (π
31
)
∗
(π
−1
32
(Z
) ∩ π
−1
21
(Z)) ⊂ X
1
X
3
. (2.9)
• Extend by linearity to correspondences of the form Z =
j
n
j
Z
j
and
Z
=
r
n
r
Z
r
.
Because we are working with algebraic varieties as opposed to smooth man
ifolds, one in general cannot simply deform them to be in general position
so that the intersection π
−1
32
(Z
) ∩ π
−1
21
(Z) is transverse. Thus, one needs
to use some more sophisticated tools like excess intersection to deal with
intersection products in nontransverse situations, as one does in algebraic
geometry to deﬁne the product in the Chow ring, and use an equivalence
relation on the set of algebraic cycles that preserves the intersection prod
uct and allows for diﬀerent choices of representatives within the same class.
Reducing the size of the groups of algebraic cycles by an equivalence rela
tion is also desirable to obtain groups that are, at least in some cases, ﬁnite
dimensional.
There is more than one possible choice of an equivalence relation on
algebraic cycles, and the properties of the resulting category of motives
change according to which equivalence relation one considers. We only
mention here numerical and homological equivalence. We refer the reader to
[Andr´e (2004)] for a discussion of rational equivalence and the corresponding
properties. Imposing the numerical equivalence relation means identifying
two cycles Z
1
∼
num
Z
2
if their intersection products agree, #(Z
1
∩ Z) =
#(Z
2
∩ Z), for arbitrary subvarieties Z of complementary dimension. The
homological equivalence requires the choice of a Weil cohomology H
∗
and
prescribes that a cycle Z ∈ Z
i
(X Y ) is homologically trivial, Z ∼
hom
0,
if its image under the cycle map γ
i
X×Y
: Z
i
(X Y ) → H
2i
(X Y ) is
trivial, γ
i
X×Y
(Z) = 0. Although this second relation depends on the choice
of a Weil cohomology, it is part of Grothendieck’s standard conjectures
that numerical and homological equivalence agree, thus making the latter
independent of the choice of the cohomology. One denotes by Corr
i
∼
(XY )
the abelian group generated by the algebraic cycles of codimension i in
X Y modulo the equivalence relation ∼. This is compatible with the
composition of correspondences deﬁned using the ﬁber product. We also
32 Feynman Motives
use the notation
C
0
∼
(X, Y ) := Corr
dim(X)
∼
(Y X) ⊗
Z
Q. (2.10)
One can then deﬁne a category of motives as follows.
Deﬁnition 2.2.2. The category /
eﬀ
num
(K) of eﬀective numerical pure mo
tives has as objects pairs (X, p), where X is a smooth projective variety
over K and p ∈ C
0
num
(X, X) is an idempotent, p
2
= p. The morphisms are
given by
Hom
M
eff
num
(K)
((X, p), (Y, q)) = qC
0
∼
(X, Y )p. (2.11)
The reason for introducing projectors p
2
= p is in order to be able to
“cut out” pieces of the cohomology of a variety X, for example the piece
that corresponds to a certain degree only. It is in general very diﬃcult to
give an explicit description in terms of algebraic cycles of a projector that
isolates a given component of the cohomology of a variety. Notice also that,
since p ∈ C
0
∼
(X, X), the property of being idempotent also depends on the
choice of the equivalence relation. The set of morphisms here are Qvector
spaces instead of abelian groups (the category is Qlinear).
An important recent result [Jannsen (1992)] shows that the category
/
eﬀ
num
(K) is a semisimple abelian category. The semisimplicity property
means that there exists a set o of objects such that for all X ∈ o one
has Hom(X, X) = Q and Hom(X, Y ) = 0 for all X ,= Y ∈ o (simple
objects), and such that every object X of the category has a decomposition
X = ⊕
i
X
i
with summands X
i
∈ o.
The category /
eﬀ
num
(K) has a tensor product that comes form the prod
uct of varieties. This makes it into a tensor category. A desirable property
for a tensor category is a duality that makes it into a rigid tensor category.
Recall ﬁrst that a category ( where the set Hom
C
(X, Y ) is a Kvector
space, for any X, Y ∈ Obj((), is called Klinear. A tensor category over K
is a Klinear category (, with a bifunctor ⊗ : ( ( → ( and an object
1 ∈ Obj(() for which there are functorial isomorphisms
a
X,Y,Z
: X ⊗(Y ⊗Z) → (X ⊗Y ) ⊗Z
c
X,Y
: X ⊗Y → Y ⊗X
l
X
: X ⊗1 → X and r
X
: 1 ⊗X → X.
These satisfy consistency conditions expressed in the form of a triangle,
a pentagon, and a hexagon diagram, see [Andr´e (2004)] for details. One
Motives and periods 33
also generally assumes that commutativity holds, which means that c
Y,X
=
c
−1
X,Y
, although this can be replaced by a signed version.
Deﬁnition 2.2.3. A Klinear tensor category ( is rigid if there is a duality
∨ : ( → (
op
such that, for all X ∈ Obj(() the functor ⊗X
∨
is left adjoint
to ⊗ X and X
∨
⊗ is right adjoint to X ⊗ . Moreover, there are an
evaluation : X ⊗X
∨
→ 1 (the unit object for tensor product) and a unit
δ : 1 → X
∨
⊗X satisfying
( ⊗1) ◦ (1 ⊗δ) = 1
X
, (1 ⊗) ◦ (δ ⊗1) = 1
X
∨.
One also assumes that Hom
C
(1, 1) = K.
Recall here that given two functors F : ( → (
and G : (
→ ( satisfying
Hom
C
(F(X), Y ) = Hom
C
(X, G(Y )),
for all X ∈ Obj(() and Y ∈ Obj((
), with the identiﬁcation realized by a
natural isomorphism, one says that F is leftadjoint to G and G is right
adjoint to F.
The category of numerical eﬀective pure motives does not have a du
ality. One can see that, to have such a duality, one would need to take
the transpose of a correspondence Z ⊂ X Y , but this would have
the eﬀect of changing the codimension. Namely, if Z ∈ C
0
num
(X, Y ) =
Corr
dim(X)
num
(Y X)⊗
Z
Q, then the transpose Z
†
is in Corr
dim(X)
num
(XY )⊗
Z
Q,
but the latter is not C
0
num
(Y, X) = Corr
dim(Y )
num
(XY )⊗
Z
Q. Thus, to have a
duality one needs to include among morphisms also cycles of other codimen
sions. This can be achieved by introducing new objects, the Tate motives
Q(n). One deﬁnes the Tate motive Q(1) to be the formal inverse of the
Lefschetz motive L. This is the motive whose cohomological realization is
H
2
(P
1
). This means that one writes the motive of P
1
as 1⊕L, where 1 is the
motive of a point, which corresponds to the cohomology H
0
(P
1
). Similarly,
one writes the motive of P
n
as 1 ⊕ L ⊕ L
2
⊕ ⊕ L
n
, which corresponds
to the decomposition of P
n
into cells A
0
∪ A
1
∪ ∪ A
n
, and the corre
sponding cohomology ring H
∗
(P
n
; Z) = Z[u]/u
n+1
with the generator u in
degree 2. As an object in the category /
num
(K) (see Deﬁnition 2.2.4 be
low), the Lefschetz motive is given by the triple L = (Spec(K), id, −1), with
1 = (Spec(K), id, 0) being the motive of a point. Introducing the formal
inverse Q(1) = L
−1
= (Spec(K), id, 1), with Q(n) = Q(1)
⊗n
and Q(0) = 1,
has the eﬀect of shifting the codimensions of cycles in the deﬁnitions of
morphisms in the category of motives in the following way.
34 Feynman Motives
Deﬁnition 2.2.4. The category /
num
(K) has objects given by triples
(X, p, m) with (X, p) ∈ Obj(/
eﬀ
num
(K)) an eﬀective numerical motive and
m ∈ Z. The morphisms are given by
Hom
M
num
(K)
((X, p, m), (Y, q, n)) = qC
n−m
(X, Y )p, (2.12)
where
C
m−n
(X, Y ) = Corr
dim(X)+n−m
num
(Y X) ⊗
Z
Q (2.13)
The tensor product is then given by
(X, p, m) ⊗(Y, q, n) = (X Y, p ⊗q, n +m), (2.14)
and the duality is of the form
(X, p, m)
∨
= (X, p
†
, dim(X) −m). (2.15)
One obtains in this way a rigid tensor category. As we discuss in ¸2.9
below, the resulting category is in fact more than a rigid tensor category:
it has enough structure to be identiﬁed uniquely as the category of ﬁnite
dimensional linear representations of a group (more precisely of an aﬃne
group scheme) via the Tannakian formalism, see ¸2.9 below.
2.3 Mixed motives and triangulated categories
Having just quickly mentioned how one constructs a good category of pure
motives, we jump ahead and leave the world of smooth projective varieties
to venture into the more complicated and more mysterious realm of mixed
motives. One should think of mixed motives as objects that come endowed
with ﬁltrations whose graded pieces are pure motives. The diﬃcult nature
of mixed motives therefore lies in the nontrivial extensions of pure mo
tives. From the cohomological point of view, instead of working with Weil
cohomologies, one considers a diﬀerent set of properties that a universal
cohomology theory should satisfy in this case, where the varieties involved
may be noncompact or singular. The main properties that one expects
such a cohomology theory to satisfy become of the following form:
• Homotopy invariance: the projection π : X A
1
→ X induces an
isomorphism
π
∗
: H
∗
(X)
→ H
∗
(X A
1
). (2.16)
Motives and periods 35
• Mayer–Vietoris sequence: if U, V are an open covering of X then there
is a long exact cohomology sequence
→ H
i−1
(U ∩ V ) → H
i
(X) → H
i
(U) ⊕H
i
(V ) → H
i
(U ∩ V ) →
(2.17)
• Poincar´e duality is replaced by the duality between cohomology H
∗
and cohomology with compact support, H
∗
c
.
• The K¨ unneth formula still holds.
This replaces the notion of a Weil cohomology with “mixed Weil cohomolo
gies”. Ideally, one would like to have an abelian category //(K) of mixed
motives that has realization functors to all the mixed Weil cohomologies.
This is (at present) not available, although after a long series of develop
ments and deep results in the ﬁeld (see [Voevodsky (2000)], [Levine (1998)],
[Hanamura (1995)]) one has at least a weaker version available, in the form
of a triangulated category of mixed motives T/(K).
Triangulated categories still allow for several techniques from homolog
ical algebra, in particular in the form of long exact cohomology sequences.
Deﬁnition 2.3.1. A triangulated category T is an additive category en
dowed with an automorphism T : T → T (the shift functor) and a family
of distinguished triangles
X → Y → Z → T(X) (2.18)
with the following properties:
• X
id
→ X → 0 → T(X) is a distinguished triangle for all X ∈ Obj(T ).
• For all f ∈ Hom
T
(X, Y ), there exists an object C(f) ∈ Obj(T ) (the
mapping cone) with a morphism C(f) → T(X), such that X
f
→ Y →
C(f) → T(X) is a distinguished triangle.
• A triangle isomorphic to a distinguished triangle is distinguished (where
isomorphisms of triangles are isomorphisms of the objects such that the
resulting diagram commutes).
• Rotations Y → Z → T(X) → T(Y ) and T
−1
(Z) → X → Y → Z of a
distinguished triangle X → Y → Z → T(X) are distinguished.
• Given distinguished triangles X → Y → Z → T(X) and X
→ Y
→
Z
→ T(X
) and morphisms f : X → X
and h : Y → Y
such that
the square commutes, there exists a morphism Z → Z
such that all
squares commute, with T(f) : T(X) → T(X
) completing the diagram.
• Given X
f
→ Y → C(f) → T(X) and Y
h
→ Z → C(h) → T(Y ), and
the composite X
h◦f
→ Z → C(h ◦ f) → T(X), these three triangles can
36 Feynman Motives
be placed on the vertices and edges of an octahedron with morphisms
completing the picture so that all the faces are commutative diagrams.
One typically writes X[1] for T(X) and X[n] = T
n
(X) for n ∈ Z.
One more property of triangulated categories that follows from these
axioms, and which we will need explicitly later, is the following.
Lemma 2.3.2. Given a triangulated category T , a full subcategory T
is a
triangulated subcategory if and only if it is invariant under the shift T of
T and for any distinguished triangle X → Y → Z → X[1] for T where X
and Y are in T
there is an isomorphism Z · Z
where Z
is also an object
in T
.
A full triangulated subcategory T
⊂ T is thick if it is closed under
direct sums.
It is sometimes possible to extract an abelian category from a triangu
lated category using the method of tstructures.
Deﬁnition 2.3.3. A tstructure on a triangulated category T consists of
two full subcategories T
≤0
and T
≥0
with the following property. Upon
denoting by T
≥n
= T
≥0
[−n] and T
≤n
= T
≤0
[−n], one has
T
≤−1
⊂ T
≤0
, and T
≥1
⊂ T
≥0
with the property that
Hom
T
(X, Y ) = 0, ∀X ∈ Obj(T
≤0
), ∀Y ∈ Obj(T
≥1
),
and such that, for all Y ∈ Obj(T ), there exists a distinguished triangle
X → Y → Z → X[1]
with X ∈ Obj(T
≤0
) and Z ∈ Obj(T
≥1
).
When one has a tstructure on a triangulated category, one can con
sider the full subcategory T
0
= T
≤0
∩ T
≥0
. This is called the heart of
the tstructure and it is known to be an abelian category, see [Be˘ılinson,
Bernstein, Deligne (1982)], [Kashiwara and Schapira (1990)].
Theoretical physicists have recently become familiar with the formal
ism of tstructures in triangulated categories in the context of homological
mirror symmetry, see for instance [Bridgeland (2009)].
Motives and periods 37
2.4 Motivic sheaves
We also mention brieﬂy, for later use, a recent construction due to [Arapura
(2008)] of a category of motivic sheaves over a base scheme S, modeled on
Nori’s approach to the construction of categories of mixed motives. We will
return in Chapter 7 to discuss how this ﬁts with the motives associated to
Feynman diagrams.
The category of motivic sheaves constructed in [Arapura (2008)] is based
on Nori’s construction of categories of motives via representations of graphs
made up of objects and morphisms (cf. [Brugui`eres (2004)]).
According to [Arapura (2008)], one constructs a category of motivic
sheaves over a scheme S, by taking as vertices of the corresponding graph
objects of the form
(f : X → S, Y, i, w) , (2.19)
where f : X → S is a quasiprojective morphism, Y ⊂ X is a closed
subvariety, i ∈ N, and w ∈ Z. One thinks of such an object as determining a
motivic version h
i
S
(X, Y )(w) of the local system given by the (Tate twisted)
ﬁberwise cohomology of the pair H
i
S
(X, Y ; Q) = R
i
f
∗
j
!
Q
XY
, where j =
j
XY
: X Y → X is the open inclusion, i.e. the sheaf deﬁned by
U → H
i
(f
−1
(U), f
−1
(U) ∩ Y ; Q).
The edges are given by the geometric morphisms:
• morphisms of varieties over S,
(f
1
: X
1
→ S, Y
1
, i, w) → (f
2
: X
2
→ S, Y
2
= F(Y ), i, w), (2.20)
with f
2
◦ F = f
1
;
• the connecting morphisms
(f : X → S, Y, i + 1, w) → (f[
Y
: Y → S, Z, i, w), for Z ⊂ Y ⊂ X;
(2.21)
• the twisted projection morphisms
(f : XP
1
→ S, Y P
1
∪X¦0¦, i +2, w+1) → (f : X → S, Y, i, w).
(2.22)
The product in this category of motivic sheaves is given by the ﬁber product
(X → S, Y, i, w) (X
→ S, Y
, i
, w
) =
(X
S
X
→ S, Y
S
X
∪ X
S
Y
, i +i
, w +w
).
(2.23)
38 Feynman Motives
2.5 The Grothendieck ring of motives
An invariant of algebraic varieties and motives that contains a lot of in
formation and can be used to evaluate the motivic complexity of a given
algebraic variety is the class in the Grothendieck ring. This is constructed
by breaking up the variety into pieces, by simple cutandpaste operations
of the scissors congruence type. The operation of assigning to a variety its
class in the Grothendieck ring behaves like a universal Euler characteristic.
Deﬁnition 2.5.1. Let 1
K
denote the category of algebraic varieties over
a number ﬁeld K. The Grothendieck ring K
0
(1
K
) is the abelian group
generated by isomorphism classes [X] of varieties, with the relation
[X] = [Y ] + [X Y ], (2.24)
for Y ⊂ X closed. It is made into a ring by the product [X Y ] = [X][Y ].
Let /(K) denote, as above, the abelian category of (numerical) pure
motives over K. As above we write the objects of /(K) in the form
(X, p, m), with X a smooth projective variety over K, p = p
2
∈ End(X)
a projector, and m ∈ Z the Tate twist. Let K
0
(/(K)) denote the
Grothendieck ring of the abelian category /(K).
We’ll see that the class of a variety, or of a motive, in the corresponding
Grothendieck ring, can be thought of as a “universal Euler characteristic”.
To this purpose, we ﬁrst recall the notion of additive invariants of varieties.
Deﬁnition 2.5.2. An additive invariant is a map χ : 1
K
→ R, with values
in a commutative ring R, satisfying the following properties.
• Isomorphism invariance: χ(X) = χ(Y ) if X
∼
= Y are isomorphic.
• Inclusionexclusion: χ(X) = χ(Y ) +χ(X Y ), for a closed embedding
Y ⊂ X.
• Multiplicative property: χ(X Y ) = χ(X)χ(Y ).
The topological Euler characteristic clearly satisﬁes all of these proper
ties and is in fact the prototypical example of such an invariant. Counting
points over ﬁnite ﬁelds is another invariant that behaves in the same way
and factors through the Grothendieck ring.
By the deﬁnition of the relations on the Grothendieck ring one sees that
assigning an additive invariant with values in R is equivalent to assigning
a ring homomorphism χ : K
0
(1
K
) → R.
Motives and periods 39
The results of [Gillet and Soul´e (1996)] show that there exists an additive
invariant χ
mot
: 1
K
→ K
0
(/(K)). This assigns to a smooth projective
variety X the class χ
mot
(X) = [(X, id, 0)] ∈ K
0
(/(K)). The value assigned
to a more general variety is more complicated to describe and it is given
in [Gillet and Soul´e (1996)] in terms of an object W(X) in the category
of complexes over /(K). (Technically, here the category /(K) is taken
with the rational equivalence relation on algebraic cycles.) It is sometime
referred to as “motivic Euler characteristic”.
If L denotes the class L = [A
1
] ∈ K
0
(1
K
) then its image in K
0
(/(K)) is
the Lefschetz motive L = Q(−1) = [(Spec(K), id, −1)]. Since the Lefschetz
motive is invertible in K
0
(/(K)), its inverse being the Tate motive Q(1),
the ring homomorphism χ
mot
: 1
K
→ K
0
(/(K)) induces a ring homomor
phism
χ
mot
: K
0
(1
K
)[L
−1
] → K
0
(/(K)). (2.25)
This in particular means that one can work either in the Grothendieck
ring of varieties or of motives, being able to map the ﬁrst to the latter
through χ
mot
. In the following, we will usually refer to the Grothendieck
ring K
0
(1
K
). A particularly interesting subring is the one generated by the
Lefschetz motive L = [A
1
]. This is a polynomial ring Z[L] ⊂ K
0
(1
K
), or
a ring of Laurent polynomials Z[L, L
−1
] ⊂ K
0
(/(K)), which corresponds
to the classes of varieties that motivically are mixed Tate motives. See the
comment at the end of ¸2.6 below on the relation between the information
given by knowing that the class in the Grothendieck ring lies in the subring
Z[L] ⊂ K
0
(1
K
) and the question of whether the motive, as an object of the
triangulated category T/(K) is a mixed Tate motive.
2.6 Tate motives
Let T/(K) be the Voevodsky triangulated category of mixed motives over
a ﬁeld K [Voevodsky (2000)]. The triangulated category T/T (K) of mixed
Tate motives is the full triangulated thick subcategory of T/(K) generated
by the Tate objects Q(n). It is known that, over a number ﬁeld K, there
is a canonical tstructure on T/T (K) and one can therefore construct an
abelian category /T (K) of mixed Tate motives (see [Levine (1998)]).
For the kind of applications we are going to see in the rest of this book,
we will be especially interested in the question of when a given mixed mo
tive is mixed Tate. In suﬃciently simple cases, one can try to answer this
question using the properties of T/T (K) as a triangulated subcategory of
40 Feynman Motives
the triangulated category T/(K) of mixed motives. Although the avail
able explicit constructions of T/(K) given in [Voevodsky (2000)], [Levine
(1998)], [Hanamura (1995)], are technically very complicated and certainly
beyond the scope of this book, often when dealing with very speciﬁc con
crete examples, one does not need the full strength of the general construc
tion, but one can directly use the formal properties that the triangulated
category T/(K) satisﬁes. In particular, in order to understand whether
the motive of a given variety is mixed Tate, one can try to ﬁnd a geometric
decomposition, often in the form of a stratiﬁcation, of the variety, in such
a way that one has enough control on the strata to build the motive out
of pieces that are mixed Tate. More precisely, in such an approach one
wants to make repeated use of two properties of the triangulated category
T/(K):
• (Proposition 4.1.4 of [Voevodsky (2000)]) Over a ﬁeld of characteristic
zero, a closed embedding Y ⊂ X determines a distinguished triangle in
T/(K)
m(Y ) → m(X) → m(X Y ) → m(Y )[1], (2.26)
where the notation m(X) here stands for the motive with compact
support, denoted by C
c
∗
(X) in [Voevodsky (2000)].
• (Corollary 4.1.8 of [Voevodsky (2000)]) Homotopy invariance implies
the identity of motives
m(X A
1
) = m(X)(1)[2], (2.27)
where [n] is the shift, and (n) is the Tate twist by Q(n).
The distinguished triangle (2.26) implies that, since T/T (K) is a tri
angulated subcategory of T/(K), if two of the terms are known to be in
T/T (K) then the third one also is. This can often be used as a way to
reduce a motive to simpler building blocks for which it may be easier to
answer the question of whether they are mixed Tate or not, and then derive
information about the original motive via the distinguished triangles and
homotopy invariance property.
We reported above the distinguished triangles and homotopy invariance
properties for the motives with compact support in Voevodsky’s terminol
ogy. One can obtain analogous statements for motives without compact
support using the duality in T/(K) between mixed motives and mixed
motives with compact support, which replaces Poincar´e duality that holds
for Weil cohomologies in the pure motives setting.
Motives and periods 41
We give here an example from [Aluﬃ and Marcolli (2009a)], which will
be relevant to Feynman integrals (see ¸3.13 below), namely the case of de
terminant hypersurfaces. It is a well known fact that, motivically, these
varieties are mixed Tate (see [Belkale and Brosnan (2003a)]), but we il
lustrate here explicitly how one can argue that this is the case using only
the two properties listed above of the triangulated category of mixed Tate
motives.
Deﬁnition 2.6.1. Let
ˆ
T
⊂ A
2
be the hypersurface deﬁned by the van
ishing of the determinant
ˆ
T
= ¦x = (x
ij
) ∈ A
2
[ det(x) = 0¦. (2.28)
Let T
⊂ P
2
−1
be the corresponding projective hypersurface deﬁned by
the homogeneous polynomial det(x).
We have the following result on the motivic nature of the determinant
hypersurfaces. We formulate it in terms of hypersurface complements, as
that will be the natural geometric object to consider in the case of the Feyn
man integrals. Clearly, if the result holds for the hypersurface complement
it also holds for the hypersurface itself, by the ﬁrst property above on the
distinguished triangle of a closed embedding, and the fact that aﬃne and
projective spaces are Tate motives.
Theorem 2.6.2. The determinant hypersurface complement A
2
ˆ
T
de
ﬁnes an object in the category T/T of mixed Tate motives.
Proof. Since
ˆ
T
consists of matrices of rank < , in order to describe
the locus A
2
ˆ
T
we need to parameterize matrices of rank exactly equal
to , which can be done in the following way. Fix an dimensional vector
space E.
– Denote by J
1
the variety E ¦0¦;
– Note that J
1
is equipped with a trivial vector bundle E
1
= E J
1
,
and with a line bundle S
1
:= L
1
⊆ E
1
whose ﬁber over v
1
∈ J
1
consists
of the line spanned by v
1
;
– Let J
2
⊆ E
1
be the complement E
1
L
1
;
– Note that J
2
is equipped with a trivial vector bundle E
2
= E J
2
,
and two line subbundles of E
2
: the pullback of L
1
(still denoted L
1
)
and the linebundle L
2
whose ﬁber over v
2
∈ J
2
consists of the line
spanned by v
2
;
– By construction, L
1
and L
2
span a rank2 subbundle S
2
of E
2
;
42 Feynman Motives
– Let J
3
⊆ E
2
be the complement E
2
S
2
; and so on.
We ﬁrst show that the variety J
constructed as above is isomorphic
to the locus we want, namely A
2
ˆ
D
.
To see this, ﬁrst observe that at the kth step, the procedure described
above produces a variety J
k
, endowed with k line bundles L
1
, . . . , L
k
span
ning a rankk subbundle S
k
of the trivial vector bundle E
k
:= E J
k
. If
S
k
E
k
, deﬁne J
k+1
:= E
k
S
k
. Let E
k+1
= E J
k+1
, and deﬁne
the line subbundles L
1
, . . . , L
k
to be the pullbacks of the similarly named
line bundles on J
k
. Let L
k+1
be the line bundle whose ﬁber over v
k+1
is
the line spanned by v
k+1
. The line bundles L
1
, . . . , L
k+1
span a rank k +1
subbundle S
k+1
of E
k+1
, and the construction can continue. The sequence
stops at the th step, where S
has rank , equal to the rank of E
, so that
E
S
= ∅.
Each variety J
k
maps to A
2
as follows: a point of J
k
determines k
vectors v
1
, . . . , v
k
, and can be mapped to the matrix whose ﬁrst k rows
are v
1
, . . . , v
k
resp. (and the remaining rows are 0). By construction, this
matrix has rank exactly k. Conversely, any such rank k matrix is the image
of a point of J
k
, by construction.
Also notice that, in the construction described above, the bundle S
k
over J
k
is in fact trivial. An explicit trivialization is given by the map
K
k
J
k
α
→ S
k
deﬁned by α : ((c
1
, . . . , c
r
), (v
1
, . . . , v
r
)) → c
1
v
1
+ +c
r
v
r
,
where points of J
k
are parameterized by ktuples of vectors v
1
, . . . , v
k
spanning S
k
⊆ K
J
k
= E
k
.
We then apply the two properties of the triangulated category of mixed
motives described above, namely the existence of distinguished triangles
(2.26) associated to closed embeddings Y ⊂ X and the homotopy invari
ance (2.27) inductively to the loci constructed above, all of which are va
rieties deﬁned over Q. At the ﬁrst step, single points obviously belong to
the category of mixed Tate motives and we are taking the complement J
1
of a point in an aﬃne space, which gives a mixed Tate motive by the ﬁrst
observation above on distinguished triangles associated to closed embed
dings. At the next step one considers the complement of the line bundle
S
1
inside the trivial vector bundle E
1
over J
1
. Again, both m(S
1
) and
m(E
1
) are mixed Tate motives, since both are products by aﬃne spaces,
since we know that the bundles S
k
are trivial. Therefore m(E
1
S
1
) is also
mixed Tate. The same argument shows that, for all 1 ≤ k ≤ , the motive
m(E
k
S
k
) is mixed Tate, by repeatedly using the triviality of S
k
and the
two properties of T/T
Q
recalled above.
Motives and periods 43
The result for the projective hypersurface T
is then the same, since
ˆ
T
is the aﬃne cone over T
and the same two properties of the triangulated
category of mixed motives used above show that if
ˆ
T
is mixed Tate so
is T
. One can use the same geometric construction described above to
compute explicitly the class of
ˆ
T
in the Grothendieck ring of varieties.
One ﬁnds the following expression [Aluﬃ and Marcolli (2009a)].
Theorem 2.6.3. The class in the Grothendieck ring of varieties of the
aﬃne determinant hypersurface is
[A
2
ˆ
T
] = L
(
2
)
i=1
(L
i
−1) (2.29)
where L is the class of A
1
. In the projective case, the class is
[P
2
−1
T
] = L
(
2
)
i=2
(L
i
−1). (2.30)
Proof. Using the same argument as in Theorem 2.6.2, one shows induc
tively that the class of J
k
is given by
[J
k
] = (L
−1)(L
−L)(L
−L
2
) (L
−L
k−1
)
= L
(
k
2
)
(L
−1)(L
−1
−1) (L
−k+1
−1).
(2.31)
This suﬃces to obtain the expression (2.29). The formula (2.30) is then
obtained from (2.29) using the fact that
ˆ
T
is the aﬃne cone over T
.
Notice that, by Theorem 2.6.2, we already know that the class [A
2
ˆ
T
]
will be in the mixed Tate part Z[L] of the Grothendieck ring of varieties
K
0
(1
Q
). In general, computing the class in the Grothendieck ring may
be easier than knowing explicitly the motive in T/
Q
. If one only knows
that the class belongs to the Tate part Z[L] of the Grothendieck ring, this
does not imply that the motive itself will be in the subcategory T/T
Q
of mixed Tate motives, as in principle there may be cancellations that
happen in the Grothendieck ring that do not take place at the level of the
motives themselves. However, knowing that the class in the Grothendieck
ring is a function of the Lefschetz motive L is a very good indication of
the possible mixed Tate nature of the motive. In fact, as observed for
instance in [Andr´e (2009)], the Tate conjecture predicts that knowing the
number N
p
of points of the reduction mod p for almost all p would suﬃce
to determine the motive. For mixed Tate motives the numbers N
p
are
polynomial in p, and assuming the conjecture holds, knowing that this
44 Feynman Motives
is the case would conversely imply that the motive is mixed Tate. The
role in this of the class in the Grothendieck ring comes from the fact that
the counting of points N
p
over ﬁnite ﬁelds has the properties of an Euler
characteristic (as we saw above, an additive invariant), hence it factors
through the Grothendieck ring. Thus, assuming the conjecture, knowing
that the class in the Grothendieck ring is mixed Tate would suﬃce. Even
without these general considerations and the role of the Tate conjecture,
in the best cases, the same type of geometric information on stratiﬁcations
that one can use to compute the class in the Grothendieck ring K
0
(1
K
)
may already contain enough information to run an argument similar to the
one used in Theorem 2.6.2, directly at the level of objects in T/
K
, as we
have seen in the example of the determinant hypersurfaces.
2.7 The algebra of periods
Betti and de Rham cohomology are two of the possible Weil cohomologies
for smooth projective varieties. They are related by the period isomorphism
H
∗
dR
(X) ⊗
K
C · H
∗
B
(X) ⊗
Q
C, (2.32)
for varieties deﬁned over a number ﬁeld K. The isomorphism is induced
(over aﬃne pieces) by the pairing of diﬀerential forms and cycles
ω ⊗γ →
_
γ
ω, (2.33)
for
ω ∈ Ker(d : Ω
i
(X) → Ω
i+1
(X))
and
γ ∈ Ker(∂
i
: C
i
(X) → C
i−1
(X)),
where one identiﬁes Betti cohomology H
∗
B
(X) = Hom(H
∗
(X), Q) =
H
∗
sing
(X, Q) with the dual of singular homology H
i
(X, Z) =
Ker(∂
i
)/Im(∂
i+1
). When one is given a basis of H
∗
dR
(X) and of H
∗
B
(X),
the period isomorphism is speciﬁed by a period matrix.
Periods are typically transcendental numbers. From the number
theoretic point of view they are, arguably, the most interesting class of
numbers beyond the algebraic ones. In fact, while being transcendental,
they are obtained from a procedure that uses only algebraic data: an alge
braic diﬀerential form on an algebraic variety, paired with a cycle deﬁned
Motives and periods 45
by algebraic relations. Let us denote by T
C
the Qsubalgebra of C gener
ated by periods. An interesting problem is understanding the structure of
this algebra: what relations (over Q) exist between periods.
Following [Kontsevich and Zagier (2001)], one can deﬁne a Qalgebra T
as the algebra generated by “formal periods”, that is, equivalence classes
of elements of the form
[(X, D, ω, γ)], (2.34)
consisting of a smooth (aﬃne) variety X deﬁned over Q, a normal cross
ings divisor D in X, an algebraic diﬀerential form ω ∈ Ω
dim(X)
(X)
(not necessarily closed), and a class in the relative homology γ ∈
H
dim(X)
(X(C), D(C), Q). The equivalence in (2.34) is given by the fol
lowing relations
• Linearity: the expressions (X, D, ω, γ) are Qlinear in ω and γ.
• Change of variable formula: given a morphism f : (X, D) → (X
, D
)
and ω
∈ Ω
dim(X
)
(X
) and γ ∈ H
dim(X)
(X(C), D(C), Q) one has
[(X, D, f
∗
(ω), γ)] = [(X
, D
, ω, f
∗
(γ))]. (2.35)
• Stokes’ formula:
[(X, D, dω, γ)] = [(
˜
D,
˜
D
(1)
, ω, ∂γ)], (2.36)
where
˜
D is the normalization of D and
˜
D
(1)
is the normalization of
∪
i,j
(D
i
∩ D
j
) where D
i
are the components of the divisor D.
Kontsevich conjectured that the map T → T
C
⊂ C given by
[(X, D, ω, γ)] →
_
γ
ω (2.37)
is injective. This means that all possible nontrivial relations between pe
riods come from either a change of variable formula or Stokes’ theorem.
2.8 Mixed Tate motives and the logarithmic extensions
An interesting class of mixed Tate motives is associated to logarithm and
polylogarithm functions [Be˘ılinson and Deligne (1994)], see also [Ayoub
(2007)], [Bloch (1997)].
46 Feynman Motives
The extensions Ext
1
DM(K)
(Q(0), Q(1)) of Tate motives are given by the
Kummer motives M = [Z
u
→ G
m
] with u(1) = q ∈ K
∗
. This extension has
period matrix of the form
_
1 0
log q 2πi
_
. (2.38)
When, instead of working with motives over the base ﬁeld K, one works
with the relative setting of motivic sheaves over a base scheme S, instead
of the Tate motives Q(n) one considers the Tate sheaves Q
S
(n). These
correspond to the constant sheaf with the motive Q(n) over each point
s ∈ S. In the case where S = G
m
, there is a natural way to assemble the
Kummer motives into a unique extension in Ext
1
DM(G
m
)
(Q
G
m
(0), Q
G
m
(1)).
This is the Kummer extension
Q
G
m
(1) → / →Q
G
m
(0) →Q
G
m
(1)[1], (2.39)
where over the point s ∈ G
m
one is taking the Kummer extension M
s
=
[Z
u
→ G
m
] with u(1) = s. Because of the logarithm function log(s) that
appears in the period matrix for this extension, the Kummer extension
(2.39) is also referred to as the logarithmic motive, and denoted by / = Log,
cf. [Ayoub (2007)] [Be˘ılinson and Deligne (1994)].
Let T/(G
m
) be the Voevodsky category of mixed motives (motivic
sheaves) over the multiplicative group G
m
. When working with Q
coeﬃcients, so that one can include denominators in the deﬁnition of pro
jectors, one can then consider the logarithmic motives Log
n
, deﬁned by
Log
n
= Sym
n
(/), (2.40)
where the symmetric powers of an object in T/
Q
(G
m
) are deﬁned as
Sym
n
(X) =
1
#Σ
n
σ∈Σ
n
σ(X
n
). (2.41)
The polylogarithms appear naturally as period matrices for extensions
involving the symmetric powers Log
n
= Sym
n
(/), in the form [Bloch
(1997)]
0 → Log
n−1
(1) → L
n
→Q(0) → 0, (2.42)
where M(1) = M ⊗Q(1) and L
1
= Log. The mixed motive L
n
has period
matrix
_
1 0
M
(n)
Li
M
Log
n−1
(1)
_
(2.43)
Motives and periods 47
with
M
(n)
Li
= (−Li
1
(s), −Li
2
(s), , −Li
n
(s))
τ
, (2.44)
where τ means transpose and where
Li
1
(s) = −log(1 −s), and Li
n
(s) =
_
s
0
Li
n−1
(u)
du
u
,
equivalently deﬁned (on the principal branch) using the power series
Li
n
(s) =
k
s
k
k
n
,
and with
M
Log
n
(1)
=
_
_
_
_
_
_
_
_
2πi 0 0 0
2πi log(s) (2πi)
2
0 0
2πi
log
2
(s)
2!
(2πi)
2
log(s) (2πi)
3
0
.
.
.
.
.
.
.
.
.
.
.
.
2πi
log
n
(s)
n!
(2πi)
2
log
n−1
(s)
(n−1)!
(2πi)
3
log
n−2
(s)
(n−2)!
(2πi)
n
_
_
_
_
_
_
_
_
.
(2.45)
The period matrices for the motives Log
n
correspond to the description
of Log
n
as the extension of Q(0) by Log
n−1
(1), i.e. to the distinguished
triangles in T/(G
m
) of the form
Log
n−1
(1) → Log
n
→Q(0) → Log
n−1
(1)[1]. (2.46)
The motives Log
n
form a projective system under the canonical maps
β
n
: Log
n+1
→ Log
n
given by the composition of the morphisms Sym
n+m
(/) → Sym
n
(/) ⊗
Sym
m
(/), as in [Ayoub (2007)], Lemma 4.35, given by the fact that
Sym
n+m
(/) is canonically a direct factor of Sym
n
(/) ⊗ Sym
m
(/), and
the map Sym
m
(/) → Q(0) of (2.46), in the particular case m = 1. Let
Log
∞
denote the promotive obtained as the projective limit
Log
∞
= lim
←−
n
Log
n
. (2.47)
The analog of the period matrix (2.45) then becomes the inﬁnite matrix
M
Log
∞
(1)
=
_
_
_
_
_
_
_
_
_
_
_
_
2πi 0 0 0
2πi log(s) (2πi)
2
0 0
2πi
log
2
(s)
2!
(2πi)
2
log(s) (2πi)
3
0
.
.
.
.
.
.
.
.
.
.
.
.
2πi
log
n
(s)
n!
(2πi)
2
log
n−1
(s)
(n−1)!
(2πi)
3
log
n−2
(s)
(n−2)!
(2πi)
n
.
.
.
.
.
.
.
.
.
.
.
.
_
_
_
_
_
_
_
_
_
_
_
_
.
(2.48)
48 Feynman Motives
This can best be expressed in terms of mixed Hodge structures. These
comprise the analytic side of the theory of motives and they provide one of
the possible realizations of categories of mixed motives, hence another way
to test complexity of various types of mixed motives. One has the following
deﬁnition as in [Steenbrink (1994)].
Deﬁnition 2.8.1. A Qmixed Hodge structure consists of the data M =
(V, W
•
, F
•
) of a ﬁnitedimensional Qvector space, with an increasing
ﬁltration W
•
V (weight ﬁltration) and a decreasing ﬁltration F
•
V
C
for
V
C
= V ⊗
Q
C (the Hodge ﬁltration), satisfying
gr
W
n
V
C
= F
p
gr
W
n
V
C
⊕F
n−p+1
gr
W
n
V
C
, (2.49)
where
¯
F
•
is the complex conjugate of F
•
with respect to V
R
⊂ V
C
.
The Hodge structure (V, W
•
, F
•
) is pure of weight n if W
n
= V and
W
n−1
= 0.
A morphism M
1
→ M
2
of QMHS is a linear map φ : V
1
→ V
2
such
that φ(W
k
V
1
) ⊂ W
k
V
2
and φ(F
p
V
1,C
) ⊂ F
p
V
2,C
. Tensoring of QMHS is
also deﬁned, by considering V
1
⊗
Q
V
2
with ﬁltrations
W
m
(V
1
⊗
Q
V
2
) =
i+j=m
W
i
(V
1
) ⊗
Q
W
j
(V
2
),
F
p
(V
1
⊗
Q
V
2
) =
r+s=p
F
r
(V
1,C
) ⊗
C
F
s
(V
2,C
).
A Qmixed Hodge structure M = (V, W
•
, F
•
) is mixed Tate if
gr
W
2p
M = ⊕Q(−p), and gr
W
2p−1
M = 0.
The condition (2.49) is equivalent to a Hodge decomposition
gr
W
n
V
C
= ⊕
p+q=n
V
p,q
, (2.50)
V
p,q
= F
p
gr
W
p+q
V
C
∩ F
q
gr
W
p+q
V
C
. (2.51)
An example of pure Hodge structure of weight −2n is the Tate Hodge
structure with V = (2πi)
n
Q. These correspond to the pure Tate motives.
These deﬁne, by tensoring, Tate twists on Qmixed Hodge structures.
For Hodge–Tate structures one has an associated period matrix obtained
as in ¸2 of [Goncharov (1999)], as the composition S
−1
HT
◦ S
W
of an isomor
phism S
HT
: ⊕
p
gr
W
2p
(V
Q
)⊗
Q
C → V
C
and a splitting of the weight ﬁltration,
which is an isomorphism S
W
: ⊕
p
gr
W
2p
(V
Q
) ⊗
Q
C → V
C
, expressed in a
chosen basis of gr
W
2p
V
Q
.
Motives and periods 49
In the example above, the mixed Hodge structure associated to the
motives Log
n
is the one that has as the weight ﬁltrations W
−2k
the range
of multiplication by the matrix M
Log
n
deﬁned as in (2.45) on vectors in Q
n
with the ﬁrst k − 1 entries equal to zero, while the Hodge ﬁltration F
−k
is given by the range of multiplication of M
Log
n
on vectors of C
n
with the
entries from k + 1 to n equal to zero [Bloch (1997)].
2.9 Categories and Galois groups
We review here a type of structure on a category which is more rigid than
that of an abelian category we encountered before, and which is suﬃcient to
identify it uniquely as a category of ﬁnite dimensional linear representations
of a group, namely the notion of Tannakian category, see [Saavedra Rivano
(1972)], [Deligne and Milne (1982)].
Deﬁnition 2.9.1. A klinear rigid abelian tensor category ( is a Tannakian
category if it admits a ﬁber functor, i.e. an exact faithful tensor functor,
ω : ( → Vect
K
, with target the category of ﬁnitedimensional vector spaces
over a ﬁeld extension K of k. A Tannakian category ( is neutral if K = k.
Recall here that a functor ω : ( → (
is faithful if, for all X, Y ∈ Obj((),
the mapping ω : Hom
C
(X, Y ) → Hom
C
(ω(X), ω(Y )) is injective. If (
and (
are klinear categories, a functor ω is additive if the map of Hom
sets above is a klinear map. An additive functor ω is exact if, for any
exact sequence 0 → X → Y → Z → 0 in (, the corresponding sequence
0 → ω(X) → ω(Y ) → ω(Z) → 0 in (
is also exact. A functor ω : ( → (
between klinear tensor categories is a tensor functor if there are functorial
isomorphisms τ
1
: ω(1) → 1 and τ
X,Y
: ω(X ⊗Y ) → ω(X) ⊗ω(Y ).
The main result of the theory of Tannakian categories, of which we
are going to see an application to motives in the next section and one to
quantum ﬁeld theory in a later chapter, is the following statement that
reconstructs a group from the category.
Theorem 2.9.2. [Saavedra Rivano (1972)] Let ( be a neutral Tannakian
category, with ﬁber functor ω : ( → Vect
k
. Let G be the group of auto
morphisms (invertible natural transformations) of the ﬁber functor. Then
ω induces an equivalence of rigid tensor categories ω : ( → Rep
G
, where
Rep
G
is the category of ﬁnite dimensional linear representations of G.
50 Feynman Motives
In fact, here G is really an aﬃne group scheme in the sense that we are
going to recall in ¸5.2 below. The aﬃne group scheme G is referred to as
the Tannakian Galois group of (.
If a Tannakian category ( is semisimple, then it is known that the
Tannakian Galois group is reductive. For example, the Tannakian Galois
group of the category of pure motives is reductive. So is the one of the
Tannakian subcategory generated by the pure Tate motives Q(n). In fact,
the latter is the simplest example of a reductive group, the multiplicative
group G
m
.
2.10 Motivic Galois groups
We recalled above that, over a number ﬁeld K, it is possible to form an
abelian category /T (K) of mixed Tate motives, as the heart of a t
structure inside the triangulated subcategory T/T (K) of the triangulated
category T/(K) of mixed motives. In fact, over a number ﬁeld one has an
explicit description of the extensions in T/(K) between pure Tate motives
in terms of algebraic Ktheory as
Ext
1
DM(K)
(Q(0), Q(n)) = K
2n−1
(K) ⊗Q (2.52)
and Ext
2
DM(K)
(Q(0), Q(n)) = 0, see [Levine (1993)] and [Levine (1998)],
Appendix B. For example, the mixed Tate motives that correspond to ex
tensions in Ext
1
DM(K)
(Q(0), Q(1)) are the Kummer motives we have en
countered in the previous sections.
The resulting abelian category /T (K) of mixed Tate motives over a
number ﬁeld is in fact also a Tannakian category, with ﬁber functor ω to
Zgraded Qvector spaces, M → ω(M) = ⊕
n
ω
n
(M) with
ω
n
(M) = Hom(Q(n), Gr
w
−2n
(M)), (2.53)
where Gr
w
−2n
(M) = W
−2n
(M)/W
−2(n+1)
(M) is the graded structure asso
ciated to the ﬁnite increasing weight ﬁltration W.
The construction of a Tannakian category of mixed Tate motives
/T (O
V
) and the explicit form of its motivic Galois group are also known
for motivic sheaves over the scheme O
V
of V integers of a number ﬁeld K,
for V a set of ﬁnite places of K. This means that objects in this category are
mixed Tate motives over K that are unramiﬁed at each ﬁnite place v / ∈ V .
In this case one also has a Ktheory group parameterizing extensions in the
Motives and periods 51
form
Ext
1
DM(O
V
)
(Q(0), Q(n)) =
_
¸
¸
_
¸
¸
_
K
2n−1
(K) ⊗Q n ≥ 2
O
∗
V
⊗Q n = 1
0 n ≤ 0.
(2.54)
and Ext
2
DM(O
V
)
(Q(0), Q(n)) = 0. The aﬃne group scheme that gives the
motivic Galois group in this case is of the form U G
m
, where the reduc
tive group G
m
comes from the pure Tate motives, and U is a prounipotent
aﬃne group scheme, that takes into account the nontrivial extensions be
tween pure Tate motives. In fact, the corresponding Lie algebra Lie(U) is
freely generated by a set of homogeneous generators in degree n identiﬁed
with a basis of the dual of Ext
1
(Q(0), Q(n)), as proved in Proposition 2.3
of [Deligne and Goncharov (2005)]. There is however no canonical identi
ﬁcation between Lie(U) and the free Lie algebra generated by the graded
vector space ⊕Ext
1
(Q(0), Q(n))
∨
.
The following special case of this general result of [Deligne and Gon
charov (2005)] is especially relevant to the context of renormalization, since
it provides a motivic Galois group (noncanonically) isomorphic to the Tan
nakian Galois group of renormalization of [Connes and Marcolli (2004)],
which we review in Chapter 6.
Proposition 2.10.1. [Deligne and Goncharov (2005)] For K = Q(ζ
N
) the
cyclotomic ﬁeld of level N = 3 or N = 4 and O its ring of integers, the
motivic Galois group of the Tannakian category /T (O[1/N]) is of the form
U G
m
, where the Lie algebra Lie(U) is (noncanonically) isomorphic to
the free graded Lie algebra with one generator e
n
in each degree n ≤ −1.
Chapter 3
Feynman integrals and algebraic
varieties
In this chapter we begin to connect the two topics introduced above, pertur
bative quantum ﬁeld theory and motives. We follow ﬁrst what we referred
to as the “bottomup approach” in the introduction, by showing how to as
sociate to individual Feynman integrals an algebraic variety, in the form of
a hypersurface complement, and a (possibly divergent) period integral on a
chain with boundary on a normal crossings divisor. The motivic nature of
this period, and the question of whether this (after removing divergences)
will be a period of a mixed Tate motive can then be formulated as the ques
tion of whether a certain relative cohomology of the hypersurface comple
ment relative to the normal crossings divisor is a realization of a mixed Tate
motive. We begin by describing the parametric form of Feynman integrals
and the main properties of the two Kirchhoﬀ–Symanzik graph polynomials
involved in this formulation and the graph hypersurfaces deﬁned by the
ﬁrst polynomial. We show, as in [Aluﬃ and Marcolli (2008a)], how one can
sometimes obtain explicit computations of the classes in the Grothendieck
ring of varieties for some families of Feynman graphs, using the classical
Cremona transformation and dual graphs. These classes also satisfy a form
of deletion/contraction relation, as proved in [Aluﬃ and Marcolli (2009b)].
We then describe, following the results of [Aluﬃ and Marcolli (2008b)], a
way of formulating Feynman rules directly at the algebrogeometric level
of graph hypersurface complements. We give examples of such Feynman
rules that factor through the Grothendieck ring and others, deﬁned us
ing characteristic classes of singular varieties, that do not descend to the
level of isomorphism classes of varieties in the Grothendieck ring, although
they still satisfy an inclusionexclusion principle. We discuss the diﬀerence
between working with aﬃne or projective hypersurfaces in describing Feyn
man integrals. Finally, following [Aluﬃ and Marcolli (2009a)], we show
53
54 Feynman Motives
that one can reformulate the period computation underlying the Feynman
amplitudes in terms of a simpler hypersurface complement, using deter
minant hypersurfaces, whose mixed Tate nature is established, but with a
more complicated normal crossings divisor as the second term in the rela
tive cohomology whose motivic nature one wishes to establish. We describe
the advantages and the diﬃculties of this viewpoint.
3.1 The parametric Feynman integrals
Using the Feynman parameters introduced in ¸1.7 above, we show how to
reformulate the Feynman integral
U(Γ, p
1
, . . . , p
N
) =
_
δ(
n
i=1
v,i
k
i
+
N
j=1
v,j
p
j
)
q
1
(k
1
) q
n
(k
n
)
d
D
k
1
d
D
k
n
in the form known as the Feynman parametric representation. (We neglect
here an overall multiplicative factor in the coupling constants and powers
of 2π.)
The ﬁrst step is to rewrite the denominator q
1
q
n
of (1.37) in the
form of an integration over the topological simplex σ
n
as in (1.51), in terms
of the Feynman parameters t = (t
1
, . . . , t
n
) ∈ σ
n
.
In writing the integral (1.37) we have made a choice of an orientation
of the graph Γ, since the matrix
v,i
involved in writing the conservation
laws at vertices in (1.37) depends on the orientation given to the edges of
the graph. Now we also make a choice of a set of generators for the ﬁrst
homology group H
1
(Γ, Z), i.e. a choice of a maximal set of independent
loops in the graph, ¦l
1
, . . . , l
¦ with = b
1
(Γ) the ﬁrst Betti number.
We then deﬁne another matrix associated to the graph Γ, the circuit
matrix η = (η
ik
), with i ∈ E(Γ) and k = 1, . . . , ranging over the chosen
basis of loops, given by
η
ik
=
_
_
_
+1 if edge e
i
∈ loop l
k
, same orientation
−1 if edge e
i
∈ loop l
k
, reverse orientation
0 if edge e
i
/ ∈ loop l
k
.
(3.1)
There is a relation between the circuit matrix and the incidence matrix
of the graph, which is given as follows.
Lemma 3.1.1. The incidence matrix = (
v,i
) and the circuit matrix η =
(η
ik
) of a graph Γ satisfy the relation η = 0. This holds independently of
the choice of the orientation of the graph and the basis of H
1
(Γ, Z).
Feynman integrals and algebraic varieties 55
Proof. The result follows from the observation that, independently of
the choices of orientations, for two given edges e
i
and e
j
in a loop l
k
, both
incident to a vertex v, one has
v,i
η
ik
= −
v,j
η
jk
so that one obtains
i
v,i
η
ik
= 0,
since
v,i
η
ik
,= 0 for only two edges in the loop l
k
, with diﬀerent signs, as
above; see ¸2.2 of [Nakanishi (1971)].
We then deﬁne the Kirchhoﬀ matrix of the graph, also known as the
Symanzik matrix.
Deﬁnition 3.1.2. The Kirchhoﬀ–Symanzik matrix M
Γ
(t) of the graph Γ
is the matrix given by
(M
Γ
(t))
kr
=
n
i=1
t
i
η
ik
η
ir
. (3.2)
Equivalently, it can be written as
M
Γ
(t) = η
†
Λ(t)η,
where † is the transpose and Λ(t) is the diagonal matrix with entries
(t
1
, . . . , t
n
). We think of M
Γ
as a function
M
Γ
: A
n
→A
2
, t = (t
1
, . . . , t
n
) → M
Γ
(t) = (M
Γ
(t))
kr
(3.3)
where A denotes the aﬃne line over a ﬁeld (here mostly C or R or Q).
Deﬁnition 3.1.3. The Kirchhoﬀ–Symanzik polynomial Ψ
Γ
(t) of the graph
Γ is deﬁned as
Ψ
Γ
(t) = det(M
Γ
(t)). (3.4)
Notice that, while the construction of the matrix M
Γ
(t) depends on the
choice of an orientation on the graph Γ and of a basis of H
1
(Γ, Z), the graph
polynomial is independent of these choices.
Lemma 3.1.4. The Kirchhoﬀ–Symanzik polynomial Ψ
Γ
(t) is independent
of the choice of edge orientation and of the choice of generators for
H
1
(Γ, Z).
56 Feynman Motives
Proof. A change of orientation in a given edge results in a change of sign
in one of the columns of η = η
ik
. The change of sign in the corresponding
row of η
†
leaves the determinant of M
Γ
(t) = η
†
Λ(t)η unaﬀected. A change
of basis for H
1
(Γ, Z) changes M
Γ
(t) → AM
Γ
(t)A
−1
, where A ∈ GL
(Z) is
the matrix that gives the change of basis. The determinant is once again
unchanged.
We view it as a function Ψ
Γ
: A
n
→ A. We deﬁne the aﬃne graph
hypersurface
ˆ
X
Γ
to be the locus of zeros of the graph polynomial
ˆ
X
Γ
= ¦t ∈ A
n
[ Ψ
Γ
(t) = 0¦. (3.5)
The polynomial Ψ
Γ
is by construction a homogeneous polynomial of degree
= b
1
(Γ), hence we can view it as deﬁning a hypersurface in projective
space P
n−1
= (A
n
¦0¦)/G
m
,
X
Γ
= ¦t ∈ P
n−1
[ Ψ
Γ
(t) = 0¦, (3.6)
of which
ˆ
X
Γ
is the aﬃne cone
ˆ
X
Γ
= C(X
Γ
).
After rewriting the denominator of the integrand in (1.37) in terms of
an integration over σ
n
using the Feynman parameters, we want to replace
in the Feynman integral U(Γ, p
1
, . . . , p
N
) the variables k
i
associated to the
internal edges, and the integration in these variables, by variables x
r
asso
ciated to the independent loops in the graph and an integration only over
these variables, using the linear constraints at the vertices. We set
k
i
= u
i
+
r=1
η
ir
x
r
, (3.7)
with the constraint
n
i=1
t
i
u
i
η
ir
= 0, ∀r = 1, . . . , , (3.8)
that is, we require that the column vector Λ(t)u is orthogonal to the rows
of the circuit matrix η.
The momentum conservation conditions in the delta function in the
numerator of (1.37) give
n
i=1
v,i
k
i
+
N
j=1
v,j
p
j
= 0. (3.9)
Lemma 3.1.5. Using the change of variables (3.7) and the constraint (3.8)
one ﬁnds the conservation condition
n
i=1
v,i
u
i
+
N
j=1
v,j
p
j
= 0. (3.10)
Feynman integrals and algebraic varieties 57
Proof. This follows immediately from the orthogonality relation between
the incidence matrix and circuit matrix of Lemma 3.1.1.
The two equations (3.8) and (3.10) constitute the Krichhoﬀ laws of cir
cuits applied to the ﬂow of momentum through the Feynman graph. In
particular they determine uniquely the u
i
= u
i
(p) as functions of the exter
nal momenta. We see the explicit form of the solution in Proposition 3.1.7
below. First we give a convenient reformulation of the graph polynomial.
The graph polynomial Ψ
Γ
(t) has a more explicit combinatorial descrip
tion in terms of the graph Γ, as follows.
Proposition 3.1.6. The Kirchhoﬀ–Symanzik polynomial Ψ
Γ
(t) of (3.4) is
given by
Ψ
Γ
(t) =
T⊂Γ
e/ ∈E(T)
t
e
, (3.11)
where the sum is over all the spanning trees T of the graph Γ and for each
spanning tree the product is over all edges of Γ that are not in that spanning
tree.
Proof. The polynomial Ψ
Γ
= det(M
Γ
(t)) can equivalently be described
as the polynomial (see [Itzykson and Zuber (2006)], ¸623 and [Nakanishi
(1971)] ¸1.32)
Ψ
Γ
(t) =
S
e∈S
t
e
, (3.12)
where S ranges over all the subsets S ⊂ E
int
(Γ) of = b
1
(Γ) internal edges
of Γ, such that the removal of all the edges in S leaves a connected graph.
This can be seen to be equivalent to the formulation (3.11) in terms of
spanning trees of the graph Γ (see [Nakanishi (1971)] ¸1.3). In fact, each
spanning tree has #V (Γ) − 1 edges and is the complement of a set S as
above.
In the case of graphs with several connected components, one deﬁnes
the Kirchhoﬀ–Symanzik polynomial as in (3.11), with the sum over span
ning forests, by which we mean here collections of a spanning tree in each
connected component. This polynomial is therefore multiplicative over con
nected components.
We then have the following description of the resulting term
i
t
i
u
2
i
after the change of variables (3.7). This will be useful in Theorem 3.1.9
below.
58 Feynman Motives
Proposition 3.1.7. The term
i
t
i
u
2
i
is of the form
i
t
i
u
2
i
= p
†
R
Γ
(t)p,
where R
Γ
(t) is an N N matrix, with N = #E
ext
(Γ) with
p
†
R
Γ
(t)p =
v,v
∈V (Γ)
P
v
(D
Γ
(t)
−1
)
v,v
P
v
,
with
(D
Γ
(t))
v,v
=
n
i=1
v,i
v
,i
t
−1
i
(3.13)
and
P
v
=
e∈E
ext
(Γ),t(e)=v
p
e
. (3.14)
Proof. We give a quick sketch of the argument and we refer the reader
to [Nakanishi (1971)] and [Itzykson and Zuber (2006)], ¸623 for more
details. The result is a consequence of the fact that the u
i
, as functions
of the external momenta p = (p
j
), are determined by the Kirchhoﬀ law
(3.10). Thus, there is a matrix A
i,v
such that −
v
A
i,v
P
v
= t
i
u
i
, hence
i
t
i
u
2
i
=
i
v,v
P
v
P
v
A
i,v
A
i,v
t
−1
i
. The constraints on the u
i
given by
the Kirchhoﬀ laws also show that A
i,v
=
i,v
so that one obtains the matrix
D
Γ
(t) as in (3.13).
We set
V
Γ
(t, p) = p
†
R
Γ
(t)p +m
2
. (3.15)
In the massless case (m = 0), we will see below that this is a ratio of two
homogeneous polynomials in t,
V
Γ
(t, p)[
m=0
=
P
Γ
(t, p)
Ψ
Γ
(t, p)
, (3.16)
of which the denominator is the graph polynomial (3.4) and P
Γ
(t, p) is a
homogeneous polynomial of degree b
1
(Γ) +1. We have the following result;
see [Itzykson and Zuber (2006)], ¸623.
Proposition 3.1.8. The function V
Γ
(t, p) satisﬁes (3.16) in the massless
case, with the polynomial P
Γ
(t, p) of the form
P
Γ
(t, p) =
C⊂Γ
s
C
e∈C
t
e
, (3.17)
where the sum is over the cutsets C ⊂ Γ, i.e. the collections of b
1
(Γ) + 1
edges that divide the graph Γ into exactly two connected components Γ
1
∪Γ
2
.
Feynman integrals and algebraic varieties 59
The coeﬃcient s
C
is a function of the external momenta attached to the
vertices in either one of the two components,
s
C
=
_
_
v∈V (Γ
1
)
P
v
_
_
2
=
_
_
v∈V (Γ
2
)
P
v
_
_
2
, (3.18)
where the P
v
are deﬁned as in (3.14), as the sum of the incoming external
momenta.
Proof. We give a brief sketch of the argument and refer the reader to the
more detailed treatment in [Nakanishi (1971)]. The matrix D
Γ
(t) of (3.13)
has determinant
det(D
Γ
(t)) =
T
e∈T
t
−1
e
,
where T ranges over the spanning trees of the graph Γ. This is related to
the polynomial Ψ
Γ
(t) by
Ψ
Γ
(t) = (
e
t
e
) det(D
Γ
(t)) =
T
e/ ∈T
t
e
.
Thus, for m = 0, the function (3.15) becomes
p
†
R
Γ
(t)p =
v,v
∈V (Γ)
P
v
(D
Γ
(t)
−1
)
v,v
P
v
=
1
Ψ
Γ
(t)
C⊂Γ
s
C
e∈C
t
e
.
We can now rewrite the Feynman integral in its parametric form as
follows; see [Bjorken and Drell (1964)] ¸8 and [Bjorken and Drell (1965)]
¸18.
Theorem 3.1.9. Up to a multiplicative constant C
n,
,the Feynman integral
U(Γ, p
1
, . . . , p
N
) can be equivalently written in the form
U(Γ, p
1
, . . . , p
N
) =
Γ(n −
D
2
)
(4π)
D/2
_
σ
n
ω
n
Ψ
Γ
(t)
D/2
V
Γ
(t, p)
n−D/2
, (3.19)
where ω
n
is the volume form on the simplex σ
n
.
Proof. We ﬁrst show that we have
_
d
D
x
1
d
D
x
(
n
i=0
t
i
q
i
)
n
= C
,n
det(M
Γ
(t))
−D/2
(
n
i=0
t
i
(u
2
i
+m
2
))
−n+D/2
, (3.20)
60 Feynman Motives
where u
i
= u
i
(p) as above. In fact, after the change of variables (3.7), the
left hand side of (3.20) reads
_
d
D
x
1
d
D
x
(
n
i=0
t
i
(u
2
i
+m
2
) +
k,r
(M
Γ
)
kr
x
k
x
r
)
n
.
The integral can then be reduced by a further change of variables that
diagonalizes the matrix M
Γ
to an integral of the form
_
d
D
y
1
d
D
y
(a +
k
λ
k
y
2
k
)
n
= C
,n
a
−n+D/2
k=1
λ
−D/2
k
,
with
C
,n
=
_
d
D
x
1
d
D
x
(1 +
k
x
2
k
)
n
.
We then write det M
Γ
(t) = Ψ
Γ
(t) and we use the expression of Proposition
3.1.7 to express the term (
i
t
i
(u
2
i
+m
2
))
−n+D/2
in terms of
i
t
i
(u
2
i
+m
2
) =
i
t
i
u
2
i
+m
2
= V
Γ
(t, p),
with V
Γ
(t, p) as in (3.15).
In the massless case, using the expression (3.16) of V
Γ
(t, p) in terms
of the polynomials P
Γ
(t, p) and Ψ
Γ
(t), one writes equivalently the integral
U(Γ, p) as
U(Γ, p
1
, . . . , p
N
) =
Γ(n −
D
2
)
(4π)
D/2
_
σ
n
P
Γ
(t, p)
−n+D/2
ω
n
Ψ
Γ
(t)
−n+D(+1)/2
. (3.21)
Up to a divergent Gammafactor, one is interested in understanding the
motivic nature (i.e. the nature as a period) of the remaining integral
1(Γ, p
1
, . . . , p
N
) =
_
σ
n
P
Γ
(t, p)
−n+D/2
ω
n
Ψ
Γ
(t)
−n+D(+1)/2
. (3.22)
3.2 The graph hypersurfaces
We introduced above the projective hypersurfaces X
Γ
⊂ P
n−1
, with n =
#E
int
(Γ), and the corresponding aﬃne hypersurface
ˆ
X
Γ
⊂ A
n
.
We observe here that these hypersurfaces are in general singular and
with singular locus of positive dimension; see ¸3.5 below for more details.
The fact that we are dealing with singular hypersurfaces in projective spaces
Feynman integrals and algebraic varieties 61
implies that, when we consider motives associated to these geometric ob
jects, we will necessarily be dealing with mixed motives.
We consider only the case where the spacetime dimension D of the
quantum ﬁeld theory is suﬃciently large. This means that we look at the
parametric Feynman integrals (3.19) in the stable range where
n ≤ D/2, (3.23)
for n = #E
int
(Γ) and = b
1
(Γ). In this range, the algebraic diﬀerential
form that is integrated in the parametric representation of the Feynman
integral only has singularities along the graph hypersurface X
Γ
, while the
polynomial P
Γ
(t, p) only appears in the numerator. This range covers, in
particular, the special case of the log divergent graphs. These are the graphs
with n = D/2, for which the integral (3.22) reduces to the simpler form
1(Γ, p
1
, . . . , p
N
) =
_
σ
n
ω
n
Ψ
Γ
(t)
D/2
. (3.24)
We will not discuss the unstable range of small D, though it is often of
signiﬁcant physical interest, because of additional diﬃculties in treating
the hypersurfaces deﬁned by the second graph polynomial P
Γ
(t, p) caused
by the additional dependence on the external momenta. We only make
some general considerations about this case in ¸3.3 below.
In terms of graph hypersurfaces, the Feynman integral with the con
dition (3.23) can be regarded (modulo the problem of divergences) as a
period obtained by integrating an algebraic diﬀerential form deﬁned on the
hypersurface complement P
n−1
X
Γ
over a domain given by the topolog
ical simplex σ
n
, whose boundary ∂σ
n
is contained in the normal crossings
divisor Σ
n
given by the algebraic simplex (the union of the coordinate hy
perplanes). Thus, the motivic nature of this period should be detected by
the mixed motive
m(P
n−1
X
Γ
, Σ
n
(Σ
n
∩ X
Γ
)), (3.25)
whose realization is the relative cohomology
H
n−1
(P
n−1
X
Γ
, Σ
n
(Σ
n
∩ X
Γ
)). (3.26)
A possible strategy to showing that the Feynman integrals evaluate to num
bers that are periods of mixed Tate motives would be to show that the
motive (3.25) is mixed Tate. This is one of the main themes that we are
going to develop in the rest of the book.
It is important to stress the fact that identifying the motivic nature of
the relative cohomology above would only suﬃce up to the important issue
62 Feynman Motives
of divergences. Divergences occur where the graph hypersurface X
Γ
meets
the domain of integration σ
n
. Notice that the intersections X
Γ
∩σ
n
can only
occur on the boundary ∂σ
n
. In fact, in the interior of σ
n
all the coordinates
t
e
are strictly positive real numbers, and the graph polynomial is then
also strictly positive, Ψ
Γ
(t) > 0, as can easily be seen by the expression
(3.11). Thus, we have X
Γ
∩σ
n
⊂ X
Γ
∩Σ
n
. These intersections are usually
nonempty, so that some method of regularization needs to be introduced
to remove the corresponding divergences in the integral, before one can
identify the integration
_
σ
n
P
Γ
(t, p)
−n+D/2
ω
n
Ψ
Γ
(t)
−n+D(+1)/2
(3.27)
with a period of the motive (3.25). The regularization method proposed
in [Bloch, Esnault, Kreimer (2006)] consists of performing a number of
blowups of (strata of) the locus X
Γ
∩ Σ
n
, with the result of separating
the domain of integration and the hypersurface and therefore regularizing
the integral by replacing the original integration with one performed in the
blown up variety. This regularization procedure adds a further complication
to the problem of identifying the motivic nature of (3.25). Namely, one also
needs to know that the motive remains mixed Tate after the blowups. We
will return to this issue in more detail in ¸3.14 below, after we discuss the
approach of [Aluﬃ and Marcolli (2009a)].
We now discuss brieﬂy the diﬀerence between working with the aﬃne
or the projective hypersurfaces. While it is natural to work projectively
when one has equations deﬁned by homogeneous polynomials, there are
various reasons why the aﬃne context appears to be more natural in this
case. In particular, we argue as in [Aluﬃ and Marcolli (2008b)] that the
multiplicative property of the Feynman rules is more naturally reﬂected by
a corresponding multiplicative property of the aﬃne hypersurface comple
ments, which in the projective case is replaced by the more complicated
join operation.
As we saw in ¸1.3, the Feynman rules are multiplicative over disjoint
unions of graphs, namely
U(Γ, p) = U(Γ
1
, p
1
) U(Γ
k
, p
k
), (3.28)
for Γ = Γ
1
∪ ∪ Γ
k
a disjoint union of graphs with N
j
= #E
ext
(Γ
j
)
external edges, with external momenta
p = (p
1
, . . . , p
k
), with p
j
= (p
j,1
, . . . , p
j,N
j
).
The other “multiplicative” property that the Feynman rules satisfy is the
one we discussed in the ﬁrst chapter, which allows one to determine the
Feynman integrals and algebraic varieties 63
Feynman integrals of non1PI graphs using the integrals for 1PI graphs. In
fact, upon representing an arbitrary (ﬁnite) graph as a tree T with vertices
v ∈ V (T) replaced by 1PI graphs Γ
v
with a number of external edges
#E
ext
(Γ
v
) = val(v) equal to the valence of the vertex, one obtains the
expression for the Feynman integral of the resulting graph Γ as
V (Γ, p) =
v∈V (T),e∈E
int
(T),v∈∂(e)
V (Γ
v
, p
v
)q
e
(p
v
)
−1
δ((p
v
)
e
−(p
v
)
e
), (3.29)
where, as in (1.35) we write
V (Γ, p
1
, . . . , p
N
) = ε(p
1
, . . . , p
N
) U(Γ, p
1
, . . . , p
N
),
with ε(p
1
, . . . , p
N
) =
e∈E
ext
(Γ)
q
e
(p
e
)
−1
.
In the particular case of a massive theory m ,= 0 where one sets all the
external momenta equal to zero, this becomes an actual product
U(Γ, p)[
p=0
= U(e, p
e
)[
#E
int
(T)
p
e
=0
v∈V (T)
U(Γ
v
, p
v
)[
p
v
=0
, (3.30)
where in this case the edge propagator U(e, p
e
)[
p
e
=0
= m
−2
is just a con
stant depending on the mass parameter m ,= 0.
One can then, following [Aluﬃ and Marcolli (2008b)], deﬁne an abstract
Feynman rule in the following way.
Deﬁnition 3.2.1. An abstract Feynman rule is an assignment of an el
ement U(Γ) in a commutative ring ¹ for each ﬁnite graph Γ, with the
property that, for a disjoint union Γ = Γ
1
∪ ∪ Γ
k
this satisﬁes the mul
tiplicative property
U(Γ) = U(Γ
1
) U(Γ
k
) (3.31)
and for a non1PI graph obtained by inserting 1PI graphs Γ
v
at the vertices
of a tree T it satisﬁes
U(Γ) = U(L)
#E
int
(T)
v∈V (T)
U(Γ
v
). (3.32)
A ﬁrst diﬀerence one encounters between the projective graph hypersur
faces X
Γ
and the aﬃne
ˆ
X
Γ
, which is directly relevant to the quantum ﬁeld
theory setting, is the fact that the aﬃne hypersurface complements behave
multiplicatively as abstract Feynman rules are expected to do, while the
projective hypersurface complements do not. More precisely, we have the
following result from [Aluﬃ and Marcolli (2008b)].
Lemma 3.2.2. Let Γ = Γ
1
∪ ∪ Γ
k
be a disjoint union of ﬁnite graphs.
Then the aﬃne hypersurface complements satisfy
A
n
ˆ
X
Γ
= (A
n
1
ˆ
X
Γ
1
) (A
n
k
ˆ
X
Γ
k
), (3.33)
where n
i
= #E
int
(Γ
i
).
64 Feynman Motives
Proof. Using the deﬁnition of the graph polynomials in terms of the
determinant Ψ
Γ
(t) = det /
Γ
(t) of the Kirchhoﬀ matrix of the graph, one
sees easily that if Γ = Γ
1
∪ ∪ Γ
k
is a disjoint union, then one has
Ψ
Γ
(t) =
k
i=1
Ψ
Γ
1
(t
i,1
, . . . , t
i,n
i
),
where t = (t
e
) = (t
i,j
)
i=1,...,k,j=1,...,n
i
are the edge variables of the internal
edges of Γ. This means that the aﬃne hypersurface
ˆ
X
Γ
is a union
ˆ
X
Γ
=
k
_
i=1
(A
n
1
A
n
i−1
ˆ
X
Γ
i
A
n
i+1
A
n
k
),
which implies that the hypersurface complement is then given by the prod
uct (3.33).
The situation with the projective hypersurfaces is more complicated.
Instead of getting directly a product of the complements one gets a
torus bundle. For instance, suppose given a disjoint union of two graphs
Γ = Γ
1
∪ Γ
2
and assume, moreover, that neither graph is a forest (the
assumption is not necessary in the aﬃne case). Then the projective hy
persurface complement P
n
1
+n
2
−1
X
Γ
is a G
m
bundle over the product
(P
n
1
−1
X
Γ
1
) (P
n
2
−1
X
Γ
2
). To see this, observe that in this case the
projective hypersurface X
Γ
is a union of cones C
n
2
(X
Γ
1
) and C
n
1
(X
Γ
2
) in
P
n
1
+n
2
−1
, respectively over X
Γ
1
and X
Γ
2
with vertices P
n
2
−1
and P
n
1
−1
,
respectively. The hypothesis that neither graph is a forest ensures that
there is a regular map
P
n
1
+n
2
−1
X
Γ
→ (P
n
1
−1
X
Γ
1
) (P
n
2
−1
X
Γ
2
) (3.34)
given by
(t
1,1
: : t
1,n
1
: t
2,1
: : t
2,n
2
) → ((t
1,1
: : t
1,n
1
), (t
2,1
: : t
2,n
2
)),
(3.35)
while if either Γ
i
happened to be a forest, the corresponding X
Γ
i
would be
empty, and the map (3.35) would not be deﬁned everywhere. Under the
assumption that neither Γ
i
is a forest, the map (3.34) is surjective and the
ﬁber over a point ((t
1,1
: : t
1,n
1
), (t
2,1
: : t
2,n
2
)) is a copy of the
multiplicative group G
m
given by all the points of the form
(ut
1,1
: : ut
1,n
1
: vt
2,1
: : vt
2,n
2
), with (u : v) ∈ P
1
, uv ,= 0.
Equivalently, notice that if Γ is not a forest, then there is a regular map
A
n
ˆ
X
Γ
→P
n−1
X
Γ
,
and (3.34) is then induced by the isomorphism in (3.33). This is no longer
the case if Γ is a forest.
Feynman integrals and algebraic varieties 65
3.3 Landau varieties
For simplicity we restrict here everywhere to the case where the diﬀerential
form in the parametric Feynman integral has singularities only along the
graph hypersurface
ˆ
X
Γ
deﬁned by the vanishing of the graph polynomial
Ψ
Γ
(t). This is certainly the case within what we referred to as the “sta
ble range”, i.e. for D (considered as a variable) suﬃciently large so that
n ≤ D/2. This, however, is typically not the range of physical interest for
theories where D = 4 and the number of loops and internal edges violates
the above inequality. Away from this stable range, the second graph poly
nomial P
Γ
(t, p) also appears in the denominator, and the diﬀerential form
of the parametric Feynman integral is then deﬁned on the complement of
the hypersurface (family of hypersurfaces)
ˆ
Y
Γ
(p) = ¦t ∈ A
n
[ P
Γ
(t, p) = 0¦ (3.36)
or in projective space, since P
Γ
(, p) is homogeneous of degree b
1
(Γ) +1, on
the complement of
Y
Γ
(p) = ¦t ∈ P
n−1
[ P
Γ
(t, p) = 0¦, (3.37)
or else on the complement of the union of hypersurfaces
ˆ
Y
Γ
(p) ∪
ˆ
X
Γ
, again
depending on the values of n, and D.
We only discuss here the case where −n + D/2 ≥ 0, though some of
the same arguments extend to the other case as in [Marcolli (2008)]. The
hypersurfaces
ˆ
Y
Γ
(p) deﬁned by the vanishing of P
Γ
(p, t), for a ﬁxed value of
the external momenta p, are sometimes referred to as the Landau varieties.
The general type of arguments described above using the graph hypersur
faces
ˆ
X
Γ
should then be adapted to the case of the Landau varieties. In
such cases, however, the situation is more complicated because of the ex
plicit dependence on the additional parameters p. In particular, some of
the divergences of the integral, coming from the intersection of the hyper
surface Y
Γ
(p) with the domain of integration σ
n
, can occur in the interior
of the domain, not only on its boundary as in the case of X
Γ
; see [Bjorken
and Drell (1965)], ¸18.
For simplicity, it is also convenient to make the following assumption on
the polynomials P
Γ
(t, p) and Ψ
Γ
, so that we avoid cancellations of common
factors.
Deﬁnition 3.3.1. A 1PI graph Γ satisﬁes the generic condition on the
external momenta if, for p in a dense open set in the space of external
momenta, the polynomials P
Γ
(t, p) and Ψ
Γ
(t) have no common factor.
66 Feynman Motives
Recall that P
Γ
(t, p) is deﬁned in terms of external momenta and cut
sets through the functions P
v
and s
C
as in Proposition 3.1.8. In terms of
spanning trees, one has
P
Γ
(t, p) =
T
e
∈T
s
T,e
t
e
e∈T
c
t
e
, (3.38)
where s
T,e
= s
C
for the cutset C = T
c
∪ ¦e
¦.
The parameterizing space of the external momenta is the hyperplane in
the aﬃne space A
D·#E
ext
(Γ)
obtained by imposing the conservation law
e∈E
ext
(Γ)
p
e
= 0. (3.39)
Thus, the simplest possible conﬁguration of external momenta is the
one where one puts all the external momenta to zero, except for a pair
p
e
1
= p = −p
e
2
associated to a choice of a pair of external edges ¦e
1
, e
2
¦ ⊂
E
ext
(Γ). Let v
i
be the unique vertex attached to the external edge e
i
of
the chosen pair. We then have, in this case, P
v
1
= p = −P
v
2
. Upon writing
the polynomial P
Γ
(t, p) in the form (3.38), we obtain in this case
P
Γ
(p, t) = p
2
T
(
e
∈T
v
1
,v
2
t
e
)
e/ ∈T
t
e
, (3.40)
where T
v
1
,v
2
⊂ T is the unique path in T without backtrackings connecting
the vertices v
1
and v
2
. We use as in (3.18)
s
C
=
_
_
v∈V (Γ
1
)
P
v
_
_
2
=
_
_
v∈V (Γ
2
)
P
v
_
_
2
to get s
C
= p
2
for all the nonzero terms in this (3.40). These are all the
terms that correspond to cut sets C such that the vertices v
1
and v
2
belong
to diﬀerent components. These cut sets consist of the complement of a
spanning tree T and an edge of T
v
1
,v
2
. Consider for simplicity the case
where the polynomial Ψ
Γ
is irreducible. If we denote the linear functions
of (3.40) by
L
T
(t) = p
2
e∈T
v
1
,v
2
t
e
, (3.41)
we see that, if the polynomial Ψ
Γ
(t) divides (3.40), one would have
P
Γ
(p, t) = Ψ
Γ
(t) L(t),
for a degree one polynomial L(t), and this would give
T
(L
T
(t) −L(t))
e/ ∈T
t
e
≡ 0,
Feynman integrals and algebraic varieties 67
for all t. One then sees, for example, that the 1PI condition on the graph
Γ is necessary in order to have the condition of Deﬁnition 3.3.1. In fact,
for a graph that is not 1PI, one may be able to ﬁnd vertices and momenta
as above such that the degree one polynomials L
T
(t) are all equal to the
same L(t). Generally, the validity of the condition of Deﬁnition 3.3.1 can
be checked algorithmically to be satisﬁed for large classes of 1PI graphs,
though a complete combinatorial characterization of the graphs satisfying
these properties does not appear to be known.
3.4 Integrals in aﬃne and projective spaces
As we have seen above, the homogeneous graph polynomial Ψ
Γ
deﬁnes
either a projective hypersurface X
Γ
⊂ P
n−1
, with n = #E
int
(Γ), or an
aﬃne hypersurface
ˆ
X
Γ
⊂ A
n
, the aﬃne cone over X
Γ
. Thus, assuming
we are in the stable range of suﬃciently high D where −n +D/2 ≥ 0, for
= b
1
(Γ) the number of loops, one can regard the Feynman integral U(Γ, p)
as computed over the aﬃne hypersurface complement A
n
ˆ
X
Γ
, but one can
also reformulate it in terms of the projective hypersurface complement.
In order then to reformulate in projective space P
n−1
integrals originally
deﬁned in aﬃne space A
n
, we ﬁrst recall some basic facts about diﬀerential
forms on aﬃne and projective spaces and their relation, following mostly
[Dimca (1992)], see also [Gelfand, Gindikin, and Graev (1980)].
The projective analog of the volume form
ω
n
= dt
1
∧ ∧ dt
n
is given by the form
Ω =
n
i=1
(−1)
i+1
t
i
dt
1
∧ ∧
´
dt
i
∧ ∧ dt
n
. (3.42)
The relation between the volume form dt
1
∧ ∧dt
n
and the homogeneous
form Ω of degree n of (3.42) is given by
Ω = ∆(ω
n
), (3.43)
where ∆ : Ω
k
→ Ω
k−1
is the operator of contraction with the Euler vector
ﬁeld
E =
i
t
i
∂
∂t
i
, (3.44)
∆(ω)(v
1
, . . . , v
k−1
) = ω(E, v
1
, . . . , v
k−1
). (3.45)
68 Feynman Motives
Let ¹ = C[t
1
, . . . , t
n
] be the ring of polynomials of A
n
. Let ¹
m
denote
the subset of homogeneous polynomials of degree m. Similarly, let Ω
k
denote the ¹module of kforms on A
n
and let Ω
k
m
denote the subset of
kforms that are homogeneous of degree m.
Let f be a homogeneous polynomial function f : A
n
→ A of de
gree deg(f). Let π : A
n
¦0¦ → P
n−1
be the standard projection t =
(t
1
, . . . , t
n
) → π(t) = (t
1
: : t
n
). We denote by
ˆ
X
f
= ¦t ∈ A
n
[ f(t) = 0¦
the aﬃne hypersurface and by X
f
= ¦π(t) ∈ P
n−1
[ f(t) = 0¦ the projective
hypersurface, and by
ˆ
T(f) = A
n
ˆ
X
f
and T(f) = P
n−1
X
f
the hyper
surface complements. With the notation introduced above, we can always
write a form ω ∈ Ω
k
(T(f)) as
ω =
η
f
m
, with η ∈ Ω
k
mdeg(f)
. (3.46)
We then have the following characterization of the pullback along π :
ˆ
T(f) → T(f) of forms on T(f).
Proposition 3.4.1. (see [Dimca (1992)], p.180 and [Dolgachev (1982)])
Given ω ∈ Ω
k
(T(f)), the pullback π
∗
(ω) ∈ Ω
k
(
ˆ
T(f)) is characterized by
the properties of being invariant under the G
m
action on A
n
¦0¦ and of
satisfying ∆(π
∗
(ω)) = 0, where ∆ is the contraction (3.45) with the Euler
vector ﬁeld E of (3.44).
Thus, since the sequence
0 → Ω
n
∆
→ Ω
n−1
∆
→
∆
→ Ω
1
∆
→ Ω
0
→ 0
is exact at all but the last term, one can write
π
∗
(ω) =
∆(η)
f
m
, with η ∈ Ω
k
mdeg(f)
. (3.47)
In particular, any (n −1)form on T(f) ⊂ P
n−1
can be written as
hΩ
f
m
, with h ∈ ¹
mdeg(f)−n
(3.48)
and with Ω = ∆(dt
1
∧ ∧ dt
n
) the (n − 1)form (3.42), homogeneous of
degree n.
We then obtain the following result formulating integrals in aﬃne spaces
in terms of integrals of pullbacks of forms from projective spaces.
Proposition 3.4.2. Let ω ∈ Ω
k
mdeg(f)
be a closed kform, which is ho
mogeneous of degree mdeg(f), and consider the form ω/f
m
on A
n
. Let
Feynman integrals and algebraic varieties 69
σ ⊂ A
n
¦0¦ be a kdimensional domain with boundary ∂σ ,= ∅. Then the
integration of ω/f
m
over σ satisﬁes
mdeg(f)
_
σ
ω
f
m
=
_
∂σ
∆(ω)
f
m
+
_
σ
df ∧
∆(ω)
f
m+1
. (3.49)
Proof. Recall that we have ([Dimca (1992)], [Dolgachev (1982)])
d
_
∆(ω)
f
m
_
= −
∆(d
f
ω)
f
m+1
, (3.50)
where, for a form ω that is homogeneous of degree mdeg(f),
d
f
ω = f dω −mdf ∧ ω. (3.51)
Thus, we have
d
_
∆(ω)
f
m
_
= −
∆(dω)
f
m
+m
∆(df ∧ ω)
f
m+1
. (3.52)
Since the form ω is closed, dω = 0, and we have
∆(df ∧ ω) = deg(f) f ω −df ∧ ∆(ω), (3.53)
we obtain from the above
d
_
∆(ω)
f
m
_
= mdeg(f)
ω
f
m
−
df ∧ ∆(ω)
f
m+1
. (3.54)
By Stokes’ theorem we have
_
∂σ
∆(ω)
f
m
=
_
σ
d
_
∆(ω)
f
m
_
.
Using (3.54) this gives
_
∂σ
∆(ω)
f
m
= mdeg(f)
_
σ
ω
f
m
−
_
σ
df ∧ ∆(ω)
f
m+1
.
which gives (3.49).
In the range −n + D( + 1)/2 ≥ 0 we consider the Feynman integral
U(Γ, p) as deﬁned on the aﬃne hypersurface complement A
n
ˆ
X
Γ
and we
use the result above to reformulate the parametric Feynman integrals in
terms of integrals of forms that are pullbacks to A
n
¦0¦ of forms on a
hypersurface complement in P
n−1
. For simplicity, we remove here the diver
gent Γfactor from the parametric Feynman integral and we concentrate on
the residue given by the integration over the simplex σ as in (3.55) below.
70 Feynman Motives
Proposition 3.4.3. Under the generic condition on the external momenta,
and assuming that −n +D/2 ≥ 0, the parametric Feynman integral
U(Γ, p) =
_
σ
V
−n+D/2
Γ
ω
n
Ψ
D/2
Γ
, (3.55)
with V
Γ
(t, p) = P
Γ
(t, p)/Ψ
Γ
(t, p), can be computed as
U(Γ, p) =
1
C(n, D, )
__
∂σ
π
∗
(η) +
_
σ
df ∧
π
∗
(η)
f
_
, (3.56)
where π : A
n
¦0¦ → P
n−1
is the projection and η is the form on the
hypersurface complement T(f) in P
n−1
with
π
∗
(η) =
∆(ω)
f
m
, (3.57)
on A
n
, where f = Ψ
Γ
and m = −n+D(+1)/2, and ω = P
Γ
(t, p)
−n+D/2
ω
n
with ω
n
= dt
1
∧ ∧dt
n
the volume form on A
n
. The coeﬃcient C(n, D, )
in (3.56) is given by
C(n, D, ) = (−n +D( + 1)/2). (3.58)
Proof. Consider on A
n
the form given by ∆(ω)/f
m
, with f, m, and ω
as above. We assume the condition of Deﬁnition 3.3.1, i.e. for a generic
choice of the external momenta the polynomials P
Γ
and Ψ
Γ
have no common
factor. First notice that, since the polynomial Ψ
Γ
is homogeneous of degree
and P
Γ
is homogeneous of degree +1, the form ∆(ω)/f
m
is G
m
invariant
on A
n
¦0¦. Moreover, since it is of the form α = ∆(ω)/f
m
, it also
satisﬁes ∆(α) = 0, hence it is the pullback of a form η on T(f) ⊂ P
n−1
.
Also notice that the domain of integration σ ⊂ A
n
, given by the simplex
σ = σ
n
= ¦
i
t
i
= 1, t
i
≥ 0¦, is contained in a fundamental domain of the
action of the multiplicative group C
∗
on C
n
¦0¦.
Applying the result of Proposition 3.4.2 above, we obtain
_
σ
P
Γ
(t, p)
−n+D/2
dt
1
∧ ∧ dt
n
Ψ
−n+D(+1)/2
Γ
=
_
σ
ω
f
m
=
1
mdeg(f)
(
_
∂σ
∆(ω)
f
m
+
_
σ
df ∧
∆(ω)
f
m+1
)
= C(n, D, )
−1
__
∂σ
P
Γ
(t, p)
a
∆(ω
n
)
Ψ
m
Γ
+
_
σ
df ∧
P
Γ
(t, p)
a
∆(ω
n
)
Ψ
m+1
Γ
_
,
with a = −n +D/2 and m = −n +D( +1)/2. The coeﬃcient C(n, D, )
is given by C(n, D, ) = mdeg(f), with m and f as above, hence it is given
by (3.58).
Feynman integrals and algebraic varieties 71
3.5 Nonisolated singularities
The graph hypersurfaces X
Γ
⊂ P
n−1
deﬁned by the vanishing of the poly
nomial Ψ
Γ
(t) = det(M
Γ
(t)) usually have nonisolated singularities. This
can easily be seen by the following observation.
Lemma 3.5.1. Let Γ be a graph with deg Ψ
Γ
> 2. The singular locus of
X
Γ
is given by the intersection of cones over the hypersurfaces X
Γ
e
, for
e ∈ E(Γ), where Γ
e
is the graph obtained by removing the edge e of Γ. The
cones C(X
Γ
e
) do not intersect transversely.
Proof. First observe that, since X
Γ
is deﬁned by a homogeneous equation
Ψ
Γ
(t) = 0, with Ψ
Γ
a polynomial of degree m, the Euler formula mΨ
Γ
(t) =
e
t
e
∂
∂t
e
Ψ
Γ
(t) implies that ∩
e
Z(∂
e
Ψ
Γ
) ⊂ X
Γ
, where Z(∂
e
Ψ
Γ
) is the zero
locus of the t
e
derivative. Thus, the singular locus of X
Γ
is just given by the
equations ∂
e
Ψ
Γ
= 0. The variables t
e
appear in the polynomial Ψ
Γ
(t) only
with degree zero or one, hence the polynomial ∂
e
Ψ
Γ
consists of only those
monomials of Ψ
Γ
that contain the variable t
e
, where one sets t
e
= 1. The
resulting polynomial is therefore of the form Ψ
Γ
e
, where Γ
e
is the graph
obtained from Γ by removing the edge e. In fact, one can see in terms of
spanning trees that, if T is a spanning tree containing the edge e, then T e
is no longer a spanning tree of Γ
e
, so the corresponding terms disappear
in passing from Ψ
Γ
to Ψ
Γ
e
, while if T is a spanning tree of Γ which does
not contain e, then T is still a spanning tree of Γ
e
and the corresponding
monomial m
T
of Ψ
Γ
e
is the same as the monomial m
T
in Ψ
Γ
without the
variable t
e
. Thus, the zero locus Z(Ψ
Γ
e
) ⊂ P
n−1
is a cone C(X
Γ
e
) over
the graph hypersurface X
Γ
e
⊂ P
n−2
with vertex at the coordinate point
v
e
= (0, . . . , 0, 1, 0, . . . 0) with t
e
= 1. To see that these cones do not
intersect transversely, notice that, in the case where deg Ψ
Γ
> 2, given any
two C(X
Γ
e
) and C(X
Γ
e
), the vertex of one cone is contained in the graph
hypersurface spanning the other cone.
In fact, the hypersurfaces X
Γ
tend to have singularity loci of low codi
mension. As we observe again later, this has important consequences from
the motivic viewpoint. Moreover, it makes it especially useful to adopt
methods from singularity theory to study these hypersurfaces. For exam
ple, we are going to see in the following sections how characteristic classes
of singular varieties can be employed in this context. These provide a way
of measuring how singular a hypersurface is, as we discuss more in detail
in ¸3.9 below. Ongoing work [Bergbauer and Rej (2009)] gives an analysis
72 Feynman Motives
of the singular locus of the graph hypersurfaces using a formula for the
Kirchhoﬀ polynomials Ψ
Γ
under insertion of subgraphs at vertices.
3.6 Cremona transformation and dual graphs
A useful tool to study the algebraic geometry of the projective graph hy
persurfaces X
Γ
is the Cremona transformation, which for planar graphs
relates the hypersurface of a graph with that of its dual. This was pointed
out in [Bloch (2007)] and we illustrate the principle here in the particular
completely explicit case of the “banana graphs” computed in [Aluﬃ and
Marcolli (2008a)]. Even for graphs that are nonplanar, one can still use
the image of the graph hypersurface under the Cremona transformation
to derive useful information on its motivic nature, as shown in the recent
results of [Bloch (2008)]. We follow [Aluﬃ and Marcolli (2008a)].
Deﬁnition 3.6.1. The standard Cremona transformation of P
n−1
is the
map
( : (t
1
: : t
n
) →
_
1
t
1
: :
1
t
n
_
. (3.59)
Let ((() denote the closure of the graph of (. Then ((() is a subvariety
of P
n−1
P
n−1
with projections
((()
π
1
π
2
?
?
?
?
?
?
?
P
n−1
C
// _ _ _ _ _ _
P
n−1
(3.60)
We introduce the following notation. Let o
n
⊂ P
n−1
be deﬁned by the
ideal
1
S
n
= (t
1
t
n−1
, t
1
t
n−2
t
n
, . . . , t
1
t
3
t
n
, t
2
t
n
). (3.61)
Also, let L be the hyperplane deﬁned by the equation
L = ¦(t
1
: : t
n
) ∈ P
n−1
[ t
1
+ +t
n
= 0¦. (3.62)
The Cremona transformation is a priori deﬁned only away from the
algebraic simplex of coordinate axes
Σ
n
:= ¦(t
1
: : t
n
) ∈ P
n−1
[
i
t
i
= 0¦ ⊂ P
n−1
, (3.63)
Feynman integrals and algebraic varieties 73
though, as we now show, it is in fact well deﬁned also on the general point
of Σ
n
, its locus of indeterminacies being only the singularity subscheme o
n
of the algebraic simplex Σ
n
.
Lemma 3.6.2. Let (, (((), o
n
, and L be as above.
(1) The indeterminacy locus of ( is o
n
.
(2) π
1
: ((() →P
n−1
is the blowup along o
n
.
(3) L intersects every component of o
n
transversely.
(4) Σ
n
cuts out a divisor with simple normal crossings on L.
Proof. The components of the map deﬁning ( given in (3.59) can be
rewritten as
(t
1
: : t
n
) → (t
2
t
n
: t
1
t
3
t
n
: : t
1
t
n−1
). (3.64)
Using (3.64), the map π
1
: ((() → P
n
may be identiﬁed with the blowup
of P
n
along the subscheme o
n
deﬁned by the ideal 1
S
n
of (3.61) generated
by the partial derivatives of the equation of the algebraic simplex (i.e. the
singular locus of Σ
n
). The properties (3) and (4) then follow.
Properties (3) and (4) of the proposition above only play a role implicitly
in the proof of Proposition 3.7.1 below, so we leave here the details to the
reader.
An example where it is easier to visualize the eﬀect of the Cremona
transformation is the case where n = 3. In this case, one can view the
algebraic simplex Σ
3
in the projective plane P
2
as a union of three lines,
each two intersecting in a point. These three points, the zerodimensional
strata of Σ
3
, are blown up in ((() and each replaced by a line. The proper
transforms of the 1dimensional components of Σ
3
are then blown down by
the other map π
2
of (3.60), so that the resulting image is again a triangle
of three lines in P
2
. The Cremona transformation (the horizontal rational
map in the diagram (3.60)) is an isomorphism of the complements of the
two triangles.
74 Feynman Motives
The dual graph Γ
∨
of a planar graph Γ is deﬁned by choosing an em
bedding ι : Γ → S
2
in the sphere and then taking as vertices v ∈ V (Γ
∨
) a
point in each connected component of S
2
ι(Γ),
#V (Γ
∨
) = b
0
(S
2
ι(Γ)), (3.65)
and edges E(Γ
∨
) given by adding an edge between two vertices v, w ∈
V (Γ
∨
) for each edge of Γ that is in the boundary between the regions of
S
2
ι(Γ) containing the points v and w. Thus, the dual graph has
#E(Γ
∨
) = #E(Γ). (3.66)
Notice that it is somewhat misleading to talk about the dual graph Γ
∨
,
since in fact this depends not only on Γ itself, but also on the choice of the
embedding ι : Γ → S
2
. One can give examples of diﬀerent embeddings ι
1
,
ι
2
of the same graph Γ, for which the corresponding dual graphs Γ
∨
1
and
Γ
∨
2
are topologically inequivalent, as the following ﬁgure shows.
Feynman integrals and algebraic varieties 75
dual graph
dual graph
It was shown in [Bloch (2007)] that the Cremona transformation relates
the graph hypersurfaces of Γ and its dual Γ
∨
in the following way (see also
[Aluﬃ and Marcolli (2008a)]).
Lemma 3.6.3. Suppose given a planar graph Γ with #E(Γ) = n, with dual
graph Γ
∨
. Then the graph polynomials satisfy
Ψ
Γ
(t
1
, . . . , t
n
) = (
e∈E(Γ)
t
e
) Ψ
Γ
∨(t
−1
1
, . . . , t
−1
n
), (3.67)
hence the graph hypersurfaces are related by the Cremona transformation (
of (3.59),
((X
Γ
∩ (P
n−1
Σ
n
)) = X
Γ
∨ ∩ (P
n−1
Σ
n
). (3.68)
Proof. This follows from the combinatorial identity
Ψ
Γ
(t
1
, . . . , t
n
) =
T⊂Γ
e/ ∈E(T)
t
e
= (
e∈E(Γ)
t
e
)
T⊂Γ
e∈E(T)
t
−1
e
= (
e∈E(Γ)
t
e
)
T
⊂Γ
∨
e/ ∈E(T
)
t
−1
e
= (
e∈E(Γ)
t
e
)Ψ
Γ
∨(t
−1
1
, . . . , t
−1
n
).
The third equality uses the fact that #E(Γ) = #E(Γ
∨
) and #V (Γ
∨
) =
b
0
(S
2
Γ), so that deg Ψ
Γ
+deg Ψ
Γ
∨ = #E(Γ), and the fact that there is a
bijection between complements of spanning tree T in Γ and spanning trees
T
in Γ
∨
obtained by shrinking the edges of T in Γ and taking the dual graph
of the resulting connected graph. Written in the coordinates (s
1
: : s
n
)
of the target P
n−1
of the Cremona transformation, the identity (3.67) gives
Ψ
Γ
(t
1
, . . . , t
n
) = (
e∈E(Γ
∨
)
s
−1
e
)Ψ
Γ
∨(s
1
, . . . , s
n
)
from which (3.68) follows.
76 Feynman Motives
3.7 Classes in the Grothendieck ring
An example of an inﬁnite family of Feynman graphs for which the classes
in the Grothendieck ring K
0
(1) can be computed explicitly was considered
in [Aluﬃ and Marcolli (2008a)]. These are known as the banana graphs:
for n ≥ 2 the graph Γ
n
in this family has two vertices of valence n and n
parallel edges between them.
The banana graph Γ
n
has graph polynomial
Ψ
Γ
(t) = t
1
t
n
(
1
t
1
+ +
1
t
n
).
Thus, the parametric integral in this case is
_
σ
n
(t
1
t
n
)
(
D
2
−1)(n−1)−1
ω
n
Ψ
Γ
(t)
(
D
2
−1)n
,
up to an overall multiplicative factor proportional to the sum of external
momenta at a vertex.
We gave in [Aluﬃ and Marcolli (2008a)] an explicit formula for the class
in the Grothendieck ring of the graph hypersurfaces for the banana graphs.
This is given by the following result.
Proposition 3.7.1. The class in the Grothendieck ring of the hypersurface
X
Γ
n
of the nth banana graph is explicitly given by the formula
[X
Γ
n
] =
L
n
−1
L −1
−
(L −1)
n
−(−1)
n
L
−n(L −1)
n−2
. (3.69)
In particular, this formula shows that, in this example, the class [X
Γ
n
]
manifestly lies in the mixed Tate part Z[L] of the Grothendieck ring. We
sketch the proof here and refer the reader to [Aluﬃ and Marcolli (2008a)]
for more details.
Proof. The formula (3.69) is proved using the method of Cremona trans
formation and dual graphs described above. In fact, one observes easily
Feynman integrals and algebraic varieties 77
that, in the case of the banana graphs, the dual graph Γ
∨
n
is simply a poly
gon with n sides, and therefore the associated graph hypersurface X
Γ
∨
n
= L
is just a hyperplane in P
n−1
.
One then checks that the complement in this hyperplane of the algebraic
simplex has class
[L (L ∩ Σ
n
)] = [L] −[L ∩ Σ
n
] =
T
n−1
−(−1)
n−1
T + 1
where T = [G
m
] = [A
1
] − [A
0
] is the class of the multiplicative group.
Moreover, one ﬁnds that X
Γ
n
∩Σ
n
= o
n
, the scheme of singularities of Σ
n
.
The latter has class
[o
n
] = [Σ
n
] −nT
n−2
.
This then gives
[X
Γ
n
] = [X
Γ
n
∩ Σ
n
] + [X
Γ
n
(X
Γ
n
∩ Σ
n
)],
where one uses the Cremona transformation to identify
[X
Γ
n
] = [o
n
] + [L (L ∩ Σ
n
)].
In particular, since the class in the Grothendieck ring is a universal
Euler characteristic, this calculation also yields as a consequence the value
for the topological Euler characteristic of the (complex) variety X
Γ
n
(C),
which is of the form
χ(X
Γ
n
) = n + (−1)
n
,
for n ≥ 3, as one can see from the fact that the Euler characteristic is a ring
homomorphism from K
0
(1) to Z and that in (3.69) L = [A
1
] has Euler char
acteristic χ(A
1
) = 1. A diﬀerent computation of the Euler characteristic
χ(X
Γ
n
), based on the use of characteristic classes of singular varieties, is also
given in [Aluﬃ and Marcolli (2008a)] along with the more complicated ex
plicit computation of the Chern–Schwartz–MacPherson characteristic class
of the hypersurfaces X
Γ
n
.
More generally, the use of the Cremona transformation method can lead
to interesting results even in the case where the graphs are not necessarily
planar. In fact, one can still consider a dual hypersurface X
∨
Γ
obtained
as the image of X
Γ
under the Cremona transformation and still have an
isomorphism of X
Γ
and X
∨
Γ
away from the algebraic simplex Σ
n
,
X
Γ
(X
Γ
∩ Σ
n
)
C
· X
∨
Γ
(X
∨
Γ
∩ Σ
n
),
78 Feynman Motives
although in nonplanar cases X
∨
Γ
is no longer identiﬁed with the graph
hypersurface of a dual graph X
Γ
∨. By applying this method to the complete
graph Γ
N
on N vertices, Bloch proved in [Bloch (2008)] a very interesting
result showing that, although the individual hypersurfaces X
Γ
of Feynman
graphs are not always mixed Tate motives, if one sums the classes over
graphs with a ﬁxed number of vertices one obtains a class that is always in
the Tate part Z[L] of the Grothendieck ring. More precisely, it is shown in
[Bloch (2008)] that the class
S
N
=
#V (Γ)=N
[X
Γ
]
N!
#Aut(Γ)
,
where the sum is over all graphs with a ﬁxed number of vertices, is always in
Z[L]. This shows that there will be very interesting cancellations between
the classes [X
Γ
] at least for a suﬃciently large number of loops, where one
knows by the general result of [Belkale and Brosnan (2003a)] that non
mixedTate contributions will eventually appear.
This result ﬁts in well with the general fact that, in quantum ﬁeld
theory, individual Feynman graphs do not represent observable physical
processes and only sums over graphs (usually with ﬁxed external edges
and external momenta) can be physically meaningful. Although in the
more physical setting it would be natural to sum over graphs with a given
number of loops and with given external momenta, rather than over a
ﬁxed number of vertices, the result of [Bloch (2008)] suggests that a more
appropriate formulation of the conjecture on Feynman integrals and motives
should perhaps be given directly in terms that involve the full expansion of
perturbative quantum ﬁeld theory, with sums over graphs, rather than in
terms of individual graphs. This approach via families of graphs also ﬁts
in well with the treatment of gauge theories and Slavonov–Taylor identities
within the context of the Connes–Kreimer approach to renormalization via
Hopf algebras, as in [van Suijlekom (2007)], [van Suijlekom (2006)] and
also [Kreimer and van Suijlekom (2009)] in the setting of Dyson–Schwinger
equations.
3.8 Motivic Feynman rules
We have discussed above a deﬁnition of “abstract Feynman rules” based on
the multiplicative properties (3.31) and (3.32) that express the expectation
value of Feynman integrals associated to arbitrary Feynman graphs of a
Feynman integrals and algebraic varieties 79
given scalar quantum ﬁeld theory ﬁrst in terms of connected graphs and
then in terms of 1PI graphs.
This more abstract deﬁnition of Feynman rules allows one to make sense
of algebrogeometric and motivic U(Γ), with the same formal properties as
the original Feynman integrals with respect to the combinatorics of graphs.
As we discuss later on in the book, this will be closely related to deﬁning
Feynman rules in terms of characters of the Connes–Kreimer Hopf algebra
of Feynman graphs and to the abstract form of the Birkhoﬀ factorization
that gives the renormalization procedure for Feynman integrals.
One can ﬁrst observe, as in [Aluﬃ and Marcolli (2008b)], that the hy
persurface complement A
n
ˆ
X
Γ
where the parametric Feynman integral is
computed behaves itself like a Feynman integral, in the sense that it satis
ﬁes the multiplicative properties (3.31) and (3.32). More precisely, we have
seen in Lemma 3.2.2 above that for a disjoint union of graphs Γ = Γ
1
∪ Γ
2
the hypersurface complement splits multiplicatively as
A
n
ˆ
X
Γ
= (A
n
1
ˆ
X
Γ
1
) (A
n
2
ˆ
X
Γ
2
).
Lemma 3.2.2 shows that the hypersurface complement A
n
ˆ
X
Γ
sat
isﬁes the multiplicative property (3.31), when we think of the Cartesian
product (A
n
1
ˆ
X
Γ
1
) (A
n
2
ˆ
X
Γ
2
) as deﬁning the product operation in a
suitable commutative ring. We will specify the ring more precisely below
as a reﬁnement of the Grothendieck ring of varieties, namely the ring of
immersed conical varieties introduced in [Aluﬃ and Marcolli (2008b)]. One
can similarly see that the second property of Feynman rules, the multiplica
tive property (3.32) reducing connected graphs to 1PI graphs and inverse
propagators, is also satisﬁed by the hypersurface complements.
Lemma 3.8.1. For a connected graph Γ that is obtained by inserting 1PI
graphs Γ
v
at the vertices of a ﬁnite tree T, the hypersurface complement
satisﬁes
A
n
ˆ
X
Γ
= A
#E(T)
v∈V (T)
(A
#E
int
(Γ
v
)
ˆ
X
Γ
v
). (3.70)
Proof. If Γ is a forest with n = #E(Γ), then A
n
ˆ
X
Γ
= A
n
. Also
notice that if Γ
1
and Γ
2
are graphs joined at a single vertex, then the same
argument given in Lemma 3.2.2 for the disjoint union still gives the desired
result. Thus, in the case of a connected graphs Γ obtained by inserting 1PI
graphs at vertices of a tree, these two observations combine to give
(A
n
1
ˆ
X
Γ
v
1
) (A
n
#V (T)
ˆ
X
Γ
v
#V (T)
) A
#E(T)
.
80 Feynman Motives
In the above, one can consider the hypersurface complements A
n
ˆ
X
Γ
modulo certain equivalence relations. For example, if these varieties are
considered up to isomorphism, and one imposes the inclusion–exclusion
relation [X] = [X Y ] +[Y ] for closed subvarieties Y ⊂ X, one can deﬁne
a Feynman rule with values in the Grothendieck ring of varieties
U(Γ) := [A
n
ˆ
X
Γ
] = [A
n
] −[
ˆ
X
Γ
] ∈ K
0
(1). (3.71)
This satisﬁes (3.31) and (3.32) as an immediate consequence of Lemma
3.2.2 and Lemma 3.8.1.
The relation between the classes in the Grothendieck ring of the aﬃne
and projective hypersurface complements is given by the formula
[A
n
ˆ
X
Γ
] = (L −1)[P
n−1
X
Γ
],
which expresses the fact that
ˆ
X
Γ
is the aﬃne cone over X
Γ
.
One can also impose a weaker equivalence relation, by identifying vari
eties not up to isomorphism, but only up to linear changes of coordinates
in an ambient aﬃne space. This leads to the following reﬁnement of the
Grothendieck ring of varieties ([Aluﬃ and Marcolli (2008b)]).
Deﬁnition 3.8.2. The ring of immersed conical varieties T
K
is generated
by classes [V ] of equivalence under linear coordinate changes of varieties
V ⊂ A
N
deﬁned by homogeneous ideals (hence the name “conical”) im
mersed in some arbitrarily large aﬃne space, with the inclusionexclusion
and product relations
[V ∪ W] = [V ] + [W] −[V ∩ W]
[V ] [W] = [V W].
By imposing equivalence under isomorphisms one falls back on the usual
Grothendieck ring K
0
(1).
Thus, one can deﬁne an ¹valued algebrogeometric Feynman rule, for
a given commutative ring ¹, as in [Aluﬃ and Marcolli (2008b)] in terms of
a ring homomorphism I : T → ¹ by setting
U(Γ) := I([A
n
]) −I([
ˆ
X
Γ
])
and by taking as value of the inverse propagator
U(L) = I([A
1
]).
This then satisﬁes both (3.31) and (3.32). The ring T is then the receptacle
of the universal algebrogeometric Feynman rule given by
U(Γ) = [A
n
ˆ
X
Γ
] ∈ T.
Feynman integrals and algebraic varieties 81
A Feynman rule deﬁned in this way is motivic if the homomorphism I :
T → ¹ factors through the Grothendieck ring K
0
(1
K
). In this case the
inverse propagator is the Lefschetz motive (or the propagator is the Tate
motive).
The reason for working with T
K
instead is that it allowed us in [Aluﬃ
and Marcolli (2008b)] to construct invariants of the graph hypersurfaces
that behave like algebrogeometric Feynman rules and that measure to some
extent how singular these varieties are, and which do not factor through
the Grothendieck ring, since they contain speciﬁc information on how the
ˆ
X
Γ
are embedded in the ambient aﬃne space A
#E
int
(Γ)
.
3.9 Characteristic classes and Feynman rules
In the case of compact smooth varieties, the Chern classes of the tangent
bundle can be written as a class c(V ) = c(TV ) ∩ [V ] in homology whose
degree of the zero dimensional component satisﬁes the Poincar´e–Hopf theo
rem
_
c(TV ) ∩[V ] = χ(V ), which gives the topological Euler characteristic
of the smooth variety.
Chern classes for singular varieties that generalize this property were
introduced independently, following two diﬀerent approaches, in [Schwartz
(1965)] and [MacPherson (1974)]. The two deﬁnitions were later proved to
be equivalent.
The approach followed by Marie H´el`ene Schwartz generalized the deﬁ
nition of Chern classes as the homology classes of the loci where a family
of k +1 vector ﬁelds become linearly dependent (for the lowest degree case
one reads the Poincar´e–Hopf theorem as saying that the Euler characteris
tic measures where a single vector ﬁeld has zeros). In the case of singular
varieties a generalization is obtained, provided that one assigns some radial
conditions on the vector ﬁelds with respect to a stratiﬁcation with good
properties.
The approach followed by Robert MacPherson is instead based on func
toriality. The functor F of constructible functions assigns to an algebraic
variety V the Zmodule F(V ) spanned by the characteristic functions 1
W
of subvarieties W ⊂ V and to a proper morphism f : V → V
the map
f
∗
: F(V ) → F(V
) deﬁned by f
∗
(1
W
)(p) = χ(W ∩ f
−1
(p)), with χ the
Euler characteristic.
A conjecture of Grothendieck–Deligne predicted the existence of a
unique natural transformation c
∗
between the functor F and the homol
82 Feynman Motives
ogy (or Chow group) functor, which in the smooth case agrees with
c
∗
(1
V
) = c(TV ) ∩ [V ].
MacPherson constructed this natural transformation in terms of data
of Mather classes and local Euler obstructions and deﬁned in this way
what is usually referred to now as the Chern–Schwartz–MacPherson (CSM)
characteristic classes of (singular) algebraic varieties c
SM
(X) = c
∗
(1
X
).
It is convenient here to work inside an ambient space, for instance an
ambient projective space, so that one can regard c
∗
(1
X
) as taking values
in the homology (or Chow group) of the ambient P
N
, so that the charac
teristic class c
SM
(X) = c
∗
(1
X
) can be computed for X an arbitrary locally
closed subset of P
N
, without requiring a compactness hypothesis. This is
important in order to have an inclusionexclusion relation for these classes,
as explained below.
The results of [Aluﬃ (2006)] show that, in fact, it is possible to compute
these classes without having to use Mather classes and Euler obstructions
that are usually very diﬃcult to compute. Most notably, these characteristic
classes satisfy an inclusionexclusion formula for a closed Y ⊂ X,
c
SM
(X) = c
SM
(Y ) +c
SM
(X Y ),
but are not invariant under isomorphism, hence they are naturally deﬁned
on classes in T
K
but not on K
0
(1
K
).
For a smooth locally closed X in some ambient projective space P
N
,
the class c
SM
(X) can be obtained by a computation using Chern classes of
sheaves of diﬀerentials with logarithmic poles in a resolution of the closure
¯
X of X. Then inclusionexclusion gives a way to compute any embedded
c
SM
without recourse to MacPherson’s Euler obstructions or ChernMather
classes.
The CSM classes give good information on the singularities of a variety:
for example, in the case of hypersurfaces with isolated singularities, they
can be expressed in terms of Milnor numbers.
To construct a Feynman rule out of these Chern classes, one uses the
following procedure, [Aluﬃ and Marcolli (2008b)]. Given a variety
ˆ
X ⊂ A
N
,
one can view it as a locally closed locus in P
N
, hence one can apply to
its characteristic function 1
ˆ
X
the natural transformation c
∗
that gives an
element in the Chow group A(P
N
) or in the homology H
∗
(P
N
). This gives
as a result a class of the form
c
∗
(1
ˆ
X
) = a
0
[P
0
] +a
1
[P
1
] + +a
N
[P
N
].
One then deﬁnes an associated polynomial given by
G
ˆ
X
(T) := a
0
+a
1
T + +a
N
T
N
.
Feynman integrals and algebraic varieties 83
It is in fact independent of N as it stops in degree equal to dim
ˆ
X. It is by
construction invariant under linear changes of coordinates. It also satisﬁes
an inclusionexclusion property coming from the fact that the classes c
SM
satisfy inclusionexclusion, namely
G
ˆ
X∪
ˆ
Y
(T) = G
ˆ
X
(T) +G
ˆ
Y
(T) −G
ˆ
X∩
ˆ
Y
(T)
It is a more delicate result to show that it is multiplicative, namely that
the following holds.
Theorem 3.9.1. The polynomials G
ˆ
X
(T) satisfy
G
ˆ
X×
ˆ
Y
(T) = G
ˆ
X
(T) G
ˆ
Y
(T).
Proof. The proof of this fact is obtained in [Aluﬃ and Marcolli (2008b)]
using an explicit formula for the CSM classes of joins in projective spaces,
where the join J(X, Y ) ⊂ P
m+n−1
of two X ⊂ P
m−1
and Y ⊂ P
n−1
is
deﬁned as the set of
(sx
1
: : sx
m
: ty
1
: : ty
n
), with (s : t) ∈ P
1
,
and is related to the product in aﬃne spaces by the property that the
product
ˆ
X
ˆ
Y of the aﬃne cones over X and Y is the aﬃne cone over
J(X, Y ). One has
c
∗
(1
J(X,Y )
) = ((f(H) +H
m
)(g(H) +H
n
) −H
m+n
) ∩ [P
m+n−1
],
where c
∗
(1
X
) = H
n
f(H) ∩ [P
n+m−1
] and c
∗
(1
Y
) = H
m
g(H) ∩ [P
n+m−1
],
as functions of the hyperplane class H in P
n+m−1
. We refer the reader to
[Aluﬃ and Marcolli (2008b)] for more details.
The resulting multiplicative property of the polynomials G
ˆ
X
(T) shows
that one has a ring homomorphism I
CSM
: T →Z[T] deﬁned by
I
CSM
([
ˆ
X]) = G
ˆ
X
(T)
and an associated Feynman rule
U
CSM
(Γ) = C
Γ
(T) = I
CSM
([A
n
]) −I
CSM
([
ˆ
X
Γ
]).
This is not motivic, i.e. it does not factor through the Grothendieck
ring K
0
(1
K
), as can be seen by the example given in [Aluﬃ and Marcolli
(2008b)] of two graphs (see the ﬁgure below) that have diﬀerent U
CSM
(Γ),
C
Γ
1
(T) = T(T + 1)
2
C
Γ
2
(T) = T(T
2
+T + 1)
but the same hypersurface complement class in the Grothendieck ring,
[A
n
ˆ
X
Γ
i
] = [A
3
] −[A
2
] ∈ K
0
(1).
84 Feynman Motives
There is an interesting positivity property that the CSM classes of the
graph hypersurfaces X
Γ
appear to satisfy, namely the coeﬃcients of all the
powers H
k
in the expression of the CMS class of X
Γ
as a polynomial in
H are positive. This was observed in [Aluﬃ and Marcolli (2008a)] on the
basis of numerical computations of these classes, done using the program
of [Aluﬃ (2003)], for sample graphs. As the graphs that are accessible
to computer calculations of the corresponding CSM classes are necessarily
combinatorially simple (and all planar), it is at present not known whether
the observed positivity phenomenon holds more generally. This property
may be related to other positivity phenomena observed for CSM classes of
similarly combinatorial objects such as Schubert varieties [Aluﬃ and Mi
halcea (2006)]. We see in the last chapter that, in the case of quantum ﬁeld
theories on noncommutative spacetimes and the corresponding modiﬁcation
of the graph polynomials, positivity fails in nonplanar cases.
3.10 Deletioncontraction relation
We report here brieﬂy the recent results of [Aluﬃ and Marcolli (2009b)] on
deletion–contraction relations for motivic Feynman rules.
We have seen above that one can construct a polynomial invariant C
Γ
(T)
of graphs that is an algebrogeometric Feynman rule in the sense of [Aluﬃ
and Marcolli (2008b)], which does not factor through the Grothendieck
ring of varieties but satisﬁes an inclusion–exclusion principle. The polyno
mial C
Γ
(T) is deﬁned in terms of CSM classes of the graph hypersurface
complement.
Several well known examples of polynomial invariants of graphs are ex
amples of Tutte–Grothendieck invariants and are obtained as specializations
of the Tutte polynomial of graphs, and they all have the property that they
satisfy deletion–contraction relations, which make it possible to compute
the invariant indictively from simpler graphs. These relations express the
invariant of a given graph in terms of the invariants of the graphs obtained
by deleting or contracting edges.
Feynman integrals and algebraic varieties 85
The results of [Aluﬃ and Marcolli (2009b)] investigate to what extent
motivic and algebrogeometric Feynman rules satisfy deletion–contraction
relations. Since such relations make it possible to control the invariant
for more complicated graphs in terms of combinatorially simpler ones, the
validity or failure of such relations can provide useful information that can
help to identify where the motivic complexity of the graph hypersurfaces
increases beyond the mixed Tate case (see ¸3.12 below).
In particular, it was shown in [Aluﬃ and Marcolli (2009b)] that the poly
nomial C
Γ
(T) is not a specialization of the Tutte polynomial. However, a
form of deletion–contraction relation does hold for the motivic Feynman
rule [A
n
ˆ
X
Γ
] ∈ K
0
(1). The latter, unlike the relation satisﬁed by Tutte–
Grothendieck invariants, contains a term expressed as the class of an inter
section of hypersurfaces which does not appear to be easily controllable in
terms of combinatorial information alone. However, this form of deletion–
contraction relations suﬃces to obtain explicit recursive relation for certain
operations on graphs, such as replacing an edge by multiple parallel copies.
We start by recalling the case of the Tutte polynomial of graphs and
the form of the deletion–contraction relation in that well known classical
case. The Tutte polynomial T
Γ
∈ C[x, y] of ﬁnite graphs Γ is completely
determined by the following properties:
• If e ∈ E(Γ) is neither a looping edge nor a bridge the deletion–
contraction relation holds:
T
Γ
(x, y) = T
Γe
(x, y) +T
Γ/e
(x, y). (3.72)
• If e ∈ E(Γ) is a looping edge then
T
Γ
(x, y) = yT
Γ/e
(x, y).
• If e ∈ E(Γ) is a bridge then
T
Γ
(x, y) = xT
Γe
(x, y)
• If Γ has no edges then T
Γ
(x, y) = 1.
Here we call an edge e a “bridge” if the removal of e disconnects the
graph Γ, and a “looping edge” if e starts and ends at the same vertex.
Tutte–Grothendieck invariants are specializations of the Tutte polynomial.
Among them one has well known invariants such as the chromatic polyno
mial and the Jones polynomial.
A ﬁrst observation of [Aluﬃ and Marcolli (2009b)] is that the Tutte
polynomial can be regarded as a Feynman rule, in the sense that it has the
86 Feynman Motives
right multiplicative properties over disjoint unions and over decompositions
of connected graphs in terms of trees and 1PI graphs.
Proposition 3.10.1. The Tutte polynomial invariant deﬁnes an abstract
Feynman rule with values in the polynomial ring C[x, y], by assigning
U(Γ) = T
Γ
(x, y), with inverse propagator U(L) = x. (3.73)
Proof. One can ﬁrst observe that the properties listed above for the Tutte
polynomial determine the closed form
T
Γ
(x, y) =
γ⊂Γ
(x −1)
#V (Γ)−b
0
(Γ)−(#V (γ)−b
0
(γ))
(y −1)
#E(γ)−#V (γ)+b
0
(γ)
,
(3.74)
where the sum is over subgraphs γ ⊂ Γ with vertex set V (γ) = V (Γ) and
edge set E(γ) ⊂ E(Γ). This can be written equivalently as
T
Γ
(x, y) =
γ⊂Γ
(x −1)
b
0
(γ)−b
0
(Γ)
(y −1)
b
1
(γ)
.
This is sometime referred to as the “sum over states” formula for the Tutte
polynomial.
The multiplicative property under disjoint unions of graphs is clear from
the closed expression (3.74), since for Γ = Γ
1
∪Γ
2
we can identify subgraphs
γ ⊂ Γ with V (γ) = V (Γ) and E(γ) ⊂ E(Γ) with all possible pairs of
subgraphs (γ
1
, γ
2
) with V (γ
i
) = V (Γ
i
) and E(γ
i
) ⊂ E(Γ
i
), with b
0
(γ) =
b
0
(γ
1
) + b
0
(γ
2
), #V (Γ) = #V (Γ
1
) + #V (Γ
2
), and #E(γ) = #E(γ
1
) +
#E(γ
2
). Thus, we get
T
Γ
(x, y) =
γ=(γ
1
,γ
2
)
(x −1)
b
0
(γ
1
)+b
0
(γ
2
)−b
0
(Γ)
(y −1)
b
1
(γ
1
)+b
1
(γ
2
)
= T
Γ
1
(x, y) T
Γ
2
(x, y).
The property for connected and 1PI graphs follows from the fact that,
when writing a connected graph in the form Γ = ∪
v∈V (T)
Γ
v
, with Γ
v
1PI
graphs inserted at the vertices of the tree T, the internal edges of the
tree are all bridges in the resulting graph, hence the property of the Tutte
polynomial for the removal of bridges gives
T
Γ
(x, y) = x
#E
int
(T)
T
Γ∪
e∈E
int
(T)
e
(x, y).
Then one obtains an abstract Feynman rule with values in ¹ = C[x, y] of
the form (3.73).
Feynman integrals and algebraic varieties 87
It follows that Tutte–Grothendieck invariants such as the chromatic and
Jones polynomials can also be regarded as Feynman rules.
One can then observe, however, that the algebro–geometric Feynman
rule
C
Γ
(T) = I
CSM
([A
n
ˆ
X
Γ
])
that we discussed in the previous section is not a specialization of the Tutte
polynomial.
Proposition 3.10.2. The polynomial invariant C
Γ
(T) is not a specializa
tion of the Tutte polynomial.
Proof. We show that one cannot ﬁnd functions x = x(T) and y = y(T)
such that
C
Γ
(T) = T
Γ
(x(T), y(T)).
First notice that, if e ∈ E(Γ) is a bridge, the polynomial C
Γ
(T) satisﬁes
the relation
C
Γ
(T) = (T + 1)C
Γe
(T). (3.75)
In fact, (T +1) is the inverse propagator of the algebrogeometric Feynman
rule U(Γ) = C
Γ
(T) and the property of abstract Feynman rules for 1PI
graphs connected by a bridge gives (3.75). In the case where e ∈ E(Γ) is a
looping edge, we have
C
Γ
(T) = T C
Γ/e
(T). (3.76)
In fact, adding a looping edge to a graph corresponds, in terms of graph
hypersurfaces, to taking a cone on the graph hypersurface and intersecting
it with the hyperplane deﬁned by the coordinate of the looping edge. This
implies that the universal algebrogeometric Feynman rule with values in
the Grothendieck ring T of immersed conical varieties satisﬁes
U(Γ) = ([A
1
] −1)U(Γ/e)
if e is a looping edge of Γ and U(Γ) = [A
n
ˆ
X
Γ
] ∈ T. The property
(3.76) then follows since the image of the class [A
1
] is the inverse propagator
(T +1). (See ¸3.9 above and Proposition 2.5 and ¸2.2 of [Aluﬃ and Marcolli
(2008b)].)
This implies that, if C
Γ
(T) has to be a specialization of the Tutte poly
nomial, the relations for bridges and looping edges imply that one has to
identify x(T) = T + 1 and y(T) = T. However, this is not compatible
with the behavior of the invariant C
Γ
(T) on more complicated graphs. For
example, for the triangle graph one has C
Γ
(T) = T(T + 1)
2
while the
specialization T
Γ
(x(T), y(T)) = (T + 1)
2
+ (T + 1) +T.
88 Feynman Motives
One can then investigate what kind of deletion–contraction relations
hold in the case of motivic and algebro–geometric Feynman rules. An
answer is given in [Aluﬃ and Marcolli (2009b)] for the motivic case, that
is, for the Feynman rule
U(Γ) = [A
n
ˆ
X
Γ
] ∈ K
0
(1).
Theorem 3.10.3. Let Γ be a graph with n > 1 edges.
(1) If e ∈ E(Γ) is neither a bridge nor a looping edge then the following
deletion–contraction relation holds:
[A
n
´
X
Γ
] = L [A
n−1
(
´
X
Γe
∩
´
X
Γ/e
)] −[A
n−1
´
X
Γe
]. (3.77)
(2) If the edge e is a bridge in Γ, then
[A
n
´
X
Γ
] = L [A
n−1
´
X
Γe
] = L [A
n−1
´
X
Γ/e
]. (3.78)
(3) If e is a looping edge in Γ, then
[A
n
´
X
Γ
] = (L−1) [A
n−1
´
X
Γe
] = (L−1) [A
n−1
´
X
Γ/e
]. (3.79)
(4) If Γ contains at least two loops, then the projective version of (3.77)
also holds, in the form
[P
n−1
X
Γ
] = L [P
n−2
(X
Γe
∩ X
Γ/e
)] −[P
n−2
X
Γe
]. (3.80)
(5) Under the same hypotheses, the Euler characteristics satisfy
χ(X
Γ
) = n +χ(X
Γe
∩ X
Γ/e
) −χ(X
Γe
). (3.81)
Proof. (1) and (4): Let Γ be a graph with n ≥ 2 edges e
1
, . . . , e
n−1
, e =
e
n
, with (t
1
: . . . : t
n
) the corresponding variables in P
n−1
. As above we
consider the Kirchhoﬀ polynomial Ψ
Γ
and the aﬃne and projective graph
hypersurfaces
ˆ
X
Γ
⊂ A
n
and X
Γ
⊂ P
n−1
. We assume deg Ψ
Γ
= > 0.
We consider an edge e that is not a bridge or looping edge. Then the
polynomials for the deletion Γ e and contraction Γ/e of the edge e = e
n
are both nontrivial and given by
F :=
∂Ψ
Γ
∂t
n
= Ψ
Γe
and G := Ψ
Γ
[
t
n
=0
= Ψ
Γ/e
. (3.82)
The Kirchhoﬀ polynomial Ψ
Γ
satisﬁes
Ψ
Γ
(t
1
, . . . , t
n
) = t
n
F(t
1
, . . . , t
n−1
) +G(t
1
, . . . , t
n−1
). (3.83)
Given a projective hypersurface Y ⊂ P
N−1
we denote by
ˆ
Y the aﬃne
cone in A
N
, and we use the notation
¯
Y for the projective cone in P
N
.
Feynman integrals and algebraic varieties 89
It is proved in [Aluﬃ and Marcolli (2009b)], Theorem 3.3, in a more gen
eral context that includes the case under consideration, that the projection
from the point (0 : : 0 : 1) induces an isomorphism
X
Γ
(X
Γ
∩ X
Γe
)
∼
−→P
n−2
X
Γe
, (3.84)
where X
Γe
is the hypersurface in P
n−2
deﬁned by the polynomial F. In
fact, the projection P
n−1
P
n−2
from p = (0 : : 0 : 1) acts as
(t
1
: : t
n
) → (t
1
: : t
n−1
).
If F is constant (that is, if deg ψ = 1), then X
Γe
= X
Γe
= ∅ and the
statement is trivial. Thus, assume deg F > 0. In this case, Ψ
Γ
(p) = F(p) =
0, hence p ∈ X
Γ
∩ X
Γe
, and hence p ,∈ X (X
Γ
∩ X
Γe
). Therefore, the
projection restricts to a regular map
X
Γ
(X
Γ
∩ X
Γe
) →P
n−2
.
The image is clearly contained in P
n−2
X
Γe
, and the statement is that
this map induces an isomorphism
X
Γ
(X
Γ
∩ X
Γe
)
∼
−→P
n−2
X
Γe
.
Let q = (q
1
: : q
n−1
). The line from p to q is parametrized by
(q
1
: : q
n−1
: t).
We show that this line meets X
Γ
(X
Γ
∩X
Γe
) transversely at one point.
The intersection with X
Γ
is determined by the equation
tF(q
1
: : q
n−1
) +G(q
1
: : q
n−1
) = 0.
Since F(q) ,= 0, this is a polynomial of degree exactly 1 in t, and determines
a reduced point, as needed.
It then follows that, in the Grothendieck ring K
0
(1) we have identities
[P
n−1
X
Γ
] = [P
n−1
(X
Γ
∩ X
Γe
)] −[P
n−2
X
Γe
]. (3.85)
If Γ has at least two loops, so that deg Ψ
Γ
> 1, then we also have
[P
n−1
X
Γ
] = L [P
n−2
(X
Γe
∩ X
Γ/e
)] −[P
n−2
X
Γe
], (3.86)
where L = [A
1
] is the Lefschetz motive and X
Γ/e
is the hypersurface G = 0.
Indeed, the ideal of X
Γ
∩ X
Γe
is
(ψ, F) = (t
n
F +G, F) = (F, G).
This means that
X
Γ
∩ X
Γe
= X
Γe
∩ X
Γ/e
. (3.87)
90 Feynman Motives
If deg X
Γ
> 1, then F is not constant, hence X
Γe
,= ∅. It then follows
that X
Γe
∩ X
Γ/e
contains the point p = (0 : : 0 : 1). The ﬁbers of the
projection
P
n−1
(X
Γe
∩ X
Γ/e
) →P
n−2
X
Γe
with center p are then all isomorphic to A
1
, and it follows that
[P
n−1
(X
Γe
∩ X
Γ/e
)] = L [P
n−2
(X
Γe
∩ X
Γ/e
)].
The aﬃne version (3.77) follows by observing that, if deg X
Γ
> 1, then
deg F > 0, hence
X
Γ
∩ X
Γe
and
´
X
Γe
∩
´
X
Γ/e
contain the origin. In this case, the classes in the Grothendieck ring satisfy
[A
N
X
Γ
∩ X
Γe
] = (L −1) [P
N−1
(X
Γ
∩ X
Γe
)]
[A
N
(
´
X
Γe
∩
´
X
Γ/e
)] = (L −1) [P
N−1
(X
Γe
∩ X
Γ/e
)].
Then (3.77) follows from the formula for the projective case, while it can
be checked directly for the remaning case with deg X
Γ
= 1.
(2) If e is a bridge, then Ψ
Γ
does not depend on the variable t
e
and
F ≡ 0. The equation for X
Γe
is Ψ
Γ
= 0 again, but viewed in one fewer
variables. The equation for X
Γ/e
is the same.
(3) If e is a looping edge, then Ψ
Γ
is divisible by t
e
, so that G ≡ 0. The
equation for X
Γ/e
is obtained by dividing Ψ
Γ
through by t
e
, and one has
X
Γe
= X
Γ/e
.
(5) Under the hypothesis that deg X
Γ
> 1, the statement about the
Euler characteristics follows from (3.80). One has from (3.84) the identity
χ(P
n−1
(X
Γ
∪ X
Γe
)) = 0.
In the case where deg X
Γ
= 1 one has χ(X
Γ
) = n −1.
This general result on deletioncontraction relations for motivic Feyn
man rules was applied in [Aluﬃ and Marcolli (2009b)] to analyze some
operations on graphs, which have the property that the problem of describ
ing the intersection X
Γe
∩ X
Γ/e
can be bypassed and the class of more
complicated graphs can be computed inductively only in terms of combi
natorial data.
One such operation replaces a chosen edge e in a graph Γ with m parallel
edges connecting the same two vertices ∂(e). Another operation consists of
replacing an edge with a chain of triangles (referred to as a “lemon graph”).
Feynman integrals and algebraic varieties 91
The latter operation can easily be generalized to chains of polygons as
explained in [Aluﬃ and Marcolli (2009b)].
Assume that e is an edge of Γ, and denote by Γ
me
the graph obtained
from Γ by replacing e by m parallel edges. (Thus, Γ
0e
= Γe, and Γ
e
= Γ.)
Theorem 3.10.4. Let e be an edge of a graph Γ. Let T = L −1 = [G
m
] ∈
K
0
(1).
(1) If e is a looping edge, then
m≥0
U(Γ
me
)
s
m
m!
= e
Ts
U(Γ e). (3.88)
(2) If e is a bridge, then
m≥0
U(Γ
me
)
s
m
m!
=
_
T
e
Ts
−e
−s
T + 1
+s e
Ts
+ 1
_
U(Γ e). (3.89)
(3) If e is not a bridge nor a looping edge, then
m≥0
U(Γ
me
)
s
m
m!
=
e
Ts
−e
−s
T + 1
U(Γ)
+
e
Ts
+Te
−s
T + 1
U(Γ e)
+
_
s e
Ts
−
e
Ts
−e
−s
T + 1
_
U(Γ/e).
(3.90)
The result is proved in [Aluﬃ and Marcolli (2009b)] by ﬁrst deriving the
relation in the case of doubling an edge and then repeating the construction.
We do not report here the details of the proof, and we refer the readers to
the paper for more details, but it is worth pointing out that the main step
that makes it possible to obtain in this case a completely explicit formula
in terms of U(Γ), U(Γ e) and U(Γ/e) is a cancellation that occurs in the
calculation of the classes in the operation of doubling an edge. In fact, one
ﬁrst expresses the class U(Γ
2e
) as
U(Γ
2e
) = L [A
n
(
ˆ
X
Γ
∩
ˆ
X
Γ
o
)] −U(Γ),
where n = #E(Γ) and Γ
o
denotes the graph obtained by attaching a looping
edge named e to Γ/e. In the notation we used earlier in this section, the
equation for Γ
o
is given by the polynomial t
e
G. Using inclusion–exclusion
for classes in the Grothendieck ring, it is then shown in [Aluﬃ and Marcolli
(2009b)] that one has
[
ˆ
X
Γ
∩
ˆ
X
Γ
o
] = [
ˆ
X
Γ/e
] + (L −1) [
ˆ
X
Γe
∩
ˆ
X
Γ/e
].
92 Feynman Motives
Thus, one obtains
U(Γ
2e
) = (L −2) U(Γ) + (L −1) U(Γ e) +L U(Γ/e).
This expression no longer contains classes of intersections of graph hyper
surfaces, and can be iterated to obtain the general formula for U(Γ
me
).
It is shown in [Aluﬃ and Marcolli (2009b)] that the recursive relation
obtained in this way for the operation of multiplying edges is very sim
ilar in form to the recursive relation that the Tutte polynomial satisﬁes
under the same operation. In fact, they both are solutions to a universal
recursive relation with diﬀerent initial conditions. One also derives from
the same universal recursion a conjectural form for the recursion relation,
under the operation of multiplying edges Γ → Γ
me
, for the polynomial
invariant C
Γ
(T) = I
CSM
([A
n
ˆ
X
Γ
]).
The operation of replacing an edge with a chain of triangles, or “lemon
graph” depicted in the ﬁgure produces, similarly, a recursive formula for
the motivic Feynman rule U(Γ
Λ
m
), where Γ
Λ
m
is the graph obtained from Γ
by replacing an edge e by a lemon graph Λ
m
.
!
e
Proposition 3.10.5. Let e be an edge of a graph Γ, and assume that e
is neither a bridge nor a looping edge. Let Γ
Λ
m
be the “lemonade graph”
obtained by building an mlemon fanning out from e. Then
m≥0
U(Γ
Λ
m
)s
m
=
1
1 −T(T + 1)s −T(T + 1)
2
s
2
_
(1 −(T + 1)s) U(Γ) + (T + 1)Ts U(Γ e) + (T + 1)
2
s U(Γ/e)
_
.
This result is proven in [Aluﬃ and Marcolli (2009b)] by ﬁrst obtaining
a recursive formula for the classes U(Λ
m
) of the lemon graphs. This is a
succession of operations where one ﬁrst doubles an edge and then splits
the added edge by inserting a vertex. The ﬁrst operation is covered by
the expression derived before for the class U(Γ
2e
) while splitting an edge
with a vertex corresponds to taking a cone over the corresponding graph
Feynman integrals and algebraic varieties 93
hypersurface. Thus, both operations are well understood at the level of
classes in the Grothendieck ring and they give the recursive relation
U(Λ
m+1
) = T(T + 1)U(Λ
m
) +T(T + 1)
2
U(Λ
m−1
),
for m ≥ 1, which gives then
U(Λ
m
) = (T + 1)
m+1
m
i=0
_
m−i
i
_
T
m−i
.
It is interesting to observe that the sequence a
m
= U(Λ
m
) is a divisibility
sequence, in the sense that U(Λ
m−1
) divides U(Λ
n−1
) if m divides n. This is
the property satisﬁed, for instance, by the sequence of Fibonacci numbers.
(See [Aluﬃ and Marcolli (2009b)], Corollary 5.12 for more details.)
Returning to the classes U(Γ
Λ
m
), one then shows that these also satisfy
a relation of the form
U(Γ
Λ
m
) = f
m
(T)U(Γ) +g
m
(T)U(Γ e) +h
m
(T)U(Γ/e),
where the f
m
, g
m
, h
m
satisfy the same recursion that gives U(Λ
m
), but
with diﬀerent seeds given by
_
¸
¸
_
¸
¸
_
f
0
(T) = 1 , f
1
(T) = T
2
−1
g
0
(T) = 0 , g
1
(T) = T(T + 1)
h
0
(T) = 0 , h
1
(T) = (T + 1)
2
.
For more details we refer the reader to [Aluﬃ and Marcolli (2009b)].
3.11 Feynman integrals and periods
Numerical calculations performed in [Broadhurst and Kreimer (1997)] gave
very strong evidence for a relation between residues of Feynman integrals
and periods of mixed Tate motives. In fact, the numerical evidence in
dicates that the values computed in the log divergent case are Qlinear
combinations of multiple zeta values, which, as we recalled in the previous
chapter, can be realized as periods of mixed Tate motives.
Multiple zeta values are real numbers obtained by summing convergent
series of the form
ζ(n
1
, . . . , n
r
) =
0<k
1
<k
2
<···<k
r
1
k
n
1
1
k
n
r
r
, (3.91)
where the n
i
are positive integers with n
r
≥ 2. There are ways to real
ize multiple zeta values as periods of mixed Tate motives; see [Goncharov
94 Feynman Motives
(2001)], [Goncharov and Manin (2004)], [Terasoma (2002)]. In [Goncharov
and Manin (2004)], for instance, as well as in [Brown (2006)], multiple zeta
values are explicitly realized as periods on the moduli spaces of curves. It is
the occurrence of multiple zeta values in Feynman integral computations,
along with the fact that these numbers arise as periods of mixed Tate mo
tives, that gave rise to the idea of a deeper relation between perturbative
quantum ﬁeld theory and motives.
Among the examples of Feynman graphs computed in [Broadhurst and
Kreimer (1997)], the wheel with n spokes graphs give simple zeta values
ζ(2n − 3). It is also shown in [Broadhurst and Kreimer (1997)] that the
nonplanar graph given by the complete bipartite graph K
3,4
evaluates to
the double zeta value ζ(5, 3).
Recent results of [Brown (2009a)] for multiloop Feynman diagrams pro
vide a new method for evaluating the case of primitive divergent graphs (i.e.
those that do not contain any smaller divergent subgraph, and are there
fore primitive elements in the Connes–Kreimer Hopf algebra). This shows
the occurrence of values at roots of unity of multiple polylogarithms, in
addition to the multiple zeta values observed in [Broadhurst and Kreimer
(1997)]. The multiple polylogarithm function, introduced in [Goncharov
(2001)] is deﬁned as
Li
n
1
,...,n
r
(x
1
, . . . , x
r
) =
0<k
1
<k
2
<···<k
r
x
k
1
1
x
k
r
r
k
n
1
1
k
n
r
r
. (3.92)
It is absolutely convergent for [x
i
[ < 1 and it extends analytically to an open
domain in C
r
as a multivalued holomorphic function. As shown in [Gon
charov (2001)], the multiple polylogarithms are also associated to objects
in the category of mixed Tate motives.
3.12 The mixed Tate mystery
The results recalled in the previous section suggest that, in its strongest
possible form, one could formulate a conjecture as follows.
Conjecture 3.12.1. Are all residues of Feynman graphs of perturbative
scalar ﬁeld theories (possibly after a suitable regularization and renormal
ization procedure is used) periods of mixed Tate motives?
In this form the conjecture is somewhat vaguely formulated, but it does
not matter since it is unlikely to hold true in such generality. One can
Feynman integrals and algebraic varieties 95
however reﬁne the question and formulate it more precisely in the following
form.
Question 3.12.2. Under what conditions on the graphs, the scalar ﬁeld
theory, and the renormalization procedure are the residues of Feynman
graphs periods of mixed Tate motives?
As we have seen above, using the Feynman parameters, one can write
the Feynman integrals in a form that looks exactly like a period of an
algebraic variety obtained as a hypersurface complement, modulo the issue
of divergences which needs to be handled separately. In the stable range for
D, this hypersurface is the graph variety X
Γ
, while in the unstable range
it is the Landau variety Y
Γ
, or the union of the two.
Kontsevich formulated the conjecture that the graph hypersurfaces X
Γ
themselves may always be mixed Tate motives, which would imply Con
jecture 3.12.1. Although numerically this conjecture was at ﬁrst veriﬁed
up to a large number of loops in [Stembridge (1998)], it was later dis
proved in [Belkale and Brosnan (2003a)]. They proved that the varieties
X
Γ
can be arbitrarily complicated as motives: indeed, the X
Γ
generate the
Grothendieck ring of varieties.
The fact that the classes [X
Γ
] can be arbitrarily far from the mixed Tate
part Z[L] of the Grothendieck ring K
0
(1) seems to indicate at ﬁrst that
Conjecture 3.12.1 cannot hold. However, that is not necessarily the case. In
fact, only a part of the cohomology of the complement of the hypersurface
X
Γ
is involved in the period computation of the Feynman integral. As
we have seen earlier in this chapter, ignoring divergences, one is in fact
considering only a certain relative cohomology group, namely
H
n−1
(P
n−1
X
Γ
, Σ
n
(Σ
n
∩ X
Γ
)), (3.93)
where n = #E
int
(Γ) and Σ
n
= ¦t ∈ P
n−1
[
i
t
i
= 0¦ denotes the normal
crossings divisor given by the union of the coordinate hyperplanes. In
fact, if one ignores momentarily the issue of divergences, the evaluation of
the integral
_
σ
n
P
Γ
(p, t)
−n+D/2
Ψ
Γ
(t)
−n+(+1)D/2
ω
n
(3.94)
can be seen as a pairing of a diﬀerential form on the hypersurface comple
ment with a chain σ
n
with boundary ∂σ
n
⊂ Σ
n
contained in the normal
crossings divisor. Thus, one is working with a relative cohomology, and we
have seen that the question of whether the integral above (i.e. the residue of
96 Feynman Motives
the Feynman graph after dropping the divergent Gamma factor in the para
metric Feynman integral) evaluates to a period of a mixed Tate motive can
be addressed in terms of the question of whether the relative cohomology
(3.93) is a realization of a mixed Tate motive
m(P
n−1
X
Γ
, Σ
n
X
Γ
∩ Σ
n
). (3.95)
This could still be the case even though the motive of X
Γ
itself need not
be mixed Tate. However, this observation makes it clear that, if something
like Conjecture 3.12.1 holds, it is due to a very subtle interplay between
the geometry of the graph hypersurface X
Γ
and the locus of intersection
X
Γ
∩Σ
n
. Notice that this locus is also the one that is responsible for the di
vergences of the integral (3.94), so that the situation is further complicated,
when one tries to eliminate divergences, by possibly performing blowups
of this locus. Even without introducing immediately the further diﬃculties
related to eliminating divergences, the main problem remains that as the
graphs increase in combinatorial complexity, it becomes extemely diﬃcult
to control the motivic nature of (3.95). Thus, at present, not only is the
conjecture in its stronger form 3.12.1 still open, but not much is understood
on speciﬁc conditions that would ensure that (3.95) is in fact mixed Tate,
according to the formulation of Question 3.12.2.
Some new work of Francis Brown, which became available in the preprint
[Brown (2009b)] while this book was going to print, identiﬁes speciﬁc fam
ilies of graphs for which (3.95) is shown to give a mixed Tate motive and
for which the corresponding Feynman integrals evaluate to multiple zeta
values.
In fact, a weaker statement than the relative cohomology (3.93) being
mixed Tate may still suﬃce to prove that the residue of the Feynman in
tegral is a period of a mixed Tate motive, since the period only captures
a piece of the relative cohomology (3.93), so that the question on the mo
tivic nature of this relative cohomology provides a suﬃcient condition for
a mixed Tate period, but not a necessary one.
Very nice recent results of [Doryn (2008)] compute the middle coho
mology of the graph hypersurfaces for a larger class of graphs than those
originally considered in the work of [Bloch, Esnault, Kreimer (2006)], con
sisting of so called “zigzag graphs”. This identiﬁes in a signiﬁcant class of
examples a mixed Tate piece of the cohomology.
As we mentioned in relation to the recent result of [Bloch (2008)] on
sums of graphs with ﬁxed number of vertices, one should also take into ac
count that the possible occurrence of nonmixed Tate periods in the Feyn
man amplitudes for more complicated graphs may cancel out in sums over
Feynman integrals and algebraic varieties 97
graphs with ﬁxed external structure or ﬁxed number of vertices and leave
only a mixed Tate contribution.
One should also keep in mind that, in the case of divergent Feynman
integrals, handling divergences in diﬀerent ways might aﬀect the nature of
the resulting periods, after subtraction of inﬁnities.
3.13 From graph hypersurfaces to determinant hypersur
faces
An attempt to provide speciﬁc conditions, as stated in Question 3.12.2
above, that would ensure that certain Feynman integrals evaluate to periods
of mixed Tate motives was developed in [Aluﬃ and Marcolli (2009a)].
One can use the properties of periods, in particular the change of vari
able formula, to recast the computation of a given integral
_
σ
ω computing
a period in a diﬀerent geometric ambient variety, whose motivic nature is
easier to control. As long as no information is lost in the period computa
tion, one can hope to obtain in this way some suﬃcient conditions that will
ensure that the periods associated to Feynman integrals are mixed Tate.
A period
_
σ
ω associated to the data (X, D, ω, σ) of a variety X, a
divisor D, a diﬀerential form ω on X, and an integration domain σ with
boundary ∂σ ⊂ D can be computed equivalently after a change of variables
obtained by mapping it via a morphism f of varieties to another set of data
(X
, D
, ω
, σ
), with ω = f
∗
(ω
) and σ
= f
∗
(σ).
In the case of parametric Feynman integrals one can proceed in the fol
lowing way. The matrix M
Γ
(t) associated to a Feynman graph Γ determines
a linear map of aﬃne spaces
Υ : A
n
→A
2
, Υ(t)
kr
=
i
t
i
η
ik
η
ir
(3.96)
such that the aﬃne graph hypersurface is obtained as the preimage
ˆ
X
Γ
= Υ
−1
(
ˆ
T
)
under this map of the determinant hypersurface,
ˆ
T
= ¦x = (x
ij
) ∈ A
2
[ det(x
ij
) = 0¦.
One can give explicit combinatorial conditions on the graph that ensure
that the map Υ is an embedding. As shown in [Aluﬃ and Marcolli (2009a)],
for any 3edgeconnected graph with at least 3 vertices and no looping edges,
which admit a closed 2cell embedding of face width at least 3, the map Υ
98 Feynman Motives
is injective. To see why this is the case, one can start with the following
observation on the properties of the matrix M
Γ
(t) that deﬁnes the map Υ
of (3.96).
Lemma 3.13.1. The matrix M
Γ
(t) = η
†
Λη deﬁning the map Υ has the
following properties.
• For i ,= j, the corresponding entry is the sum of ±t
k
, where the t
k
correspond to the edges common to the ith and jth loop, and the sign
is +1 if the orientations of the edges both agree or both disagree with
the loop orientations, and −1 otherwise.
• For i = j, the entry is the sum of the variables t
k
corresponding to the
edges in the ith loop (all taken with sign +).
Similarly, for a speciﬁc edge e, with t
e
the corresponding variable, one has
the following.
• The variable t
e
appears in η
†
Λη if and only if e is part of at least one
loop.
• If e belongs to a single loop
i
, then t
e
only appears in the diagonal entry
(i, i), added to the variables corresponding to the other edges forming
the loop
i
.
• If there are two loops
i
,
j
containing e, and not having any other edge
in common, then the ±t
e
appears by itself at the entries (i, j) and (j, i)
in the matrix η
†
Λη.
In the following, we denote by Υ
i
the composition of the map Υ with
the projection to the ith row of the matrix η
†
Λη, viewed as a map of the
variables corresponding only to the edges that belong to the ith loop in
the chosen bases of the ﬁrst homology of the graph Γ.
Lemma 3.13.2. If Υ
i
is injective for i ranging over a set of loops such
that every edge of Γ is part of a loop in that set, then Υ is itself injective.
Proof. Let (t
1
, . . . , t
n
) = (c
1
, . . . , c
n
) be in the kernel of τ. Since each
(i, j) entry in the target matrix is a combination of edges in the ith loop,
the map τ
i
must send to zero the tuple of c
j
’s corresponding to the edges
in the ith loop. Since we are assuming τ
i
to be injective, that tuple is
the zerotuple. Since every edge is in some loop for which τ
i
is injective, it
follows that every c
j
is zero, as needed.
One can then give conditions based on properties of the graph that
Feynman integrals and algebraic varieties 99
ensure that the components Υ
i
are injective. This is done in [Aluﬃ and
Marcolli (2009a)] as follows.
Lemma 3.13.3. The map Υ
i
is injective if the following conditions are
satisﬁed:
• For every edge e of the ith loop, there is another loop having only e in
common with the ith loop, and
• The ith loop has at most one edge not in common with any other loop.
Proof. In this situation, all but at most one edge variable appear by
themselves as an entry of the ith row, and the possible last remaining
variable appears summed together with the other variables. More explicitly,
if t
i
1
, . . . , t
i
v
are the variables corresponding to the edges of a loop
i
, up
to rearranging the entries in the corresponding row of η
†
Λη and neglecting
other entries, the map Υ
i
is given by
(t
i
1
, . . . , t
i
v
) → (t
i
1
+ +t
i
v
, ±t
i
1
, . . . , ±t
i
v
)
if
i
has no edge not in common with any other loop, and
(t
i
1
, . . . , t
i
v
) → (t
i
1
+ +t
i
v
, ±t
i
1
, . . . , ±t
i
v−1
)
if
i
has a single edge t
v
not in common with any other loop. In either case
the map Υ
i
is injective, as claimed.
Every (ﬁnite) graph Γ may be embedded in a compact orientable surface
of ﬁnite genus. The minimum genus of an orientable surface in which Γ may
be embedded is the genus of Γ. Thus, Γ is planar if and only if it may be
embedded in a sphere, if and only if its genus is 0.
Deﬁnition 3.13.4. An embedding of a graph Γ in an orientable surface
S is a 2cell embedding if the complement of Γ in S is homeomorphic to a
union of open 2cells (the faces, or regions determined by the embedding).
An embedding of Γ in S is a closed 2cell embedding if the closure of every
face is a disk.
It is known that an embedding of a connected graph is minimal genus
if and only if it is a 2cell embedding ([Mohar and Thomassen (2001)],
Proposition 3.4.1 and Theorem 3.2.4). We discuss below conditions on the
existence of closed 2cell embeddings, cf. [Mohar and Thomassen (2001)],
¸5.5.
For our purposes, the advantage of having a closed 2cell embedding for
a graph Γ is that the faces of such an embedding determine a choice of loops
100 Feynman Motives
of Γ, by taking the boundaries of the 2cells of the embedding together with
a basis of generators for the homology of the Riemann surface in which the
graph is embedded.
Lemma 3.13.5. A closed 2cell embedding ι : Γ → S of a connected graph
Γ on a surface of (minimal) genus g, together with the choice of a face of
the embedding and a basis for the homology H
1
(S, Z) determine a basis of
H
1
(Γ, Z) given by 2g + f − 1 loops, where f is the number of faces of the
embedding.
Proof. Orient (arbitrarily) the edges of Γ and the faces, and then add
the edges on the boundary of each face with sign determined by the orien
tations. The fact that the closure of each face is a 2disk guarantees that
the boundary is nullhomotopic. This produces a number of loops equal to
the number f of faces. It is clear that these f loops are not independent:
the sum of any f − 1 of them must equal the remaining one, up to sign.
Any f −1 loops, however, will be independent in H
1
(Γ). Indeed, these f −1
loops, together with 2g generators of the homology of S, generate H
1
(Γ).
The homology group H
1
(Γ) has rank 2g + f − 1, as one can see from the
Euler characteristic formula
b
0
(S) −b
1
(S) +b
2
(S) = 2 −2g
= χ(S) = v −e +f = b
0
(Γ) −b
1
(Γ) +f = 1 − +f,
so there will be no other relations.
One refers to the chosen one among the f faces as the “external face”
and the remaining f −1 faces as the “internal faces”.
Thus, given a closed 2cell embedding ι : Γ → S, we can use a basis of
H
1
(Γ, Z) constructed as in Lemma 3.13.5 to compute the map Υ and the
maps Υ
i
of (3.13.1). We then have the following result.
Lemma 3.13.6. Assume that Γ is closed2cell embedded in a surface.
With notation as above, assume that
• any two of the f faces have at most one edge in common.
Then the f −1 maps Υ
i
, deﬁned with respect to a choice of basis for H
1
(Γ)
as in Lemma 3.13.5, are all injective. If further
• every edge of Γ is in the boundary of two of the f faces,
then Υ is injective.
Feynman integrals and algebraic varieties 101
Proof. The injectivity of the f −1 maps Υ
i
follows from Lemma 3.13.3.
If is a loop determined by an internal face, the variables corresponding
to edges in common between and any other internal loop will appear as
(±) individual entries on the row corresponding to . Since has at most
one edge in common with the external region, this accounts for all but at
most one of the edges in . By Lemma 3.13.3, the injectivity of Υ
i
follows.
Finally, as shown in Lemma 3.13.2, the map Υ is injective if every edge is
in one of the f −1 loops and the f −1 maps Υ
i
are injective. The stated
condition guarantees that the edge appears in the loops corresponding to
the faces separated by that edge. At least one of them is internal, so that
every edge is accounted for.
Recall the notions of connectivity of graphs that we discussed in the
ﬁrst chapter in relation to the 1PI condition, Deﬁnition 1.5.1.
We also recall the notion of face width for an embedding of a graph in
a Riemann surface.
Deﬁnition 3.13.7. Let ι : Γ → S be a given embedding of a graph Γ on
a Riemann surface S. The face width fw(Γ, ι) is the largest number k ∈ N
such that every noncontractible simple closed curve in S intersects Γ at
least k times. When S is a sphere, hence ι : Γ → S is a planar embedding,
one sets fw(Γ, ι) = ∞.
For a graph Γ with at least 3 vertices and with no looping edges, the
condition that an embedding ι : Γ → S is a closed 2cell embedding is
equivalent to the properties that Γ is 2vertexconnected and that the em
bedding has face width fw(Γ, ι) ≥ 2; see Proposition 5.5.11 of [Mohar
and Thomassen (2001)]. It is not known whether every 2vertexconnected
graph Γ admits a closed 2cell embedding. The “strong orientable embed
ding conjecture” states that this is the case, namely, that every 2vertex
connected graph Γ admits a closed 2cell embedding in some orientable
surface S, of face width at least two (see [Mohar and Thomassen (2001)],
Conjecture 5.5.16).
Theorem 3.13.8. The following conditions imply the injectivity of Υ.
(1) Let Γ be a graph with at least 3 vertices and with no looping edges,
which is closed2cell embedded in an orientable surface S. Then, if
any two of the faces have at most one edge in common, the map Υ is
injective.
102 Feynman Motives
(2) Let Γ be a 3edgeconnected graph, with at least 3 vertices and no looping
edges, admitting a closed2cell embedding ι : Γ → S with face width
fw(Γ, ι) ≥ 3. Then the maps Υ
i
, Υ are all injective.
Proof. (1) It suﬃces to show that, under these conditions on the graph
Γ, the second condition of Lemma 3.13.6 is automatically satisﬁed, so that
only the ﬁrst condition remains to be checked. That is, we show that every
edge of Γ is in the boundary of two faces. Assume an edge is not in the
boundary of two faces. Then that edge must bound the same face on both
of its sides. The closure of the face is a cell, by assumption. Let γ be a
path from one side of the edge to the other. Since γ splits the cell into two
connected components, it follows that removing the edge splits Γ into two
connected components, hence Γ is not 2edgeconnected. However, the fact
that Γ has at least 3 vertices and no looping edges and it admits a closed
2cell embedding implies that Γ is 2vertexconnected, hence in particular
it is 1PI by Lemma 1.5.5, and this gives a contradiction.
(2) The previous result shows that the second condition stated in
Lemma 3.13.6 is automatically satisﬁed, so the only thing left to check
is that the ﬁrst condition stated in Lemma 3.13.6 holds. Assume that two
faces F
1
, F
2
have more than one edge in common. Since F
1
, F
2
are (path)
connected, there are paths γ
i
in F
i
connecting corresponding sides of the
edges. With suitable care, it can be arranged that γ
1
∪ γ
2
is a closed path
γ meeting Γ in 2 points. Since the embedding has face width ≥ 3, γ must
be nullhomotopic in the surface, and in particular it splits it into two con
nected components. Therefore Γ is split into two connected components by
removing the two edges, hence Γ cannot be 3edgeconnected.
A more geometric description of these combinatorial properties in terms
of wheel neighborhoods at the vertices of the graph is discussed in [Aluﬃ and
Marcolli (2009a)]. Further, it is shown there that the injectivity property
of the map Υ extends from the classes of graphs described in the previous
theorem to other graphs obtained from these by simple combinatorial op
erations such as attaching a looping edge at a vertex of a given graph or
subdividing edges by adding intermediate valence two vertices.
The combinatorial conditions of Theorem 3.13.8 are fairly natural from
a physical viewpoint. In fact, 2edgeconnected is just the usual 1PI condi
tion, while 3edgeconnected or 2PI is the next strengthening of this con
dition (the 2PI eﬀective action is often considered in quantum ﬁeld the
ory), and the face width condition is also the next strengthening of face
width 2, which a well known combinatorial conjecture on graphs [Mohar
Feynman integrals and algebraic varieties 103
and Thomassen (2001)] expects should simply follow for graphs that are 2
vertexconnected. (The latter condition is a bit more than 1PI: for graphs
with at least two vertices and no looping edges it is equivalent to all the
splittings of the graph at vertices also being 1PI, as we showed in Lemma
1.5.5.) The condition that the graph has no looping edges is only a tech
nical device for the proof. In fact, it is then easy to show (see [Aluﬃ and
Marcolli (2009a)]) that adding looping edges does not aﬀect the injectivity
of the map Υ.
When the map Υ is injective, it is possible to rephrase the computation
of the parametric Feynman integral as a period of the complement of the
determinant hypersurface
ˆ
T
⊂ A
2
. Notice also that if the map Υ is
injective then one has a well deﬁned map P
n−1
→P
2
−1
, which is otherwise
not everywhere deﬁned. Motivically, when the map Υ : A
n
→ A
2
is
injective, the complexity of Feynman integrals of the graph Γ as a period is
controlled by the motive m(A
2
ˆ
T
,
ˆ
Σ
Γ
(
ˆ
T
∩
ˆ
Σ
Γ
)), where
ˆ
Σ
Γ
is a normal
crossings divisor in A
2
such that Υ(∂σ
n
) ⊂
ˆ
Σ
Γ
. We recall below how to
construct the divisor
ˆ
Σ
Γ
.
One can in fact rewrite the Feynman integral (as usual up to a divergent
Γfactor) in the form
U(Γ) =
_
Υ(σ
n
)
T
Γ
(x, p)
−n+D/2
ω
Γ
(x)
det(x)
−n+(+1)D/2
,
for a polynomial T
Γ
(x, p) on A
2
that restricts to P
Γ
(t, p), and with ω
Γ
(x)
the image of the volume form. Let then
ˆ
Σ
Γ
be a normal crossings divisor in
A
2
, which contains the boundary of the domain of integration, Υ(∂σ
n
) ⊂
ˆ
Σ
Γ
. The question on the motivic nature of the resulting period can then
be reformulated (again modulo divergences) in this case as the question of
whether the motive
m(A
2
ˆ
T
,
ˆ
Σ
Γ
(
ˆ
Σ
Γ
∩
ˆ
T
)) (3.97)
is mixed Tate.
The advantage of moving the period computation via the map Υ =
Υ
Γ
from the hypersurface complement A
n
ˆ
X
Γ
to the complement of the
determinant hypersurface A
2
ˆ
T
is that, unlike what happens with the
graph hypersurfaces, it is well known that the determinant hypersurface
ˆ
T
is a mixed Tate motive, as we have already seen in detail in Theorems 2.6.2
and 2.6.3 above.
One then sees that, in this approach, the diﬃculty has been moved
from understanding the motivic nature of the hypersurface complement to
104 Feynman Motives
having some control on the other term of the relative cohomology, namely
the normal crossings divisor
ˆ
Σ
Γ
and the way it intersects the determinant
hypersurface.
If the motive of
ˆ
Σ
Γ
(
ˆ
Σ
Γ
∩
ˆ
T
) is mixed Tate, then knowing that A
2
ˆ
T
is always mixed Tate, the fact that mixed Tate motives form a triangulated
subcategory of the triangulated category of mixed motives would show that
the motive (3.97) whose realization is the relative cohomology would also
be mixed Tate.
Thus, Question 3.12.2 can be reformulated, under the combinatorial
conditions given above for the embedding, as a question on when the mo
tive of
ˆ
Σ
Γ
(
ˆ
Σ
Γ
∩
ˆ
T
) is mixed Tate. This requires, ﬁrst of all, a better
description of the divisor
ˆ
Σ
Γ
that contains the boundary ∂(Υ(σ
n
)).
A ﬁrst observation in [Aluﬃ and Marcolli (2009a)] is that one can use
the same normal crossings divisor
ˆ
Σ
,g
for all graphs Γ with a ﬁxed number
of loops and a ﬁxed genus (that is, the minimal genus of an orientable
surface in which the graph can be embedded). This divisor is given by a
union of linear spaces.
Proposition 3.13.9. There exists a normal crossings divisor
ˆ
Σ
,g
⊂ A
2
,
which is a union of N =
_
f
2
_
linear spaces,
ˆ
Σ
,g
:= L
1
∪ ∪ L
N
, (3.98)
such that, for all graphs Γ with loops and genus g closed 2cell embedding,
the preimage under Υ = Υ
Γ
of the union
ˆ
Σ
Γ
of a subset of components of
ˆ
Σ
,g
is the algebraic simplex Σ
n
in A
n
. More explicitly, the components of
the divisor
ˆ
Σ
,g
can be described by the N =
_
f
2
_
equations
_
x
ij
= 0 1 ≤ i < j ≤ f −1
x
i1
+ +x
i,f−1
= 0 1 ≤ i ≤ f −1,
(3.99)
where f = − 2g + 1 is the number of faces of the embedding of the graph
Γ on a surface of genus g.
Proof. The polynomial det M
Γ
(t) does not depend on the choice of ori
entation for the loops of Γ. Thus, we can make the following convenient
choice of these orientations. We have chosen a closed 2cell embedding of Γ
into an orientable surface of genus g. Such an embedding has f faces, with
= 2g+f −1. We can arrange M
Γ
(t) so that the ﬁrst f −1 rows correspond
to the f −1 loops determined by the ‘internal’ faces of the embedding. On
each face, we choose the positive orientation (counterclockwise with respect
to an outgoing normal vector). Then each edgevariable in common between
Feynman integrals and algebraic varieties 105
two faces i, j will appear with a minus sign in the entries (i, j) and (j, i) of
M
Γ
(t). These entries are both in the ( −2g) ( −2g) upperleft minor.
We have in fact seen that the injectivity of an ( −2g) ( −2g) minor
of the matrix M
Γ
suﬃces to control the injectivity of the map Υ, and we
can arrange so that the minor is the upperleft part of the ambient
matrix. Then the hyperplanes in A
n
associated to the coordinates t
i
can be
obtained by pulling back linear spaces along this minor. On the diagonal
of the (f − 1) (f − 1) submatrix we ﬁnd all edges making up each face,
with a positive sign. It follows that the pullbacks of the equations (3.99)
produce a list of all the edge variables, possibly with redundancies. The
components of
ˆ
Σ
,g
that form the divisor
ˆ
Σ
Γ
are selected by eliminating
those components of
ˆ
Σ
,g
that contain the image of the graph hypersurface
(i.e. coming from the zero entries of the matrix M
Γ
(t)).
A second observation of [Aluﬃ and Marcolli (2009a)] is then that, using
inclusionexclusion, it suﬃces to show that arbitrary intersections of the
components L
i
of
ˆ
Σ
,g
have the property that
(∩
i∈I
L
i
)
ˆ
T
(3.100)
is mixed Tate. In fact, this would then imply that also the locus
ˆ
Σ
Γ
(
ˆ
T
∩
ˆ
Σ
Γ
)
is mixed Tate, by repeatedly using inclusion–exclusion.
Notice that the intersection ∩
i∈I
L
i
is a linear subspace of codimension
#I in A
2
. In general, the intersection of a linear subspace with the de
terminant is not mixed Tate. For example, the intersection of a general
A
3
with
ˆ
T
3
is a cone over a genus1 curve. In fact, working projectively,
T
3
is a degree3 hypersurface in P
8
, with singularities in codimension > 1.
Therefore, the intersection with a general P
2
is a nonsingular cubic curve,
therefore a curve of genus = 1. The aﬃne version is a cone over this. Thus,
in order to understand under what conditions the locus (3.100) will be
mixed Tate, we have to understand in what sense the intersections ∩
i∈I
L
i
are special.
The following characterization, which is given in [Aluﬃ and Marcolli
(2009a)], leads to a reformulation of the problem in terms of certain “man
ifolds of frames”.
Lemma 3.13.10. Let E be a ﬁxed dimensional vector space. Every I ⊆
¦1, . . . , N¦ determines a choice of linear subspaces V
1
, . . . , V
of E with the
property that
∩
k∈I
L
k
= ¦(v
1
, . . . , v
) ∈ A
2
[ ∀i, v
i
∈ V
i
¦. (3.101)
106 Feynman Motives
Here, we denote an matrix in A
2
by its rowvectors v
i
∈ E.
Further, dimV
i
≥ i −1. Further still, there exists a basis (e
1
, . . . , e
) of
E such that each space V
i
is the span of a subset (of cardinality ≥ i −1) of
the vectors e
j
.
Proof. Recall (Proposition 3.13.9) that the components L
k
of
ˆ
Σ
,g
consist
of matrices for which either the (i, j) entry x
ij
equals 0, for 1 ≤ i < j ≤
−2g, or
x
i1
+ +x
i,−2g
= 0
for 1 ≤ i ≤ −2g. Thus, each L
k
consists of tuples (v
1
, . . . , v
) for which
exactly one row v
i
belongs to a ﬁxed hyperplane of E, and more precisely
to one of the hyperplanes
x
1
+ +x
−2g
= 0, x
2
= 0, . . . x
−2g
= 0. (3.102)
The statement follows by choosing V
i
to be the intersection of the hy
perplanes corresponding to the L
k
in the ith row, among those listed in
(3.102). Since there are at most − 2g − i + 1 hyperplanes L
k
in the ith
row,
dimV
i
≥ −( −2g −i + 1) = 2g +i −1 ≥ i −1 .
Finally, to obtain the basis (e
1
, . . . , e
) mentioned in the statement, simply
choose the basis dual to the basis (x
1
+ +x
−2g
, x
2
, . . . , x
) of the dual
space to E.
A suﬃcient condition that would ensure that the locus
ˆ
Σ
Γ
(
ˆ
Σ
Γ
∩
ˆ
T
)
is mixed Tate is then described in terms of manifolds of frames.
Deﬁnition 3.13.11. For a given ambient space V = A
and an assigned
collection of linear subspaces V
i
, i = 1, . . . , , the manifold of frames
F(V
1
, . . . , V
) is deﬁned as the locus
F(V
1
, . . . , V
) := ¦(v
1
, . . . , v
) ∈ A
2
[ v
k
∈ V
k
¦ ∩ (A
2
ˆ
T
). (3.103)
More generally, F(V
1
, . . . , V
r
) ⊂ V
1
V
r
denotes the locus of rtuples
of linearly independent vectors in a given vector space V , where each v
i
is
constrained to belong to the given subspace V
i
.
The previous characterization of the locus (3.100), together with the use
of inclusion–exclusion arguments, shows that if, for a ﬁxed , the manifold
of frames is mixed Tate for all choices of the subspaces V
i
, this suﬃces to
guarantee that the motive
m(A
2
ˆ
T
,
ˆ
Σ
,g
(
ˆ
Σ
,g
∩
ˆ
T
)) (3.104)
Feynman integrals and algebraic varieties 107
is also mixed Tate. When this is the case, the result then implies that,
modulo divergences, the residues of Feynman graphs Γ with loops and
genus g, for which the injectivity condition on Υ holds, are all periods of
mixed Tate motives.
In the case of two and three loops, one can verify explicitly that the
manifolds of frames F(V
1
, V
2
) and F(V
1
, V
2
, V
3
) are mixed Tate. This is
done in [Aluﬃ and Marcolli (2009a)] by exhibiting an explicit stratiﬁca
tion, from which, as in the case of the determinant hypersurface discussed
above, one can show that the motive deﬁnes an object in the category of
mixed Tate motives as a subtriangulated category of the category of mixed
motives. One can also compute explicitly the corresponding classes in the
Grothendieck ring, as a function of the Lefschetz motive L. They are of the
following form.
Proposition 3.13.12. For given subspaces V
1
, V
2
, the manifold of frames
F(V
1
, V
2
) is a mixed Tate motive, whose class in the Grothendieck ring is
[F(V
1
, V
2
)] = L
d
1
+d
2
−L
d
1
−L
d
2
−L
d
12
+1
+L
d
12
+L,
with d
i
= dim(V
i
) and d
ij
= dim(V
i
∩ V
j
).
Proof. We want to parameterize all pairs (v
1
, v
2
) of vectors such that
v
1
∈ V
1
, v
2
∈ V
2
, and dim¸v
1
, v
2
) = 2. This locus can be decomposed into
the two (possibly empty) pieces
(1) v
1
∈ V
1
(V
1
∩ V
2
), and v
2
∈ V
2
¦0¦;
(2) v
1
∈ (V
1
∩ V
2
) ¦0¦, and v
2
∈ V
2
¸v
1
).
This exhausts all the possible ways of obtaining linearly independent vectors
with the ﬁrst one in V
1
and the second one in V
2
and the manifold of frames
F(V
1
, V
2
) is the union of these two loci. The ﬁrst case describes the locus
(V
1
(V
1
∩ V
2
)) (V
2
¦0¦)
which is clearly mixed Tate, being obtained by taking products and com
plements of aﬃne spaces. Its class in the Grothendieck ring is of the form
(L
d
1
−L
d
12
)(L
d
2
−1).
The locus deﬁned by the second case can be described equivalently by
the following procedure. Consider the projective space P(V
1
∩ V
2
), and
the trivial bundles 1
12
⊆ 1
2
with ﬁber V
1
∩ V
2
⊆ V
2
. The tautological line
bundle O
12
(−1) over P(V
1
∩V
2
) sits inside 1
12
, hence inside 1
2
. The desired
pairs of vectors (v
1
, v
2
) are obtained by choosing a point p ∈ P(V
1
∩ V
2
), a
108 Feynman Motives
vector v
1
,= 0 in the ﬁber of O
12
(−1) over p, and a vector v
2
in the ﬁber of
1
2
O
12
(−1) over p. This again deﬁnes a mixed Tate locus, using the same
two properties of the existence of distinguished triangles in the category of
mixed motives associated to closed embeddings and homotopy invariance,
as in Theorems 2.6.2 and 2.6.3. The homotopy invariance property, which
is formulated for products, extends in fact to the locally trivial case of
vector bundles or projective bundles, see [Voevodsky (2000)]. The class in
the Grothendieck ring, in this case, is then given by the expression
(L
d
12
−1)(L
d
2
−L).
Thus, the class of F(V
1
, V
2
) is the sum of these two classes,
[F(V
1
, V
2
)] = (L
d
1
+d
2
−L
d
1
−L
d
2
+d
12
+L
d
12
) +(L
d
2
+d
12
−L
d
12
+1
−L
d
2
+L)
= L
d
1
+d
2
−L
d
1
−L
d
2
−L
d
12
+1
+L
d
12
+L.
The case of three subspaces already requires a more delicate analysis
of the contribution of the various strata, as in ¸6.2 of [Aluﬃ and Marcolli
(2009a)], which we reproduce here below.
Theorem 3.13.13. Let V
1
, V
2
, V
3
be assigned subspaces of a vector space
V . Then the manifold of frames F(V
1
, V
2
, V
3
) is a mixed Tate motive, with
class in the Grothendieck ring given by
[F(V
1
, V
2
, V
3
)] = (L
d
1
−1)(L
d
2
−1)(L
d
3
−1)
−(L −1)((L
d
1
−L)(L
d
23
−1) + (L
d
2
−L)(L
d
13
−1) + (L
d
3
−L)(L
d
12
−1)
+(L −1)
2
(L
d
1
+d
2
+d
3
−D
−L
d
123
+1
) + (L −1)
3
,
where d
i
= dim(V
1
), d
ij
= dim(V
i
∩ V
j
), d
ijk
= dim(V
i
∩ V
j
∩ V
k
), and
D = D
ijk
= dim(V
i
+V
j
+V
k
).
Proof. This time one can proceed in the following way to obtain a strati
ﬁcation. One can look for a stratiﬁcation ¦S
α
¦ of V
3
which is ﬁner than the
one induced by the subspace arrangement V
1
∩ V
3
, V
2
∩ V
3
and such that,
for v
3
in S
α
, the class F
α
:= [F(π(V
1
), π(V
2
))], with π : V → V
:= V/¸v
3
)
the projection, depends only on α and not on the chosen vector v
3
∈ S
α
.
In other words, we want to construct a stratiﬁcation of V
3
such that the
dimensions of the spaces π(V
1
), π(V
2
) and π(V
1
∩ V
2
) are constant along
the strata. The following ﬁve loci deﬁne such a stratiﬁcation of V
3
¦0¦:
(1) S
123
:= (V
1
∩ V
2
∩ V
3
) ¦0¦;
Feynman integrals and algebraic varieties 109
(2) S
13
:= (V
1
∩ V
3
) (V
1
∩ V
2
∩ V
3
);
(3) S
23
:= (V
2
∩ V
3
) (V
1
∩ V
2
∩ V
3
);
(4) S
(12)3
:= ((V
1
+V
2
) ∩ V
3
) ((V
1
∪ V
2
) ∩ V
3
);
(5) S
3
:= V
3
((V
1
+V
2
) ∩ V
3
).
The dimensions are indeed constant along the strata and given by
dimπ(V
1
) dimπ(V
2
) dim(π(V
1
) ∩ π(V
2
))
S
123
d
1
−1 d
2
−1 d
12
−1
S
13
d
1
−1 d
2
d
12
S
23
d
1
d
2
−1 d
12
S
(12)3
d
1
d
2
d
12
+ 1
S
3
d
1
d
2
d
12
This information can then be translated into explicit classes [F
α
] in the
Grothendieck ring associated to the strata S
α
so that the class of the frame
manifold is of the form
[F(V
1
, V
2
, V
3
)] =
α
L
s
α
[F
α
][S
α
], (3.105)
where the number s
α
is the number of subspaces V
i
, i = 1, 2 containing
the vector v
3
∈ S
α
. This is also independent of the choice of v
3
and only
dependent on α by the properties of the stratiﬁcation. For the table of
dimensions obtained above the corresponding classes [F
α
] are as follows.
[F
α
]
S
123
L
d
1
+d
2
−2
−L
d
1
−1
−L
d
2
−1
−L
d
12
+L
d
12
−1
+L
S
13
L
d
1
+d
2
−1
−L
d
1
−1
−L
d
2
−L
d
12
+1
+L
d
12
+L
S
23
L
d
1
+d
2
−1
−L
d
1
−L
d
2
−1
−L
d
12
+1
+L
d
12
+L
S
(12)3
L
d
1
+d
2
−L
d
1
−L
d
2
−L
d
12
+2
+L
d
12
+1
+L
S
3
L
d
1
+d
2
−L
d
1
−L
d
2
−L
d
12
+1
+L
d
12
+L
One then obtains the following values for the class of the stratum [S
α
]
and the number s
α
.
[S
α
] s
α
S
123
L
d
123
−1 2
S
13
L
d
13
−L
d
123
1
S
23
L
d
23
−L
d
123
1
S
(12)3
L
d
1
+d
2
+d
3
−D−d
12
−L
d
13
−L
d
23
+L
d
123
0
S
3
L
d
3
−L
d
1
+d
2
+d
3
−D−d
12
0
110 Feynman Motives
Using the formula (3.105) then gives the stated result. Although the
stratiﬁcation is used there only to compute the class in the Grothendieck
group, an argument similar to the one of Theorems 2.6.2 and 2.6.3 can be
used to show that the same stratiﬁcations used to compute [F(V
1
, V
2
)] and
[F(V
1
, V
2
, V
3
)] also show that F(V
1
, V
2
) and F(V
1
, V
2
, V
3
) deﬁne objects in
the triangulated category of mixed Tate motives.
One can use these explicit computations of classes of manifolds of frames
for two and three subspaces to obtain speciﬁc information about the motives
m(A
2
ˆ
T
,
ˆ
Σ
Γ
(
ˆ
Σ
Γ
∩
ˆ
T
))
for speciﬁc Feynman graphs with up to three loops. For example, we report
here brieﬂy the case of the wheelwiththreespokes graph (the 1skeleton
of a tetrahedron) described in [Aluﬃ and Marcolli (2009a)].
5
t
4
t
6
t
2
t
1
t
3
1
2
3
t
This graph has matrix M
Γ
(t) given by
_
_
t
1
+t
2
+t
5
−t
1
−t
2
−t
1
t
1
+t
3
+t
4
−t
3
−t
2
−t
3
t
2
+t
3
+t
6
_
_
.
Labeling entries of the matrix as x
ij
, we can obtain t
1
, . . . , t
6
as pullbacks
of the following:
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
t
1
= −x
12
t
2
= −x
13
t
3
= −x
23
t
4
= x
21
+x
22
+x
23
t
5
= x
11
+x
12
+x
13
t
6
= x
31
+x
32
+x
33
Thus, we can consider as
ˆ
Σ
Γ
the normal crossings divisor deﬁned by the
equation
x
12
x
13
x
23
(x
11
+x
12
+x
13
)(x
21
+x
22
+x
23
)(x
31
+x
32
+x
33
) = 0.
Feynman integrals and algebraic varieties 111
In order to compute the class in the Grothendieck ring of the intersection
ˆ
Σ
Γ
∩ (A
9
ˆ
T
3
), one can use inclusionexclusion and compute the classes
for all the intersections of subsets of components of the divisor
ˆ
Σ
Γ
. This
divisor has 6 components, hence there are 2
6
= 64 such intersections. Each
of them determines a choice of three subspaces V
1
, V
2
, V
3
corresponding
to the linearly independent vectors given by the rows of the matrix (x
ij
)
in A
9
. All of the corresponding classes [F(V
1
, V
2
, V
3
)], for each of the 64
possibilities, were computed explicitly in [Aluﬃ and Marcolli (2009a)]. We
do not reproduce them all here, but we only give the ﬁnal result, which
shows that the resulting class is then of the form
[
ˆ
Σ
Γ
(
ˆ
T
3
∩
ˆ
Σ
Γ
)] = L(6L
4
−3L
3
+ 2L
2
+ 2L −1)(L −1)
3
.
By arguing as in Theorem 2.6.2, and using the same stratiﬁcation of the
complement A
9
ˆ
T
3
induced by the divisor
ˆ
Σ
Γ
as in the computation of
the class above, one can improve the result from a statement about the
class in the Grothendieck ring being a polynomial function of the Lefschetz
motive L to one about the motive m(
ˆ
Σ
Γ
(
ˆ
T
3
∩
ˆ
Σ
Γ
)) itself being an object
in the triangulated subcategory T/T
Q
of mixed Tate motives inside the
triangulated category T/
Q
of mixed motives. Then this fact, together with
the fact that the hypersurface complement itself is a mixed Tate motive,
and the distinguished triangle corresponding to the long exact cohomology
sequence, suﬃce to show that the mixed motive
m(A
9
ˆ
T
3
,
ˆ
Σ
Γ
(
ˆ
Σ
Γ
∩
ˆ
T
3
))
is mixed Tate for the case of the wheel with three spokes. The result for
other graphs with three loops can be derived from the same analysis used
for the wheel with three spokes, by restricting only to certain components
of the divisor, as explained in [Aluﬃ and Marcolli (2009a)].
For the cases of manifolds of frames F(V
1
, . . . , V
r
) with more than three
subspaces, it would seem at ﬁrst that one should be able to establish an
inductive argument that would take care of the cases of more subspaces,
but the combinatorics of the possible subspace arrangements quickly be
comes very diﬃcult to control. In fact, in general it seems unlikely that
the frame manifolds F(V
1
, . . . , V
) will continue to be mixed Tate for large
and for arbitrarily complex subspace arrangements. The situation may
in fact be somewhat similar to that of “Murphy’s law in algebraic geome
try” described in [Vakil (2006)], where a suﬃciently general case may be as
bad as possible, while speciﬁc cases one can explicitly construct are much
better behaved. The connection can probably be made more precise, given
112 Feynman Motives
that both the result of [Belkale and Brosnan (2003a)] on the general mo
tivic properties of the graph hypersurfaces X
Γ
and Murphy’s law result of
[Vakil (2006)] are based on the same universality result for matroid repre
sentations of [Mn¨ev (1988)]. Another observation from the point of view
of matroids is that, while the result of [Belkale and Brosnan (2003a)] is
nonconstructive, in the sense that it shows that the graph hypersurfaces
generate the Grothendieck ring of motives but it does not provide an ex
plicit construction of matroids on which to test the mixed Tate property, it
may be that the loci
ˆ
Σ
Γ
∩(A
2
ˆ
T
) considered here may provide a possible
way to make that general result more explicit, by constructing explicit ma
troids, along the lines of [Gelfand and Serganova (1987)]. A reformulation
of the problem of understanding when frame manifolds are mixed Tate is
given in [Aluﬃ and Marcolli (2009a)] in terms of intersections of unions of
Schubert cells in ﬂag varieties. This version of the problem suggests a pos
sible connection to Kazhdan–Lusztig theory [Kazhdan and Lusztig (1980)].
3.14 Handling divergences
All the considerations above on the motivic nature of the relative coho
mology involved in the computation of the parametric Feynman integral,
either in the hypersurface complement P
n−1
X
Γ
or in the complement of
the determinant hypersurface P
2
−1
T
, assume that the integral is con
vergent and therefore directly deﬁnes a period. In other words, the issue
of divergences is not dealt with in this approach. However, one knows very
well that most Feynman integrals are divergent, even when in the para
metric form one removes the divergent Gamma factor and only looks at
the residue, and that some regularization and renormalization procedure is
needed to handle these divergences.
In terms of the geometry, after removing the divergent Gamma factor,
the source of other possible divergences in the parametric Feynman integral
is the locus of intersection of the graph hypersurface X
Γ
with the domain
of integration σ
n
.
Notice that the poles of the integrand that fall inside the integration
domain σ
n
occur necessarily along the boundary ∂σ
n
, since in the interior
the graph polynomial Ψ
Γ
takes only strictly positive real values.
Thus, one needs to modify the integrals suitably in such a way as to
eliminate, by a regularization procedure, the intersections X
Γ
∩∂σ
n
, or (to
Feynman integrals and algebraic varieties 113
work in algebrogeometric terms) the intersections X
Γ
∩Σ
n
which contains
the former.
There are diﬀerent possible ways to achieve such a regularization pro
cedure. We mention here three possible approaches. We discuss the third
one in more detail in the following chapter.
One method was developed in [Belkale and Brosnan (2003b)] in the log
arithmically divergent case where n = D/2, that is, when the polynomial
P
Γ
(t, p) is not present and only the denominator Ψ
Γ
(t)
D/2
appears in the
parametric Feynman integral. As we have already mentioned, using dimen
sional regularization, one can, in this case, rewrite the Feynman integral in
the form of a local Igusa Lfunction
I(s) =
_
σ
f(t)
s
ω,
for f = Ψ
Γ
. They prove that this Lfunction has a Laurent series expansion
where all the coeﬃcients are periods. In this setting, the issue of eliminat
ing divergences becomes similar to the techniques used, for instance, in
the context of log canonical thresholds. The result was more recently ex
tended algorithmically to the nonlogdivergent case [Bogner and Weinzierl
(2007a)], [Bogner and Weinzierl (2007b)].
Another method, used in [Bloch, Esnault, Kreimer (2006)], consists of
eliminating the divergences by separating Σ
n
and X
Γ
performing a series of
blowups. Similarly, working in the setting of the determinant hypersurface
complement A
2
ˆ
T
one can perform blowups to separate the locus of
intersection of the determinant hypersurface
ˆ
T
with the divisor
ˆ
Σ
,g
. To
make sure that this operation does not change the motivic properties of the
resulting period integrals, one needs to ensure that blowups along
ˆ
Σ
,g
∩
ˆ
T
maintain the mixed Tate properties if we know that (3.104) is mixed Tate.
For blowups performed over a smooth locus one has projective bundles,
for which one knows that the mixed Tate property is maintained, since
one has an explicit formula for the motive of a projective bundle in the
Voevodsky category, which is a sum of Tate twisted copies of the motive of
the base, but when the locus is singular, as is typically the case here, then
the required analysis is more delicate.
Yet another method was proposed in [Marcolli (2008)], based on defor
mations instead of resolutions. By considering the graph hypersurface X
Γ
as the special ﬁber X
0
of a family X
s
of varieties deﬁned by the level sets
f
−1
(s), for f = Ψ
Γ
: A
n
→A
1
, one can form a tubular neighborhood
D
(X) = ∪
s∈∆
∗
X
s
,
114 Feynman Motives
for ∆
∗
a punctured disk of radius , and a circle bundle π
: ∂D
(X) → X
.
One can then regularize the Feynman integral by integrating “around the
singularities” in the ﬁber π
−1
(σ ∩ X
). The regularized integral has a
Laurent series expansion in the parameter .
Motivically, one should take into account the fact that the presence of
singularities in the hypersurface X
Γ
increases the likelihood that the part
of the cohomology involved in the period computation can be mixed Tate.
In fact, if these were smooth hypersurfaces, they would typically not be
mixed Tate even for a small number of loops, while it is precisely because
of the fact that they are highly singular that the X
Γ
continue to be mixed
Tate for a fairly large number of loops and even though one knows that
eventually one will run into graphs for which X
Γ
is no longer mixed Tate,
the fact that they continue to be singular in low codimension stil makes
it possible to have a signiﬁcant part of the cohomology that will still be a
realization of a mixed Tate motive.
The use of deformations X
to regularize the integral, which we describe
more in detail in the following chapter, is natural from the point of view of
singularity theory, where one works with Milnor ﬁbers of singularites. The
singularities of the hypersurfaces X
Γ
are nonisolated, but many techniques
of singularity theory are designed to cover also this more general case.
From this viewpoint, it is not clear though how to perform the regu
larization and subtraction of divergences in a way that does not alter the
motivic properties. In fact, while the special singular ﬁber may be mixed
Tate, the generic ﬁber in a family used as deformation can be a general
hypersurface that will not be motivically mixed Tate any longer. However,
there are notions such as a motivic tubular neighborhood [Levine (2005)]
that may be useful in this context, as well as the theory of limiting mixed
Hodge structures for a degeneration of a family of algebraic varieties. It is
not yet clear how to use this type of methods to obtain motivic information
on Feynman integrals after subtraction of divergences.
In general, as we discuss more at length in Chapter 4.2, a regularization
procedure for Feynman integrals replaces a divergent integral with a func
tion of some regularization parameters (such as the complexiﬁed dimension
of DimReg, or the deformation parameter in the example here above) in
which the resulting function has a Laurent series expansion around the pole
that corresponds to the divergent integral originally considered. One then
uses a procedure of extraction of ﬁnite values to eliminate the polar parts
of these Laurent series in a way that is consistent over graphs, that is, a
renormalization procedure.
Feynman integrals and algebraic varieties 115
3.15 Motivic zeta functions and motivic Feynman rules
As we mentioned brieﬂy in describing the Grothendieck ring as a universal
Euler characteristic, the counting of points of a variety over ﬁnite ﬁelds
is an example of an additive invariant that, as such, factors through the
Grothendieck ring. The behavior of the number of points over ﬁnite ﬁelds
is one of the properties of a variety that reveal its motivic nature. For
example, if a variety, say deﬁned over Q, is motivically mixed Tate, then
the number of points of the reductions mod p is polynomial in p.
The information on the number of points over ﬁnite ﬁelds is conveniently
packaged in the form of a zeta function
Z
X
(t) = exp
_
n
#X(F
q
n)
n
t
n
_
(3.106)
Notice that this can be written equivalently in terms of symmetric products
as
Z
X
(t) =
n≥0
#s
n
(X)(F
q
) t
n
,
where s
n
(X) denotes the nth symmetric power of X.
The fact that the counting of the number of points behaves like an Euler
characteristic, hence it descends from the Grothendieck ring, suggested that
the zeta function itself may be lifted at the motivic level. This was done in
[Kapranov (2000)], where the motivic zeta function is deﬁned as
Z
X
(t) =
n≥0
[s
n
(X)] t
n
(3.107)
where [s
n
(X)] are the classes in the Grothendieck ring and the zeta function
can be seen as an element of K
0
(1
K
)[[t]]. For example, one has
Z
P
1(t) = (1 −t)
−1
(1 −Lt)
−1
.
Kapranov proved that, when X is the motive of a curve, then the zeta
function is a rational function, in the sense that, given a motivic measure
µ : K
0
(/) → ¹, the zeta function Z
X,µ
(T) ∈ ¹[[T]] is a rational function
of T. Later, it was proved in [Larsen and Lunts (2003)] that in general this
is not true in the case of algebraic surfaces.
One recovers the zeta function of the variety, for a ﬁnite ﬁeld K = F
q
, by
applying a “counting of points” homomorphism K
0
(1
K
) →Z to the motivic
zeta function. One obtains other kinds of zeta functions by applying other
“motivic measures” µ : K
0
(1
K
) → ¹, which give
Z
X,R
(t) =
n≥0
µ(s
n
(X)) t
n
∈ ¹[[t]].
116 Feynman Motives
It is well known that the motivic zeta function contains all the informa
tion on the behavior of the number of points of reductions mod p, hence
potentially the information on the mixed Tate nature of a variety. A recent
survey of these ideas is given in [Andr´e (2009)], where it is also observed
that the motivic zeta function should play a useful role in the case of mo
tives associated to Feynman graphs.
There are other reasons why one can expect that, indeed, Kapranov’s
motivic zeta functions may be especially suitable tools to investigate mo
tivic properties of Feynman integrals. One is, for example, the observation
made at the end of [Marcolli (2008)] which suggests that the function Ψ
s
Γ
,
for s a complex variable, that appears in the dimensional regularization of
parametric Feynman integrals, regarded as a zeta function
Z
Γ
(T) :=
n≥0
log
n
Ψ
Γ
n!
T
n
= Ψ
T
Γ
,
may be related to a motivic zeta function
Z
Log,Γ
(T) :=
n≥0
s
n
(Log
Γ
) T
n
,
where the motive Log
Γ
is the pullback of the logarithmic motive (2.39), via
the map Ψ
Γ
.
Another reason is the formulation of parametric Feynman integrals as
local Igusa Lfunctions given in [Belkale and Brosnan (2003b)]. There is a
motivic Igusa Lfunction constructed by [Denef and Loeser (1998)], which
may provide the right tool for a motivic formulation of the dimensionally
regularized parametric Feynman integrals.
In quantum ﬁeld theory it is customary to consider the full partition
function of the theory, arranged in an asymptotic series by loop number,
instead of looking only at the contribution of individual Feynman graphs.
Besides the loop number = b
1
(Γ), other suitable gradings δ(Γ) are given
by the number n = #E
int
(Γ) of internal edges, or by #E
int
(Γ) − b
1
(Γ) =
#V (Γ) − b
0
(Γ), the number of vertices minus the number of connected
components. As shown in [Connes and Marcolli (2008)] p.77, these all
deﬁne gradings on the Connes–Kreimer Hopf algebra of Feynman graphs.
When one considers motivic Feynman rules, these partition functions
appear to be interesting analogs of the motivic zeta functions considered
in [Kapranov (2000)], [Larsen and Lunts (2003)]. For instance, one can
Feynman integrals and algebraic varieties 117
consider a partition function given by the formal series
Z(t) =
N≥0
δ(Γ)=N
U(Γ)
#Aut(Γ)
t
N
, (3.108)
where δ(Γ) is any one of the gradings described above and where U(Γ) =
[A
n
ˆ
X
Γ
] ∈ K
0
(1
K
). Given a motivic measure µ : K
0
(1
K
) → ¹, this gives
a zeta function with values in ¹[[t]] of the form
Z
R
(t) =
N≥0
δ(Γ)=N
µ(U(Γ))
#Aut(Γ)
t
N
.
In the case of the sums over graphs
S
N
=
#V (Γ)=N
[X
Γ
]
N!
#Aut(Γ)
considered in [Bloch (2008)], one ﬁnds that cancellations between the classes
occur in the Grothendieck ring K
0
(1
K
) and the resulting class S
N
is always
in the Tate subring Z[L] even if the individual terms [X
Γ
] may be nonmixed
Tate. Similarly, it is possible that interesting cancellations of a similar
nature may occur in suitable evaluations of the zeta functions (3.108).
Chapter 4
Feynman integrals and Gelfand–Leray
forms
We focus in this chapter, following [Marcolli (2008)], on the similarities
between the parametric form of the Feynman integrals and the oscillatory
integrals used in singularity theory to describe integrals of holomorphic
forms over vanishing cycles of a singularity and to relate these to mixed
Hodge structures, see [Arnold, Goryunov, Lyashko, Vasilev (1998)] and
Vol.II of [Arnold, GuseinZade, Varchenko (1988)]. In particular, we show
that dimensionally regularized parametric Feynman integrals can be related
to the Mellin transform of a Gelfand–Leray form, whose Fourier transform
is an oscillatory integral of the sort considered in the context of singularity
theory. We show that one can deﬁne in this way a regularization procedure
for the parametric Feynman integrals, based on the use of Leray cocycles,
aimed at eliminating the divergences coming from the intersections X
Γ
∩σ
n
,
as mentioned in §3.14 above.
4.1 Oscillatory integrals
An oscillatory integral is an expression of the form
I(α) =
R
n
e
iαf(x)
φ(x) dx
1
· · · dx
n
, (4.1)
where f : R
n
→ R and φ : R
n
→ R are smooth functions and α ∈ R
∗
+
is a
real parameter. It is well known that, if the phase f(x) is an analytic func
tion in a neighborhood of a critical point x
0
, then (4.1) has an asymptotic
development for α → ∞ given by a series
I(α) ∼ e
iαf(x
0
)
u
n−1
k=0
a
k,u
(φ) α
u
(log α)
k
, (4.2)
119
120 Feynman Motives
where u runs over a ﬁnite set of arithmetic progressions of negative rational
numbers depending only on the phase f(x), and the a
k,u
are distributions
supported on the critical points of the phase, cf. §2.6.1, Vol.II of [Arnold,
GuseinZade, Varchenko (1988)].
It is also well known that the integral (4.1) can be reformulated in
terms of onedimensional integrals using the Gelfand–Leray forms, deﬁned
as follows.
Deﬁnition 4.1.1. Let X
t
be the level hypersurfaces X
t
= {x f(x) = t}.
Then the Gelfand–Leray form ω
f
(x, t) of f is the unique (n − 1)form on
the level hypersurface X
t
with the property that
df ∧ ω
f
(x, t) = dx
1
∧ · · · ∧ dx
n
. (4.3)
Notice that there is an ambiguity given by diﬀerent possible choices of
an (n − 1)form satisfying (4.3), but their restrictions to X
t
all agree so
that the Gelfand–Leray form on X
t
is well deﬁned.
More generally, given a form α, we say that α admits a Gelfand–Leray
form with respect to f if there is a diﬀerential form, which we denote by
α/df, satisfying
df ∧
α
df
= α. (4.4)
As above, the restriction of the form (α/df)(x, t) to X
t
is uniquely deﬁned.
The form ω
f
(x, t) above is the Gelfand–Leray form of the volume form
α = dx
1
∧ · · · ∧ dx
n
, hence one writes it with the notation
ω
f
(x, t) =
dx
1
∧ · · · ∧ dx
n
df
. (4.5)
The Gelfand–Leray form is given by the Poincar´e residue
α
df
= Res
=0
α
f −
. (4.6)
The Gelfand–Leray function is the associated function
J(t) :=
X
t
φ(x)ω
f
(x, t). (4.7)
The oscillatory integral, expressed in terms of Gelfand–Leray forms, is
then written as
I(α) =
R
e
iαt
X
t
(R)
φ(x)ω
f
(x, t)
dt, (4.8)
Feynman integrals and Gelfand–Leray forms 121
where X
t
(R) ⊂ R
n
is the level set X
t
(R) = {x ∈ R
n
: f(x) = t} and
ω
f
(x, t) is the Gelfand–Leray form of f. For more details, see §2.6 and
§2.7, Vol.II [Arnold, GuseinZade, Varchenko (1988)].
Notice that, up to throwing away a set of measure zero, we can assume
here that the integration is over the values t ∈ R such that the level set X
t
is a smooth hypersurface.
4.2 Leray regularization of Feynman integrals
We consider here the case of complex hypersurfaces X ⊂ A
n
= C
n
, with
deﬁning polynomial equation f = 0 (where f will be the graph polynomial
Ψ
Γ
) and the hypersurface complement D(f) ⊂ A
n
, such that the restriction
of f to the interior of the domain of integration σ
n
⊂ A
n
takes strictly
positive real values.
Deﬁnition 4.2.1. The Leray coboundary L(σ) of a kchain σ in X is a
(k +1)chain in D(f) obtained by considering a tubular neighborhood of X
in A
n
, in the following way. If X is a smooth hypersurface, the boundary
of its tubular neighborhood is a circle bundle over X. One considers the
preimage of σ under the projection map as a chain in D(f).
The Leray coboundary L(σ) is a cycle if σ is a cycle, and if one changes
σ by a boundary then L(σ) also changes by a boundary.
Lemma 4.2.2. Let σ
be a kchain in X
= {t ∈ A
n
f(t) = } and let
L(σ
) be its Leray coboundary in D(f − ). Then, for a form α ∈ Ω
k
that
admits a Gelfand–Leray form, one has
1
2πi
L(σ())
df ∧
α
f −
=
σ()
α, (4.9)
where
d
d
σ()
α =
σ()
dα
df
−
∂σ()
α
df
. (4.10)
Proof. First let us show that if α has a Gelfand–Leray form then dα also
does. We have a form α/df such that
df ∧
α
df
= α.
Its diﬀerential gives
dα = d
df ∧
α
df
= −df ∧ d
α
df
.
122 Feynman Motives
Thus, the form
dα
df
= −d
α
df
is a Gelfand–Leray form for dα.
Then we proceed to prove the ﬁrst statement. One can write
1
2πi
L(σ())
df ∧
α
f −
=
1
2πi
γ
σ(s)
α
ds
s −
,
where γ
∼
= S
1
is the boundary of a small disk centered at ∈ C. This can
then be written as
=
1
2πi
γ
σ()
α
ds
s −
+
1
2πi
γ
σ(s)
α
ds
s −
−
1
2πi
γ
σ()
α
ds
s −
.
The last term can be made arbitrarily small, so one gets (4.9). To obtain
(4.10) notice that
1
2πi
d
d
L(σ())
df ∧
α
f −
=
1
2πi
L(σ())
df ∧
α
(f −)
2
.
One then uses
d
α
f −
=
dα
f −
−
α
(f −)
2
to rewrite the above as
1
2πi
L(σ())
dα
f −
−
L(σ())
d
α
f −
=
1
2πi
L(σ())
df ∧
dα
df
f −
−
1
2πi
L(∂σ())
α
f −
=
1
2πi
L(σ())
df ∧
dα
df
f −
−
1
2πi
L(∂σ())
df ∧
α
df
f −
,
where dα/df is a Gelfand–Leray form such that
df ∧
dα
df
= dα,
and α/df is a Gelfand–Leray form with the property that
df ∧
α
df
= α.
This then gives by (4.9)
d
d
σ()
α =
σ()
dα
df
−
∂σ()
α
df
.
This completes the proof.
Feynman integrals and Gelfand–Leray forms 123
The formulation given in Proposition 3.4.3 of the parametric Feynman
integrals suggests a regularization procedure based on Leray coboundaries,
which has the eﬀect of replacing a divergent integral with a meromorphic
function of a regularization parameter , similarly to what happens in the
case of DimReg. We know that the divergences of the parametric Feynman
integral appear at the intersections X
Γ
∩σ
n
, so we can concentrate only on
the part of the integral that is supported near this locus.
As mentioned brieﬂy in §3.14 above, one considers a neighborhood
D
(X) of a hypersurface X in a projective space P
n−1
, given by the level
sets
D
(X) = ∪
s∈∆
∗
X
s
, (4.11)
where X
s
= {tf(t) = s} and ∆
∗
⊂ C
∗
is a small punctured disk of radius
> 0. The boundary ∂D
(X) is given by
∂D
(X) = ∪
s∈∂∆
∗
X
s
. (4.12)
It is a circle bundle over the generic ﬁber X
, with projection
π
: ∂D
(X) → X
.
Given a domain of integration σ, possibly with boundary, we consider
the intersection σ∩D
(X). This contains the locus σ∩X which is of interest
to us in terms of divergences in the Feynman integral. We let L
(σ) denote
the set
L
(σ) = π
−1
(σ ∩ X
). (4.13)
This satisﬁes L
(∂σ) = ∂L
(σ).
We consider forms
π
∗
(η
m
) =
∆(ω)
f
m
, (4.14)
as in Proposition 3.4.3, where X = {f = 0} with order of vanishing m =
−n+D(+1)/2, and ω = P
Γ
(t, p)
−n+D/2
ω
n
, with ω
n
the standard volume
form. We can then regularize the Feynman integral, written in the form of
Proposition 3.4.3, in the following way.
Deﬁnition 4.2.3. The Leray regularized Feynman integral is obtained
from (3.56) by replacing the part
∂σ
n
∩D
(X
Γ
)
π
∗
(η
m
) +
σ
n
∩D
(X
Γ
)
df ∧
π
∗
(η
m
)
f
(4.15)
of (3.56) with the integral
L
(∂σ
n
)
π
∗
(η
m−1
)
f −
+
L
(σ
n
)
df ∧
π
∗
(η
m
)
f −
, (4.16)
where f = Ψ
Γ
and m = −n + D( + 1)/2.
124 Feynman Motives
Thus, the Leray regularization introduced here consists in replacing the
integral over σ
n
∩D
(X
Γ
) with an integral over L
(σ
n
) (σ
n
∩X
Γ,
) ×S
1
,
which avoids the locus σ
n
∩ X
Γ
where the divergence can occur by going
around it along a circle of small radius > 0. Here X
Γ,
is the level set
of f = Ψ
Γ
.
Lemma 4.2.4. The Leray regularization of the Feynman integral (3.56)
can be equivalently written in the form
U(Γ)
=
1
C(n, D, )
∂σ
n
∩D
(X
Γ
)
c
π
∗
(η
m
) +
σ
n
∩D
(X
Γ
)
c
df ∧
π
∗
(η
m
)
f
+
2πi
C(n, D, )
∂σ
n
∩X
Γ,
π
∗
(η
m−1
)
df
+
σ
n
∩X
Γ,
π
∗
(η
m
)
,
(4.17)
with π
∗
(η
m
) = ∆(ω)/f
m
as in (4.14) and Proposition 3.4.3 and D
(X
Γ
)
c
denoting the complement of D
(X
Γ
).
Proof. The result follows directly from Proposition 3.4.3 and Lemma
4.2.2 applied to (4.16).
We now study the dependence on the parameter > 0 of the Leray
regularized Feynman integral (4.16), i.e. of an integral of the form
I
:=
∂σ∩X
π
∗
(η
m−1
)
df
+
σ∩X
π
∗
(η
m
). (4.18)
Theorem 4.2.5. The function I
of (4.18) is inﬁnitely diﬀerentiable in
. Moreover, it extends to a holomorphic function for ∈ ∆
∗
⊂ C, a
small punctured disk, with a pole of order at most m at = 0, with m =
−n + D( + 1)/2.
Proof. To prove the diﬀerentiability of I
, let us write
A
(η) =
σ∩X
π
∗
(η), (4.19)
with π
∗
(η) as in (3.57). By Lemma 4.2.2 above, and the fact that dπ
∗
(η) =
0, we obtain
d
d
A
(η) = −
∂σ∩X
π
∗
(η)
df
, (4.20)
where X
= {f = } and π
∗
(η)/df is the Gelfand–Leray form of π
∗
(η).
Thus, we can write
I
= A
(η
m
) −
d
d
A
(η
m−1
).
Feynman integrals and Gelfand–Leray forms 125
Thus, to check the diﬀerentiability in the variable to all orders of I
is
equivalent to checking that of A
. We then deﬁne Υ : Ω
n
→ Ω
n
by setting
Υ(α) = d
α
df
, (4.21)
where α/df is a Gelfand–Leray form for α. In turn, the nform Υ(α) also
has a Gelfand–Leray form, which we denote by
δ(α) =
Υ(α)
df
=
d
α
df
df
. (4.22)
We then prove that, for k ≥ 2,
d
k
d
k
A
= −
∂σ∩X
δ
k−1
π
∗
(η)
df
. (4.23)
This follows by induction. In fact, we ﬁrst see that
d
2
d
2
A
= −
d
d
∂σ∩X
π
∗
(η)
df
which, applying Lemma 4.2.2 gives
= −
∂σ∩X
d
π
∗
(η)
df
df
.
Assuming then that
d
k
d
k
A
= −
∂σ∩X
δ
k−1
π
∗
(η)
df
we obtain again by a direct application of Lemma 4.2.2
d
k+1
d
k+1
A
= −
∂σ∩X
d
δ
k−1
π
∗
(η)
df
df
df
= −
∂σ∩X
δ
k
π
∗
(η)
df
.
This proves diﬀerentiability to all orders.
Notice then that, while the expression (4.16) used in Deﬁnition 4.2.3
is, a priori, only deﬁned for > 0, the equivalent expression given in the
second line of (4.17) and in (4.18) is clearly deﬁned for any complex ∈ ∆
∗
in a punctured disk around = 0 of suﬃciently small radius. It can then be
seen that the expression (4.18) depends holomorphically on the parameter
by the general argument on holomorphic dependence on parameters given
in Part III, §10.2 of Vol.II of [Arnold, GuseinZade, Varchenko (1988)].
Finally, to see that I
has a pole of order at most m at = 0, notice
that the form π
∗
(η
m
) of (4.14) is given by ∆(ω)/f
m
and has a pole of order
at most m at X.
126 Feynman Motives
Notice that the regularization procedure described here is natural from
the point of view of singularity theory and oscillatory integrals, but cannot
be used directly to derive information at the motivic level. To be able to
do so one would have to ﬁrst derive an algebraic version of the procedure
described here, possibly involving a motivic version of the notion of “tubular
neighborhood”, see e.g. [Levine (2005)], applied to the locus Σ ∩ X
Γ
, with
Σ a normal crossings divisor containing the boundary ∂σ
n
of the domain
of integration, such as Σ = Σ
n
the union of the coordinate hyperplanes.
Algebrogeometric notions of tubular neighborhood were also elaborated
upon in Grothendieck’s “Esquisse d’un Programme”. The role of such
notion in the algebrogeometric theory of Feynman integrals awaits further
investigation.
Chapter 5
Connes–Kreimer theory in a nutshell
So far we focused on individual Feynman graphs and associated algebro
geometric constructions. However, it is well known in quantum ﬁeld the
ory that the process of removing divergences from Feynman integrals via
renormalization cannot be achieved without taking into account the nested
structure of divergences, i.e. the presence of subgraphs inside a given Feyn
man graph that already contribute divergences to the Feynman integral.
A process of subtracting divergences that takes into account this hierar
chy of nested subdivergences was developed in [Bogolyubov and Parasiuk
(1957)] and later completed by [Hepp (1966)] and [Zimmermann (1969)].
A very elegant geometric formulation of the BPHZ renormalization proce
dure was given in [Connes and Kreimer (2000)] and [Connes and Kreimer
(2001)] in terms of a Hopf algebra of Feynman graphs, where the coproduct
encodes the information on the subdivergences, and the BPHZ procedure
becomes an extraction of ﬁnite values via the Birkhoﬀ factorization of loops
in a prounipotent Lie group of characters of the Hopf algebra of Feynman
graphs. In this chapter we give a very brief account of the main steps in
the Connes–Kreimer theory. For brevity, we skip several important tech
nical points, such as how to deal with external momenta and we also do
not reproduce full proofs of the main results. In fact, these are already
described in full detail in the original papers of Connes and Kreimer and
also in the ﬁrst chapter of the book [Connes and Marcolli (2008)], so we do
not duplicate them here. We restrict to the essential skeleton of what we
need in order to continue with our narrative in the following chapters.
127
128 Feynman Motives
5.1 The Bogolyubov recursion
The main steps of the Bogolyubov–Parasiuk–Hepp–Zimmermann renormal
ization procedure (BPHZ) are summarized as follows. (Again, for more
details the reader is invited to look at Chapter 1 on [Connes and Marcolli
(2008)]).
5.1.1 Step 1: Preparation
One replaces the Feynman integral U(Γ) of (1.55) by the expression
¯
R(Γ) = U(Γ) +
γ∈V(Γ)
C(γ)U(Γ/γ). (5.1)
Here we suppress the dependence on z, µ and the external momenta p for
simplicity of notation. The expression (5.1) is to be understood as a sum
of Laurent series in z, depending on the extra parameter µ. The coeﬀcients
C(γ), themselves Laurent series, are deﬁned recursively in (5.2) below.
The sum in (5.1) is over the set 1(Γ) of (not necessarily connected)
proper nonempty subgraphs γ ⊂ Γ with the property that the residue
Res(γ
i
) of each component γ
i
of the graph γ has valence equal to one of the
monomials in the Lagrangian L(φ). Here the residue is deﬁned as follows.
Deﬁnition 5.1.1. The residue Res(Γ) of a connected Feynman graph Γ is
the Feynman graph consisting of a single vertex and as many half edges
attached to it as the number of external edges of Γ, i.e. the graph obtained
by contracting all the internal edges of Γ to a single vertex.
5.1.2 Step 2: Counterterms
These are the expressions by which the Lagrangian needs to be modiﬁed
to cancel the divergence produced by the graph Γ. They are deﬁned as the
polar part of the Laurent series
¯
R(Γ),
C(Γ) = −T(
¯
R(Γ)). (5.2)
Here T denotes the operator of projection of Laurent series onto their polar
part. Notice how the counterterms C(Γ) are deﬁned here by a recursion,
where in the right hand side of (5.2) only terms C(γ) for strictly smaller
graphs γ ⊂ Γ are involved.
We discuss here brieﬂy a property of the operator T that will be useful
later. We’ll see in ¸5.5 below that this property corresponds to the fact
Connes–Kreimer theory in a nutshell 129
that the diﬀerential ﬁeld / of convergent Laurent series with the operator
T has the structure of a Rota–Baxter algebra.
Lemma 5.1.2. Let / be the ﬁeld of convergent Laurent series (germs of
meromorphic functions at the origin). Let T be, as above, the operator that
projects a Laurent series onto its polar part. This satisﬁes the identity
T(f
1
)T(f
2
) = T(f
1
T(f
2
)) +T(T(f
1
)f
2
) −T(f
1
f
2
). (5.3)
Proof. If f
1
=
∞
k=−N
1
a
k
z
k
and f
2
=
∞
r=−N
2
b
r
z
r
then
T(f
1
f
2
) =
−1
m=−(N
1
+N
2
)
k+r=m
a
k
b
r
z
m
.
On the other hand, we also have
T(f
1
)T(f
2
) = (
−1
k=−N
1
a
k
z
k
)(
−1
r=−N
2
b
r
z
r
).
One then ﬁnds the identity (5.3), since all the terms that appear in T(f
1
f
2
)
are accounted for with multiplicity one by summing
T(T(f
1
)f
2
) = T((
−1
k=−N
1
a
k
z
k
)(
∞
r=−N
2
b
r
z
r
))
and
T(f
1
T(f
2
)) = T((
∞
k=−N
1
a
k
z
k
)(
−1
r=−N
2
b
r
z
r
))
and then subtracting T(f
1
)T(f
2
), which has the eﬀect of removing the terms
that had been counted twice in the previous sum and leaving exactly the
terms that constitute the polar part T(f
1
f
2
).
5.1.3 Step 3: Renormalized values
One can then extract a ﬁnite value from the integral U(Γ) by removing the
polar part, not of U(Γ) itself but of its preparation. Namely, one sets
R(Γ) =
¯
R(Γ) +C(Γ) = U(Γ) +C(Γ) +
γ∈V(Γ)
C(γ)U(Γ/γ), (5.4)
which we also write equivalently as
R(Γ) = U(Γ) +
γ∈V(Γ)
C(γ)U(Γ/γ) −T
U(Γ) +
γ∈V(Γ)
C(γ)U(Γ/γ)
,
(5.5)
130 Feynman Motives
that is, the minimal subtraction of the polar part not of the original Laurent
series U(Γ) but of the corrected one
¯
R(Γ).
The main result of BPHZ is that, unlike what happens typically with
the Laurent series U(Γ) alone, the coeﬃcient of pole of the corrected Lau
rent series
¯
R(Γ) provided by the Bogolyubov preparation is always a local
expression in the variables of the external momenta, hence one that can be
corrected either by altering the coeﬃcients of the already existing terms in
the Lagrangian or by a adding a ﬁnite number of new polynomial terms to
the Lagrangian. Namely, the following statement is the ﬁnal result of the
combined [Bogolyubov and Parasiuk (1957)], [Hepp (1966)] and [Zimmer
mann (1969)].
Theorem 5.1.3. The coeﬃcient of pole of
¯
R(Γ) is local.
Here local means that it can be expressed as a local function in the mo
mentum variables (involving only polynomials and derivatives). In fact, if
one wants to be able to remove the polar part by correcting the Lagrangian
by adding suitable counterterms, these themselves have to be local functions
as the terms already present in the Lagrangian are.
The BPHZ procedure also goes under the name of Bogolyubov recursion,
because of the recursive nature of the deﬁnition of the counterterms C(Γ) in
(5.2). A very nice conceptual understanding of the BPHZ renormalization
procedure with the DimReg+MS regularization was obtained by Connes
and Kreimer [Connes and Kreimer (2000)], [Connes and Kreimer (2001)],
based on a reformulation of the BPHZ procedure in geometric terms. A
detailed account of the Connes–Kreimer theory was given in Chapter 1
of [Connes and Marcolli (2008)], so we limit ourselves to a more sketchy
overview here and refer the reader to that more detailed treatment.
5.2 Hopf algebras and aﬃne group schemes
In the following we work with algebras deﬁned over a ﬁeld Kof characteristic
zero. In fact, for our main application, K will be either Q or C. We also
restrict our attention here to commutative Hopf algebras, because the Hopf
algebras that appear in applications to the theory of renormalization in
quantum ﬁeld theory are usually of this kind.
Deﬁnition 5.2.1. A commutative Hopf algebra over K is a commutative
Kalgebra H endowed with an algebra homomorphism ∆ : H → H ⊗ H,
Connes–Kreimer theory in a nutshell 131
the coproduct, an algebra homomorphism ε : H → K, the counit, and an
algebra antihomomorphism S : H → H, the antipode. These satisfy the
coassociativity
(∆⊗id)∆ = (id ⊗∆)∆
as morphisms H → H ⊗
K
H ⊗
K
H, the compatibility of the coproduct ∆
with the counit ε,
(id ⊗ε)∆ = id = (ε ⊗id)∆
as morphisms H → H, and the compatibility of comultiplication ∆, an
tipode S, and multiplication µ,
µ(id ⊗S)∆ = µ(S ⊗id)∆ = 1 ε
as morphisms H →H.
Notice that in general the coproduct will not be cocommutative.
A useful notion we refer to often in the following, which also already
appeared in ¸2.9 and ¸2.10, in the context of the Tannakian formalism and
motivic Galois groups, is that of aﬃne group scheme. As we recall brieﬂy
here (see also [Waterhouse (1979)]) this is a functor from algebras to groups
naturally associated to a commutative Hopf algebra.
Proposition 5.2.2. A commutative Hopf algebra H over a ﬁeld K of char
acteristic zero determines a covariant functor G from the category /
K
of
commutative algebras over K to the category of groups, by setting
G(A) = Hom
A
K
(H, A), (5.6)
for A ∈ Obj(/
K
).
Proof. We need to show that the set G(A) is a group. Elements of G(A)
are homomorphisms of Kalgebras
φ : H →A, φ(xy) = φ(x)φ(y), ∀x, y ∈ H, φ(1) = 1 .
There is a product operation on G(A) deﬁned by the coproduct ∆ of H
φ
1
∗ φ
2
(x) = ¸φ
1
⊗φ
2
, ∆(x)) . (5.7)
It is associative because of the coassociativity of ∆. There is a unit element
in G(A) which is determined by the counit ε of H by
1
G(A)
: H
ε
→K →A.
The property that this is a unit with respect to the multiplication in G(A)
come from the compatibility relation between counit and coproduct in the
132 Feynman Motives
axioms of Hopf algebra. Similarly, there is an inverse φ
−1
= φ◦S, deﬁned by
precomposing a morphism φ : H →A with the antipode S : H →H. The
fact that it satisﬁes the correct properties with respect to the multiplication
and unit,
φ ∗ φ
−1
= φ
−1
∗ φ = 1,
to be the inverse in the group G(A) follows from the remaining axiom of the
Hopf algebra structure that gives the compatibility between the counit, co
multiplication, and antipode. The map φ →φ
−1
is an antihomomorphism,
(φ
1
∗ φ
2
)
−1
= φ
−1
2
∗ φ
−1
1
, because S is an antihomomorphism of the Hopf
algebra H.
An aﬃne group scheme is a representable covariant functor from the
category of commutative algebras over the ﬁeld K to the category of groups.
So the above shows that a commutative Hopf algebra determines an aﬃne
group scheme. Any such representable functor will be of the form (5.6), for
some commutative algebra H, which in turn inherits the structure of Hopf
algebra from the group structure of G(A).
The simplest examples of aﬃne group schemes are the additive and the
multiplicative group, G
a
and G
m
, respectively, which correspond to the
Hopf algebras
• The additive group G
a
corresponds to the Hopf algebra H = K[t] with
coproduct ∆(t) = t ⊗1+1⊗t, counit ε(t) = 0 and antipode S(t) = −t.
• The multiplicative group G
m
corresponds to the Hopf algebra H =
K[t, t
−1
], with coproduct ∆(t) = t ⊗ t, counit ε(t) = 1 and antipode
S(t) = t
−1
.
The main examples in the following will be prounipotent aﬃne group
schemes. These are projective limits of unipotent algebraic groups, where
by algebraic groups we mean here subgroups of the aﬃne group scheme
GL
n
. These prounipotent cases arise from Hopf algebras that are graded,
H = ⊕
n≥0
H
n
and connected, H
0
= K, as the Connes–Kreimer Hopf al
gebra is. We discuss more the example of GL
n
as an aﬃne group scheme
in Proposition 5.4.2 below. For more detailed information on aﬃne group
schemes, we refer the reader to [Demazure and Grothendieck (1970)] and
[Waterhouse (1979)].
Connes–Kreimer theory in a nutshell 133
5.3 The Connes–Kreimer Hopf algebra
The Hopf algebra of Feynman graphs introduced in [Connes and Kreimer
(2000)] depends on the choice of the physical theory, in the sense that it
involves only those graphs that are Feynman graphs for the speciﬁed La
grangian L(φ), and, as we see below, the coproduct formula also involves
only those subgraphs whose residue graph corresponds to one of the mono
mials in the Lagrangian.
As an algebra H
CK
it is the free commutative algebra with generators
the 1PI Feynman graphs Γ of the speciﬁed physical theory. It is graded by
loop number, or by the number of internal lines, with
deg(Γ
1
Γ
n
) =
i
deg(Γ
i
), deg(1) = 0.
This grading corresponds to the order in the perturbative expansion.
The most interesting part of the structure is the coproduct, which is
given on generators by
∆(Γ) = Γ ⊗1 + 1 ⊗Γ +
γ∈V(Γ)
γ ⊗Γ/γ. (5.8)
The sum is over the class 1(Γ) of proper subgraphs of Γ with the property
that the quotient graph Γ/γ is still a 1PI Feynman graph of the same theory.
The subgraphs in 1(Γ) are not necessarily connected. In the case of
several connected components, the quotient graph Γ/γ is understood as
the graph obtained by shrinking each component of γ ⊂ Γ to a vertex.
The condition that Γ/γ is a 1PI Feynman graph can be formulated in
terms of residue graphs Res(γ) of the sugraphs γ ⊂ Γ. For a given graph Γ
the residue Res(Γ) is the graph that is obtained by keeping all the external
edges of Γ and shrinking all internal edges to a single vertex. The valence
of this vertex is then equal to the number of external edges. The class of
subgraphs 1(Γ) can then be charcterized by requiring that the components
of γ ⊂ Γ are 1PI Feynman graphs with the property that the residue Res(Γ)
has valence corresponding to one of the monomials in the Lagrangian of the
theory. This in fact ensures that the corresponding vertex in the quotient
graph Γ/γ has an admissible valence for it to be a Feynman graph. The fact
that Γ/γ is 1PI then follows from the fact that Γ and all the components
of γ are 1PI.
The antipode is deﬁned inductively by
S(X) = −X −
S(X
)X
,
134 Feynman Motives
where X is an element with coproduct ∆(X) = X⊗1+1⊗X+
X
⊗X
,
where all the X
and X
have lower degrees.
The aﬃne group scheme dual to the Connes–Kreimer Hopf algebra is
referred to as the group of diﬀeographisms. Its set of complex points
G(C) = Hom(H
CK
, C)
is a prounipotent complex Lie group G(C). It was proved in [Connes
and Kreimer (2000)] that this group acts by local diﬀeomorphisms on the
coupling constants of the theory.
A variant of the Connes–Kreimer Hopf algebra, which is useful in
algebrogeometric applications, was considered in [Bloch and Kreimer
(2008)], see also [Kreimer (2009)]. It is referred to as the core Hopf al
gebra and it is simply deﬁned by considering as generators all 1PI graphs
without restriction on the valence of vertices. This would correspond in
physical terms to arbitrarily large powers φ
k
in the Lagrangian. Corre
spondingly, the coproduct is taken over all the subgraphs whose connected
components are 1PI, without restriction on the residue graphs.
The algebrogeometric Feynman rules discussed above and introduced in
[Aluﬃ and Marcolli (2008b)] deﬁne ring homomorphisms from the core Hopf
algebra to the Grothendieck ring of varieties, or to the ring of immersed
conical varieties deﬁned in [Aluﬃ and Marcolli (2008b)].
A further important remark about the Connes–Kreimer Hopf algebra:
the version we recalled brieﬂy here is only the combinatorial part of a larger
Hopf algebra, whose generators are pairs (Γ, σ) of a 1PI graph and a dis
tribution on the space of test functions of the external momenta. This
continuous version of the the Hopf algebra of renormalization is necessary
when taking into account the external momenta of the graphs in the ex
traction of subdivergences. This additional structure is described in detail
in [Connes and Kreimer (2000)] as well as in Chapter 1 of [Connes and
Marcolli (2008)].
Finally, the Connes–Kreimer construction of the Hopf algebra is given
for scalar quantum ﬁeld theories. Generalizations to vector valued ﬁelds do
not present any problem, while extending the same setting to gauge theories
and chiral theories is more subtle. The results were recently generalized to
gauge theories in [van Suijlekom (2007)], [van Suijlekom (2006)], [van Sui
jlekom (2008)], by showing that the Ward identities deﬁne Hopf ideals. The
case of chiral theories runs into a well known problem of reconciling the use
of dimensional regularization with the chirality operator γ
5
, see [Jegerlehner
(2001)]. A way to describe the prescription of [Breitenlohner and Maison
Connes–Kreimer theory in a nutshell 135
(1977)] to express the chirality operator γ
5
within DimReg in terms of
noncommutative geometry was given in the unpublished [Connes and Mar
colli (2005b)]; see also Chapter 1 of [Connes and Marcolli (2008)]. In this
setting the operation of dimensional regularization is then expressed as a
cup product of spectral triples. Recently, a version of the Connes–Kreimer
renormalization as Birkhoﬀ factorization, based on zeta function regulariza
tion instead of DimReg and adapted to curved backgrounds was developed
in [Agarwala (2009)]. An extension of the parametric Feynman integrals to
theories with bosonic and fermionic ﬁelds, using periods of supermanifolds,
was given in [Marcolli and Rej (2008)], though the most appropriate form
of the Connes–Kreimer renormalization procedure in that context remains
to be properly discussed. Another interesting recent result is a categori
ﬁcation of the Connes–Kreimer Hopf algebra obtained in [Kremnizer and
Szczesny (2008)], where the dual Hopf algebra is identiﬁed with the Hall
algebra of a ﬁnitary abelian category.
5.4 Birkhoﬀ factorization
We give an algebraic formulation of Birkhoﬀ factorization in the group
scheme G of a commutative Hopf algebra in terms of the Hopf algebra H
and its characters.
Deﬁnition 5.4.1. Suppose given an algebra homomorphism φ ∈
Hom(H, /), where / is the ﬁeld of convergent Laurent series, i.e. germs
of meromorphic functions at z = 0. Then φ has a Birkhoﬀ factorization if
there exist algebra homomorphisms φ
+
∈ Hom(H, O), with O the complex
algebra of convergent power series, and φ
−
∈ Hom(H, Q), with values in
the polynomial algebra Q = C[z
−1
], such that
φ = (φ
−
◦ S) φ
+
, (5.9)
where S is the antipode in the Hopf algebra and is the product in G
dual to the coproduct in the Hopf algebra. The factorization (5.9), if it
exists, is not unique. To make it unique one also requires the normalization
condition
−
◦ φ
−
= , (5.10)
where
−
: C[z
−1
] →C is the augmentation and is the counit of the Hopf
algebra H.
136 Feynman Motives
The characters φ ∈ Hom(H, /) can be equally interpreted as loops de
ﬁned on an inﬁnitesimal punctured disk ∆
∗
around the origin z = 0 in C
∗
and with values in the prounipotent complex Lie group G(C) = Hom(H, C)
of algebra homomorphisms from H to C, so that the recursive formula for
the factorization of φ gives the Birkhoﬀ factorization of loops written in the
more traditional form
γ(z) = γ
−
(z)
−1
γ
+
(z), (5.11)
with the normalization condition γ
−
(∞) = 1.
Not all commutative Hopf algebras admit a Birkhoﬀ factorization for
arbitrary elements φ ∈ Hom(H, /). The following example illustrates a case
where there are elements in Hom(H, /) that do not admit a factorization
of the form (5.9).
Proposition 5.4.2. Consider the commutative Hopf algebra H =
C[x
ij
, t]/(det(x
ij
) −t) with the coproduct
∆(x
ij
) =
k
x
ik
⊗x
kj
.
Then there exist elements φ ∈ Hom(H, /) that do not admit a factorization
of the form (5.9).
Proof. The aﬃne group scheme dual to the Hopf algebra H =
C[x
ij
, t]/(det(x
ij
) − t) with the given coproduct is the group GL
n
. To see
that there exist elements of φ ∈ Hom(H, /) that do not admit a Birkhoﬀ
factorization in the form stated above, we can give an equivalent formula
tion in terms of loops in a complex Lie group. We can use the equivalent
description of characters φ ∈ Hom(H, /) in terms of loops γ : ∆
∗
→ GL
n
,
as above. Then, it suﬃces to notice that, if we use the function γ(z) to de
ﬁne the transition function of a holomorphic vector bundle over the 2sphere
P
1
(C), the existence of a factorization (5.11) shows that the resulting bun
dle is trivial. However, one knows that there are nontrivial holomorphic
vector bundles on P
1
(C), so it suﬃces to take one such nontrivial vector
bundle E and a transition function γ(z) relating the trivializations of E
on the two contractible sets ∆ and P
1
(C) ∆ to see that this loop cannot
admit a factorization of the form (5.11).
One of the main results of the Connes–Kreimer theory is a recursive
formula for the Birkhoﬀ factorization in the case of graded connected com
mutative Hopf algebras. We review this result in the next section.
Connes–Kreimer theory in a nutshell 137
5.5 Factorization and RotaBaxter algebras
Let H be a graded connected commutative Hopf algebra H = ⊕
n≥0
H
n
,
where the H
n
are ﬁnite dimensional. The connectedness condition means
that H
0
= C, for an algebra deﬁned over the ﬁeld of complex numbers.
Notice that here the condition that His a commutative graded Hopf algebra
does not mean that it is “gradedcommutative”: the commutators are not
graded commutators but ordinary commutators. It behaves like the case of
a graded–commutative Hopf algebra where the graded pieces are nonzero
only in even degrees.
Let / denote the ﬁeld of convergent Laurent series. We consider, as
above, elements φ ∈ Hom(H, /) and we ask whether all such elements
admit a Birkhoﬀ factorization as in (5.9). The Connes–Kreimer formula
[Connes and Kreimer (2000)] gives a positive answer, for the class of Hopf
algebras mentioned above, in the form of an explicit recursive formula. In
fact, for later use, we can formulate their result in a slightly more general
form, as was done for instance in [EbrahimiFard, Guo, Kreimer (2004)],
using the formalism of Rota–Baxter algebras.
To this purpose we recall some preliminary notions on Rota–Baxter
algebras and operators.
Deﬁnition 5.5.1. A commutative Rota–Baxter algebra of weight λ is a
commutative algebra / endowed with a linear operator R satisfying the
Rota–Baxter identity
R(x)R(y) = R(xR(y)) +R(R(x)y) +λR(xy). (5.12)
Such an operator is called a Rota–Baxter operator of weight λ.
The prototype of a Rota–Baxter operator is integration, for which the
Rota–Baxter identity is the integration by parts formula.
Lemma 5.5.2. The pair (/, T) of the ﬁeld of convergent Laurent series
with the projection on the polar part is a Rota–Baxter algebra of weight
λ = −1.
Proof. This is an immediate consequence of Lemma 5.1.2.
Lemma 5.5.3. Given a commutative unital Rota–Baxter algebra (/, R) of
weight λ = −1, one obtains two commutative unital algebras /
±
by taking
as /
−
the image R(/) with a unit adjoined (i.e. the smallest unital algebra
containing R(/) as an ideal) and /
+
= (1 −R)(/).
138 Feynman Motives
Proof. These are indeed algebras. In fact, for x, y ∈ /
−
we have x =
R(a) and y = R(b) for some a, b ∈ /. The product R(a)R(b) = R(aR(b) +
R(a)b + λab) is also in the range of R, hence in /
−
. Similarly, given
x, y ∈ /
+
we have x = a −R(a) and y = b −R(b) for some a, b ∈ /. Then
xy = (a −R(a))(b −R(b)) = ab −aR(b) −bR(a) +R(a)R(b)
= ab −aR(b) −bR(a) +R(aR(b) +R(a)b +λab).
Thus, for λ = −1, one obtains
xy = (ab −aR(b) −bR(a)) −R(ab −aR(b) −bR(a)) ∈ /
+
.
Just as Deﬁnition 5.4.1 formulates the Birkhoﬀ factorization for algebra
homomorphisms φ ∈ Hom(H, /) we can similarly consider the same prob
lems for algebra homomorphisms from the Hopf algebra H with values in
any Rota–Baxter algebra (/, R) of weight λ = −1. Given φ ∈ Hom(H, /),
we look for algebra homomorphisms φ
±
∈ Hom(H, /
±
) with the algebras
/
±
obtained as in Lemma 5.5.3, satisfying the factorization
φ = (φ
−
◦ S) φ
+
, (5.13)
as in (5.9), where here we use the fact that /
±
⊂ / so that the equality
above is understood as the equality
φ(X) = (φ
−
◦ S)(X) φ
+
(X), ∀X ∈ H
viewed as an equality of elements in /. The normalization condition is
also formulated in this case as
−
◦ φ
−
= , where
−
: /
−
→ C is the
augmentation map and is the counit of the Hopf algebra H.
It was proved in [Connes and Kreimer (2000)], and reformulated in the
RotaBaxter context in [EbrahimiFard, Guo, Kreimer (2004)], that if H
is a graded connected commutative Hopf algebra and (/, R) is a unital
Rota–Baxter algebra of weight λ = −1, then any homomorphism of uni
tal commutative algebras φ ∈ Hom(H, /) admits a Birkhoﬀ factorization
(5.13) where the homomorphisms φ
±
∈ Hom(H, /
±
) are deﬁned induc
tively by the formulae
φ
−
(X) = −R
φ(X) +
φ
−
(X
)φ(X
)
, (5.14)
where ∆(X) = X ⊗ 1 + 1 ⊗ X +
X
⊗ X
, with X
and X
of lower
degree, and
φ
+
(X) = (1 −R)
φ(X) +
φ
−
(X
)φ(X
)
. (5.15)
Connes–Kreimer theory in a nutshell 139
The RotaBaxter property of the operator R with weight −1 is needed to
show that the φ
±
deﬁned recursively in this way are algebra homomor
phisms, φ
±
(XY ) = φ
±
(X)φ
±
(Y ).
Thus, the results of [Connes and Kreimer (2000)] and [Connes and
Kreimer (2001)] show that, in terms of Birkhoﬀ factorization, the formulae
of the BPHZ renormalization procedure (5.2), (5.4), (5.5) are exactly the re
cursive formulae for the Birkhoﬀ factorization of characters of the Connes–
Kreimer Hopf algebra with values in a Rota–Baxter algebra, namely
U
−
(Γ) = −T
U(Γ) +
γ⊂Γ
U
−
(γ)U(Γ/γ)
, (5.16)
U
+
(Γ) = (1 −T)
U(Γ) +
γ⊂Γ
U
−
(γ)U(Γ/γ)
. (5.17)
This is just the same as the formulae (5.14) and (5.14), given above for any
algebra homomorphism from a graded connected commutative Hopf algebra
H to a commutative Rota–Baxter algebra (/, R) of weight −1, applied to
the case where H = H
CK
is the Connes–Kreimer Hopf algebra and where
the RotaBaxter algebra is given by Laurent series in the DimReg variable
z and R = T is the projection onto the polar part, and the homomorphism
U is the usual Feynman amplitude regularized using DimReg. Then the
C(Γ) = U
−
(Γ) are the counterterms and the R(Γ) = U
+
(Γ)[
z=0
are the
renormalized values.
5.6 Motivic Feynman rules and RotaBaxter structure
We have seen, following [Aluﬃ and Marcolli (2008b)], that one can de
ﬁne motivic Feynman rules in terms of the class [A
n
ˆ
X
Γ
] of the graph
hypersurface complement in the Grothendieck ring of varieties K
0
(1).
A variant of such motivic Feynman rules is obtained by setting
U(Γ) =
[A
n
ˆ
X
Γ
]
L
n
, (5.18)
with values in the ring K
0
(1)[L
−1
], where one inverts the Lefschetz motive,
or equivalently, via the additive invariant of [Gillet and Soul´e (1996)], with
values in the Grothendieck ring of motives K
0
(/). Dividing by L
n
has
the eﬀect of normalizing the “motivic Feynman rule” [A
n
ˆ
X
Γ
] by the
value it would have if Γ were a forest on the same number of edges. For
140 Feynman Motives
the original Feynman integrals this would measure the amount of linear
dependence between the edge momentum variables created by the presence
of the interaction vertices.
In the ring K
0
(1)[L
−1
] or K
0
(/) one can still consider the Rota–Baxter
structure given by the operator R of projection onto the polar part in the
variable L. The renormalized Feynman rule
U
+
(Γ) = (1 −R)
U(Γ) +
γ⊂Γ
U
−
(γ)U(Γ/γ)
then gives a class in the Grothendieck ring of varieties K
0
(1). More inter
estingly, the renormalized value of the Feynman integral, in this case, gives
some information on the birational geometry of the graph hypersurface.
In fact, we ﬁrst recall the following useful result of [Larsen and Lunts
(2003)]. We work here with the Grothendieck ring K
0
(1
C
) for varieties over
the ﬁeld of complex numbers.
It is proved in [Larsen and Lunts (2003)] that the quotient of the
Grothendieck ring K
0
(1
C
) by the ideal generated by L = [A
1
] is isomorphic
as a ring to Z[SB], the ring of the multiplicative monoid SB of stable bi
rational equivalence classes of varieties in 1
C
. Recall that two (irreducible)
varieties X and Y are stably birationally equivalent if X P
n
and Y P
m
are birationally equivalent for some n, m ≥ 0.
We then see that the renormalized value is in this case
U
+
(Γ)[
L=0
= (1 −R)
U(Γ) +
γ⊂Γ
U
−
(γ)U(Γ/γ)
[
L=0
∈ Z[SB]. (5.19)
Notice that the parts of [A
n
ˆ
X
Γ
], [A
n
1
ˆ
X
γ
] and [A
n
2
ˆ
X
Γ/γ
] that are
contained in the ideal (L) ⊂ K
0
(1
0
) contribute cancellations to the L
n
in
the denominator. It is possible that this invariant and the Birkhoﬀ factor
ization of U(Γ) may help to detect the presence of nonmixedTate strata in
the graph hypersurface X
Γ
coming from the contributions of hypersurfaces
of smaller graphs γ ⊂ Γ or quotient graphs Γ/γ.
The class of a graph hypersurface X
Γ
in the ring Z[SB] of stable bi
rational equivalence classes detects some of the properties of the graph
Γ. For example, it was shown in [Aluﬃ and Marcolli (2008b)] that the
class is trivial, [X
Γ
]
SB
= 1, whenever the graph is not 1PI. In fact, if
a graph is not 1PI then in the Grothendieck ring one has [A
n
ˆ
X
Γ
] =
L [A
n−1
ˆ
X
Γ
], where Γ
is the graph obtained by removing a disconnect
ing edge of Γ. Thus, [A
n
ˆ
X
Γ
]
SB
= 0 ∈ Z[SB]. Using then the fact that
Connes–Kreimer theory in a nutshell 141
[A
n
ˆ
X
Γ
] = L
n
− 1 − (L − 1)[X
Γ
] in the Grothendieck ring, one obtains
[A
n
ˆ
X
Γ
]
SB
= −1 + [X
Γ
]
SB
= 0.
It would be interesting to see if there are other natural Rota–Baxter
structures on the Grothendieck ring of varieties, or in the ring of immersed
conical varieties, that can be used in deﬁning a renormalized version of
motivic and algebrogeometric Feynman rules.
Chapter 6
The Riemann–Hilbert correspondence
This chapter is dedicated to the Riemann–Hilbert correspondence of renor
malization, as derived in [Connes and Marcolli (2004)]. A very detailed
account of this topic is available in [Connes and Marcolli (2008)], hence the
treatment given here will be shorter and will only aim at summarizing the
main steps involved in the main result of [Connes and Marcolli (2004)]. The
interested reader is encouraged to look at [Connes and Marcolli (2006b)]
and at the ﬁrst chapter of [Connes and Marcolli (2008)] for more extensive
details, and to [Connes and Marcolli (2005a)] for a shorter survey.
The purpose of this approach is to parameterize the data of divergences
in perturbative scalar quantum ﬁeld theories in terms of suitable gauge
equivalence classes of diﬀerential systems with irregular singularities. These
are then classiﬁed in terms of the Riemann–Hilbert correspondence, which
means that one shows that these diﬀerential systems form a Tannakian
category, hence they can be classiﬁed in terms of ﬁnite dimensional linear
representations of an associated aﬃne group scheme. The Riemann–Hilbert
correspondence is the equivalence of categories between the analytic data
of diﬀerential systems and the representation theoretic data.
6.1 From divergences to iterated integrals
The ﬁrst step in reformulating the divergences of renormalization in terms
of diﬀerential systems consists of writing the negative piece γ
−
(z) of the
Birkhoﬀ factorization as an iterated integral depending on a single element
β in the Lie algebra Lie(G) of the aﬃne group scheme dual to the Connes–
Kreimer Hopf algebra. This is a way of formulating what is known in physics
as the ’t Hooft–Gross relations. These express the fact that counterterms
143
144 Feynman Motives
only depend on the beta function of the theory, that is, the inﬁnitesimal
generator of the renormalization group ﬂow.
The starting point is the observation that, in the Birkhoﬀ factorization,
there is a dependence on a mass (energy) scale µ, inherited from the same
dependence of the dimensionally regularized Feynman integrals U
µ
(Γ), dis
cussed in (1.55) above. Thus, we have
γ
µ
(z) = γ
−
(z)
−1
γ
µ,+
(z),
where one knows by reasons of dimensional analysis (see [Collins (1986)])
that the negative part is independent of µ.
This negative part γ
−
(z) can then be written as a time ordered expo
nential
γ
−
(z) = Te
−
1
z
∞
0
θ
−t
(β)dt
= 1 +
∞
n=1
d
n
(β)
z
n
,
where
d
n
(β) =
s
1
≥s
2
≥···≥s
n
≥0
θ
−s
1
(β) θ
−s
n
(β)ds
1
ds
n
,
and where β ∈ Lie(G) is the beta function, that is, the inﬁnitesimal gen
erator of renormalization group ﬂow, and the action θ
t
is induced by the
grading of the Hopf algebra by
θ
u
(X) = u
n
X, for u ∈ G
m
, and X ∈ H, with deg(X) = n,
with generator the grading operator Y (X) = nX. This result follows from
the analysis of the renormalization group given in [Connes and Kreimer
(2000)] [Connes and Kreimer (2001)], after the recursive formula for the
coeﬃcients d
n
is explicitly solved, so as to give the time ordered exponential
above.
The loop γ
µ
(z) collects all the unrenormalized values U
µ
(Γ) of the Feyn
man integrals in the form of an algebra homomorphism φ
µ
: H → C(¦z¦)
from the Connes–Kreimer Hopf algebra to the ﬁeld of germs of meromorphic
functions, given on generators by φ
µ
(Γ) = U
µ
(Γ).
The renormalization group analysis of [Connes and Kreimer (2001)]
shows that γ
µ
(z) satisﬁes the scaling property
γ
e
t
µ
(z) = θ
tz
(γ
µ
(z)) (6.1)
in addition to the property that its negative part is independent of µ,
∂
∂µ
γ
−
(z) = 0. (6.2)
The Riemann–Hilbert correspondence 145
The Birkhoﬀ factorization is then written in [Connes and Marcolli (2004)]
in terms of iterated integrals as
γ
µ,+
(z) = Te
−
1
z
−z log µ
0
θ
−t
(β)dt
θ
z log µ
(γ
reg
(z)).
Thus γ
µ
(z) is speciﬁed by β up to an equivalence given by the regular term
γ
reg
(z). The equivalence corresponds to “having the same negative part of
the Birkhoﬀ factorization”.
6.2 From iterated integrals to diﬀerential systems
Having written the counterterms as iterated integrals has the advantage
that it makes it possible to reformulate these data in terms of diﬀerential
systems. In fact, one can use the fact that iterated integrals are uniquely
solutions of certain diﬀerential equations.
An iterated integral (or timeordered exponential) g(b) = Te
b
a
α(t)dt
is
the unique solution of a diﬀerential equation dg(t) = g(t)α(t)dt with initial
condition g(a) = 1. In particular, suppose given the diﬀerential ﬁeld (/, δ),
where / = C(¦z¦) is the ﬁeld of convergent Laurent series and δ(f) = f
,
and an aﬃne group scheme G. One then has a logarithmic derivative
G(/) ÷ f → D(f) = f
−1
δ(f) ∈ Lie(G(/)),
and one can consider diﬀerential equations of the form D(f) = ω, for a ﬂat
Lie(G(C))valued connection ω, singular at z = 0 ∈ ∆
∗
. The existence of
solutions is ensured by the condition of trivial monodromy on ∆
∗
M(ω)() = Te
1
0
∗
ω
= 1, ∈ π
1
(∆
∗
).
In [Connes and Marcolli (2004)] these diﬀerential systems are consid
ered up to a gauge equivalence relation by a regular gauge transformation,
D(fh) = Dh+h
−1
Df h, for h ∈ G(C¦z¦), with C¦z¦ the ring of convergent
power series (germs of holomorphic functions). This gauge equivalence is
the same thing as the requirement that solutions have the same negative
piece of the Birkhoﬀ factorization,
ω
= Dh +h
−1
ωh ⇔ f
ω
−
= f
ω
−
,
where D(f
ω
) = ω and D(f
ω
) = ω
.
The fact that, by dimensional analysis, counterterms are independent
of the energy scale corresponds to the fact that the diﬀerential systems
are given by connections satisfying an equisingularity condition, as we re
call below. More precisely, in [Connes and Marcolli (2004)] a geometric
146 Feynman Motives
reformulation of the two conditions (6.1), (6.2) is given in terms of prop
erties of connections on a principal Gbundle P = B G over a ﬁbration
G
b
→ B → ∆, where z ∈ ∆ is the complexiﬁed dimension of DimReg and
the ﬁber µ
z
∈ G
m
over z corresponds to the changing mass scale. The
multiplicative group acts by
u(b, g) = (u(b), u
Y
(g)) ∀u ∈ G
m
.
The two conditions (6.1) and (6.2) correspond to the properties that the
ﬂat connection on P
∗
is equisingular, that is, it satisﬁes:
• Under the action of u ∈ G
m
the connection transforms like
(z, u(v)) = u
Y
((z, u)).
• If γ is a solution in G(C(¦z¦)) of the equation Dγ = , then the
restrictions along diﬀerent sections σ
1
, σ
2
of B with σ
1
(0) = σ
2
(0) have
“the same type of singularities”, namely
σ
∗
1
(γ) ∼ σ
∗
2
(γ),
where f
1
∼ f
2
means that f
−1
1
f
2
∈ G(C¦z¦), regular at zero.
The ﬁrst property of equisingular connections corresponds to the scaling
property (6.1) for the solutions of the corresponding diﬀerential systems,
while the second property expresses geometrically the independence of the
counterterms from the mass scale, that is, equation (6.2).
6.3 Flat equisingular connections and vector bundles
The setting described above still depends on the choice of a given physical
theory through the group G of diﬀeographisms, dual to the Connes–Kreimer
Hopf algebra of the theory.
The aﬃne group scheme G is uniquely determined by its Tannakian
category of ﬁnite dimensional linear representations Rep
G
. Thus, without
loss of information, one can replace singular connections on the principal
Gbundle P = B G with singular connections on associated vector bun
dles E = B V obtained by ﬁnite dimensional linear representations of
G in GL(V ). This provides a natural larger ambient category where the
diﬀerent diﬀeographism groups G of diﬀerent physical theories can be com
pared. It is obtained by working with all ﬂat equisingular vector bundles.
Among them one will identify a subcategory of those that are obtained
The Riemann–Hilbert correspondence 147
from representations of a given G, thus reconstructing the data associated
to divergences of a particular physical theory.
It is proved in [Connes and Marcolli (2004)] (see also Chapter 1 of
[Connes and Marcolli (2008)]) that ﬂat equisingular vector bundles form a
Tannakian category. We recall here brieﬂy the main steps of the construc
tion of this category.
The objects Obj(c) are pairs Θ = (V, [∇]), where V is a ﬁnite dimen
sional Zgraded vector space, out of which one forms a bundle E = B V .
The vector space has a ﬁltration W
−n
(V ) = ⊕
m≥n
V
m
induced by the
grading and a G
m
action also coming from the grading. The class [∇] is
an equivalence class of equisingular connections on the vector bundle E,
singular over the preimage in B of the point 0 ∈ ∆, which are compati
ble with the ﬁltration and trivial on the induced graded spaces Gr
W
−n
(V ),
up to the equivalence relation of Wequivalence. This is deﬁned by
T ◦ ∇
1
= ∇
2
◦ T for some T ∈ Aut(E) which is compatible with the
ﬁltration and trivial on Gr
W
−n
(V ). Here the condition that the connections
∇ are equisingular means that they are G
m
invariant and that restrictions
of solutions to sections of B with the same σ(0) are Wequivalent. The
morphisms Hom
E
(Θ, Θ
) are linear maps T : V → V
that are compatible
with the grading, and such that on E ⊕ E
the following connections are
Wequivalent:
∇
0
0 ∇
W−equiv
·
∇
T∇−∇
T
0 ∇
. (6.3)
Notice that, as usual, the diﬃcult part in deﬁning a good categorical
setting is to have the correct notion of morphisms. The reason why (6.3) is
the correct notion here is that one needs to bypass a well known obstacle,
which also arises when dealing with categories of Hodge structures, namely
the fact that ﬁltered spaces do not form an abelian category, so the ﬁrst
guess about the choice of morphisms, namely linear maps intertwining the
connections and compatible with the ﬁltrations, has to be corrected to
account for this problem.
6.4 The “cosmic Galois group”
The proof given in [Connes and Marcolli (2004)] and [Connes and Marcolli
(2008)] of the fact that the resulting category c is Tannakian directly ex
hibits an equivalence of categories with Rep
U
∗ for an aﬃne group scheme
U
∗
= U G
m
where U is dual, under the relation U(A) = Hom(H
U
, A),
148 Feynman Motives
to the Hopf algebra H
U
= U(L)
∨
dual (as Hopf algebra) to the universal
enveloping algebra of the free graded Lie algebra L = T(e
−1
, e
−2
, e
−3
, ).
The key step of the proof is showing that, for any given object Θ =
(V, [∇]) in c there exists unique ρ ∈ Rep
U
∗ such that
Dρ(γ
U
)
W−equiv
· ∇,
where γ
U
is the universal singular frame deﬁned as
γ
U
(z, v) = Te
−
1
z
v
0
u
Y
(e)
du
u
,
where the element e =
∞
n=1
e
−n
is the generator of a 1parameter subgroup
of U
∗
which lifts the renormalization group at the level of the group U
∗
.
Notice that the inﬁnite sum is well deﬁned since the Lie algebra is pro
nilpotent.
The group homomorphism U → G that realizes the ﬁnite dimensional
linear representations of G with equisingular connections as a subcategory
of c is obtained by mapping the generators e
−n
→ β
n
to the nth graded
piece of the beta function of the theory, seen as an element β =
n
β
n
in
the Lie algebra Lie(G).
Thus, the universal singular frame can be regarded as a universal source
for the counterterms for all given physical theories, as it maps to the correct
counterterms under the map of the generators e
−n
to the graded compo
nents of the beta function.
It is shown in Proposition 1.98 of [Connes and Marcolli (2008)] that
the universal singular frame is given, in terms of the generators of the Lie
algebra, in the form
γ
U
(−z, v) =
n≥0,k
j
>0
e
−k
1
e
−k
n
k
1
(k
1
+k
2
) (k
1
+k
2
+ +k
n
)
v
j
k
j
z
−n
.
The occurrence of these coeﬃcients, which also appear in the local index
formula of [Connes and Moscovici (1995)], has a natural explanation in
terms of the combinatorics of Dynkin idempotents and Dynkin operators
for free Lie algebras, as shown in [EbrahimiFard, GraciaBond
´
ia, Patras
(2007)].
Cartier conjectured the existence of a group, closely related to symme
tries of multiple zeta values, that would act on all the coupling constants
of scalar quantum ﬁeld theories, generalizing the renormalization group ac
tion. He referred to it as the “cosmic Galois group”. The result of [Connes
and Marcolli (2004)] reported here above gives a positive answer to Cartier’s
The Riemann–Hilbert correspondence 149
conjecture. In fact, the Tannakian Galois group U
∗
= UG
m
obtained as
above acts on the coupling constants via its map to the Connes–Kreimer
group of diﬀeographisms given by the β function, which in turn acts on the
coupling constants by the result of [Connes and Kreimer (2000)]. The rela
tion between the aﬃne group scheme U
∗
and symmetries of multiple zeta
values can be seen via the fact that the same group appears as a motivic
Galois group for a category of mixed Tate motives via the result of [Deligne
and Goncharov (2005)].
Chapter 7
The geometry of DimReg
Usually, as we have seen in the ﬁrst chapter of this book, the procedure of
dimensional regularization of divergent Feynman integrals, is deﬁned as a
purely formal procedure based on analytic continuation to a complex vari
able of the integral of a Gaussian in dimension D, as a function of D. This
means that, while this provides an eﬃcient recipe for analytically contin
uing Feynman integrals to a complexiﬁed dimension, typically no attempt
is made to make sense of an underlying geometry in the complexiﬁed di
mension, that would justify the formal procedures used in dimensional reg
ularization. We describe in this chapter two diﬀerent possible approaches
to a geometric description of dimensional regularization. The ﬁrst is based
on motivic notions and it applies to individual motives (motivic sheaves)
associated to given Feynman graphs, in the way we have seen through the
previous chapters. This ﬁrst method was discussed in [Marcolli (2008)].
The second method is based on the use of techniques from noncommuta
tive geometry and was ﬁrst discussed in Chapter 1 of [Connes and Marcolli
(2008)] and in the unpublished manuscript [Connes and Marcolli (2005b)].
We concentrate here especially on the later part of [Connes and Marcolli
(2005b)], which was not reported in [Connes and Marcolli (2008)], and
which makes contact with the motivic notions considered here, through the
language of mixed Hodge structures.
7.1 The motivic geometry of DimReg
Let us consider again the mixed Tate motives (motivic sheaves) deﬁned
by the logarithmic extension Log of (2.39) and its symmetric powers Log
n
of (2.40), as objects in the triangulated category T/
Q
(G
m
) of motivic
sheaves over G
m
.
151
152 Feynman Motives
As we have seen, the period matrix (2.48) for the promotive Log
∞
of (2.47) deﬁnes the mixed Hodge structure associated to the motive. In
this Hodge realization, the H
0
piece corresponds to the ﬁrst column of the
matrix M
Log
n
, where the kth entry corresponds to the kth graded piece
of the weight ﬁltration.
Let us consider then the corresponding grading operator, that multi
plies the kth entry by T
k
. One can then associate to the h
0
piece of the
Log
∞
motive the following formal expression that corresponds in the period
matrix (2.48) to the H
0
part in the MHS realization:
Q
k
log
k
(s)
k!
T
k
=: Q s
T
. (7.1)
The formal expression (7.1) has in fact an interpretation in terms of
periods. This follows from a well known result (cf. e.g. [Goncharov (2001)],
Lemma 2.10) expressing the powers of the logarithm in terms of iterated
integrals. For iterated integrals we use the notation as in [Goncharov (2001)]
_
b
a
ds
s
◦
ds
s
◦ ◦
ds
s
=
_
a≤s
1
≤···≤s
n
≤b
ds
1
s
1
∧ ∧
ds
n
s
n
. (7.2)
We also denote by Λ
a,b
(n) the domain
Λ
a,b
(n) = ¦(s
1
, . . . , s
n
) [ a ≤ s
1
≤ ≤ s
n
≤ b¦. (7.3)
Lemma 7.1.1. The expression (7.1) is obtained as rational multiples of
the pairing
s
T
=
_
Λ
1,s
(∞)
η(T), (7.4)
with Λ
1,s
(∞) = ∪
n
Λ
1,s
(n) and the form
η(T) :=
n
ds
1
s
1
∧ ∧
ds
n
s
n
T
n
. (7.5)
Proof. The result follows from the basic identity (cf. [Goncharov (2001)],
Lemma 2.10)
_
Λ
a,b
(n)
ds
1
s
1
∧ ∧
ds
n
s
n
=
log
_
b
a
_
n
n!
. (7.6)
Observe then that, in terms of the motivic sheaves deﬁned as in [Arapura
(2008)] (see the discussion in ¸2.4 above), one can associate to the graph
The geometry of DimReg 153
hypersurfaces X
Γ
and the divisors Σ
n
with ∂σ
n
⊂ Σ
n
objects in the Arapura
category of motivic sheaves over G
m
, by considering the morphism
Ψ
Γ
: A
#E
Γ
ˆ
X
Γ
→G
m
(7.7)
deﬁned by the graph polynomial Ψ
Γ
(s) = det(M
Γ
(s)).
Deﬁnition 7.1.2. The category of Feynman motivic sheaves, for a ﬁxed
scalar quantum ﬁeld theory, is the subcategory of the Arapura category of
motivic sheaves over G
m
spanned by the objects of the form
(Ψ
Γ
: A
n
ˆ
X
Γ
→G
m
,
ˆ
Σ
n
(
ˆ
Σ
n
∩
ˆ
X
Γ
), n −1, n −1), (7.8)
with n = #E
int
(Γ) and where Γ ranges over the Feynman graphs of the
given scalar ﬁeld theory, and where
ˆ
Σ
n
= ¦t ∈ A
n
[
i
t
i
= 0¦ is the union
of the coordinate hyperplanes.
The above correspond to the local systems
H
n−1
G
m
(A
n
ˆ
X
Γ
,
ˆ
Σ
n
(
ˆ
Σ
n
∩
ˆ
X
Γ
), Q(n −1)). (7.9)
We can then consider the pullback of the logarithmic motive Log ∈
T/(G
m
) by this morphism, as in the construction of the logarithmic spe
cialization system given in [Ayoub (2007)]. This gives a motive
Log
Γ
:= Ψ
∗
Γ
(Log) ∈ T/(U
Γ
), (7.10)
where U
Γ
= A
#E
Γ
ˆ
X
Γ
.
In terms of periods, for the product of motivic sheaves deﬁned by (2.23)
one has the following.
Lemma 7.1.3. Suppose given σ ⊂ X and σ
⊂ X
, deﬁning relative ho
mology cycles for (X, Σ) and (X
, Σ
), respectively. One then has, for the
ﬁber product (2.23), the period pairing
_
σ×
S
σ
π
∗
X
(ω) ∧ π
∗
X
(η) =
_
σ
ω ∧ f
∗
f
∗
(η), (7.11)
where f : σ → S and f
: σ
→ S are the restrictions of the maps X → S
and X
→S.
Proof. First recall that, when integrating a diﬀerential form over a ﬁber
product, one has the formula
_
X×
S
X
π
∗
X
(ω) ∧ π
∗
X
(η) =
_
X
ω ∧ (π
X
)
∗
π
∗
X
(η) =
_
X
ω ∧ f
∗
f
∗
(η), (7.12)
154 Feynman Motives
which corresponds to the diagram
X
S
X
π
X
zzv
v
v
v
v
v
v
v
v
π
X
$$
I
I
I
I
I
I
I
I
I
X
f
$$
I
I
I
I
I
I
I
I
I
I
X
f
zzt
t
t
t
t
t
t
t
t
t
S
(7.13)
Suppose then given σ ⊂ X such that ∂σ ⊂ Σ and σ
⊂ X
with ∂σ
⊂ Σ
.
One has
∂(σ
S
σ
) = ∂σ
S
σ
∪ σ
S
∂σ
⊂ Σ
S
X
∪ X
S
Σ
,
so that σ
S
σ
deﬁnes a relative homology class in
(X
S
X
, Σ
S
X
∪ X
S
Σ
).
Given elements [ω] ∈ H
·
S
(X, Σ) and [η] ∈ H
·
S
(X
, Σ
), we then apply the
formula (7.12) to the integration on σ
S
σ
and obtain (7.11).
In these terms, the procedure of dimensional regularization can then be
thought of as follows. Consider again the logarithmic (pro)motive, viewed
itself as a motivic sheaf X
Log
∞
→ G
m
over G
m
. One can then take the
product of a Feynman motive
(Ψ
Γ
: A
n
ˆ
X
Γ
→G
m
, Σ
n
(Σ
n
∩
ˆ
X
Γ
), n −1, n −1), (7.14)
by the (pro)motive
(X
∞
Log
→G
m
, Λ
∞
, 0, 0), (7.15)
where Λ
∞
is such that the domain of integration Λ
1,t
(∞) of the period
computation of Lemma 7.1.1 deﬁnes a cycle. The product is then given by
a ﬁber product as in (2.23), namely
Ψ
∗
Γ
(Log
∞
) = (A
n
ˆ
X
Γ
)
G
m
X
Log
∞
//
X
Log
∞
A
n
ˆ
X
Γ
Ψ
Γ
//
G
m
(7.16)
We then have the following interpretation of the dimensionally regular
ized Feynman integrals.
Proposition 7.1.4. The dimensionally regularized Feynman integrals are
periods on the product, in the category of motivic sheaves enlarged to in
clude projective limits, of the Feynman motive (7.14) by the logarithmic
promotive Log
∞
of (7.15).
The geometry of DimReg 155
Proof. Consider the product (7.16), with the two projections
π
X
: Ψ
∗
Γ
(Log
∞
) →A
n
ˆ
X
Γ
π
L
: Ψ
∗
Γ
(Log
∞
) →X
Log
∞
and the form π
∗
X
(α) ∧ π
∗
L
(η(T)), where α is the form involved in the para
metric Feynman integral, and η(T) is the form on X
Log
∞
that gives the
period (7.4). The period computation of Lemma 7.1.1 gives
Ψ
∗
Γ
_
_
Λ
1,s
(∞)
η(T)
_
=
_
Λ
1,Ψ
Γ
(t)
(∞)
η(T) =
n
log(Ψ
Γ
(t))
n
n!
T
n
= Ψ
Γ
(t)
T
.
(7.17)
We then have, by (7.11),
_
σ
n
×
G
m
Λ
1,Ψ
Γ
(t)
(∞)
π
∗
X
(α) ∧ π
∗
L
(η(T)) =
_
σ
n
α ∧ (π
X
)
∗
π
∗
L
(η(T)) =
_
σ
n
Ψ
T
Γ
α.
This is the dimensionally regularized integral, up to replacing the formal
variable T of (7.1) with the complex DimReg variable z.
7.2 The noncommutative geometry of DimReg
We now discuss a diﬀerent approach to the geometry of DimReg, based on
[Connes and Marcolli (2005b)], using spectral triples and noncommutative
geometry and we discuss its points of contact with motivic notions.
The notion of metric space in noncommutative geometry is provided
by spectral triples, see [Connes (1995)]. These consist of data of the form
X = (/, H, T), with / an associative involutive algebra represented as an
algebra of bounded operators on a Hilbert space H, together with a self
adjoint operator T on H, with compact resolvent, and with the property
that the commutators [a, T] are bounded operators on H, for all a ∈ /. A
spectral triple is even if there is a linear involution, γ
2
= 1, on the Hilbert
space H satisfying [a, γ] = 0 and Dγ +γD = 0. Assume for simplicity that
the operator D is invertible. Then the spectral triple is ﬁnitely summable
if, for some real s, the operator [D[
s
is of trace class, Tr([D[
s
) < ∞.
This structure generalizes the data of a compact Riemannian spin mani
fold, with the (commutative) algebra of smooth functions, the Hilbert space
of square integrable spinors and the Dirac operator. It makes sense, how
ever, for a wide range of examples that are not ordinary manifolds, such as
quantum groups, fractals, noncommutative tori, etc.
156 Feynman Motives
For such spectral triples there are various diﬀerent notions of dimension.
The most sophisticated one is the dimension spectrum which is not a single
number but a subset of the complex plane consisting of all poles of the
family of zeta functions associated to the spectral triple,
Dim = ¦s ∈ C[ζ
a
(s) = Tr(a[D[
−s
) have poles ¦.
These are points where one has a well deﬁned integration theory on the
noncommutative space, the analog of a volume form, given in terms of a
residue for the zeta functions.
It is shown in [Connes and Marcolli (2005b)], [Connes and Marcolli
(2008)] that there exists a (type II) spectral triple X
z
with the properties
that the dimension spectrum is Dim = ¦z¦ and that one recoves the DimReg
prescription for the Gaussian integration in the form
Tr(e
−λD
2
z
) = π
z/2
λ
−z/2
.
The operator D
z
is of the form D
z
= ρ(z)F[Z[
1/z
, where Z = F[Z[ is a
selfadjoint operator aﬃliated to a type II
∞
von Neumann algebra ^ and
ρ(z) = π
−1/2
(Γ(1 + z/2))
1/z
, with the spectral measure Tr(χ
[a,b]
(Z)) =
1
2
_
[a,b]
dt, for the type II trace.
The technical aspects of the diﬀerence between ordinary spectral triples
and type II spectral triples are beyond the purpose of this book, and we
simply refer the reader to the relevant literature, especially [Carey, Phillips,
Rennie, Sukochev (2006a)] and [Carey, Phillips, Rennie, Sukochev (2006b)].
The ordinary spacetime over which the quantum ﬁeld theory is con
structed can itself be modeled as a (commutative) spectral triple
X = (/, H, T) = ((
∞
(X), L
2
(X, S), / D
X
)
and one can take a product X X
z
given by the cup product of spectral
triples (adapted to the type II case)
(/, H, T) ∪ (/
z
, H
z
, D
z
) = (/⊗/
z
, H⊗H
z
, T ⊗1 +γ ⊗D
z
).
This agrees with what is known in physics as the Breitenlohner–Maison
prescription to resolve the problem of the compatibility of the chirality
operator γ
5
with the DimReg procedure [Breitenlohner and Maison (1977)].
The Breitenlohner–Maison prescription consists of changing the usual Dirac
operator to a product, which is indeed of the form as in the cup product of
spectral triples,
T ⊗1 +γ ⊗D
z
.
The geometry of DimReg 157
It is shown in [Connes and Marcolli (2005b)] and [Connes and Marcolli
(2008)] that an explicit example of a space X
z
that can be used to perform
dimensional regularization geometrically can be constructed from the ad`ele
class space, the noncommutative space underlying the spectral realization of
the Riemann zeta function in noncommutative geometry (see e.g. [Connes
(1999)] and Chapters 2 and 4 of [Connes and Marcolli (2008)]), by taking
the crossed product of the partially deﬁned action
^ = L
∞
(
ˆ
Z R
∗
) GL
1
(Q)
and the trace
Tr
N
(f) =
_
ˆ
Z×R
∗
f(1, a) da,
with the operator
Z(1, ρ, λ) = λ, Z(r, ρ, λ) = 0, r ,= 1 ∈ Q
∗
.
The type II
∞
factor obtained in this way is dual (in the sense of the duality
introduced in [Connes (1973)]) to the type III
1
factor associated to the
KMS state at “critical temperature” β = 1 on the Bost–Connes system,
see [Bost and Connes (1995)], [Connes and Marcolli (2006a)], and Chapter
3 of [Connes and Marcolli (2008)].
This setting for the geometrization of DimReg provides a way of com
puting the anomalous graphs of quantum ﬁeld theory in terms of index co
cycles in noncommutative geometry, as explained in [Connes and Marcolli
(2005b)] and Chapter 1 of [Connes and Marcolli (2008)]. Here we concen
trate on a related aspect that brings us back to motivic notions, namely an
interesting formal analogy between the evanescent gauge potentials in the
noncommutative geometry of DimReg and the complexes of vanishing and
nearby cycles in algebraic geometry.
Suppose given a spectral triple (/, H, T), which may correspond, as
above, to the ordinary spacetime over which a given quantum ﬁeld theory
is considered. We assume that it is an even spectral triple, so that one has a
chirality operator γ, which is the usual γ
5
in dimension four. Then consider
the algebra
˜
/ generated by / and γ, endowed with a Z/2Z grading with
deg(a) = 1 for a ∈ / and deg(γ) = −1. The usual commutators are then
replaced by graded commutators
[T, a]
−
:= Ta −(−1)
deg(a)
aT (7.18)
We consider then the spectral triple where the Hilbert space and Dirac
operator are given by the product with X
z
,
(H⊗H
z
, T ⊗1 +γ ⊗T
z
)
158 Feynman Motives
and the algebra is
˜
/, acting by a ⊗1. We use the notation
T
= T
z
,
¯
T = T ⊗1,
ˆ
T = γ ⊗T
T
=
¯
T +
ˆ
T, (7.19)
as in [Connes and Marcolli (2005b)]. The lack of invariance of T
under the
chiral gauge transformation γ gives rise to new evanescent gauge potentials
coming from the presence of a nontrivial commutator
[T
, γ]
−
= 2γ
ˆ
T, (7.20)
which disappears when one is back in the case z = 0. These evanescent
gauge potentials ﬁgure prominently in the computations of anomalies in
[Connes and Marcolli (2005b)].
We now show that these evanescent gauge potentials also give rise to
a complex reminiscent of vanishing cycles. We use an approach based on
polarized Hodge–Lefschetz modules, which we have already described in a
very similar form in [Consani and Marcolli (2006)] in the context of the
archimedean cohomology of [Consani (1998)] and its relation to noncom
mutative geometry, [Consani and Marcolli (2004)].
The two geometric settings we will compare are the following. On the
algebrogeometric side, we consider the case of a geometric degeneration of
a family X of smooth algebraic varieties over a disk ∆ ⊂ C. Here X is a
complex analytic manifold of dimension dim
C
X = n+1 and f : X →∆ is a
ﬂat, proper morphism with projective ﬁbers. We assume that the map f is
smooth on X
∗
= XY and that Y is a divisor with normal crossings in X.
On the side of noncommutative geometry, we consider the noncommutative
spaces (/
, H
, D
) obtained by taking the cup product of a spectral triple
(/, H, D) with a noncommutative space X
z
in “complexiﬁed dimension”
z ∈ ∆, with the algebra /
=
˜
/ as above.
We shall see that the complex of gauge potentials on (/
, H
, D
) be
haves in many ways like a complex of forms with logarithmic poles associ
ated to the family X of smooth algebraic varieties over the disk z ∈ ∆. In
particular, in the algebrogeometric case, it is known that the special ﬁber
f
−1
(0) of f : X →∆ carries a mixed Hodge structure, [Steenbrink (1976)],
[Guill´en and Navarro Aznar (1990)].
To describe the analogous structure associated to the complex of gauge
potentials in the case of a noncommutative space and its deformation to
complexiﬁed dimension, we take the point of view of Saito’s polarized
Hodge–Lefschetz modules, [Saito (1988)] and [Guill´en and Navarro Aznar
(1990)], formulated as in [Consani and Marcolli (2006)].
The geometry of DimReg 159
Let L = ⊕
i,j∈Z
L
i,j
be a ﬁnite dimensional bigraded real vector space.
Let
1
,
2
be endomorphisms of L, with [
1
,
2
] = 0 and
1
: L
i,j
→L
i+2,j
,
2
: L
i,j
→L
i,j+2
. (7.21)
The data (L,
1
,
2
) deﬁne a bigraded Lefschetz module if
i
1
: L
−i,j
→L
i,j
and
j
2
: L
i,−j
→L
i,j
(7.22)
are isomorphisms for i > 0 and j > 0, respectively.
In the case of a geometric degeneration, with the generic ﬁber a compact
K¨ahler manifold, one can obtain such a structure from the action of the
Lefschetz operator on the primitive part of the cohomology (induced by
wedging with the K¨ahler form) and of the monodromy: the Lefschetz and
the log of the monodromy give the endomorphisms
1
,
2
.
Bigraded Lefschetz modules correspond bijectively to ﬁnite dimensional
representations of SL(2, R) SL(2, R). We will use the following notation:
χ(λ) :=
_
λ 0
0 λ
−1
_
, u(s) :=
_
1 s
0 1
_
, w :=
_
0 1
−1 0
_
, (7.23)
for λ ∈ R
∗
and s ∈ R, and
u =
_
0 1
0 0
_
=
d
ds
u(s)[
s=0
.
In these terms, the representation σ = σ
(L,
1
,
2
)
satisﬁes
dσ(u, 1) =
1
, dσ(1, u) =
2
, σ(χ(λ), χ(t)) = λ
i
t
j
x, ∀x ∈ L
i,j
.
The data (L,
1
,
2
) deﬁne a bigraded Hodge–Lefschetz module if all the
L
i,j
have a real Hodge structure and
1
,
2
are morphisms of Hodge struc
tures.
A polarization on (L,
1
,
2
) is a bilinear form
ψ : L ⊗L →R (7.24)
which is compatible with the Hodge structure, satisﬁes
ψ(
k
x, y) +ψ(x,
k
y) = 0, k = 1, 2, (7.25)
and is such that
ψ(, C
i
1
j
2
) (7.26)
is symmetric and positive deﬁnite on L
−i,−j
. (Here C is the Weil operator.)
The data (L,
1
,
2
, ψ) then deﬁne a bigraded polarized Hodge–Lefschetz
module.
160 Feynman Motives
It is also convenient (e.g. when working at the level of forms) to consider
diﬀerential bigraded polarized Hodge–Lefschetz modules (L,
1
,
2
, ψ, d),
where
d : L
i,j
→L
i+1,j+1
is a diﬀerential (d
2
= 0) satisfying
[
1
, d] = 0 = [
2
, d]
and
ψ(dx, y) = ψ(x, dy).
In this case, the cohomology H
∗
(L, d) inherits the structure of a bi
graded polarized Hodge–Lefschetz module. Moreover, H
∗
is identiﬁed with
Ker(), where = d
∗
d +dd
∗
, where
d
∗
= σ(w, w)
−1
◦ d ◦ σ(w, w),
for w as in (7.23) and σ = σ
(L,
1
,
2
)
, the representation of SL(2, R)
SL(2, R).
We now introduce an analog of this structure for noncommutative
spaces. We let (/, H, D) be an even ﬁnitely summable spectral triple. Let γ
be the grading operator. We consider the graded algebra
˜
/ generated by /
and by γ. Let (/
, H
, D
) be the noncommutative space obtained as the
product of (
˜
/, H, D) with the noncommutative space X
z
in “complexiﬁed
dimension” z. Namely, we have
/
=
˜
/, H
= H⊗H
and
D
= D
z
= D ⊗1 +γ ⊗D
z
,
where as above D
z
is the Dirac operator on the space X
z
. We let Ω
∗
D
(/)
denote the complex of gauge potentials of a triple (/, H, D), and H
∗
D
(/)
its cohomology.
We begin now by considering the complex Ω
m,r,k
, with diﬀerentials δ :
Ω
m,r,k
→ Ω
m+1,r,k
and δ
: Ω
m,r,k
→ Ω
m,r+1,k
. Here we take Ω
m,r,k
to be
the span of elements of the form
∇
k
(ω)D
2r
, (7.27)
with ω ∈ Ω
m
D
(
˜
/). Here we deﬁne ∇(a) = [D
2
, a] as before, for an element
a ∈ / or a ∈ [D, /], while we set ∇(aγ) := [D
2
, a]γ, for all a ∈ / or
a ∈ [D, /].
The geometry of DimReg 161
We consider a descending ﬁltration ⊃ F
p
⊃ F
p+1
⊃ on Ω
∗
D
(
˜
/)
deﬁned by setting
F
p
Ω
m
D
(
˜
/) = ⊕
t+s=m,t≥p
Ω
t,s
D
(
˜
/), (7.28)
where Ω
t,s
D
(
˜
/) = Ω
t
D
(/)θ
s
is the span of ω θ
s
, with ω ∈ Ω
t
D
(/) and θ = γ
ˆ
D.
We take
F
m+r−k
Ω
m
D
(
˜
/) = ⊕
t+s=m,s+r≤k
Ω
t
D
(/) θ
s
. (7.29)
In this way, we can endow the complex Ω
m,r,k
with a tensor product of two
“Hodge structures”, where the second is 1dimensional with F
−r
D
(C) :=
C D
2r
and F
−(r+1)
D
(C) = 0. We have
Ω
m,r,k
= F
m+r−k
Ω
m
D
(
˜
/) ⊗
C
F
−r
D
(C). (7.30)
The index of the resulting ﬁltration is i−k, where i = 2r+m, hence we then
take the Ω
m,r,k
with the conditions k ≥ 0 and k ≥ 2r +m. The diﬀerential
d = δ + δ
on this complex is the same as described above, induced by
da = [D
, a] for a ∈ ∇
k
(/), when decomposing D
= D ⊗ 1 + γ ⊗ D
z
, so
that δ
a = [
¯
D, a] and δ
(a) = [
ˆ
D, a].
Notice that the decomposition of the total diﬀerential d = δ +δ
, where
δ is essentially the original de Rham diﬀerential on Ω
∗
D
(/) and δ
acts by
wedging with the diﬀerential θ = γ
ˆ
D, resembles very closely the case of
geometric degenerations, where one also has a total diﬀerential d = δ +δ
,
with δ the usual de Rham diﬀerential and δ
given by wedging with the
form θ = f
∗
(dz/z), for f : X →∆ and z the coordinate on the base ∆ (cf.
[Steenbrink (1976)]).
We introduce endomorphisms
1
: Ω
m,r,k
→ Ω
m,r+1,k+1
and
2
:
Ω
m,r,k
→Ω
m+2,r−1,k
of (Ω, d = δ +δ
) deﬁned as follows:
1
(∇
k
(ω)D
2r
) = (∇
k
(ω)D
2r
) = ∇
k+1
(ω)D
2(r−1)
(7.31)
2
(∇
k
(ω)D
2r
) =
√
−1 ∇
k
(ω) ∧ θ
2
D
2(r+1)
=
√
−1 ∇
k
(ω θ
2
) D
2(r+1)
,
(7.32)
where for simplicity we write θ for the term θ = γ
ˆ
D.
We then have the following result.
Lemma 7.2.1. The endomorphisms
1
and
2
satisfy [
1
,
2
] = 0 and are
compatible with the diﬀerential, namely, [
1
, d] = 0 and [
2
, d] = 0.
162 Feynman Motives
Proof. It is immediate to verify that [
1
,
2
] = 0. We check compatibility
with the diﬀerential. We have [
1
, d] = 0. In fact, we can check the result
on elements of the form ∇
k
(a)D
2r
or ∇
k
(γa)D
2r
for a ∈ /. We have
(
1
δ −δ
1
)∇
k
(a)D
2r
= [D
2
, [D, ∇
k
(a)]]D
2(r−1)
−[D, ∇
k+1
(a)]D
2(r−1)
= 0,
since [D
2
, [D, b]] = −[D, [D
2
, b]], for all b ∈ ∇
k
(/). We also have
(
1
δ −δ
1
)∇
k
(γa)D
2r
= [D
2
, [
ˆ
D, γ∇
k
(a)]]D
2(r−1)
−[
ˆ
D, γ∇
k+1
(a)]D
2(r−1)
= 0,
where we use the fact that we have set
∇(γa) := γ∇(a)
and we get
[D
2
, [
ˆ
D, γb]] = [D
2
, 2γb
ˆ
D] = 2[D
2
, b]γ
ˆ
D
and
[
ˆ
D, γ[D
2
, b]] = 2[D
2
, b]γ
ˆ
D.
Similarly, we verify that [
2
, d] = 0. This can be seen easily on elements of
the form bD
2r
, since
[D, b(γ
ˆ
D)
2
] = −[D, b](γ
ˆ
D)
2
,
and on elements γbD
2r
, where
2
[
ˆ
D, γb] = 2
√
−1 bθ
3
= −[
ˆ
D,
2
(γb)].
Thus,
1
,
2
induce endomorphisms on the cohomology of the double
complex.
We introduce an SL(2, R)SL(2, R) representation (σ
1
, σ
2
) on the com
plex introduced above, which is associated to the operators
1
,
2
.
Deﬁnition 7.2.2. Assume that the spectral triple (/, H, D) is N
summable with N = 2n. Consider σ
1
and σ
2
deﬁned by
σ
1
(χ(λ)) = λ
2r+m
, σ
1
(u(s)) = exp(s
1
), σ
1
(w) = S
1
. (7.33)
σ
2
(χ(λ)) = λ
n−m
, σ
2
(u(s)) = exp(s
2
), σ
2
(w) = S
2
. (7.34)
The geometry of DimReg 163
Here the operators S
1
and S
2
are deﬁned by powers of
1
,
2
, in the following
way. We consider involutions
ˆ
S
1
: Ω
m,r,k
→Ω
m,−(r+m),k−(2r+m)
ˆ
S
2
: Ω
m,r,k
→Ω
2n−m,r−(n−m),k
of the form
ˆ
S
1
(∇
k
(ω)D
2r
) =
−(2r+m)
1
(∇
k
(ω)D
2r
)
ˆ
S
2
(∇
k
(ω)D
2r
) =
n−m
2
(∇
k
(ω)D
2r
),
(7.35)
for ω ∈ Ω
m
D
(
˜
/). We set S
1
=
√
−1
m
ˆ
S
1
.
These satisfy the following property.
Proposition 7.2.3. The data speciﬁed in Deﬁnition 7.2.2 deﬁne a repre
sentation of SL(2, R) SL(2, R) on (Ω, d).
Proof. In order to show that σ = σ
k
, for k = 1, 2, deﬁned by (7.33) and
(7.34) is indeed a representations of SL(2, R) it is suﬃcient ([Lang (1975)]
¸XI.2) to check that it satisﬁes the relations
σ(w)
2
= σ(χ(−1)),
σ(χ(λ))σ(u(s))σ(χ(λ
−1
)) = σ(u(sλ
2
)).
(7.36)
The ﬁrst relation is clearly satisﬁed and the second can be veriﬁed easily
on elements ∇
k
(ω)D
2r
, with ω ∈ Ω
m
D
(/), where we have
σ
1
(χ(λ)) σ
1
(u(s)) σ
1
(χ(λ
−1
)) ∇
k
(ω)D
2r
=
σ
1
(χ(λ))
_
1 +s
1
+
s
2
2
2
1
+
_
λ
−(2r+m)
∇
k
(ω)D
2r
=
_
1 +λ
2(r+1)+m
s
1
λ
−(2r+m)
+λ
2(r+2)+m
s
2
2
2
1
λ
−(2r+m)
+
_
∇
k
(ω)D
2r
= exp(sλ
2
1
) ∇
k
(ω)D
2r
,
and
σ
2
(χ(λ)) σ
2
(u(s)) σ
2
(χ(λ
−1
)) ∇
k
(ω)D
2r
=
σ
2
(χ(λ))
_
1 +s
2
+
s
2
2
2
2
+
_
λ
n−m
∇
k
(ω)D
2r
=
_
1 +λ
−n+m+2
s
2
λ
n−m
+λ
−n+m+4
s
2
2
2
2
λ
n−m
+
_
∇
k
(ω)D
2r
=
exp(sλ
2
2
) ∇
k
(ω)D
2r
.
164 Feynman Motives
Notice that, in the case of the representation associated to the “Lef
schetz”
2
, the analog of the Hodge on forms, which appears in σ
1
(w), is
realized by a power of
2
, by analogy to what happens in the classical case,
where the Hodge can be realized (on the primitive cohomology) by the
(n −m)th power of the Lefschetz operator.
This construction parallels exactly what happens in the construction
of the archimedean cohomology of [Consani (1998)] for the ﬁbers at in
ﬁnity of arithmetic varieties, in the form presented in [Consani and Mar
colli (2006)]. This deﬁnes a structure that is analogous to the diﬀeren
tial bigraded Hodge–Lefschetz modules, by setting L
i,j
= ⊕
k
Ω
m,r,k
with
i = 2r +m and j = −n +m.
We now discuss the polarization. Deﬁne the bilinear form ψ by setting
ψ(∇
k
(ω)D
2r
, ∇
k
(ω
)D
2r
) :=
_
−γ∇
k
(η
∗
)∇
k
(η
)D
2(r+r
)
(7.37)
where ω = ηθ
s
and ω
= η
θ
s
.
Lemma 7.2.4. The bilinear form (7.37) satisﬁes the relation (7.25).
Proof. For a = ∇
k
(η)θ
s
D
2r
and b = ∇
k
(η
)θ
s
D
2r
, we have
ψ(
1
(a), b) =
_
−γ∇
k+1
(η
∗
)∇
k
(η
) D
2(r−1+r
)
,
while
ψ(a,
1
(b)) =
_
−γ∇
k
(η)∇
k
+1
(η
) D
2(r+r
−1)
.
Thus, integration by parts gives ψ(
1
(a), b) +ψ(a,
1
(b)) = 0.
We also have
ψ(
2
(a), b) =
_
−γ(−
√
−1 ∇
k
(η
∗
))∇
k
(η
) D
2(r+1+r
)
,
and
ψ(a,
2
(b)) =
_
−γ∇
k
(η
∗
)
√
−1 ∇
k
(η
) D
2(r+r
+1)
.
Thus, we also have ψ(
2
(a), b) +ψ(a,
2
(b)) = 0.
We also have the following result, analogous to the requirement (7.26)
for polarizations of Hodge–Lefschetz modules. We work here under an
assumption of “tameness” for spectral triples deﬁned in [V´arilly and Gracia
Bond
´
ia (1993)].
The geometry of DimReg 165
Lemma 7.2.5. The bilinear form (7.37) has the property that
¸a
∗
, b) := ψ(a,
i
1
j
2
b)
agrees on L
−i,−j
with the integral with respect to the volume form of
(/, H, D). Under the assumptions of “tameness” for (/, H, D), it agrees
with the inner product on the complex of gauge potentials.
Proof. For b = ∇
k
(η
)θ
s
D
2r
in L
−i,−j
, we have
j
2
(b) = ∇
k
(η
)θ
s
+2j
D
2r+2j
.
We then have
i
1
j
2
(b) = ∇
k
+i
(η
)θ
s
+2j
D
2r+2j−2i
.
This gives
i
1
j
2
(b) = ∇
k
+2r+m
(η
)θ
s
+2(m−n)
D
−2r−2n
.
Thus, for a = ∇
k
(η)θ
s
D
2r
in L
−i,−j
, we obtain
ψ(a,
i
1
j
2
(b)) =
_
−γ∇
k
(η
∗
)∇
k
+2r+m
(η
)D
−2n
.
Recall that
_
a :=
_
−aD
−2n
is the integration with respect to the volume
form D
−2n
of the spectral triple (/, H, D). Under the assumption of “tame
ness” this satisﬁes
_
ab =
_
ba and
_
a
∗
a ≥ 0, so that ¸a
∗
, b) agrees with
the inner product of forms in Ω
∗
D
(
˜
/).
The fact that the structure described here in terms of Hodge–Lefschetz
modules parallels very closely the construction of the archimedean coho
mology of [Consani (1998)], in the form presented in [Consani and Marcolli
(2006)], suggests that the formalism of DimReg via noncommutative ge
ometry may be useful also to describe a “neighborhood” of the ﬁbers at
inﬁnity of an arithmetic variety.
In the algebrogeometric setting of a degeneration over a disk, the lo
cal monodromy plays an important role in determining the limiting mixed
Hodge structure. In fact, geometrically, the diﬀerence between the coho
mology of the generic ﬁber and of the special ﬁber is measured by the van
ishing cycles. These span the reduced cohomology
˜
H
∗
(M
z
, C) of the Milnor
ﬁber M
z
:= B
P
∩ f
−1
(z), deﬁned for P ∈ Y , B
P
a small ball around P,
and a suﬃciently small z ∈ ∆
∗
. The nearby cycles span the cohomology
H
∗
(M
z
, C). To eliminate the noncanonical dependence of everything upon
the choice of z, one usually considers all choices by passing to the universal
166 Feynman Motives
cover
˜
∆
∗
= H of the punctured disk and replacing M
z
by
˜
X
∗
= X
∆
˜
∆
∗
.
The identiﬁcation (cf. [Steenbrink (1976)])
H
m
(
˜
X
∗
, C)
→H
m
(Y, Ω
·
X/∆
(log Y ) ⊗
O
X
O
Y
) (7.38)
shows that, when working with nearby cycles, one can use a complex of
forms with logarithmic diﬀerentials. This approach, with an explicit reso
lution of the complex, was used ([Steenbrink (1976)], [Guill´en and Navarro
Aznar (1990)]) to obtain a mixed Hodge structure on H
m
(
˜
X
∗
, C) deter
mined by (H
m
(
˜
X
∗
, C), L
·
, F
·
), where F
·
is the Hodge ﬁltration and L
·
is
the Picard–Lefschetz ﬁltration associated to the local monodromy.
The analogy described above with the structure of polarized Hodge–
Lefschetz modules allows one to think of the operator
Θ(a) =
∞
n=1
(−1)
n+1
n
∇
n
(a)D
−2n
. (7.39)
as the logarithm of the local monodromy, with
1
satisfying e
Θ
= 1+
1
, and
with the induced action of (2π
√
−1)
1
on the cohomology corresponding to
the residue of the Gauss–Manin connection in the algebrogeometric setting.
The operator Θ of (7.39) is the derivation that appears in [Connes and
Moscovici (1995)] in the context of the local index formula for spectral
triples.
Chapter 8
Renormalization, singularities, and
Hodge structures
As we have seen, the graph hypersurfaces X
Γ
are typically singular with
nonisolated singularities and a singular locus of small codimension. Among
various techniques introduced for the study of nonisolated singularities, a
common procedure consists of cutting the ambient space with linear spaces
of dimension complementary to that of the singular locus of the hypersur
face (cf. e.g. [Teissier (1975)]). In this case, the restriction of the function
deﬁning the hypersurface to these linear spaces deﬁnes hypersurfaces with
isolated singularities, to which the usual invariants and constructions for
isolated singularities can be applied.
In many cases one ends up slicing the graph hypersurfaces by planes P
2
intersecting the hypersurface into a curve with isolated singular points. This
corresponds to cases where the singular locus is of (complex) codimension
one in the hypersurface. When the singular locus is of codimension two in
the hypersurface, the slicing is given by 3dimensional spaces cutting the
hypersurface into a family of surfaces in P
3
with isolated singularities.
The parametric Feynman integrals can in turn be expressed in terms of
this slicing, using projective Radon transforms, and we use the resulting
integrals on the slices as a way to relate the Feynman integrals to Hodge
structures for isolated singularities. We follow here the approach outlined
in [Marcolli (2008)].
8.1 Projective Radon transform
We recall the basic setting for integral transforms on projective spaces (cf.
¸II of [Gelfand, Gindikin, and Graev (1980)]). On any kdimensional sub
space A
k
⊂ A
n
there is a unique (up to a multiplicative constant) (k −1)
form that is invariant under the action of SL
k
. It is given as in (3.42) by
167
168 Feynman Motives
the expression
Ω
k
=
k
i=1
(−1)
i+1
t
i
dt
1
∧ ∧
´
dt
i
∧ ∧ dt
k
. (8.1)
The form (8.1) is homogeneous of degree k. Suppose given a function f on
A
n
which satisﬁes the homogeneity condition
f(λt) = λ
−k
f(t), ∀t ∈ A
n
, λ ∈ G
m
. (8.2)
Then the integrand fΩ
k
is well deﬁned on the corresponding projective
space P
k−1
⊂ P
n−1
and one deﬁnes the integral by integrating on a funda
mental domain in A
k
¦0¦, i.e. on a surface that intersects each line from
the origin once.
Suppose given linearly independent dual vectors ξ
i
∈ (A
n
)
, for i =
1, . . . , n − k. These deﬁne a kdimensional linear subspace Π = Π
ξ
⊂ A
n
by the vanishing
Π
ξ
= ¦t ∈ A
n
[ ¸ξ
i
, t) = 0, i = 1, . . . , n −k¦. (8.3)
Given a choice of a subspace Π
ξ
, there exists a (k − 1)form Ω
ξ
on A
n
satisfying
¸ξ
1
, dt) ∧ ∧ ¸ξ
n−k
, dt) ∧ Ω
ξ
= Ω
n
, (8.4)
with Ω
n
the (n −1)form of (3.42), cf. (8.1). The form Ω
ξ
is not uniquely
deﬁned on A
n
, but its restriction to Π
ξ
is uniquely deﬁned by (8.4). Then,
given a function f on A
n
with the homogeneity condition (8.2), one can
consider the integrand fΩ
ξ
and deﬁne its integral over the projective space
π(Π
ξ
) ⊂ P
n−1
as above. This deﬁnes the integral transform, that is, the
(k −1)dimensional projective Radon transform (¸II of [Gelfand, Gindikin,
and Graev (1980)]) as
T
k
(f)(ξ) =
_
π(Π
ξ
)
f(t) Ω
ξ
(t) =
_
P
n−1
f(t)
n−k
i=1
δ(¸ξ
i
, t)) Ω
ξ
(t). (8.5)
For our purposes, it is convenient to also consider the following variant
of the Radon transform (8.5).
Deﬁnition 8.1.1. Let σ ⊂ A
n
be a compact region that is contained in a
fundamental domain of the action of G
m
on A
n
¦0¦. The partial (k −1)
dimensional projective Radon transform is given by the expression
T
σ,k
(f)(ξ) =
_
σ∩π(Π
ξ
)
f(t) Ω
ξ
(t) =
_
σ∩π(Π
ξ
)
f(t)
n−k
i=1
δ(¸ξ
i
, t)) Ω
ξ
(t), (8.6)
where one identiﬁes σ with its image π(σ) ⊂ P
n−1
.
Renormalization, singularities, and Hodge structures 169
Let us now return to the parametric Feynman integrals we are consid
ering.
Proposition 8.1.2. The parametric Feynman integral can be reformulated
as
U(Γ) =
Γ(k −
D
2
)
(4π)
D/2
_
T
σ
n
,k
(f
Γ
)(ξ) ¸ξ, dt), (8.7)
where ξ is an (n − k)frame in A
n
and T
σ
n
,k
(f) is the Radon transform,
with σ
n
the simplex
i
t
i
= 1, t
i
≥ 0, and with
f
Γ
(t) =
V
Γ
(t, p)
−k+D/2
Ψ
Γ
(t)
D/2
. (8.8)
Proof. We write the parametric Feynman integral as
U(Γ) = (4π)
−D/2
_
A
n
χ
+
(t)
e
−V
Γ
(t,p)
Ψ
Γ
(t)
D/2
dt
1
dt
n
, (8.9)
where χ
+
(t) is the characteristic function of the domain R
n
+
.
Given a choice of an (n − k)frame ξ, we can then write the Feynman
integrals in the form
U(Γ) = (4π)
−D/2
_
_
_
Π
ξ
χ
+
(t)
e
−V
Γ
(t,p)
Ψ
Γ
(t)
D/2
ω
ξ
_
¸ξ, dt), (8.10)
where ¸ξ, dt) is a shorthand notation for
¸ξ, dt) = ¸ξ
1
, dt) ∧ ∧ ¸ξ
n−k
, dt)
and ω
ξ
satisﬁes
¸ξ, dt) ∧ ω
ξ
= ω
n
= dt
1
∧ ∧ dt
n
. (8.11)
We then follow the same procedure we used in the derivation of the
parametric form of the Feynman integral, applied to the integral over Π
ξ
and we write it in the form
_
Π
ξ
χ
+
(t)
e
−V
Γ
(t,p)
Ψ
Γ
(t)
D/2
ω
ξ
(t) =
Γ(k −
D
2
)
_
Π
ξ
δ(1 −
i
t
i
)
ω
ξ
(t)
Ψ
Γ
(t)
D/2
V
Γ
(t, p)
k−D/2
.
(8.12)
The function f
Γ
(t) of (8.8) satisﬁes the scaling property (8.2) and the
integrand
ω
ξ
(t)
Ψ
Γ
(t)
D/2
V
Γ
(t, p)
k−D/2
170 Feynman Motives
is therefore G
m
invariant, since the form ω
ξ
is homogeneous of degree k.
Moreover, the domain σ
n
of integration is contained in a fundamental do
main for the action of G
m
. Thus, we can reformulate the integral (8.12) in
projective space, in terms of the Radon transform, as
Γ(k −
D
2
) (4π)
−D/2
_
T
σ
n
,k
(f
Γ
)(ξ) ¸ξ, dt), (8.13)
where T
σ
n
,k
(f
Γ
) is the Radon transform over the simplex σ
n
, as in Deﬁni
tion 8.1.1.
In the following, we will then consider integrals of the form
U(Γ)
ξ
= T
σ
n
,k
(f
Γ
)(ξ) =
_
Π
ξ
δ(1 −
i
t
i
)
ω
ξ
(t)
Ψ
Γ
(t)
D/2
V
Γ
(t, p)
k−D/2
(8.14)
=
_
σ
ξ
ω
ξ
(t)
Ψ
Γ
(t)
D/2
V
Γ
(t, p)
k−D/2
as well as their dimensional regularizations
U(Γ)
ξ
(z) =
_
σ
ξ
ω
ξ
(t)
Ψ
Γ
(t)
(D+z)/2
V
Γ
(t, p)
k−(D+z)/2
, (8.15)
where Π
ξ
is a generic linear subspace of dimension equal to the codimension
of the singular locus of the hypersurface X
Γ
∪ Y
Γ
, with X
Γ
= ¦Ψ
Γ
(t) = 0¦
and Y
Γ
= ¦P
Γ
(t, p) = 0¦ as before, and σ
ξ
= σ
n
∩ Π
ξ
.
So far we have always made the simplifying assumption, in dealing with
the parametric Feynman integrals, that the spacetime dimension D is in the
“stable range” where n ≤ D/2, so that the integral lives on the hypersur
face complement of the graph hypersurface X
Γ
and the zeros of the second
graph polynomial P
Γ
(t, p) do not contribute to singularities of the diﬀer
ential form. Here the corresponding condition would then be k ≤ D/2,
which is also satisﬁed in the original stable range, so we continue to restrict
only to that case.
8.2 The polar ﬁltration and the Milnor ﬁber
Algebraic diﬀerential forms ω ∈ Ω
k
(T(f)) on a hypersurface complement
can always be written in the form ω = η/f
m
as in (3.46), for some m ∈ N
and some η ∈ Ω
k
mdeg(f)
. The minimal m such that ω can be written in the
form ω = η/f
m
is called the order of the pole of ω along the hypersurface
X and is denoted by ord
X
(ω). The order of the pole induces a ﬁltration,
Renormalization, singularities, and Hodge structures 171
called the polar ﬁltration, on the de Rham complex of diﬀerential forms on
the hypersurface complement. One denotes by P
r
Ω
k
P
n
⊂ Ω
k
P
n
the subspace
of forms of order ord
X
(ω) ≤ k − r + 1, if k − r + 1 ≥ 0, or P
r
Ω
k
= 0 for
k −r +1 < 0. The polar ﬁltration P
•
is related to the Hodge ﬁltration F
•
by P
r
Ω
m
⊃ F
r
Ω
m
, by a result of [Deligne and Dimca (1990)].
Proposition 8.2.1. Under the generic condition on the external momenta,
and in the stable range n ≤ D/2, the forms
Ω
ξ
Ψ
D/2
Γ
V
k−D/2
Γ
(8.16)
span subspaces P
r,k
ξ
of the polar ﬁltration P
r
Ω
k−1
P
n−1
of a hypersurface com
plement (f) ⊂ P
n−1
, where f = Ψ
Γ
and for the index r of the ﬁltration
in the range r ≤ 2k −D( + 1)/2.
Proof. We are assuming that P
Γ
and Ψ
Γ
have no common factor, for
generic external momenta. We also have that k ≤ D/2, so that also
k −D( + 1)/2 < 0. In this case, we write (8.16) in the form
∆(α)
f
m
=
P
−k+D/2
Γ
Ω
ξ
Ψ
−k+D(+1)/2
Γ
, (8.17)
where
α = P
−k+D/2
Γ
ω
ξ
, and f = Ψ
Γ
and m = −k +D( +1)/2. (8.18)
We are considering here the polar ﬁltration on forms on the complement of
the hypersurface X
Γ
deﬁned by Ψ
Γ
= 0. Varying the external momenta p
correspondingly varies the polynomial P
Γ
(t, p). We then obtain, for generic
ξ and varying p, a subspace of P
r,k
ξ
of the ﬁltration P
r
Ω
k−1
P
n−1
, for all r ≤
2k −D( + 1)/2.
Suppose then that k = codimSing(X), where Sing(X) is the singular
locus of the hypersurface X = ¦f = 0¦, with f as in Proposition 8.2.1
above. In this case, for generic ξ, the linear space Π
ξ
cuts the singular
locus Sing(X) transversely and the restriction X
ξ
= X ∩ Π
ξ
has isolated
singularities.
Recall that, in the case of isolated singularities, there is an isomorphism
between the cohomology of the Milnor ﬁber F
ξ
of X
ξ
and the total co
homology of the Koszul–de Rham complex of forms (3.46) with the total
diﬀerential d
f
ω = f dω −mdf ∧ ω as in (3.51).
The explict isomorphism is given by the Poincar´e residue map and can
be written in the form
[ω] → [j
∗
∆(ω
ξ
)], (8.19)
172 Feynman Motives
where j : F
ξ
→ Π
ξ
is the inclusion of the Milnor ﬁber in the ambient space
(see [Dimca (1992)], ¸6).
Let M(f) be the Milnor algebra of f, i.e. the quotient of the polynomial
ring in the coordinates of the ambient projective space by the ideal of the
derivatives of f. When f has isolated singularities, the Milnor algebra is
ﬁnite dimensional. One denotes by M(f)
m
the homogeneous component of
degree m of M(f).
It then follows from the identiﬁcation (8.19) above ([Dimca (1992)],¸6.2)
that, in the case of isolated singularities, a basis for the cohomology H
r
(F
ξ
)
of the Milnor ﬁber, with r = dimΠ
ξ
−1 is given by elements of the form
ω
α
=
t
α
∆(ω
ξ
)
f
m
, with t
α
∈ M(f)
mdeg(f)−k
, (8.20)
where f is the restriction to Π
ξ
of the function of Ψ
Γ
. We then have the
following consequence of Proposition 8.2.1.
Corollary 8.2.2. For n ≤ D/2 and for a generic (n − k)frame ξ with
n − k = dimSing(X), with X = X
Γ
, and for a ﬁxed generic choice of the
external momenta p under the generic assumption of Deﬁnition 3.3.1, the
Feynman integrand (8.16) of (8.14) deﬁnes a cohomology class in H
r
(F
ξ
),
with r = dimΠ
ξ
−1 and F
ξ
⊂ Π
ξ
the Milnor ﬁber of the hypersurface with
isolated singularities X
ξ
= X ∩ Π
ξ
⊂ Π
ξ
.
Proof. By Proposition 8.2.1, in the stable range for D, the form (8.16)
can be written as
h∆(ω
ξ
)
f
m
, (8.21)
where f = Ψ
Γ
and h = P
−k+D/2
Γ
. Let 1
ξ
denote the ideal of derivatives of
the restriction f[
Π
ξ
of f to Π
ξ
. Then let
h
ξ
= h mod 1
ξ
. (8.22)
For a ﬁxed generic choice of the external momenta, this deﬁnes an element
in the Milnor algebra M(f[
Π
ξ
), which lies in the homogeneous component
M(f[
Π
ξ
)
mdeg(f)−k
, for m = −k +D( +1)/2. Thus, the form (8.21) deﬁnes
a class in the cohomology H
r
(F
ξ
) with r = dimΠ
ξ
−1.
We have not discussed so far the dependence of our slicing method upon
the choice of the slice. This will be treated in ¸8.4 below, where we show
that one can include the choice of slices as additional data in the Hopf
algebra combinatorics, so as to consider all slices simultaneously, with a
Renormalization, singularities, and Hodge structures 173
combinatorial organizing principle. In particular, notice that in terms of the
Radon transform decomposition (8.10) only the general slices matter as the
special slices (for instance those where Π
ξ
is contained in the hypersurface
X
Γ
) form a set of measure zero in the space of (n −k)frames.
It is also worth pointing out that the construction described in this
section can also be formulated without slicing. Namely, the operations of
taking the ideal of derivatives of f, ﬁnding the class of the numerator mod
ulo these derivatives, and then ranging over all choices of external momenta
to get a subspace can be performed without the slicing operation. What
is missing in this case is the interpretation of the resulting space as the
cohomology of a Milnor ﬁber.
8.3 DimReg and mixed Hodge structures
We assume here to be in the case of isolated singularities, possibly after
replacing the original Feynman integrals with their slices along planes Π
ξ
of
dimension complementary to that of the singular locus of the hypersurface,
as above.
The cohomological Milnor ﬁbration has ﬁber over given by the complex
vector space H
k−1
(F
, C), where the Milnor ﬁber F
of X
ξ
is homotopically
a bouquet of µ spheres S
k−1
, with k = dimΠ
ξ
− 1 and with µ the Milnor
number of the isolated singularity. A holomorphic kform α = hω
ξ
/f
m
determines a section of the cohomological Milnor ﬁbration by taking the
classes
_
α
df
[
F
_
∈ H
k−1
(F
, C). (8.23)
We then have the following results ([Arnold, GuseinZade, Varchenko
(1988)], Vol.II ¸13). The asymptotic formula for the Gelfand–Leray func
tions implies that the function of obtained by pairing the section (8.23)
with a locally constant section of the homological Milnor ﬁbration has an
asymptotic expansion
__
α
df
_
, δ
_
∼
λ,r
a
r,λ
r!
λ
log()
r
, (8.24)
for → 0, where δ() ∈ H
k−1
(F
, Z). Moreover, there exist classes
η
α
r,λ
() ∈ H
k−1
(F
, C) (8.25)
such that the coeﬃcients a
r,λ
of (8.24) are given by
¸η
α
r,λ
(), δ()) = a
r,λ
. (8.26)
174 Feynman Motives
Thus, one deﬁnes the “geometric section” associated to the holomorphic
kform α as
σ(α) :=
r,λ
η
α
r,λ
()
λ
log()
r
r!
. (8.27)
The order of the geometric section σ(α) is deﬁned as being the smallest λ
in the discrete set Ξ ⊂ R such that η
α
0,λ
,= 0. One denotes it by λ
α
. The
principal part of σ(α) is then deﬁned as
σ
max
(α)() :=
λ
α
_
η
α
0,λ
α
+ +
log()
k−1
(k −1)!
η
α
k−1,λ
α
_
, (8.28)
where one knows that
η
α
r,λ
= ^
r
η
α
0,λ
, (8.29)
where ^ is the nilpotent operator given by the logarithm of the unipotent
monodromy, given by
^ = −
1
2πi
log T
with log T =
r≥1
(−1)
r+1
(T −id)
r
/r.
The asymptotic mixed Hodge structure on the ﬁbers of the cohomological
Milnor ﬁbration constructed by Varchenko ([Varchenko (1980)], [Varchenko
(1981)]) has as the Hodge ﬁltration the subspaces F
r
⊂ H
k−1
(F
, C) deﬁned
by
F
r
= ¦[α/df] [ λ
α
≤ k −r −1¦ (8.30)
and as weight ﬁltration W
⊂ H
k−1
(F
, C) the ﬁltration associated to the
nilpotent monodromy operator ^. This mixed Hodge structure has the
same weight ﬁltration as the limiting mixed Hodge structure constructed
by Steenbrink ([Steenbrink (1976)], [Steenbrink (1977)]), but the Hodge
ﬁltration is diﬀerent, though the two agree on the graded pieces of the
weight ﬁltration.
We show that, upon varying the choice of the external momenta p and
of the spacetime dimension D, the corresponding Feynman integrands, in a
neighborhood of an isolated singular point of X
Γ
∩Π
ξ
, determine a subspace
of the cohomology H
k−1
(F
ξ
, C) of the Milnor ﬁber of X
Γ
∩Π
ξ
. This inherits
a Hodge and a weight ﬁltration from the Milnor ﬁber cohomology with its
asymptotic mixed Hodge structure. We concentrate again only on the case
where k −D/2 ≤ 0, so that we can consider, for ﬁxed k, arbitrarily large
values of D ∈ N.
Renormalization, singularities, and Hodge structures 175
Proposition 8.3.1. Consider Feynman integrals, sliced along a linear
space Π
ξ
as in (8.14). We write the integrand in the form
α
ξ
=
hΩ
ξ
f
m
, (8.31)
with
_
¸
¸
_
¸
¸
_
h = P
−k+D/2
Γ
f = Ψ
Γ
m = −k +D( + 1)/2,
(8.32)
as in (8.18), with k −D/2 ≤ 0. Upon varying the external momenta p in
P
Γ
(p, t) and the spacetime dimension D ∈ N, with k −D/2 ≤ 0, the forms
α
ξ
as above determine a subspace
H
k−1
Feynman
(F
, C) ⊂ H
k−1
(F
, C),
of the ﬁbers of the cohomological Milnor ﬁbration, spanned by elements of
the form (8.31), where the polynomials h = h
T,v,w,p
are of the form
h(t) =
−k+D/2
i=1
L
T
i
(t)
e/ ∈T
i
t
e
, (8.33)
where the T
i
are spanning trees and the L
T
i
(t) are the linear functions of
(3.41).
Proof. Consider the explicit expression of the polynomial P
Γ
(t, p) as a
function of the external momenta, through the coeﬃcients s
C
of (3.18). One
can see that, by varying arbitrarily the external momenta, subject to the
global conservation law (3.39), one can reduce to the simplest possible case,
where all external momenta are zero except for a pair of opposite momenta
P
v
1
= p = −P
v
2
associated to a pair of external edges attached to a pair of
vertices v
1
, v
2
. In such a case, the polynomial P
Γ
(t, p) becomes of the form
(3.40). Thus, when considering powers P
Γ
(t, p)
−k+D/2
for varying D, we
obtain all polynomials of the form (8.33).
We denote by H
k−1
Feynman
(F
, C) the subspace of the cohomology
H
k−1
(F
, C) of the Milnor ﬁber spanned by the classes [α
ξ
/df] with α
ξ
of the form (8.31), with h of the form (8.33), considered modulo the
ideal generated by the derivatives of f = Ψ
Γ
and localized at an iso
lated singular point, i.e. viewed as elements in the Milnor algebra M(f).
The subspace H
k−1
Feynman
(F
, C) inherits a Hodge and a weight ﬁltration
176 Feynman Motives
F
•
∩ H
k−1
Feynman
and W
•
∩ H
k−1
Feynman
from the asymptotic mixed Hodge
structure of Varchenko on H
k−1
(F
, C). It is an interesting problem to
see whether the subspace H
k−1
Feynman
recovers the full H
k−1
(F
, C) and if
(F
•
∩H
k−1
Feynman
, W
•
∩H
k−1
Feynman
) still give a mixed Hodge structure, at least
for some classes of graphs Γ.
8.4 Regular and irregular singular connections
An important and still mysterious aspect of the motivic approach to
Feynman integrals and renormalization is the problem of reconciling the
Riemann–Hilbert correspondence of perturbative renormalization formu
lated in [Connes and Marcolli (2004)] (see Chapter 1 of [Connes and Mar
colli (2008)]), based on equivalence classes of certain irregular singular con
nections, with the setting of motives (especially mixed Tate motives) and
mixed Hodge structures, which are naturally related to regular singular
connections. The irregular singular connections of [Connes and Marcolli
(2004)] have values in the Lie algebra of the Connes–Kreimer group of dif
feographisms and are deﬁned on a ﬁbration over a punctured disk with
ﬁber the multiplicative group, respectively representing the complex vari
able z of dimensional regularization and the energy scale µ (or rather µ
z
)
upon which the dimensionally regularized Feynman integrals depend. On
the other hand, in the case of hypersurfaces in projective spaces, the nat
ural associated regular singular connection is the Gauss–Manin connection
on the cohomology of the Milnor ﬁber and the Picard–Fuchs equation for
the vanishing cycles. We sketch here a relation between this regular sin
gular connection and the irregular equisingular connections of [Connes and
Marcolli (2004)].
We explain here, following [Marcolli (2008)], how regular singular
Gauss–Manin connections associated to singularities of individual graph
hypersurfaces X
Γ
can be assembled, over diﬀerent graphs, to give rise to
an irregular singular connection of the kind used in [Connes and Marcolli
(2004)].
In the following we let
_
ω
i
df
_
i = 1, . . . , µ (8.34)
be a basis for the vanishing cohomology bundle, written with the same
notation we used above for the Gelfand–Leray form. Then the Gauss–
Manin connection on the vanishing cohomology bundle, which is deﬁned by
Renormalization, singularities, and Hodge structures 177
the integer cohomology lattice in each real cohomology ﬁber, acts on the
basis (8.34) by
∇
GM
s
_
ω
i
df
_
s
=
j
p
ij
(s)
_
ω
j
df
_
s
, (8.35)
where the p
ij
(s) are holomorphic away from s = 0 and have a pole at s = 0.
The Gauss–Manin connection is regular singular and its monodromy agrees
with the monodromy of the singularity (see [Arnold, Goryunov, Lyashko,
Vasilev (1998)], ¸2.3). Given a covariantly constant section δ(s) of the
vanishing homology bundle, the function
I(s) =
_
_
δ(s)
ω
1
df
, . . . ,
_
δ(s)
ω
µ
df
_
(8.36)
is a solution of the regular singular Picard–Fuchs equation
d
ds
I(s) = P(s)I(s), with P(s)
ij
= p
ij
(s). (8.37)
Similarly, suppose given a holomorphic nform ω and let ω/df be the
corresponding Gelfand–Leray form, deﬁning a section [ω/df] of the vanish
ing cohomology bundle. Let δ
1
, . . . , δ
µ
be a basis of the vanishing homology,
δ
i
(s) ∈ H
n−1
(F
s
, Z). Then the function
I(s) =
_
_
δ
1
(s)
ω
df
, . . . ,
_
δ
µ
(s)
ω
df
_
(8.38)
satisﬁes a regular singular order diﬀerential equation
I
()
(s) +p
1
(s)I
(−1)
(s) + +p
(s)I(s) = 0, (8.39)
where the order is bounded above by the multiplicity of the critical point
(see [Arnold, GuseinZade, Varchenko (1988)], ¸12.2.1). One refers to
(8.39), or to the equivalent system of regular singular homogeneous ﬁrst
order equations
d
ds
1(s) = T(s)1(s), (8.40)
with
1
r
(s) = s
r−1
I
(r−1)
(s), (8.41)
as the Picard–Fuchs equation of ω. For the relation between Picard–Fuchs
equations and mixed Hodge structures see ¸12 of [Arnold, GuseinZade,
Varchenko (1988)] and [Kulikov (1998)].
178 Feynman Motives
We also recall some of the properties of the irregular singular, equi
singular connections of [Connes and Marcolli (2004)] which we need here.
We let g denote the Lie algebra g = Lie(G). Let K denote the ﬁeld of
germs of meromorphic functions at z = 0. We also let B denote a ﬁbration
over an inﬁnitesimal disk ∆
∗
with ﬁber the multiplicative group G
m
and we
denote by P the principal Gbundle P = BG. We consider Lie(G)valued
ﬂat connections ω that are equisingular. As we recalled earlier, this means
that they satisfy the following conditions.
• The connections satisfy ω(z, λu) = λ
Y
ω(z, u), for λ ∈ G
m
, with Y the
grading operator.
• Solutions of Dγ = ω, have the property that their pullbacks σ
∗
(γ) ∈
G(K) along any section σ : ∆ → B with ﬁxed value σ(0) have the same
negative piece of the Birkhoﬀ factorization σ
∗
(γ)
−
.
The ﬁrst condition and the ﬂatness condition imply that the connection
ω(z, u) can be written in the form
ω(z, u) = u
Y
(a(z)) dz +u
Y
(b(z))
du
u
, (8.42)
where a(z) and b(z) are elements of g(K) satisfying the ﬂatness condition
db
dz
−Y (a) + [a, b] = 0. (8.43)
Recall that the Lie bracket in the Lie algebra Lie(G) of the Connes–
Kreimer Hopf algebra is obtained by assigning (see [Connes and Kreimer
(2000)], [Connes and Kreimer (2001)])
[Γ, Γ
] =
v∈V (Γ)
Γ ◦
v
Γ
−
v
∈V (Γ
)
Γ
◦
v
Γ, (8.44)
where Γ◦
v
Γ
denotes the graph obtained by inserting Γ
into Γ at the vertex
v ∈ V (Γ) and the sum is over all vertices where an insertion is possible.
Counting all possible ways in which a given graph can be inserted into
another at a speciﬁed vertex attaches multiplicities to each insertion that
appear in the form of symmetry factors.
The equisingularity condition, which determines the behavior of pull
backs of solutions along sections of the ﬁbration G
m
→ B → ∆, can be
checked by writing the equation Df = ω in the more explicit form
γ
−1
dγ
dz
= a(z), and γ
−1
Y (γ) = b(z). (8.45)
Renormalization, singularities, and Hodge structures 179
When one interprets elements γ ∈ G(K) as algebra homomorphisms φ ∈
Hom(H, K), one can write the above equivalently in the form
(φ ◦ S) ∗
dφ
dz
= a, and (φ ◦ S) ∗ Y (φ) = b, (8.46)
where S is the antipode in H and ∗ is the product dual to the coproduct
in the Hopf algebra. This means, on generators Γ of H,
¸(φ ◦ S) ⊗
dφ
dz
, ∆(Γ)) = a
Γ
, and ¸(φ ◦ S) ⊗Y (φ), ∆(Γ)) = b
Γ
, (8.47)
where
∆(Γ) = Γ ⊗1 + 1 ⊗Γ +
γ
γ ⊗Γ/γ
with the sum over subdivergences, and the antipode is given inductively by
S(X) = −X −
S(X
)X
, (8.48)
for ∆(X) = X ⊗1 +1 ⊗X +
X
⊗X
, with X
and X
of lower degree.
We now show how to produce a ﬂat connection of the desired form
(8.42), with irregular singularities, starting from the graph hypersuraces
X
Γ
, a consistent choice of slicing Π
ξ
, and the regular singular Picard–Fuchs
equation associated to the resulting isolated singularities of X
Γ
∩ Π
ξ
.
We begin by introducing a small modiﬁcation of the Hopf algebra and
coproduct, which accounts for the fact of having to choose a slicing Π
ξ
. This
is similar to what happens when one enriches the discrete Hopf algebra by
adding the data of the external momenta.
Let o
Γ
denote the manifold of planes Π
ξ
in A
#E(Γ)
with dimΠ
ξ
≤
codimSing(X
Γ
). We can write o
Γ
as a disjoint union
o
Γ
=
codimSing(X
Γ
)
_
m=1
o
Γ,m
, (8.49)
where o
Γ,m
is the manifold of mdimensional planes in A
#E(Γ)
. We denote
by (
∞
(o
Γ
) the space of test functions on o
Γ
and by (
−∞
c
(o
Γ
) its dual space
of distributions.
Lemma 8.4.1. Suppose given a subgraph γ ⊂ Γ. Then the choice of a
distribution σ ∈ (
−∞
c
(o
Γ
) induces distributions σ
γ
∈ (
−∞
c
(o
γ
) and σ
Γ/γ
∈
(
−∞
c
(o
Γ/γ
).
180 Feynman Motives
Proof. Given γ ⊂ Γ, neglecting external edges, we can realize the aﬃne
X
γ
as a hypersurface inside a linear subspace A
#E(γ)
⊂ A
#E(Γ)
and sim
ilarly for the aﬃne X
Γ/γ
, seen as a hypersurface inside a linear subspace
A
#E(Γ/γ)
⊂ A
#E(Γ)
, where we simply identify the edges of γ or Γ/γ with a
subset of the edges of the original graph Γ.
One then has a restriction map T
γ
: o
Γ,γ
→ o
γ
, where o
Γ,γ
⊂ o
Γ
is the
union of the components o
Γ,m
of o
Γ
with m ≤ codimSing(X
γ
),
o
Γ,γ
=
codimSing(X
γ
)
_
m=1
o
Γ,m
, (8.50)
which is given by
T
γ
(Π
ξ
) = Π
ξ
∩ A
#E(γ)
. (8.51)
This induces a map T
γ
: (
∞
(o
γ
) → (
∞
(o
Γ
) given by
T
γ
(f)(Π
ξ
) =
_
f(T
γ
(Π
ξ
)) Π
ξ
∈ o
Γ,γ
0 otherwise.
(8.52)
In turn, this deﬁnes a map T
γ
: (
−∞
c
(o
Γ
) → (
−∞
c
(o
γ
), at the level of
distributions, by
T
γ
(σ)(f) = σ(T
γ
(f)). (8.53)
The argument for Γ/γ is analogous. One sets σ
γ
= T
γ
(σ
Γ
) and σ
Γ/γ
=
T
Γ/γ
(σ
Γ
).
We then enrich the original Hopf algebra H by adding the datum of the
slicing Π
ξ
. We consider the commutative algebra
˜
H = Sym((
−∞
c
(o)), (8.54)
where o = ∪
Γ
o
Γ
, endowed with the coproduct
∆(Γ, σ) = (Γ, σ) ⊗1 + 1 ⊗(Γ, σ) +
γ
(γ, σ
γ
) ⊗(Γ/γ, σ
Γ/γ
). (8.55)
The following result is then proved by the same argument used in
[Connes and Marcolli (2008)], Theorem 1.27.
Lemma 8.4.2. The coproduct (8.55) is coassociative and
˜
H is a Hopf al
gebra.
Renormalization, singularities, and Hodge structures 181
We then proceed as follows. We pass to the projective instead of the
aﬃne formulation and we ﬁx a small neighborhood of an isolated singular
point of X
Γ
∩ Π
ξ
, for Π
ξ
a linear space of dimension at most equal to the
codimension of Sing(X
Γ
). Suppose given a holomorphic kform α
ξ
on Π
ξ
.
Then there exists an associated regular singular Picard–Fuchs equation
J
()
Γ,ξ
(s) +p
1
(s)J
(−1)
Γ,ξ
(s) + +p
(s)J
Γ,ξ
(s) = 0, (8.56)
with the property that any solution J
Γ,ξ
(s) is a linear combination of the
functions
J
Γ,ξ,i
(s) =
_
δ
i
(s)
α
ξ
df
, (8.57)
where δ
1
, . . . , δ
µ
is a basis of locally constant sections of the homological
Milnor ﬁbration, δ
i
(s) ∈ H
k−1
(F
s
, Z), and α
ξ
/df is the Gelfand–Leray form
associated to the holomorphic kform α
ξ
.
This depends on the choice of a singular point and can be localized in a
small neighborhood of the singular point in X
Γ
∩ Π
ξ
. In fact, introducing
a cutoﬀ χ
ξ
that is supported near the singularities of X
Γ
∩ Π
ξ
amounts
to adding the expressions (8.57) for the diﬀerent singular points. Thus,
to simplify notation, we can just assume that we have a single expression
J
Γ,ξ
(s) at a unique isolated critical point.
We then have the following result, which constructs irregular singular
connections from solutions of the regular singular Picard–Fuchs equation.
Theorem 8.4.3. Any solution J
Γ,ξ
of the regular singular Picard–Fuchs
equation (8.56) determines a ﬂat g(K)valued connection ω(z, u) of the form
(8.42). Moreover, if the kform α
ξ
is given by P
Γ
Ω
ξ
as above, then the
connection is equisingular.
Proof. We consider the Mellin transform
T
Γ,ξ
(z) =
_
∞
0
s
z
J
Γ,ξ
(s) ds. (8.58)
As in ¸7 of [Arnold, GuseinZade, Varchenko (1988)], the function T
Γ,ξ
(z)
admits an analytic continuation to a meromorphic function with poles at
points z = −(λ+1) with λ ∈ Ξ
Γ,ξ
a discrete set in R of points related to the
multiplicities of an embedded resolution of the singular point of X
Γ
∩ Π
ξ
.
We look at the function T
Γ,ξ
(z) in a small neighborhood of a chosen point
z = −D. It has an expansion as a Laurent series, with a pole at z = −D if
−D ∈ Ξ
Γ,ξ
.
182 Feynman Motives
After a change of variables on the complex coordinate z, so that we have
z ∈ ∆
∗
a small neighborhood of z = 0, we deﬁne
φ
µ
(Γ, σ)(z) := µ
−z b
1
(Γ)
σ
_
T
Γ,ξ
(−
D +z
2
)
_
, (8.59)
where we consider T
Γ,ξ
as a function of ξ to which we apply the distribution
σ. More precisely, after identifying T
Γ,ξ
with its Laurent series expansion,
we apply σ to the coeﬃcients seen as functions of ξ. This deﬁnes an algebra
homomorphism φ
µ
∈ Hom(
˜
H, K), by assigning the values (8.59) on genera
tors. Here µ is the mass scale. The homomorphism φ deﬁned by (8.59) can
be equivalently described as a family of
˜
G(C)valued loops γ
µ
: ∆
∗
→
˜
G(C),
depending on the mass scale µ. Here
˜
G denotes the aﬃne group scheme
dual to the commutative Hopf algebra
˜
H. The dependence on µ of (8.59)
implies that γ
µ
satisﬁes the scaling property
γ
e
t
µ
(z) = θ
tz
(γ
µ
(z)), (8.60)
where θ
t
is the oneparameter family of automorphisms of
˜
H generated by
the grading,
d
dt
θ
t
[
t=0
= Y . Then one sets
a
µ
(z) := (φ
µ
◦ S) ∗
d
dz
φ
µ
, and b
µ
(z) := (φ
µ
◦ S) ∗ Y (φ
µ
), (8.61)
where S and ∗ are the antipode of
˜
H and the product dual to the coproduct
∆ of (8.55). These deﬁne elements a
µ
, b
µ
Ω
1
(g(K)), which one can use to
deﬁne a connection ω(z, u) of the form (8.42). More precisely, for µ = e
t
,
one has
γ
−1
µ
d
dz
γ
µ
= θ
t
(γ
−1
d
dz
γ) = u
Y
(a(z)),
where we set u
Y
= e
tY
and then extend the resulting expression to
u ∈ G
m
(C) = C
∗
. Similarly, we get γ
−1
µ
Y (γ
µ
) = u
Y
(b(z)). Thus, the
connection ω(z, u) deﬁned in this way satisﬁes by construction the ﬁrst
condition of the equisingularity property, namely ω(z, λu) = λ
Y
ω(z, u), for
all λ ∈ G
m
. One can see that the connection is ﬂat since we have
d
dz
b
µ
(z) −Y (a
µ
(z)) =
dγ
−1
µ
(z)
dz
Y (γ
µ
(z)) +γ
−1
µ
(z)
d
dz
(Y (γ
µ
(z)))
−Y (γ
−1
µ
(z))
d
dz
γ
µ
(z) −γ
−1
µ
(z)
d
dz
(Y (γ
µ
(z)))
= −γ
−1
µ
(z)
d
dz
(γ
µ
(z))γ
−1
µ
(z)Y (γ
µ
(z)) −γ
−1
µ
(z)Y (γ
µ
(z))γ
−1
µ
(z)
Renormalization, singularities, and Hodge structures 183
= −[a(z), b(z)].
The second condition of equisingularity is the property that, in the
Birkhoﬀ factorization
γ
µ
(z) = γ
µ,−
(z)
−1
γ
µ,+
(z),
the negative part satisﬁes
d
dµ
γ
µ,−
(z) = 0.
By dimensional analysis on the counterterms, in the case of dimensional
regularization and minimal subtraction, it is possible to show (see [Collins
(1986)] ¸5.8.1) that the counterterms obtained by the BPHZ procedure
applied to the Feynman integral U
µ
(Γ)(z) do not depend on the mass pa
rameter µ. This means, as shown in [Connes and Kreimer (2000)] (see also
Proposition 1.44 of [Connes and Marcolli (2008)]), that the Feynman in
tegrals U
µ
(Γ)(z) deﬁne a G(C)valued loop γ
µ
(z) with the property that
∂
µ
γ
µ,−
(z) = 0. The integrals (8.58) considered here correspond to slices
along a linear space Π
ξ
of the Feynman integrals, localized by a cutoﬀ χ
ξ
near the singular points. The explicit dependence on µ in the Feynman
integrals is as in (8.59), which is unchanged with respect to that of the
original dimensionally regularized Feynman integrals. Thus, the same ar
gument as in [Collins (1986)] ¸5.8.1 and Proposition 1.44 of [Connes and
Marcolli (2008)] applies to this case to show that ∂
µ
γ
µ,−
(z) = 0.
Chapter 9
Beyond scalar theories
In this ﬁnal chapter we give a few examples of how the theme of Feynman
integrals and motives, based on the parametric representation of Feynman
integrals and the associated algebraic varieties, can be extended beyond
the original setting of scalar quantum ﬁeld theories. We concentrate here
only on two simple kinds of extensions, where one can already see some
new feature appearing. The ﬁrst extension we discuss is to theories with
fermionic ﬁelds, based on the results of [Marcolli and Rej (2008)]. In this
case one can again express a parametric form of the Feynman integral
in terms of periods, this time on a supermanifold, with the determinant
deﬁning the graph polynomial replaced by a Berezinian, which is a natural
analog in the supergeometry context. The second case we discuss is that
of scalar quantum ﬁeld theories, and in particular the φ
4
theory, over a
spacetime that is deformed to a noncommutative manifold with a Moyal
product. In this case it is also known that a parametric formulation of
Feynman integrals exists, and we report some observations of [Aluﬃ and
Marcolli (2008a)] on computations of characteristic classes for the resulting
graph hypersurfaces.
9.1 Supermanifolds
We describe here an extension of the parametric Feynman integral and
its motivic interpretation to the case of theories with both fermionic and
bosonic variables, where we can reformulate the parametric integral as a
period of a supermanifold, deﬁned in terms of Berezinians instead of de
terminants and fermionic integration. The results reported here are from
[Marcolli and Rej (2008)].
We report a few basic facts about the geometry of supermanifolds, fol
185
186 Feynman Motives
lowing [Manin (1988)].
By a complex supermanifold one understands a datum X = (X, A)
with the following properties: A is a sheaf of supercommutative rings
on X; (X, O
X
) is a complex manifold, where O
X
= A/N, with N the
ideal of nilpotents in A; the quotient E = N/N
2
is locally free over O
X
and A is locally isomorphic to the exterior algebra Λ
•
O
X
(E), where the
grading is the Z/2Zgrading by odd/even degrees. The supermanifold is
split if the isomorphism A
∼
= Λ
•
O
X
(E) is global. Thus, in local coordi
nates, a supermanifold is parameterized by bosonic and fermionic variables
(x
0
, . . . , x
n
; θ
1
, . . . , θ
m
), where the x
i
are ordinary, commuting, scalar vari
ables, while the θ
j
are anticommuting Grassmann variables. The bosonic
variables x
i
generate the order zero, or even degree, part and the fermionic
variables θ
j
the order one, or odd degree part with respect to the Z/2Z
grading.
A simple example is given by the projective superspaces. The complex
projective superspace P
nm
is the supermanifold (X, A) with X = P
n
the
usual complex projective space and
A = Λ
•
(C
m
⊗
C
O(−1)),
with the exterior powers Λ
•
graded by odd/even degree. It is a split super
manifold.
A morphism F : X
1
→ X
2
of supermanifolds X
i
= (X
i
, A
i
), i = 1, 2,
consists of a pair F = (f, f
#
) of a morphism of the underlying complex
manifolds f : X
1
→ X
2
together with a morphism f
#
: A
2
→ f
∗
A
1
of
sheaves of supercommutative rings with the property that at each point
x ∈ X
1
the induced morphism f
#
x
: (A
2
)
f(x)
→ (A
1
)
x
satisﬁes f
#
x
(m
f(x)
) ⊂
m
x
, on the maximal ideals of germs of sections vanishing at the point (cf.
[Manin (1988)], §4.1).
In particular, an embedding of complex supermanifolds is a morphism
F = (f, f
#
) as above, with the property that f : X
1
→ X
2
is an embedding
and f
#
: A
2
→ f
∗
A
1
is surjective. As in ordinary geometry, we deﬁne the
ideal sheaf of X
1
to be the kernel
I
X
1
:= Ker(f
#
: A
2
→ f
∗
A
1
). (9.1)
An equivalent characterization of an embedding of supermanifold is
given as follows. If we denote by E
i
, for i = 1, 2 the holomorphic vec
tor bundles on X
i
such that O(E
i
) = E
i
= N
i
/N
2
i
, with the notation as
above, then an embedding F : X
1
→ X
2
is an embedding f : X
1
→ X
2
such
Beyond scalar theories 187
that the induced morphism of vector bundles f
∗
: E
2
→ E
1
is surjective.
Thus, we say that Y = (Y, B) is a closed subsupermanifold of X = (X, A)
when there exists a closed embedding Y ⊂ X and the pullback of E
A
under
this embedding surjects to E
B
.
An open submanifold U = (U, B) → X = (X, A) is given by an open em
bedding U → X of the underlying complex manifolds and an isomorphism
of sheaves A
U
∼
= B. When Y ⊂ X is a closed embedding and U = X Y ,
the ideal sheaf of Y satisﬁes I
Y

U
= A
U
.
A subvariety in superprojective space is a supermanifold
X = (X ⊂ P
n
, (Λ
•
(C
m
⊗
C
O(−1))/I)
X
), (9.2)
where I = I
X
is an ideal generated by ﬁnitely many homogeneous
polynomials of given Z/2parity. In other words, if we denote by
(x
0
, . . . , x
n
, θ
1
, . . . , θ
m
) the bosonic and fermionic coordinates of P
nm
, then
a projective subvariety can be obtained by assigning a number of equations
of the form
Ψ
ev/odd
(x
0
, . . . , x
n
, θ
1
, . . . , θ
m
) =
i
1
<···<i
k
P
I
(x
0
, . . . , x
n
)θ
i
1
· · · θ
i
k
= 0,
(9.3)
where I = (i
1
, . . . , i
k
), and the P
I
(x
0
, . . . , x
n
) are homogeneous polynomials
in the bosonic variables.
In [Marcolli and Rej (2008)] an analog of the Grothendieck ring of va
rieties is considered for supermanifolds, deﬁned as follows.
Deﬁnition 9.1.1. Let SV
C
be the category of complex supermanifolds
with morphisms deﬁned as above. Let K
0
(SV
C
) denote the free abelian
group generated by the isomorphism classes of objects X ∈ SV
C
subject
to the following relations. Let F : Y → X be a closed embedding of
supermanifolds. Then
[X] = [Y] + [X Y], (9.4)
where X Y is the supermanifold
X Y = (X Y, A
X

XY
). (9.5)
Here the notation A
X
means the following, see [Kashiwara and Schapira
(1990)] §II.2.3. Given a locally closed subset Y ⊂ X and a sheaf A on X,
there exists a unique sheaf A
Y
with the property that
A
Y

Y
= A
Y
and A
Y

XY
= 0. (9.6)
188 Feynman Motives
In the case where Y is closed, this satisﬁes A
Y
= i
∗
(A
Y
) where i : Y → X
is the inclusion. When Y is open it satisﬁes A
Y
= Ker(A → j
∗
(A
XY
)),
where in this case j : X Y → X is the closed embedding of the comple
ment.
In particular, in the case where A = O
X
is the structure sheaf of X,
the relation (9.4) reduces to the usual relation
[X] = [Y ] + [X Y ]. (9.7)
in the Grothendieck group of ordinary varieties, for a closed embedding
Y ⊂ X.
Proposition 9.1.2. The Grothendieck ring K
0
(SV
C
) of supervarieties is
a polynomial ring over the Grothendieck ring of ordinary varieties of the
form
K
0
(SV
C
) = K
0
(V
C
)[T], (9.8)
where T = [A
01
] is the class of the aﬃne superspace of dimension (0, 1).
Proof. We show that all supermanifolds decompose in K
0
(SV
C
) as a
ﬁnite combination of split supermanifolds, and in fact of supermanifolds
where the vector bundle E with O(E) = E = N/N
2
is trivial. This is a
consequence of the d´evissage of coherent sheaves, see [Shafarevich (1996)],
§3.3, Propositions 1–3. Namely, for any coherent sheaf A over a Noetherian
reduced irreducible scheme there exists a dense open set U such that A
U
is free. The relation (9.4) then ensures that, given a pair X = (X, A) and
the sequence of sheaves
0 → i
!
(A
U
) → A → j
∗
(A
Y
) → 0,
associated to the open embedding U ⊂ X with complement Y = X U,
the class [X, A] satisﬁes
[X, A] = [U, A
U

U
] + [Y, A
Y

Y
].
The sheaf A
Y
on X, which has support Y , has a chain of subsheaves A
Y
⊃
A
1
⊃ · · · ⊃ A
k
= 0 such that each quotient A
i
/A
i+1
is coherent on Y ,
[Shafarevich (1996)], §3.3, Proposition 3. Thus, one can ﬁnd a stratiﬁcation
where on each open stratum the supermanifold is split and with trivial
vector bundle. The supermanifold X = (X, A) decomposes as a sum of the
corresponding classes in the Grothendieck group. It is then clear that the
product makes K
0
(SV
C
) into a ring with
[X][Y] = [X ×Y],
Beyond scalar theories 189
and the stated result then follows. The fact that it is indeed a polynomial
ring in the class [A
01
] without further relation is a consequence of the fact
that A
01
is parameterized by a single Grassmann variable and that the
A
0n
are nonisomorphic supermanifolds.
The fact that the vector bundle that constitutes the fermionic part of a
supermanifold is trivial when seen in the Grothendieck group is the analog
for supermanifolds of the fact that any locally trivial ﬁbration is equivalent
to a product in the Grothendieck group of ordinary varieties.
Notice that, in the supermanifold case, there are now two diﬀerent types
of Lefschetz motives, namely the bosonic one L
b
= [A
10
] and the fermionic
one L
f
= [A
01
]. The formal inverses of L
f
and L
b
correspond to two
types of Tate objects for motives of supermanifolds, respectively fermionic
and bosonic Tate motives. The contribution of the fermionic part of a
supermanifold to the class in the Grothendieck ring is always of (fermonic)
Tate type, while only the bosonic part can provide nonTate contributions.
The analog of the determinant in supergeometry is given by the
Berezinian. This is deﬁned in the following way. Suppose given a matrix
M of the form
M =
_
M
11
M
12
M
21
M
22
_
,
where the M
11
and M
22
are square matrices with entries of order zero and
the M
12
and M
21
have elements of order one. Then the Berezinian of M
is the expression
Ber(M) :=
det(M
11
−M
12
M
−1
22
M
21
)
det(M
22
)
. (9.9)
It satisﬁes Ber(MN) = Ber(M)Ber(N).
Grassmann variables integration is deﬁned in terms of the Berezinian
integral, which is determined by the property that, for a form ω =
I
f
I
(x
0
, . . . , x
n
)θ
I
in the local coordinates (x
0
, . . . , x
n
; θ
1
, . . . , θ
m
), the
integral gives (see Chapter 4, §6 of [Manin (1988)])
_
ω =
_
f
1,...,m
(x
0
, . . . , x
n
) dx
1
· · · dx
n
, (9.10)
or equivalently by the property that
_
(h
0
(x) +h
1
(x)θ) dθ = h
1
(x), (9.11)
190 Feynman Motives
for x an ordinary and θ a Grassmann variable. It is shown in [Bernˇste˘ın
and Le˘ıtes (1977)] that Grassmann integration satisﬁes a change of vari
able formula where the Jacobian of the coordinate change is given by the
Berezinian Ber(J) with J the matrix J =
∂X
α
∂Y
β
with X
α
= (x
i
, ξ
r
) and
Y
β
= (y
j
, η
s
).
There is also a well developed theory of divisors on supermanifolds,
originating from [Rosly, Schwarz, Voronov (1988)]. A Cartier divisor on
a supermanifold X of dimension (nm) is deﬁned by a collection of even
meromorphic functions φ
i
deﬁned on an open covering U
i
→ X, with φ
i
φ
−1
j
a holomorphic function on U
i
∩ U
j
nowhere vanishing on the underlying
U
i
∩U
j
. Classes of divisors correspond to equivalence classes of line bundles
and can be described in terms of integer linear combinations of (n −1m)
dimensional subvarieties Y ⊂ X.
9.2 Parametric Feynman integrals and supermanifolds
Consider now the case of Feynman propagators and Feynman diagrams
that come from theories with both bosonic and fermionic ﬁelds. It is well
known that for such theories the form of the possible fermionic terms in
the Lagrangian is severely constrained: for instance, in dimension D = 4
renormalizable interaction terms can only involve at most two fermion and
one boson ﬁeld, see [Ramond (1990)], §5.3.
The perturbative expansion for such theories corresponsingly involve
graphs Γ with two diﬀerent types of edges: fermionic and bosonic edges.
The Feynman rules assign to each bosonic edge a propagator of the usual
form we have encountered so far, and to fermionic edges a propagator
i
/ p +m
p
2
−m
2
=
i
/ p −m
. (9.12)
More generally, in cases with tensor indices, we would have / p = p
µ
γ
µ
,
involving γmatrices expressing fermions in spinorial form.
In the following we use the notation
q(p) = p
2
−m
2
, / q(p) = i(/ p +m) (9.13)
for the quadratic and linear forms that appear in the bosonic and fermionic
propagators. In the following, again just to simplify notation, we also
drop the mass terms in the propagator, i.e. we set m = 0, and ignore, for
simplicity, the resulting infrared divergence problem.
Beyond scalar theories 191
Thus, the terms of the form (q
1
· · · q
n
)
−1
, which we encountered in the
purely bosonic case, are now replaced by terms of the form
/ q
1
· · · / q
f
q
1
· · · q
n
, (9.14)
where n = #E(Γ) is the total number of edges in the graph and f = #E
f
(Γ)
is the number of fermionic edges.
We have an analog of the Feynman trick we saw in the ﬁrst chapter,
where we now introduce an extra set of Grassmann variables associated
to the fermionic edges. The derivation we present suﬀers from a kind of
“fermion doubling problem”, as each fermionic edge contributes an ordinary
integration variable, which essentially account for the denominator q
i
in
(9.12) and (9.14), as well as a pair of Grassmann variables accounting for
the numerator / q
i
in (9.12) and (9.14). Other formulations of parametric
Feynman integrals for theories with fermionic variables are possible, for
example as done in [Medvedev (1978)].
Let Q
f
denote the 2f ×2f antisymmetric matrix
Q
f
=
_
_
_
_
_
_
_
_
_
_
_
_
0 / q
1
0 0 · · · 0 0
−/ q
1
0 0 0 · · · 0 0
0 0 0 / q
2
· · · 0 0
0 0 −/ q
2
0 · · · 0 0
.
.
. · · ·
.
.
.
0 0 0 0 · · · 0 / q
f
0 0 0 0 · · · −/ q
f
0
_
_
_
_
_
_
_
_
_
_
_
_
. (9.15)
Lemma 9.2.1. Let σ
n2f
denote the superspace σ
n
× A
02f
. Then the fol
lowing identity holds:
/ q
1
· · · / q
f
q
1
· · · q
n
= K
n,f
_
σ
n2f
dt
1
· · · dt
n−1
dθ
1
· · · dθ
2f
(t
1
q
1
+· · · t
n
q
n
+
1
2
θ
t
Q
f
θ)
n−f
, (9.16)
with
K
n,f
=
2
f
(n −1)!
f
k=1
(−n +f −k + 1)
.
Proof. We ﬁrst show that the following identity holds for integration in
the Grassmann variables θ = (θ
1
, . . . , θ
2f
):
_
dθ
1
· · · dθ
2f
(1 +
1
2
θ
t
Q
f
θ)
α
=
f!
2
f
_
−α
f
_
/ q
1
· · · / q
f
. (9.17)
192 Feynman Motives
In fact, we expand using the Taylor series
(1 +x)
β
=
∞
k=0
_
β
k
_
x
k
and the identity
1
2
θ
t
Q
f
θ =
f
i=1
/ q
i
θ
2i−1
θ
2i
,
together with the fact that the degree zero variables x
i
= θ
2i−1
θ
2i
commute,
to obtain
(1 +
1
2
θ
t
Q
f
θ)
−α
=
∞
k=0
_
−α
k
_
(
f
i=1
/ q
i
θ
2i−1
θ
2i
)
k
.
The rules of Grassmann integration then imply that only the coeﬃcient of
θ
1
· · · θ
2f
remains as a result of the integration. This gives (9.17).
For simplicity of notation, we then write T = t
1
q
1
+· · · t
n
q
n
, so that we
have
_
σ
n2f
1
(t
1
q
1
+· · · t
n
q
n
+
1
2
θ
t
Q
f
θ)
n−f
dt
1
· · · dt
n−1
dθ
1
· · · dθ
2f
=
f!
2
f
_
−n +f
f
_
/ q
1
· · · / q
f
_
σ
n
T
−n+f
T
−f
dt
1
· · · dt
n−1
=
f!
2
f
_
−n +f
f
_
/ q
1
· · · / q
f
_
σ
n
dt
1
· · · dt
n−1
(t
1
q
1
+· · · +t
n
q
n
)
n
=
f!
2
f
(n −1)!
_
−n +f
f
_
/ q
1
· · · / q
f
q
1
· · · q
n
.
Consider now the case of graphs that have both bosonic and fermionic
legs. We mimic the procedure described earlier in this chapter to obtain
a parametric form for the Feynman integral. We will deal here with a
particular class of graph, which can be considered an equivalent, in the
case with fermionic variables, of the log divergent graphs, for which the
parametric form of the Feynman integral is simpliﬁed by the lack of the
second graph polynomial (see §3.1). As we are going to see below, here the
condition that characterizes the class of graphs with this property will no
longer depend only on the number of internal edges and loops but also on
a particular choice of a basis of loops.
Beyond scalar theories 193
Theorem 9.2.2. Suppose given a graph Γ with n edges, of which f are
fermionic and b = n − f bosonic. Assume that there exists a choice of a
basis for H
1
(Γ) satisfying the condition
n −
f
2
+
D
2
(
f
−
b
) = 0. (9.18)
Then the following identity holds:
_
/ q
1
· · · / q
f
q
1
· · · q
n
d
D
s
1
· · · d
D
s
b
d
D
σ
1
· · · d
D
σ
f
=
_
σ
n
Λ(t)
Ber(M(t))
D/2
dt
1
· · · dt
n
.
(9.19)
Proof. We divide the edge indices i = 1, . . . , n of the Feynman graph into
two sets i
b
= 1, . . . , n
b
and i
f
= 1, . . . , n
f
, with n = n
b
+ n
f
, respectively
labeling the bosonic and fermionic legs. Consequently, given a choice of a
basis for the ﬁrst homology of the graph, indexed as above by r = 1, . . . , ,
we replace the circuit matrix η
ir
of (3.1) with a matrix of the form
_
η
i
f
r
f
η
i
f
r
b
η
i
b
r
f
η
i
b
r
b
_
. (9.20)
Here the loop indices r = 1, . . . , are at ﬁrst divided into three
sets {1, . . . ,
ff
}, labelling the loops consisiting of only fermionic edges,
{1, . . . ,
bb
} labelling the loops consisting of only bosonic edges, and the
remaning variables {1, . . . ,
bf
= −(
ff
+
bb
)} for the loops that contain
both fermionic and bosonic edges. We then introduce two sets of momen
tum variables: ordinary variables / s
r
b
∈ A
D0
, with r
b
= 1, . . . ,
b
=
bb
+
bf
,
and Grassmann variables σ
r
f
∈ A
0D
with r
f
= 1, . . . ,
f
=
ff
+
bf
. That
is, we assign to each purely fermionic loop a Grassmann momentum vari
able, to each purely bosonic loop an ordinary momentum variable, and to
the loops containing both fermionic and bosonic legs a pair (/ s
r
, σ
r
) of an
ordinary and a Grassman variable. In (9.20) above we write r
f
and r
b
, re
spectively, for the indexing sets of these Grassmann and ordinary variables.
We then consider a change of variables
/ p
i
b
= / u
i
b
+
r
f
η
i
b
r
f
σ
r
f
+
r
b
η
i
b
r
b
/ s
r
b
, / p
i
f
= / u
i
f
+
r
f
η
i
f
r
f
σ
r
f
+
r
b
η
i
f
r
b
/ s
r
b
.
(9.21)
analogous to the one used before, where now, for reasons of homogeneity,
we need to assume that the η
ir
f
are of degree one and the η
ir
b
are of degree
zero, since the / p
i
are even (ordinary) variables.
We apply the change of variables (9.21) to the expression
i
t
i
p
2
i
+
i
f
θ
2i
f
−1
θ
2i
f
/ p
i
f
. (9.22)
194 Feynman Motives
We assume again, as in the purely bosonic case (cf. (18.35) of [Bjorken
and Drell (1965)]), the relations
i
t
i
/ u
i
η
ir
= 0
for each loop variable r = r
b
and r = r
f
.
We can then rewrite (9.22) in the form
i
t
i
u
2
i
+
i
f
θ
2i
f
−1
θ
2i
f
/ u
i
f
+
r
b
,r
b
(
i
t
i
η
ir
b
η
ir
b
)/ s
r
b
/ s
r
b
−
r
f
r
f
(
i
t
i
η
ir
f
η
ir
f
)σ
r
f
σ
r
f
+
r
b
r
f
_
(
i
t
i
η
ir
b
η
ir
f
)/ s
r
b
σ
r
f
−σ
τ
r
f
/ s
τ
r
b
(
i
t
i
η
ir
f
η
ir
b
)
_
+
r
b
(
i
f
θ
2i
f
−1
θ
2i
f
η
i
f
r
b
)/ s
r
b
+
r
f
(
i
f
θ
2i
f
−1
θ
2i
f
η
i
f
r
f
)σ
r
f
.
Notice the minus sign in front of the quadratic term in the σ
r
f
, since for
orderone variables σ
r
f
η
ir
f
= −η
ir
f
σ
r
f
. We write the above in the simpler
notation
T + / s
τ
M
b
(t)/ s −σ
τ
M
f
(t)σ +σ
τ
M
fb
(t)/ s − / s
τ
M
bf
(t)σ +N
b
(θ)/ s +N
f
(θ)σ,
(9.23)
where τ denotes transposition, / s = (/ s
r
b
), σ = (σ
r
f
), and
T =
i
t
i
u
2
i
+
i
f
θ
2i
f
−1
θ
2i
f
/ u
i
f
,
M
b
(t) =
i
t
i
η
ir
b
η
ir
b
,
M
f
(t) =
i
t
i
η
ir
f
η
ir
f
= −M
f
(t)
τ
,
M
fb
(t) =
i
t
i
η
ir
b
η
ir
f
,
N
b
(θ) =
i
f
θ
2i
f
−1
θ
2i
f
η
i
f
r
b
,
N
f
(θ) =
i
f
θ
2i
f
−1
θ
2i
f
η
i
f
r
f
.
(9.24)
Since the η
i,r
f
are of degree one and the η
i,r
b
of degree zero, the matrices
M
b
and M
f
are of degree zero, the M
bf
and M
fb
of degree one, while the
N
b
and N
f
are, respectively, of degree zero and one. Thus, the expression
(9.23) is of degree zero. Notice that, since the η
ir
f
are of order one, the
matrix M
f
(t) is antisymmetric. We also set M
bf
(t) = M
fb
(t) = M
fb
(t)
τ
.
Beyond scalar theories 195
We then consider an integral of the form
_
d
D
/ s
1
· · · d
D
/ s
b
d
D
σ
1
· · · d
D
σ
f
(
i
t
i
p
2
i
+
i
f
θ
2i
f
−1
θ
2i
f
/ p
i
f
)
n−f
=
_
d
D
/ s
1
· · · d
D
/ s
b
d
D
σ
1
· · · d
D
σ
f
R(/ s, σ, θ, t)
n−f
, (9.25)
where
R(/ s, σ, θ, t) = T + / s
τ
M
b
(t)/ s +N
b
(θ)/ s −σ
τ
M
f
(t)σ
+σ
τ
M
fb
(t)/ s − / s
τ
M
fb
(t)
τ
σ +N
f
(θ)σ
(9.26)
and where the integrations in the variables d
D
σ
i
= dσ
i1
· · · dσ
iD
are Grass
mann variables integrations satisfying the properties of the Berezinian in
tegral (9.10) and (9.11), while the integrations in d
D
/ s
i
are ordinary inte
grations.
Recall that for Grassmann variables we have the following change of
variable formula. Suppose given an invertible antisymmetric N ×N matrix
A with entries of degree zero and an Nvector J with entries of degree one.
Then we have
σ
τ
Aσ +
1
2
(J
τ
σ −σ
τ
J) = η
τ
Aη +
1
4
J
τ
A
−1
J, (9.27)
for η = σ −
1
2
A
−1
J. This follows immediately since A
τ
= −A and we have
η
τ
Aη = σ
τ
Aσ +
1
2
J
τ
σ −
1
2
σ
τ
J −
1
4
J
τ
A
−1
J.
We then use this change of variable to write
−σ
τ
M
f
(t)σ +σ
τ
M
fb
(t)/ s − / s
τ
M
fb
(t)
τ
σ +
1
2
(σ
τ
N
f
(θ) −N
f
(θ)
τ
σ) =
−η
τ
M
f
(t)η −
1
4
(M
fb
(t)/ s +
1
2
N
f
(θ))
τ
M
f
(t)
−1
(M
fb
(t)/ s +
1
2
N
f
(θ))
(9.28)
with
η = σ −
1
2
M
f
(t)
−1
_
M
fb
(t)/ s +
1
2
N
f
(θ)
_
. (9.29)
We have
1
4
(M
fb
(t)/ s +
1
2
N
f
(θ))
τ
M
f
(t)
−1
(M
fb
(t)/ s +
1
2
N
f
(θ)) =
1
4
/ s
τ
M
bf
(t)M
f
(t)
−1
M
fb
(t)/ s
196 Feynman Motives
+
1
8
(N
f
(θ)
τ
M
f
(t)
−1
M
fb
(t)/ s + / s
τ
M
bf
(t)M
f
(t)
−1
N
f
(θ))
+
1
16
N
f
(θ)
τ
M
f
(t)
−1
N
f
(θ).
We then let
U(t, θ, / s) := T +C(t, θ) + / s
τ
A
b
(t)/ s +B
b
(t, θ)/ s, (9.30)
where
A
b
(t) = M
b
(t) −
1
4
M
bf
(t)M
f
(t)
−1
M
fb
(t)
B
b
(t, θ) = N
b
(θ) −
1
4
N
f
(θ)
τ
M
f
(t)
−1
M
fb
(t)
C(t, θ) = −
1
16
N
f
(θ)
τ
M
f
(t)
−1
N
f
(θ).
(9.31)
Thus, we write the denominator of (9.25) in the form
U(t, θ, / s)
n−f
_
1 +
1
2
η
τ
X
f
(t, θ, / s)η
_
n−f
, (9.32)
where we use the notation
X
f
(t, θ, / s) := 2U(t, θ, / s)
−1
M
f
(t). (9.33)
Thus, the Grassmann integration in (9.25) gives, as in Lemma 9.2.1,
_
d
D
η
1
· · · d
D
η
f
_
1 +
1
2
η
τ
X
f
(t, θ, / s)η
_
n−f
= C
n,f,
f
2
D
f
/2
U(t, θ, / s)
D
f
/2
det(M
f
(t))
D/2
,
(9.34)
where C
n,f,
f
is a combinatorial factor, as we have already encountered in
the case of ordinary parametric Feynman integrals for ordinary variables.
We then proceed to the remaining ordinary integration in (9.25). We
have, dropping a multiplicative constant,
det(M
f
(t))
D/2
_
d
D
/ s
1
· · · d
D
/ s
b
U(t, θ, / s)
n−f+D
f
/2
. (9.35)
We use the change of variables / v = / s +
1
2
M
b
(t)
−1
N
b
(θ)
τ
. We then have
/ v
τ
A
b
(t)/ v = / s
τ
A
b
(t)/ s +
1
2
/ s
τ
B
b
(t, θ)
τ
+
1
2
B
b
(t, θ)/ s +
1
4
B
b
(t, θ)A
b
(t)
−1
B
b
(t, θ)
τ
,
(9.36)
where A
b
(t)
τ
= A
b
(t) and (B
b
(t, θ)/ s)
τ
= B
b
(t, θ)/ s.
We then rewrite (9.35) in the form
det(M
f
(t))
D/2
_
d
D
/ v
1
· · · d
D
/ v
b
(T +C −
1
4
B
b
A
−1
b
B
τ
b
+ / v
τ
A
b
/ v)
n−f+D
f
/2
. (9.37)
Beyond scalar theories 197
Set then
˜
T(t, θ) = T(t, θ) +C(t, θ) −
1
4
B
b
(t, θ)A
−1
b
(t)B
b
(t, θ)
τ
, (9.38)
so that we write the above as
det(M
f
(t))
D/2
˜
T(t, θ)
n−f+D
f
/2
_
d
D
/ v
1
· · · d
D
/ v
b
(1 + / v
τ
X
b
(t, θ)/ v)
n−f+D
f
/2
,
with
X
b
(t, θ) =
˜
T(t, θ)
−1
A
b
(t).
Then, up to a multiplicative constant, the integral gives
˜
T
−n+f−
D
f
2
+
D
b
2
det(M
f
(t))
D/2
det(A
b
(t))
D/2
. (9.39)
Consider ﬁrst the term
det(M
f
(t))
D/2
det(A
b
(t))
D/2
in (9.39) above. This can be identiﬁed with a Berezinian. In fact, we have
det(M
f
(t))
D/2
det(M
b
(t) −
1
4
M
fb
(t)M
f
(t)
−1
M
fb
(t))
D/2
= Ber(M(t))
−D/2
, (9.40)
where
M(t) =
_
M
b
(t)
1
2
M
fb
(t)
1
2
M
bf
(t) M
f
(t)
_
. (9.41)
We now look more closely at the remaining term
˜
T
−n+f−
D
f
2
+
D
b
2
in
(9.39). We know from (9.38), (9.31), and (9.24) that we can write
˜
T(t, θ)
in the form
˜
T(t, θ) =
i
u
2
i
t
i
+
j
/ u
i
θ
2j−1
θ
2j
+
i<j
C
ij
(t)θ
2i−1
θ
2i
θ
2j−1
θ
2j
, (9.42)
where the ﬁrst sum is over all edges and the other two sums are over
fermionic edges. We set λ
i
= θ
2i−1
θ
2i
. Using a change of variables
˜
λ
i
= λ
i
+
1
2
C/ u,
we rewrite the above as
˜
T(t, θ) =
i
u
2
i
t
i
−
1
4
/ u
τ
C/ u +
i<j
C
ij
η
2i−1
η
2i
η
2j−1
η
2j
,
198 Feynman Motives
with
˜
λ
i
= η
2i−1
η
2i
.
We write
ˆ
T(t) =
i
u
2
i
t
i
−
1
4
/ u
τ
C/ u
and we write
˜
T
−α
=
ˆ
T
−α
∞
k=0
_
−α
k
_
_
1
2
˜
λ
τ
C
˜
λ
ˆ
T
_
k
,
where we use the notation
1
2
˜
λ
τ
C
˜
λ =
i<j
C
ij
η
2i−1
η
2i
η
2j−1
η
2j
.
Thus, we can write the Feynman integral in the form
_
/ q
1
· · · / q
f
q
1
· · · q
n
d
D
s
1
· · · d
D
s
b
d
D
σ
1
· · · d
D
σ
f
=
κ
_
Σ
n2f
Λ(t)η
1
· · · η
2f
ˆ
T(t)
n−
f
2
+
D
2
(
f
−
b
)
Ber(M(t))
D/2
dt
1
· · · dt
n
dη
1
· · · dη
2f
, (9.43)
where Λ(t) is
ˆ
T
f/2
times the coeﬃcient of η
1
· · · η
2f
in the expansion
∞
k=0
_
−α
k
_
_
1
2
˜
λ
τ
C
˜
λ
ˆ
T
_
k
.
More explicitly, this term is of the form
Λ(t) =
C
i
1
i
2
(t) · · · C
i
f−1
i
f
(t),
over indices i
a
with i
2a−1
< i
2a
and for k = f/2. The multiplicative
constant in front of the integral on the right hand side above is given by
κ =
_
−n +f −
D
2
(
f
−
b
)
f/2
_
.
After imposing n −
f
2
+
D
2
(
f
−
b
) = 0 and performing the Grassmann
integration of the resulting term
_
σ
n2f
Λ(t)η
1
· · · η
2f
Ber(M(t))
D/2
dt
1
· · · dt
n
dη
1
· · · dη
2f
, (9.44)
the result follows directly from (9.43).
Beyond scalar theories 199
9.3 Graph supermanifolds
The result of the previous section shows that we have an analog of the
period integral
_
σ
n
dt
1
· · · dt
n
det(M
Γ
(t))
D/2
given by the similar expression
_
σ
n
Λ(t)
Ber(M(t))
D/2
dt
1
· · · dt
n
. (9.45)
Again we see that, in this case, divergences arise from the intersections
between the domain of integration given by the simplex σ
n
and the variety
deﬁned by the solutions of the equation
Ber(M(t))
D/2
Λ(t)
= 0. (9.46)
Lemma 9.3.1. For generic graphs, the set of zeros of (9.46) deﬁnes a
hypersurface in P
n
, hence a divisor in P
n−12f
of dimension (n−22f). The
support of this divisor is the same as that of the principal divisor deﬁned
by Ber(M(t)).
Proof. The generic condition on graphs is imposed to avoid the cases
with M
f
(t) ≡ 0. Thus, suppose given a pair (Γ, B) that is generic, in the
sense that M
f
(t) is not identically zero. The equation (9.46) is satisﬁed by
solutions of
det(M
b
(t) −
1
4
M
bf
(t)M
f
(t)
−1
M
fb
(t)) = 0 (9.47)
and by poles of Λ(t). (The factor 1/4 in (9.47) comes from the change of
variable formula (9.27).) Using the formulae (9.31) and (9.24) we see that
the denominator of Λ(t) is given by powers of det(M
f
(t)) and det(A
b
(t)) =
det(M
b
(t) −
1
4
M
bf
(t)M
f
(t)
−1
M
fb
(t)). Thus, the set of solutions of (9.46)
is the union of zeros and poles of Ber(M(t)). The multiplicities are given
by the powers of these determinants that appear in Λ(t)Ber(M(t))
−D/2
.
Deﬁnition 9.3.2. Let Γ be a graph with bosonic and fermionic edges and
B a choice of a basis of H
1
(Γ). We denote by X
(Γ,B)
⊂ P
n−12f
the locus
of zeros and poles of Ber(M(t)) = 0. We refer to X
(Γ,B)
as the graph
supermanifold.
200 Feynman Motives
In the degenerate cases of graphs such that M
f
(t) ≡ 0, we simply set
X
(Γ,B)
= P
n−12f
. Examples of this sort are provided by data (Γ, B) such
that there is only one loop in B containing fermionic edges. Other special
cases arise when we consider graphs with only bosonic or only fermionic
edges. In the ﬁrst case, we go back to the original calculation without
Grassmann variables and we therefore simply recover
X
(Γ,B)
= X
Γ
= {t : det(M
b
(t)) = 0} ⊂ P
n−10
.
In the case with only fermionic edges, we have det(M
b
(t) −
1
4
M
bf
(t)M
f
(t)
−1
M
fb
(t)) ≡ 0 since both M
b
(t) and M
bf
(t) are identically
zero. It is then natural to simply assume that, in such cases, the graph
supermanifold is simply given by X
(Γ,B)
= P
f−12f
.
1
t
2
t
3
t
5
t
4
t
Consider the example of the ﬁgure, with a choice of the basis of loops
satisfying the condition (9.18). It is easy to see that such a choice is possible:
it has two generators, one of them a loop made of fermionic edges and the
second a loop containing both fermionic and bosonic edges. Let us assign
the ordinary variables t
i
with i = 1, . . . , 5 to the edges as in the ﬁgure
above. We then have
M
b
(t) = t
1
+t
2
+t
3
since only the second loop in the basis contains bosonic edges, while we
have
M
bf
(t) = (t
1
+t
2
, t
1
+t
2
+t
3
)
= t
1
(1, 1) +t
2
(1, 1) +t
3
(0, 1) +t
4
(0, 0) +t
5
(0, 0)
and
M
f
(t) =
_
0 t
1
+t
2
−(t
1
+t
2
) 0
_
.
Thus, we obtain in this case
M
bf
(t)M
f
(t)
−1
M
fb
(t) =
Beyond scalar theories 201
(t
1
+t
2
, t
1
+t
2
+t
3
)
_
0
−1
t
1
+t
2
1
t
1
+t
2
0
_
_
t
1
+t
2
t
1
+t
2
+t
3
_
= (t
1
+t
2
, t
1
+t
2
+t
3
)
_
−(t
1
+t
2
+t
3
)
t
1
+t
2
1
_
= −(t
1
+t
2
+t
3
) +t
1
+t
2
+t
3
≡ 0.
Thus, in this particular example we have M
bf
(t)M
f
(t)
−1
M
fb
(t) ≡ 0 for all
t = (t
1
, . . . , t
5
), so that
Ber(M(t)) = det(M
b
(t)) det(M
f
(t))
−1
= (t
1
+t
2
+t
3
)/(t
1
+t
2
)
2
and the locus of zeros and poles X
(Γ,B)
⊂ P
58
is the union of t
1
+t
2
+t
3
= 0
and t
1
+ t
2
= 0 in P
5
(the latter counted with multiplicity two), with the
restriction of the sheaf from P
58
.
It is also shown in [Marcolli and Rej (2008)] that the universality result
of [Belkale and Brosnan (2003a)] implies a similar result for the super
manifold case, using the simple relation of Proposition 9.1.2 between the
Grothendieck ring of ordinary manifolds and of supermanifolds.
9.4 Noncommutative ﬁeld theories
Scalar quantum ﬁeld theories, especially φ
4
theories, over a noncommuta
tive spacetime have been investigated extensively in recent years, especially
in relation to their renormalization properties. Such theories arise as eﬀec
tive limits of string theory [Connes, Douglas, Schwarz (1998)], [Seiberg and
Witten (1999)], where the underlying spacetime is deformed to a Moyal
type noncommutative space.
We do not give here a detailed exposition of these theories, and we re
fer the reader to the lectures [Grosse and Wulkenhaar (2008)] for a nice
and detailed overview. We only recall that in dimension D = 4, when
the underlying R
4
is made noncommutative by deformation to R
4
θ
with the
Moyal product, the theory suﬀers from a new type of divergences, coming
from UV/IR (ultraviolet/infrared) mixing. These render the theory non
renormalizable. However, perturbative renormalization can be restored in
the φ
4
theory in diﬀerent ways. In the Grosse–Wulkenhaar model [Grosse
and Wulkenhaar (2005)], this is done by the addition of a harmonic oscilla
tor term that is quite natural from the point of view of the Moyal product,
but which breaks the translation invariance of the theory. More recently, a
202 Feynman Motives
diﬀerent approach to restoring renormalization for the φ
4
theory on Moyal
spaces, which maintains translation invariance, was introduced in [Gurau,
Magnen, Rivasseu, Tanasa (2009)]. Both models are renormalizable to all
orders.
What interests us most here is the fact that, in these theories, it is also
possible to obtain a parametric representation of Feynman integrals. Due
to the fact that the underlying spacetime coordinates no longer commute,
ordinary Feynman graphs are replaced in this context by ribbon graphs and
the planarity or nonplanarity of the ribbon graphs plays an important role
in the properties of the corresponding Feynman integrals. We concentrate
on some special examples among the parametric representation described
in [Gurau and Rivasseu (2007)], [Tanasa (2007)].
Among the family of “banana graphs”, whose motivic properties were
described in [Aluﬃ and Marcolli (2008a)], the only ones that can appear
as Feynman graphs of a φ
4
theory are the oneloop case (with two external
edges at each vertex), the twoloop case (with one external edge at each
vertex) and the threeloop case as a vacuum bubble. Excluding the vac
uum bubble because of the presence of the polynomial P
Γ
(t), we see that
the eﬀect of making the underlying spacetime noncommutative turns the
remaining two graphs into the following ribbon graphs. (Notice how the
twoloop ribbon graph now has two distinct versions, only one of which is
a planar graph.)
! !
2
3
" "
The usual Kirchhoﬀ polynomial Ψ
Γ
(t) of the Feynman graph, as well as
the polynomial P
Γ
(t, p), are replaced by new polynomials involving pairs of
spanning trees, one in the graph itself and one in another associated graph
Beyond scalar theories 203
which is a dual graph in the planar case and that is obtained from an em
bedding of the ribbon graph on a Riemann surface in the more general
case. Unlike the commutative case, these polynomials are no longer homo
geneous, hence the corresponding graph hypersurface only makes sense as
an aﬃne hypersurface. The relation of the hypersurface for the noncom
mutative case and the one of the original commutative case (also viewed as
an aﬃne hypersurface) is given by the following statement.
Proposition 9.4.1. Let
˜
Γ be a ribbon graph in the noncommutative φ
4

theory that corresponds in the ordinary φ
4
theory to a graph Γ with n inter
nal edges. Then instead of a single graph hypersurface
ˆ
X
Γ
one has a one
parameter family of aﬃne hypersurfaces
ˆ
X
˜
Γ,s
⊂ A
n
, where the parameter
s ∈ R
+
depends upon the deformation parameter θ of the noncommutative
R
4
θ
and on the parameter Ω of the harmonic oscillator term in the Grosse–
Wulkenhaar model. The hypersurface corresponding to the value s = 0 has
a singularity at the origin 0 ∈ A
n
whose tangent cone is the (aﬃne) graph
hypersurface
ˆ
X
Γ
.
Proof. This follows directly from the relation between the graph poly
nomial for the ribbon graph
˜
Γ given in [Gurau and Rivasseu (2007)] and
the Kirchhoﬀ polynomial Ψ
Γ
. It suﬃces to see that (a constant multiple
of) the Kirchhoﬀ polynomial is contained in the polynomial for
˜
Γ for all
values of the parameter s, and that it gives the part of lowest order in the
variables t
i
when s = 0.
In the speciﬁc examples of the banana graphs
˜
Γ
2
and
˜
Γ
3
of the ﬁg
ure above, the polynomials have been computed explicitly in [Gurau and
Rivasseu (2007)] and they are of the form
Ψ
˜
Γ
2
= (1 + 4s
2
)(t
1
+t
2
+t
2
1
t
2
+t
1
t
2
2
), (9.48)
where the parameter s = (4θΩ)
−1
is a function of the deformation pa
rameter θ ∈ R of the Moyal plane and of the parameter Ω in the harmonic
oscillator term in the Grosse–Wulkenhaar action functional (see [Grosse and
Wulkenhaar (2008)]). One can see the polynomial Ψ
Γ
2
(t) = t
1
+t
2
appear
ing as lowest order term. Similarly for the two graphs
˜
Γ
3
that correspond
to the banana graph Γ
3
one has ([Gurau and Rivasseu (2007)])
Ψ
˜
Γ
3
(t) = (1 + 8s
2
+ 16s
4
)(t
1
t
2
+t
2
t
3
+t
1
t
3
+t
2
1
t
2
t
3
+t
1
t
2
2
t
3
+t
1
t
2
t
2
3
)
+ 16s
2
(t
2
2
+t
2
1
t
2
3
)
(9.49)
204 Feynman Motives
for the planar case, while for the nonplanar case one has
Ψ
˜
Γ
3
(t) = (1 + 8s
2
+ 16s
4
)(t
1
t
2
+t
2
t
3
+t
1
t
3
+t
2
1
t
2
t
3
+t
1
t
2
2
t
3
+t
1
t
2
t
2
3
)
+ 4s
2
(t
2
2
+t
2
1
t
2
3
+t
2
1
+t
2
2
t
2
3
+t
2
3
+t
2
1
t
2
2
+ 1 +t
2
1
t
2
2
t
2
3
).
(9.50)
In both cases, one readily recognizes the polynomial Ψ
Γ
3
(t) = t
1
t
2
+t
2
t
3
+
t
1
t
3
as the lowest order part at s = 0. Notice how, when s = 0, one ﬁnds
other terms of order less than or equal to that of the polynomial Ψ
Γ
3
(t),
such as t
2
2
in (9.49) and 1 + t
2
1
+ t
2
2
+ t
2
3
in (9.50). Notice also how, at the
limit value s = 0 of the parameter, the two polynomials for the two diﬀerent
ribbon graphs corresponding to the third banana graph Γ
3
agree.
For each value of the parameter s = (4θΩ)
−1
one obtains in this way
an aﬃne hypersurface, which is a curve in A
2
or a surface in A
3
, and that
has the corresponding aﬃne
ˆ
X
Γ
n
as tangent cone at the origin in the case
s = 0. The latter is a line in the n = 2 case and a cone on a nonsingular
conic in the case n = 3.
Following [Aluﬃ and Marcolli (2008a)], we now look at what happens
to the invariants such as the Chern–Schwartz–MacPherson (CSM) class of
the graph hypersurfaces, when passing to the nocommutative case. An
interesting observation one obtains in this way is that the CSM classes
detect certain special values of the deformation parameter s = (4θΩ)
−1
where the hypersurface
ˆ
X
˜
Γ
3
becomes more singular and gives a quantitative
estimate of the amount by which the singularities change.
The CSM class is naturally deﬁned for projective varieties. In the case
of an aﬃne hypersurface deﬁned by a nonhomogeneous equation, one can
choose to compactify it in projective space by adding an extra variable
and making the equation homogeneous and then computing the CSM class
of the corresponding projective hypersurface. However, in doing so one
should be aware of the fact that the CSM class of an aﬃne variety, deﬁned
by choosing an embedding in a larger compact ambient variety, depends on
the choice of the embedding. An intrinsic deﬁnition of CSM classes for non
compact varieties which does not depend on the embedding was given in
[Aluﬃ (2006)]. However, for our purposes here it suﬃces to take the simpler
viewpoint of making the equation homogeneous and then computing CSM
classes with the Macaulay2 program of [Aluﬃ (2003)].
Proposition 9.4.2. Let
ˆ
X
˜
Γ
3
denote the aﬃne surface deﬁned by the equa
tion (9.49) and let
¯
X
˜
Γ
3
⊂ P
3
be the hypersurface obtained by making
the equation (9.49) homogeneous. For general values of the parameter
Beyond scalar theories 205
s = (4θΩ)
−1
the CSM class is given by
c(
¯
X
˜
Γ
3
) = 14H
3
+ 4H. (9.51)
For the special value s = 1/2 of the parameter, the CSM class becomes of
the form
c(
¯
X
˜
Γ
3
)
s=1/2
= 5H
3
+ 5H
2
+ 4H, (9.52)
while in the limit s → 0 one has
c(
¯
X
˜
Γ
3
)
s=0
= 11H
3
+ 4H. (9.53)
It is also interesting to notice that, when we consider the second equa
tion (9.50) for the nonplanar ribbon graph associated to the third banana
graph Γ
3
, we see an example where the graph hypersurfaces of the non
planar graphs of noncommutative ﬁeld theory no longer satisfy the positiv
ity property conjectured for the the graph hypersurfaces of the commutative
ﬁeld theories in [Aluﬃ and Marcolli (2008a)]. In fact, as in the case of the
equation for the planar graph (9.49), we now ﬁnd the following result.
Proposition 9.4.3. Let X
˜
Γ
3
⊂ P
3
denote the aﬃne surface deﬁned by the
equation (9.50) and let
¯
X
˜
Γ
3
⊂ P
3
be the hypersurface obtained by making
the equation (9.50) homogeneous. For general values of the parameter s =
(4θΩ)
−1
the CSM class is given by
c(
¯
X
˜
Γ
3
) = 33H
3
−9H
2
+ 6H. (9.54)
The special case s = 1/2 is given by
c(
¯
X
˜
Γ
3
)
s=1/2
= 9H
3
−3H
2
+ 6H (9.55)
Notice that, in the case of ordinary Feynman graphs of commutative
scalar ﬁeld theories, all the examples where the CSM classes of the corre
sponding hypersurfaces have been computed explicitly (either theoretically
or numerically) are planar graphs. Although it seems unlikely that pla
narity will play a role in the conjectured positivity of the coeﬃcients of the
CSM classes in the ordinary case, the example above showing that CSM
classes of graph hypersurfaces of nonplanar ribbon graphs in noncommu
tative ﬁeld theories can have negative coeﬃcients makes it more interesting
to check the case of nonplanar graphs in the ordinary case as well. It is
well known that, for an ordinary graph to be nonplanar, it or a quotient
by a subgraph has to contain a copy of either the complete graph K
5
on
ﬁve vertices or the complete bipartite graph K
3,3
on six vertices. Either
one of these graphs corresponds to a graph polynomial that is currently
206 Feynman Motives
beyond the reach of the available Macaulay2 routine and a theoretical ar
gument that provides a more direct approach to the computation of the
corresponding CSM class does not seem to be easily available. It remains
an interesting question to compute these CSM classes, especially in view of
the fact that the original computations of [Broadhurst and Kreimer (1997)]
of Feynman integrals of graphs appear to indicate that the nonplanarity
of the graph is somehow related to the presence of more interesting periods
(e.g. multiple as opposed to simple zeta values). It would be interesting to
see whether it also has an eﬀect on invariants such as the CSM class.
Bibliography
S. Agarwala, The Geometry of Renormalization. PhD Thesis, Johns Hopkins Uni
versity, 2009.
P. Aluﬃ, Limits of Chow groups, and a new construction of ChernSchwartz
MacPherson classes. Pure Appl. Math. Q. 2 (2006), no. 4, part 2, 915–941.
P. Aluﬃ, Computing characteristic classes of projective schemes. J. Symbolic
Comput. 35 (2003), no. 1, 3–19.
P. Aluﬃ, M. Marcolli, Feynman motives of banana graphs, Communications in
Number Theory and Physics, Vol.3 (2009) N.1, 1–57.
P. Aluﬃ, M. Marcolli, Algebrogeometric Feynman rules, arXiv:0811.2514.
P. Aluﬃ, M. Marcolli, Parametric Feynman integrals and determinant hypersur
faces, preprint arXiv:0901.2107, to appear in Advances in Theoretical and
Mathematical Physics.
P. Aluﬃ, M. Marcolli, Feynman motives and deletioncontraction relations,
arXiv:0907.3225.
P. Aluﬃ, L.C. Mihalcea, Chern classes of Schubert cells and varieties,
arXiv:math/0607752, to appear in Journal of Algebraic Geometry.
Y. Andr´e, Une introduction aux motifs, Societ´e Math´ematique de France, 2004.
Y. Andr´e, An introduction to motivic zeta functions of motives, arXiv:0812.3920.
D. Arapura, A category of motivic sheaves, preprint arXiv:0801.0261.
V.I. Arnold, V.V. Goryunov, O.V. Lyashko, V.A. Vasilev, Singularity theory, I,
Springer Verlag, 1998.
V.I. Arnold, S.M. GuseinZade, A.N. Varchenko, Singularities of diﬀerentiable
maps, Vol.II, Birkh¨auser, 1988.
J. Ayoub, The motivic vanishing cycles and the conservation conjecture, in “Al
gebraic Cycles and Motives”, London Mathematical Society Lecture Note
Series, Vol.343, pp.3–54, Cambridge University Press, 2007.
A. Be˘ılinson, J. Bernstein, P. Deligne, Faisceaux pervers, in “Analysis and topol
ogy on singular spaces, I” (Luminy, 1981), 5–171, Ast´erisque, 100, Soc.
Math. France, Paris, 1982.
A. Be˘ılinson, P. Deligne, Interpr´etation motivique de la conjecture de Zagier re
liant polylogarithmes et r´egulateurs. in “Motives” (Seattle, WA, 1991), 97–
121, Proc. Sympos. Pure Math., 55, Part 2, Amer. Math. Soc., Providence,
207
208 Feynman Motives
RI, 1994.
P. Belkale, P. Brosnan, Matroids, motives, and a conjecture of Kontsevich, Duke
Math. Journal, Vol.116 (2003) 147–188.
P. Belkale, P. Brosnan, Periods and Igusa local zeta functions. Int. Math. Res.
Not. 2003, no. 49, 2655–2670.
C. Bergbauer, A. Rej, Insertion of graphs and singularities of graph hypersurfaces,
preprint 2009.
I.N. Bernˇste˘ın, D.A. Le˘ıtes, Integral forms and the Stokes formula on superman
ifolds. Functional Anal. Appl. 11 (1977), no. 1, 45–47.
J. Bjorken, S. Drell, Relativistic Quantum Mechanics, McGrawHill, 1964.
J. Bjorken, S. Drell, Relativistic Quantum Fields, McGrawHill, 1965.
S. Bloch, Lectures on mixed motives. Algebraic geometry—Santa Cruz 1995, 329–
359, Proc. Sympos. Pure Math., 62, Part 1, Amer. Math. Soc., 1997.
S. Bloch, Motives associated to graphs, Japan J. Math., Vol.2 (2007) 165–196.
S. Bloch, Motives associated to sums of graphs, arXiv:0810.1313.
S. Bloch, E. Esnault, D. Kreimer, On motives associated to graph polynomials,
Commun. Math. Phys., Vol.267 (2006) 181–225.
S. Bloch, D. Kreimer, Mixed Hodge structures and renormalization in physics,
Commun. Number Theory Phys. 2 (2008), no. 4, 637–718.
N.N. Bogoliubov, O. Parasiuk,
¨
Uber die Multiplikation der Kausal Funktionen in
der Quantentheorie der Felder, Acta Mathematica 97 (1957) 227–266.
C. Bogner, S. Weinzierl, Periods and Feynman integrals, J. Math. Phys. 50 (2009),
no. 4, 042302, 16 pp.
C. Bogner, S. Weinzierl, Resolution of singularities for multiloop integrals,
arXiv:0709.4092.
J.B. Bost, A. Connes, Hecke algebras, Type III factors and phase transitions
with spontaneous symmetry breaking in number theory, Selecta Math. (New
Series) Vol.1 (1995) N.3, 411–457.
P. Breitenlohner, D. Maison, Dimensional renormalization and the action princi
ple, Comm. Math. Phys. Vol.52 (1977) N.1, 11–38.
T. Bridgeland, Spaces of stability conditions, Proc. Symposium Pure Math., Vol.
80 (2009) N.1, 1–21.
D. Broadhurst, D. Kreimer, Association of multiple zeta values with positive knots
via Feynman diagrams up to 9 loops, Phys. Lett. B, Vol.393 (1997) 403–412.
F. Brown, The massless higherloop twopoint function, Commun. Math. Phys.
Vol.287 (2009) 925–958.
F. Brown, On the periods of some Feynman integrals, arXiv:0910.0114.
F. Brown, Multiple zeta values and periods of moduli spaces M
0,n
,
arXiv:math/0606419.
A. Brugui`eres, On a Tannakian theorem due to Nori, preprint 2004.
A.L. Carey, J. Phillips, A. Rennie, F. A. Sukochev The Local Index Formula in
Semiﬁnite von Neumann Algebras I: Spectral Flow Adv. Math. 202 (2006),
no. 2, 451–516.
A.L. Carey, J. Phillips, A. Rennie, F. A. Sukochev The Local Index Formula
in Semiﬁnite von Neumann Algebras II: The Even Case Adv. Math. 202
(2006), no. 2, 517–554.
Bibliography 209
J. Collins, Renormalization, Cambridge University Press, 1986.
A. Connes, Une classiﬁcation des facteurs de type III. Ann. Sci.
´
Ecole Norm. Sup.
(4) 6 (1973), 133–252.
A. Connes, Geometry from the spectral point of view. Lett. Math. Phys. 34 (1995),
no. 3, 203–238.
A. Connes, Trace formula in noncommutative geometry and the zeros of the Rie
mann zeta function. Selecta Math. (N.S.) 5 (1999), no. 1, 29–106.
A. Connes, M. Douglas, A. Schwarz, Noncommutative geometry and matrix the
ory: compactiﬁcation on tori. JHEP 9802 (1998) 3–43.
A. Connes, D. Kreimer, Renormalization in quantum ﬁeld theory and the
Riemann–Hilbert problem I. The Hopf algebra structure of graphs and the
main theorem, Comm. Math. Phys., Vol.210 (2000) 249–273.
A. Connes, D. Kreimer, Renormalization in quantum ﬁeld theory and the
RiemannHilbert problem. II. The βfunction, diﬀeomorphisms and the
renormalization group. Comm. Math. Phys. 216 (2001), no. 1, 215–241.
A. Connes, M. Marcolli, Renormalization and motivic Galois theory. International
Math. Res. Notices (2004) N.76, 4073–4091.
A. Connes, M. Marcolli, Quantum ﬁelds and motives, Journal of Geometry and
Physics, Volume 56, (2005) N.1, 55–85.
A. Connes, M. Marcolli, Anomalies, Dimensional Regularization and non
commutative geometry, unpublished manuscript, 2005, available at
www.its.caltech.edu/~matilde/work.html
A. Connes, M. Marcolli, From physics to number theory via noncommutative ge
ometry, Part I: Quantum statistical mechanics of Qlattices, in “Frontiers in
Number Theory, Physics, and Geometry. I”, pp.269–347, Springer Verlag,
2006.
A. Connes, M. Marcolli, From Physics to Number Theory via Noncommutative
Geometry. Part II: Renormalization, the RiemannHilbert correspondence,
and motivic Galois theory, in ”Frontiers in Number Theory, Physics, and
Geometry, II” pp.617–713, Springer Verlag, 2006.
A. Connes, M. Marcolli, Noncommutative Geometry, Quantum Fields, and Mo
tives, Colloquium Publications, Vol.55, American Mathematical Society,
2008.
A. Connes, H. Moscovici, The local index formula in noncommutative geometry,
Geom. Funct. Anal. 5 (1995), no. 2, 174–243.
C. Consani, Double complexes and Euler Lfactors. Compositio Math. 111 (1998),
no. 3, 323–358.
C. Consani, M. Marcolli, Noncommutative geometry, dynamics, and ∞adic
Arakelov geometry. Selecta Math. (N.S.) 10 (2004), no. 2, 167–251.
C. Consani, M. Marcolli, Archimedean cohomology revisited, in “Noncommutative
geometry and number theory”, pp.109–140, Aspects Math., E37, Vieweg
Verlag, 2006.
P. Deligne and A. Dimca, Filtrations de Hodge et par d’ordre du pˆole pour les
hypersurfaces singuli`eres, Ann. Sci.
´
Ecole Norm. Sup. 23 (1990) 645–656.
P. Deligne, A.B. Goncharov, Groupes fondamentaux motiviques de Tate mixte,
Ann. Sci.
´
Ecole Norm. Sup. (4) 38 (2005), no. 1, 1–56.
210 Feynman Motives
P. Deligne, J.S. Milne, Tannakian categories, in “Hodge cycles, motives, and
Shimura varieties”, Lecture Notes in Mathematics, 900, pp.101–228.
SpringerVerlag, 1982.
M. Demazure, A. Grothendieck, et al. S´eminaire G´eometrie Alg´ebrique: Sch´emas
en Groupes, Lecture Notes in Mathematics, Vol. 151, 152, 153. Springer,
1970.
J. Denef, F. Loeser, Motivic Igusa zeta functions. J. Algebraic Geom. 7 (1998),
no. 3, 505–537.
A. Dimca, Singularities and topology of hypersurfaces, Universitext, Springer
Verlag, 1992.
I. Dolgachev, Weighted projective varieties, in “Group actions and vector ﬁelds”,
LNM 956, pp. 34–71, SpringerVerlag, 1982.
D. Doryn, Cohomology of graph hypersurfaces associated to certain Feynman
graphs, PhD Thesis, University of DuisburgEssen, arXiv:0811.0402.
K. EbrahimiFard, L. Guo, D. Kreimer, Spitzer’s identity and the algebraic
Birkhoﬀ decomposition in pQFT. J. Phys. A 37 (2004), no. 45, 11037–11052.
K. EbrahimiFard, J.M. GraciaBond
´
ia, F. Patras, A Lie theoretic approach to
renormalization, Commun.Math.Phys. Vol.276 (2007) 519–549.
I.M. Gel’fand, S.G. Gindikin, M.I. Graev, Integral geometry in aﬃne and pro
jective spaces. Current problems in mathematics, Vol. 16, pp. 53–226, 228,
Akad. Nauk SSSR, Vsesoyuz. Inst. Nauchn. i Tekhn. Informatsii, Moscow,
1980.
S.I. Gelfand, Yu.I. Manin, Homological algebra, Encyclopedia of Mathematical
Sciences, Vol.38, Springer Verlag, 1994.
I.M. Gelfand, V.V. Serganova, Combinatorial geometries and the strata of a torus
on homogeneous compact manifolds. Uspekhi Mat. Nauk 42 (1987), no.
2(254), 107–134, 287.
H. Gillet, C. Soul´e, Descent, motives and Ktheory. J. Reine Angew. Math. 478
(1996), 127–176.
A.B. Goncharov, Volumes of hyperbolic manifolds and mixed Tate motives. J.
Amer. Math. Soc. 12 (1999), no. 2, 569–618.
A.B. Goncharov, Multiple polylogarithms and mixed Tate motives,
math.AG/0103059.
A.B. Goncharov, Yu.I. Manin, Multiple ζmotives and moduli spaces M
0,n
. Com
pos. Math. 140 (2004), no. 1, 1–14.
H. Grosse, R. Wulkenhaar, Renormalization of noncommutative quantum ﬁeld
theory, in “An invitation to Noncommutative Geometry” pp.129–168,
World Scientiﬁc, 2008.
H. Grosse, R. Wulkenhaar, Renormalization of φ
4
theory on noncommutative R
4
in the matrix base, Commun. Math. Phys. 256 (2005) 305–374.
F. Guill´en, V. Navarro Aznar, Sur le th´eor`eme local des cycles invariants. Duke
Math. J. 61 (1990), no. 1, 133–155.
R. Gurau, J. Magnen, V. Rivasseu, A. Tanasa, A translationinvariant renormal
izable noncommutative scalar model, Commun. Math. Phys. 287 (2009)
275–290.
R. Gurau, V. Rivasseau, Parametric Representation of Noncommutative Field
Bibliography 211
Theory, Commun. Math. Phys. Vol. 272 (2007) N.3, 811–835.
M. Hanamura, Mixed motives and algebraic cycles. I. Math. Res. Letters, Vol.2
(1995) N.6, 811–821.
K. Hepp, Proof of the BogoliubovParasiuk theorem on renormalization, Comm.
Math. Phys. 2, (1966), 301–326.
C. Itzykson, J.B. Zuber, Quantum Field Theory, Dover Publications, 2006.
U. Jannsen, Motives, numerical equivalence, and semisimplicity, Invent. Math.
107 (1992) N.3, 447–452.
U. Jannsen, S. Kleiman and J.P. Serre (Eds.) Motives. Proceedings of Symposia
in Pure Mathematics, 55, Part 12. American Mathematical Society, Prov
idence, RI, 1994.
F. Jegerlehner, Facts of life with γ
5
, Eur. Phys. J. C 18 (2001) 673–679.
M. Kapranov, The elliptic curve in the Sduality theory and Eisenstein series for
KacMoody groups, arXiv:math.AG/0001005.
M. Kashiwara, P. Schapira, Sheaves on manifolds, Springer, 1990.
D. Kazhdan, G. Lusztig, Schubert varieties and Poincar´e duality. in “Geometry
of the Laplace operator” pp. 185–203, Proc. Sympos. Pure Math., XXXVI,
Amer. Math. Soc. 1980.
M. Kontsevich, D. Zagier, Periods. in “Mathematics unlimited—2001 and be
yond”, pp.771–808, Springer, 2001.
D. Kreimer, The core Hopf algebra, preprint 2009.
D. Kreimer, W.D. van Suijlekom, Recursive relations in the core Hopf algebra,
arXiv:0903.2849.
K. Kremnizer, M. Szczesny, Feynman graphs, rooted trees, and RingelHall alge
bras, arXiv:0806.1179.
V.S. Kulikov, Mixed Hodge structures and singularities. Cambridge University
Press, 1998.
S. Lang, SL
2
(R), Addison–Wesley, 1975.
M. Larsen, V. Lunts, Motivic measures and stable birational geometry, Moscow
Math. Journal, Vol.3 (2003) N.1, 85–95.
M. LeBellac, Quantum and statistical ﬁeld theory, Oxford Science Publications,
1991.
M. Levine, Tate motives and the vanishing conjectures for algebraic Ktheory,
in Algebraic Ktheory and algebraic topology (Lake Louise, AB, 1991),
167188, NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., 407, Kluwer Acad.
Publ., 1993.
M. Levine, Mixed motives, Math. Surveys and Monographs, Vol.57, American
Mathematical Society, 1998.
M. Levine, Motivic Tubular Neighborhoods, arXiv:math/0509463
R.D. MacPherson, Chern classes for singular algebraic varieties. Ann. of Math.
(2) 100 (1974), 423–432.
Yu.I. Manin, Correspondences, motifs and monoidal transformations, Math.
USSRSb. 6 (1968), 439–470.
Yu.I. Manin, Gauge ﬁeld theory and complex geometry, Springer Verlag, 1988.
M. Marcolli, Motivic renormalization and singularities, preprint arXiv:0804.4824.
M. Marcolli, A. Rej, Supermanifolds from Feynman graphs, Journal of Physics A,
212 Feynman Motives
Vol.41 (2008) 315402 (21pp). arXiv:0806.1681.
P.B. Medvedev, α representation in scalar supersymmetric theory, Teoretich
eskaya i matematicheskaya ﬁzika (transl.Institute of Theoretical and Ex
perimental Physics) Vol.35 (1978) N.1, 37–47.
B. Mohar, C. Thomassen, Graphs on Surfaces, Johns Hopkins University Press,
2001.
N. Mn¨ev, The universality theorems on the classiﬁcation problem of conﬁguration
varieties and convex polytopes varieties, in “Topology and geometry  Rohlin
seminar”. Lect. Notes Math., vol. 1346, pp. 527–543, Springer 1988.
N. Nakanishi Graph Theory and Feynman Integrals. Gordon and Breach, 1971.
J. Rammer, Quantum ﬁeld theory of nonequilibrium states, Cambridge University
Press, 2007.
P. Ramond, Field theory: a modern primer. Addison Wesley, 1990.
A.A. Rosly, A.S. Schwarz, A.A. Voronov, Geometry of superconformal manifolds,
Comm. Math. Phys., Vol.119 (1988) 129–152.
N. Saavedra Rivano, Cat´egories Tannakiennes. Lecture Notes in Mathematics,
Vol. 265. SpringerVerlag, 1972.
M. Saito, Modules de Hodge polarisables. Publ. Res. Inst. Math. Sci. 24 (1988),
no. 6, 849–995 (1989).
M. Sato, T. Miwa, M. Jimbo, T. Oshima, Holonomy structure of Landau sin
gularities and Feynman integrals, Publ. Res. Inst. Math. Sci. Volume 12,
Supplement (1976/1977) (1976), 387–439.
M.H. Schwartz, Classes caract´eristiques d´eﬁnies par une stratiﬁcation d’une
vari´et´e analytique complexe. I. C. R. Acad. Sci. Paris, Vol.260 (1965) 3262–
3264.
N. Seiberg, E. Witten, String theory and noncommutative geometry, JHEP 9909
(1999) 32–131.
I.R. Shafarevich, Basic Algebraic Geometry 2: Schemes and Complex Manifolds.
Second edition, Springer, 1996.
J. Steenbrink, Limits of Hodge structures. Invent. Math. 31 (1976), 229–257.
J. Steenbrink, Mixed Hodge structure on the vanishing cohomology. in “Real and
complex singularities” (Proc. Ninth Nordic Summer School/NAVF Sympos.
Math., Oslo, 1976), pp. 525–563. Sijthoﬀ and Noordhoﬀ, Alphen aan den
Rijn, 1977.
J. Steenbrink, A summary of mixed Hodge theory. in “Motives” (Seattle, WA,
1991), 31–41, Proc. Sympos. Pure Math., 55, Part 1, Amer. Math. Soc.,
Providence, RI, 1994.
J. Stembridge, Counting points on varieties over ﬁnite ﬁelds related to a conjecture
of Kontsevich. Ann. Comb. 2 (1998), no. 4, 365–385.
A. Tanasa, Overview of the parametric representation of renormalizable non
commutative ﬁeld theory, arXiv:0709.2270.
B. Teissier, Introduction to equisingularity problems. Algebraic geometry (Proc.
Sympos. Pure Math., Vol. 29, Humboldt State Univ., Arcata, Calif., 1974),
pp. 593–632. Amer. Math. Soc., Providence, R.I., 1975.
T. Terasoma, Mixed Tate motives and multiple zeta values. Invent. Math. 149
(2002), no. 2, 339–369.
Bibliography 213
R. Vakil, Murphy’s law in algebraic geometry: badlybehaved deformation spaces.
Invent. Math. 164 (2006), no. 3, 569–590.
W. van Suijlekom, Renormalization of gauge ﬁelds: a Hopf algebra approach.
Comm. Math. Phys. 276 (2007), no. 3, 773–798.
W. van Suijlekom, The Hopf algebra of Feynman graphs in quantum electrody
namics. Lett. Math. Phys. 77 (2006), N.3, 265–281.
W. van Suijlekom, Representing Feynman graphs on BValgebras, preprint,
arXiv:0807.0999.
A.N. Varchenko, Asymptotic behavior of holomorphic forms determines a mixed
Hodge structure. Dokl. Akad. Nauk SSSR 255 (1980), no. 5, 1035–1038.
A.N. Varchenko, Asymptotic Hodge structure on vanishing cohomology. Izv. Akad.
Nauk SSSR Ser. Mat. 45 (1981), no. 3, 540–591, 688.
J.C. V´arilly, J.M. GraciaBond
´
ia, Connes’ noncommutative diﬀerential geometry
and the standard model. J. Geom. Phys. 12 (1993), no. 4, 223–301.
V. Voevodsky, Triangulated categories of motives over a ﬁeld. in “Cycles, trans
fer and motivic homology theories”, pp.188–238, Annals of Mathematics
Studies, Vol.143, Princeton University Press, 2000.
W.C. Waterhouse, Introduction to aﬃne group schemes. Graduate Texts in Math
ematics, 66. Springer, 1979.
A. Zee, Quantum ﬁeld theory in a nutshell, Princeton University Press, 2003.
W. Zimmermann, Convergence of Bogoliubov’s method of renormalization in mo
mentum space, Comm. Math. Phys. 15, (1969), 208–234.
This page intentionally left blank This page intentionally left blank
Index
1PI, 14, 15, 63, 67, 79, 86, 101
2PI, 14, 102
action functional, 7
additive invariant, 38
adele class space, 157
aﬃne group scheme, 34, 50, 131, 148,
182
algebraic simplex, 72, 77, 95
singularities, 73
anomalies, 158
anomalous graphs, 157
antipode, 133, 179
arithmetic progression, 120
augmentation, 135
banana graphs, 76, 202
Berezinian, 185, 189, 193, 197, 199
beta function, 144, 148
Birkhoﬀ factorization, 79, 127, 135,
144, 178, 183
blowups, 62, 73, 96, 113
Bogolyubov recursion, 22, 127, 128
BostConnes system, 157
boundary, 62
BPHZ renormalization, 127, 128, 183
BreitenlohnerMaison prescription,
156
bridge, 85
categoriﬁcation, 135
category
abelian, 29
linear, 32, 49
neutral tannakian, 49
of pure motives, 32
rigid, 33, 49
semisimple, 32, 50
tannakian, 49, 50, 143, 146
tensor, 32, 49
thick, 36, 39
triangulated, 27, 35, 39, 104
triangulated subcategory, 36, 39
chains of polygons, 91
change of variables, 45, 97, 195
ChernSchwartzMacPherson classes,
77, 81, 204
positivity, 84, 205
chiral, 134
Chow group, 82
Chow ring, 31
chromatic polynomial, 85, 87
circuit matrix, 54, 57
closed 2cell embedding, 98, 99, 104
closed embedding, 38, 40, 42, 108
coherent sheaves
d´evissage, 188
cohomology
archimedean, 158, 164
Betti, 29, 44
de Rham, 29, 44
middle dimension, 96
mixed Weil, 35
primitive, 159
215
216 Feynman Motives
relative, 54, 61, 95, 103
universal, 26
vanishing, 177
Weil, 28
combinatorial factor, 4
compact support, 35, 40
complete bipartite graph, 94, 205
complete graph, 78, 205
cones, 71
connection, 146
equisingular, 146, 178, 183
ﬂat, 146, 178, 182
irregular singular, 176, 181
regular singular, 176, 181
connectivity, 14, 101
edge, 14, 98
vertex, 14
ConnesKreimer
Hopf algebra, 79, 94, 116, 133, 146,
178
combinatorial, 134
continuous, 134
Lie algebra, 178
theory, 127
constructible functions, 81
coproduct, 133, 179, 180
correspondence, 30
counterterms, 128, 145, 148, 183
coupling constant, 19, 134, 149
Cremona transformation, 53, 72
critical points, 120, 177, 181
cutset, 58, 66
cycle map, 28
cyclotomic ﬁeld, 51
deformation, 113
deletioncontraction, 84, 88
determinant hypersurface, 41, 54, 103
diﬀeographisms, 134, 146, 176
diﬀerential forms
hypersurface complement, 67
pullbacks, 68, 70, 178
with logarithmic poles, 82, 158
diﬀerential systems, 145
dimension spectrum, 156
dimensional analysis, 144
dimensional regularization, 21, 114,
119, 135, 144, 151, 154, 176
Dirac operator, 155
distinguished triangle, 35, 40
divergences, 62, 96, 112
divisibility sequence, 93
dual graph, 74
dual hypersurface, 77
duality, 33
Dynkin operator, 148
Dyson–Schwinger equation, 78
edge doubling, 91
edges
external, 4
fermionic, 191
half, 4
internal, 4
parallel, 85, 91
eﬀective action, 14
energy scale, 22, 144–146, 176, 183
equivalence
homological, 31
numerical, 31
rational, 31, 39
Euler characteristic, 27, 38, 115
motivic, 39
topological, 38, 77, 81
universal, 38, 77
Euler obstruction, 82
Euler vector ﬁeld, 68
evanescent gauge potentials, 157
expectation value, 2, 8
external momenta, 59, 63, 66, 134,
171, 175, 179
face width, 101
factorization of loops, 136
fermion doubling, 191
fermions, 185, 190
Feynman
motivic sheaves, 153
parametric representation, 54
Feynman graph, 11
residue, 96, 128, 133
Feynman parameters, 20, 54
Index 217
Feynman rules, 4, 10
abstract, 63, 79, 86
aﬃne hypersurface complement,
63, 79
algebrogeometric, 80, 83, 134
motivic, 79, 84, 88, 139
multiplicative property, 12, 62, 79
Feynman trick, 20, 191
ﬁber functor, 49, 50
ﬁber product, 30, 37, 153
Fibonacci numbers, 93
ﬁltered spaces, 147
ﬁltration, 161
ﬁnitely summable, 155
ﬂag varieties, 112
forest, 64
Fourier transform, 13
FreydMitchell theorem, 30
Galois group
cosmic, 148
motivic, 50, 51, 131, 149
Tannakian, 50
Gamma factor, 69, 96, 112
Gamma function, 20
gauge equivalence, 143, 145
gauge potentials, 160
gauge theories, 134
GaussManin connection, 166, 176
Gaussian integral, 2, 21, 156
GelfandLeray form, 120, 176, 181
GelfandLeray function, 120
generating function, 91–93
generic condition, 65, 172
genus, 99, 104
germs of meromorphic functions, 135,
144
graph hypersurface, 56
projective and aﬃne, 63
graph polynomial, 57
Grassmann integration, 190, 192, 196
Grassmann variables, 186, 189, 191
Green function, 5, 8
connected, 13
Gross’t Hooft relations, 144
Grothendieck ring, 38, 53, 76, 80, 95,
115, 134, 139
supermanifolds, 187, 201
GrothendieckDeligne conjecture, 82
Hall algebra, 135
harmonic oscillator, 203
Hodge conjecture, 29
Hodge decomposition, 48
Hodge ﬁltration, 49
Hodge structure, 147, 159
asymptotic mixed, 174
limiting mixed, 174
mixed, 48, 119, 152, 176, 177
pure, 48
Hodge structures
mixed, 151
HodgeLefschetz module, 158, 159
polarized, 158
homotopy invariance, 34, 40, 108
Hopf algebra, 130
character, 135
commutative, 130
core, 134
graded, 137, 144
Igusa Lfunction, 113, 116
incidence matrix, 10, 54, 57
inclusionexclusion, 27, 38, 82, 105,
111
insertion, 72, 178
interaction term, 2
irregular singular, 178, 179
iterated integrals, 145, 152
join, 62, 83
Jones polynomial, 85, 87
K¨ unneth formula, 28, 35
KazhdanLusztig theory, 112
Kirchhoﬀ law, 58
KirchhoﬀSymanzik matrix, 55, 98
KirchhoﬀSymanzik polynomial, 53,
55, 57, 153, 202
KMS state, 157
Kontsevich conjecture, 95
218 Feynman Motives
Koszulde Rham complex, 171
Kummer extension, 46
Kummer motive, 46, 50, 151
Lagrangian, 130, 133, 190
density, 7
Landau varieties, 65, 95
Laurent series, 22, 129, 181
Lefschetz
hard, 29
modules, 159
motive, 33, 39, 107
motive (fermionic), 189
operator, 164
weak, 28
Legendre transform, 15
lemon graph, 91, 92
Leray coboundary, 121
Leray regularization, 123
level set, 120
Lie algebra, 51, 144, 148, 176, 178
local system, 153
log canonical threshold, 113
logarithmic extension, 151
logarithmically divergent, 61, 93, 113,
192
looping edge, 85, 102
manifold of frames, 110
manifolds of frames, 105, 106
massive theory, 63
massless theory, 60, 190
Mather classes, 82
matroid, 112
MayerVietoris, 27, 35
Mellin transform, 119, 181
Milnor algebra, 172
Milnor ﬁber, 114, 172, 176, 181
Milnor number, 82
minimal subtraction, 22, 130
mirror symmetry, 36
mixed Hodge structure, 48
moduli space of curves, 94
momentum conservation, 56
monodromy, 145, 159, 165, 166
motives
eﬀective, 32
logarithmic, 46, 153
mixed, 27, 35, 61
mixed Tate, 39, 51, 61, 78, 85, 93,
94, 96, 103, 105, 110, 114,
149
Nori construction, 37
pullback, 153
pure, 26, 32, 50
Tate, 27, 33, 39, 76
motivic measure, 116
motivic sheaves, 37, 46, 151, 153
motivic zeta function, 115
Moyal product, 185, 201
multiple polylogarithms, 94
multiple zeta values, 93, 149
nearby cycles, 157, 165
nonrenormalizable theory, 24
noncommutative geometry, 151, 155,
185, 201
normal crossings, 54, 73, 95, 104
number ﬁeld, 28
orientation, 54, 55, 100, 104
oscillatory integrals, 119
parametric Feynman integral, 59, 70,
103, 119, 169, 185, 202
period isomorphism, 29, 44
period matrix, 44, 46–48
periods, 45, 61, 93, 97, 152, 199
algebra, 45
conjecture, 45
formal, 45
perturbative expansion, 3
PicardFuchs, 179
PicardFuchs equation, 176, 177, 181
PicardLefschetz ﬁltration, 166
Poincar´e duality, 28, 35
Poincar´e residue, 120, 171
Poincar´eHopf theorem, 81
polar ﬁltration, 171
polar part, 22, 128, 129, 140
polarization, 159
polylogarithms, 45, 47
Index 219
principal bundle, 146
promotive, 47, 152
pronilpotent, 148
prounipotent, 51, 132
probability amplitude, 7
projective superspace, 186
projective system, 47, 154
projector, 32, 38
propagator, 11, 79
aﬃne line, 80
fermionic, 190
Lefschetz motive, 81
Radon transform, 167
realization functors, 29
reductive, 50, 51
regular singular, 177, 179
regularization, 18, 62, 114
parameter, 19
renormalization, 18
renormalization group, 144, 148
representable functor, 132
residue, 69
ribbon graphs, 202
Riemann zeta function, 157
RiemannHilbert correspondence,
143, 176
ring of immersed conical varieties, 79,
80, 82, 134
RotaBaxter algebra, 129, 137
RotaBaxter operator, 137
scaling, 144, 146, 182
Schubert cells, 112
Schwinger parameters, 20
selfenergy graph, 22
shift functor, 35
signature
Euclidean, 7
Lorentzian, 7
singularities, 61, 65, 82, 114, 181
nonisolated, 71, 114, 167
SlavonovTaylor identities, 78
slicing, 167, 172, 179, 180
source term, 2
spanning forest, 57
spanning tree, 57, 66
special ﬁber, 158, 165
spectral triple, 135, 155
cup product, 156
local index formula, 166
tame, 164
type II, 156
splitting at a vertex, 17
stable birational equivalence, 140
stable range, 61, 65, 170
standard conjectures, 27
Stokes’ formula, 45, 69
stratiﬁcation, 81, 108, 110
strong orientable embedding
conjecture, 101
subtraction, 18
sum over states, 86
supergeometry, 185
supermanifold, 186
divisors, 190, 199
embedding, 186
morphisms, 186
split, 186
superrenormalizable, 7
symmetry factor, 4, 178
tstructure, 36, 39
heart, 36, 50
tangent cone, 203, 204
Tannakian formalism, 34, 49, 131
Tate conjecture, 43
Tate Hodge structure, 48
Tate sheaves, 46
Tate twist, 40, 48
time ordered exponential, 144
topological simplex, 21, 95, 112, 170
tubular neighborhood, 113, 121, 126
Tutte polynomial, 85, 86
TutteGrothendieck invariants, 84,
85, 87
universal enveloping algebra, 148
universal recursive relation, 92
universal singular frame, 148
unstable range, 61
220 Feynman Motives
vacuum bubble, 12
vanishing cycles, 119, 157, 165, 176
vector bundles
ﬂat equisingular, 147
holomorphic, 136, 186
vector ﬁeld, 81
von Neumann algebra, 156
Wequivalence, 147
Ward identities, 134
weight ﬁltration, 48, 50, 147, 152
wheel neighborhood, 102
wheel with spokes graph, 94, 110
Wick contraction, 5
zigzag graphs, 96
This page intentionally left blank
California Institute of Technology, USA
Matilde Marcolli
World Scientific
NEW JERSEY
•
LONDON
•
SINGAPORE
•
BEIJING
•
SHANGHAI
•
HONG KONG
•
TA I P E I
•
CHENNAI
recording or any information storage and retrieval system now known or to be invented. Covent Garden. Pte. USA.. NJ 07601 UK office: 57 Shelton Street. For photocopying of material in this volume. Suite 401402. Pte. please pay a copying fee through the Copyright Clearance Center. 222 Rosewood Drive. Singapore 596224 USA office: 27 Warren Street. Ltd. ISBN13 ISBN10 ISBN13 ISBN10 9789814271202 9814271209 9789814304481 (pbk) 9814304484 (pbk) Printed in Singapore. Hackensack. or parts thereof. MA 01923. Ltd. . London WC2H 9HE British Library CataloguinginPublication Data A catalogue record for this book is available from the British Library.Published by World Scientific Publishing Co. All rights reserved. This book. without written permission from the Publisher. Danvers. In this case permission to photocopy is not required from the publisher. FEYNMAN MOTIVES Copyright © 2010 by World Scientific Publishing Co. Inc. including photocopying. may not be reproduced in any form or by any means. 5 Toh Tuck Link. electronic or mechanical.
Yuri Manin – “(not yet quite) The Late Style” . No my — my pozdnii stil stare wei Vselennoi.Zvezda v Galaktike i voron na suku ivut. ne muqimy xekspirovskoi dilemmoi. Ona skvoz nas poet svo tosku.
This page intentionally left blank .
Preface This book originates from the notes of a course on “Geometry and Arithmetic of Quantum Fields”. the choice of material covered here focuses mostly on those aspects of the subject where I have been actively engaged in recent research work and therefore reﬂects closely my own bias and personal viewpoint. whenever possible. Also I do not try in any way to give an exhaustive viewpoint of the current status of research on the connection between quantum ﬁeld theory and motives. possibly at the cost of neglecting material that should certainly be included in a more extensive monograph. to provide an extensive list of references for the interested reader. Interested readers will have to look elsewhere for a more informative introduction to the subject (a few references are provided in the text). Many extremely interesting results are at this point available in the literature. In particular. In particular. Having just moved to Caltech and having my ﬁrst chance to oﬀer a class there. I try. I hope that having made the main focus of the lectures the yet largely unexplored relation between quantum ﬁeld theory and Grothendieck’s theory of motives in algebraic geometry may provide a suﬃciently diﬀerent viewpoint on the quantum ﬁeld theoretic notions to make the resulting combination of topics appealing to mathematicians and physicists alike. but in general I prefer to keep the text as close as possible to the very informal style of the lectures. and occasionally to summarize some of the available results. which I taught at Caltech in the fall of 2008. I will try to illustrate the fact that there are two possible vii . I decided on a topic that would fall in between mathematics and theoretical physics. Though it inevitably feels somewhat strange to be teaching Feynman diagrams at Caltech. I am not an expert in the theory of motives and this fact is clearly reﬂected in the way this text is organized.
a mixed Tate motive. but I hope that the timeliness of circulating these ideas in the community of mathematicians and physicists will somehow make up for lack of both rigor and of completeness. One then tries to understand the motivic nature of the piece of the relative cohomology of the algebraic variety involved in the computation of the period. which are suﬃciently rigid to identify it (via the Tannakian formalism) with a category of representations of an aﬃne group scheme. The bottomup approach looks at individual Feynman integrals for given Feynman graphs and. This book only partially reﬂects the state of the art on this fastmoving subject at the speciﬁc time when these lectures were delivered. quantum ﬁelds. I will also try to illustrate the points of contact between these two diﬀerent approaches and where possible new developments may arise that might eventually unify the somewhat fragmentary picture we have at the moment. though in this book I will mostly relate aspects of this approach that come from my joint work with Aluﬃ. and motives”. which one may refer to as a “bottomup” and a “topdown” approach. However. I have also included a brief and less detailed summary of this aspect of the theory in the present text. using the parametric representation in terms of Schwinger and Feynman parameters. Matilde Marcolli . is based on the formal properties that the category of mixed Tate motives satisﬁes. The bottomup approach was largely developed in the work of Bloch–Esnault–Kreimer. identiﬁes directly the Feynman integral (modulo the important issue of divergences) with an integral of an algebraic diﬀerential form on a cycle in an algebraic variety.viii Feynman Motives complementary approaches to understanding the relation between Feynman integrals and motives. for the sake of completeness. referring the readers to the more complete treatment for further information. One then approaches the question of the relation to Feynman integrals by showing that the data of Feynman integrals for all graphs and arbitrary scalar ﬁeld theories also ﬁt together to form a category with the same properties. This second approach was the focus of my joint work with Connes on renormalization and the Riemann–Hilbert correspondence and is already presented in much greater detail in our book “Noncommutative geometry. trying to identify speciﬁc conditions under which it will be a realization of a very special kind of motive. The other approach. the topdown one.
especially Domenic Denicola and Michael Smith. whom I thank for their essential contributions to the development of those ideas and results. This work was partially supported by NSF grants DMS0651925 and DMS0901221.Acknowledgments A large part of the material covered in this book is based on recent joint work with Paolo Aluﬃ and on previous joint work with Alain Connes. I thank both institutes for the hospitality and for support. Many thanks are due to Paolo Aluﬃ for his very detailed and useful feedback on the ﬁnal version of the manuscript. I thank the students of my “Geometry and Arithmetic of Quantum Fields” course. and for providing detailed comments on an earlier draft of the manuscript. I also thank Abhijnan Rej for conversations and collaboration on some of the topics in this book. I also thank Arthur Greenspoon for a ﬁnal proofreading of the text. ix . for their valuable comments and questions during the class. This book was partly written during stays of the author at MSRI and at the MPI.
This page intentionally left blank .
. . . . Perturbative quantum ﬁeld theory and Feynman diagrams 1. . The Grothendieck ring of motives . . . . . . . . . . The problem of renormalization .6 1. . . . . . . From Lagrangian to eﬀective action . Tate motives . . . . . . . . . vii ix 1 1 6 9 12 14 18 20 21 25 . . . . .5 1. . . . . . . . . . . . . . . . . .1 1. . . . The algebra of periods . . Feynman rules . .2 1. . . . . . . xi . . . . Mixed Tate motives and the logarithmic extensions Categories and Galois groups . . . . . . .1 2. . . . . . . Oneparticleirreducible graphs . . . . . . . . . . . . . . . . . . Motivic sheaves . . .3 1. . . . . . . . .4 1. Mixed motives and triangulated categories . . . . . . . . . . . . . . . . .10 The idea of motives . . . . . . . . . . . . . 25 28 34 37 38 39 44 45 49 50 2. . . .4 2. . . . . . . . . . . . .9 2. . . . . . .7 1. .Contents Preface Acknowledgments 1. . . . . . . . . . . . . .8 A calculus exercise in Feynman integrals . . . . . . . . . . .5 2. . . .3 2. . . . . . . . . Motivic Galois groups . . . . . . Gamma functions. Pure motives . . . . Motives and periods 2.8 2. . . . . . . . .6 2. Schwinger and Feynman parameters Dimensional Regularization and Minimal Subtraction . . . . .2 2. . . . . . . . . . . . . . . .7 2. . . . . . . . . . . connected graphs . . . . Simplifying graphs: vacuum bubbles. . . . . . . . . .
. . . . . . . . 72 . . . 119 Leray regularization of Feynman integrals . . The mixed Tate mystery . . . . 5. .1. .7 3. . . . . . . . Integrals in aﬃne and projective spaces . . . . . . . . .2 5. Flat equisingular connections and vector bundles The “cosmic Galois group” . . . . . Feynman integrals and Gelfand–Leray forms 4. . .4 5. .1. . . .2 Step 2: Counterterms . . . . . . . . From iterated integrals to diﬀerential systems . . . . . . . . . . . 65 . . . . . . . . . . 97 . . . Nonisolated singularities . .3 6. The Riemann–Hilbert correspondence 6. . . . . 71 . 54 . From graph hypersurfaces to determinant hypersurfaces Handling divergences . .1 Step 1: Preparation .3 Step 3: Renormalized values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10 3. . . . . . . 94 . . Hopf algebras and aﬃne group schemes The Connes–Kreimer Hopf algebra .14 3. . .3 5. . . . . . . .1 The Bogolyubov recursion . . . . . . . . . . . . .6 6. 115 119 4. .3 3. . . . . . The graph hypersurfaces . . . . . . 67 . .1 6. .1 4. . . . . . . . Landau varieties . . . . . . Characteristic classes and Feynman rules .9 3. . 84 . .xii Feynman Motives 3. . Deletioncontraction relation . . . . . . . . . 5. Feynman integrals and periods . . . .8 3. 121 127 . . 60 . . . . . . . . 53 . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 3. . 143 145 146 147 5. . . . . 76 . . Feynman integrals and algebraic varieties 3. . . . . . . . . . . . . . . structure .4 3. Classes in the Grothendieck ring . . 112 . . . . . . . . . . . . . . Motivic zeta functions and motivic Feynman rules . . . .5 5. . . 128 128 128 129 130 133 135 137 139 143 . . . . . . .2 Oscillatory integrals . . . . . . . . . . . . .5 3. . .11 3. . . 81 . . 5. . . . . . . . . . . . . . . . . . . . . . Birkhoﬀ factorization . . . . . .2 6. . . . . . . .15 The parametric Feynman integrals . . . . . . . . . . . .4 From divergences to iterated integrals . 78 .6 3. . . . . . . . . . . . . . . . . . . . . . . . . . . 93 . . . . . . . .2 3.12 3. Factorization and RotaBaxter algebras Motivic Feynman rules and RotaBaxter 5. . . . . . . Cremona transformation and dual graphs .13 3. . . . . Connes–Kreimer theory in a nutshell 5. . . . Motivic Feynman rules .1. . .
. . Noncommutative ﬁeld theories . . . and Hodge structures 8. . . . . . .1 9. . . . . . . . . . . 151 The noncommutative geometry of DimReg . . . . Beyond scalar theories 9. . .2 9. . . . . . .3 8. . . . . . . . . . . . Renormalization. . . . . . . Bibliography Index . 185 190 199 201 207 215 8. . .3 9. . 167 170 173 176 185 . .Contents xiii 7. . 155 167 . . . . .2 8. .1 8. Parametric Feynman integrals and supermanifolds Graph supermanifolds . . . singularities. . . . . . The geometry of DimReg 7. . . . . . . DimReg and mixed Hodge structures . .1 7. . . . . . . . . . . . Regular and irregular singular connections . . . 9. . . The polar ﬁltration and the Milnor ﬁber . . . . . . . . . . . . . . . . . . . . . . . . .4 Supermanifolds . . .2 151 The motivic geometry of DimReg . . . . . .4 Projective Radon transform . . . . . .
which we discuss in more detail in a later chapter. We include here also a brief discussion of some combinatorial properties of Feynman graphs. The subject is very rich and can be approached from many diﬀerent perspectives.1 A calculus exercise in Feynman integrals To understand the role of Feynman graphs in perturbative quantum ﬁeld theory. which is aimed at illustrating the main ideas in a context familiar to most readers. 1.Chapter 1 Perturbative quantum ﬁeld theory and Feynman diagrams This ﬁrst chapter gives a very brief and informal introduction to perturbative quantum ﬁeld theory. where the more serious diﬃculties of the inﬁnite dimensional setting are not present. it is useful to ﬁrst see how graphs arise in the more familiar setting of ﬁnite dimensional integrals. the reader may wish to consult [Bjorken and Drell (1964)]. For the material we are going to cover in this text. and the corresponding Feynman rules. in the setting of the Connes–Kreimer approach. A lively and very readable introduction. We start with a simple ﬁnite dimensional example. in the case of a scalar ﬁeld theory. [LeBellac (1991)]. We state the renormalization problem. by which the beginning part of this chapter is inspired. a subject to which we return later in the book from a more geometric perspective. There are many very good books on quantum ﬁeld theory. [Nakanishi (1971)]. can be found in the recent book [Zee (2003)]. [Itzykson and Zuber (2006)]. which we need in later chapters. as a convenient way of parameterizing the 1 . [Bjorken and Drell (1965)]. We then brieﬂy describe the inﬁnite dimensional setting. We also include a brief discussion of dimensional regularization (DimReg) of divergent Feynman integrals. but one can already see the logic behind the perturbative expansion as well as the roles of Feynman graphs.
As we discuss below.2 Feynman Motives terms in the integration by parts of polynomials with respect to a Gaussian measure. (1. is to treat the additional term P (x) as a perturbation of the original Gaussian integral and expand it out in Taylor series. an (1.1). (1.1). namely it counts all the diﬀerent ways of connecting in pairs the 2n linear terms x in the monomial x2n = x · x · · · x in the integral (1. It is worth pointing out that the factor (2n − 1)!! has a combinatorial meaning. J2 (1. where only the quadratic terms are present in the Lagrangian. The main idea in such cases.2) This also follows easily from (1. In this onea dimensional setting a ﬁrst example of computation of an expectation value can be given in the form x2n := R x2n e− 2 ax dx R 1 2 e− 2 ax dx 1 2 = (2n − 1)!! . The onedimensional analog of Lagrangians that include interaction terms will be integrals of the form Z(J) = R e− 2 ax 1 2 +P (x)+Jx dx.3) where (2n−1)!! = (2n−1)·(2n−3) · · · 5·3·1. One obtains (1. It all starts with the simplest Gaussian integral e− 2 ax dx = R 1 2 2π a ∞ 0 1/2 . which we’ll see applied similarly to the inﬁnite dimensional case. reducing the problem in this . In physics one refers to such pairings as Wick contractions.3) inductively d from (1.4) where P (x) is a polynomial in x of degree deg P ≥ 3.1) by repeatedly applying the operator −2 da to (1.3). Similarly. by completing the square − a 2Jx a J J2 ax2 + Jx = − (x2 − ) = − (x − )2 + 2 2 a 2 a 2a and then changing coordinates in the integral to y = x − J .1) for a > 0. which follows from the usual polar coordinates calculation ∞ −∞ e− 2 ax dx 1 2 ∞ −∞ e− 2 ay dy = 2π 1 2 e− 2 ar r dr = 1 2 2π a ∞ 0 e−u du. the Gaussian integral with source term is given by e− 2 ax R 1 2 +Jx dx = 2π a 1/2 e 2a . the analog of the Gaussian integrals in the inﬁnite dimensional setting of quantum ﬁeld theory will be the free ﬁeld case.
in the case where the polynomial P (x) consists of a single term λ k x . (1. Namely. the perturbative expansion can be written in the form Z(J) = 2π a 1/2 exp λ k! d dJ k exp J2 2a . ﬁrst observe that the term of order λα and J β in Z(J) is produced by the combination of the term of order α in the Taylor expansion . one writes Z(J) = R P (x)n n! n=0 ∞ e− 2 ax 1 2 +Jx dx.4) is deﬁned to be the series 1 n! n=0 ∞ P (x)n e− 2 ax R 1 2 +Jx dx. (1. each given by the integral of a polynomial under a Gaussian measure.8) Two examples of this kind that will reappear frequently in the inﬁnite dimensional version are the cubic case with P (x) = g x3 and the quartic 6 λ case with P (x) = 4! x4 . so that one obtains P (x) = 1 n! n=0 1 n! n=0 ∞ ∞ R λ k x k! d dJ k n e− 2 ax n 1 2 +Jx dx = λ k! e− 2 ax R 1 2 +Jx dx.Perturbative quantum ﬁeld theory and Feynman diagrams 3 way to a series of terms.7). (1. the integral above satisﬁes xk e− 2 ax R 1 2 +Jx dx = d dJ k e− 2 ax R 1 2 +Jx dx. this gives xk e− 2 ax R 1 2 +Jx dx = 2π a 1/2 d dJ k e 2a . k! one can rewrite each term in the perturbative expansion using (1. for a monomial xk . J2 Thus.6) Notice then that.5) The perturbative expansion of the integral (1.2). To see then how the combinatorics of graphs can be used as a convenient device to label the terms of diﬀerent order in λ and J in the perturbative series of Z(J).7) Using (1. Thus. (1.
each with k half edges attached. Each half edge has an end that is connected to a vertex and another end that may pair to another half edge or remain unpaired. All the resulting 2a terms will be of a similar form. of the Green functions in quantum ﬁeld theory. consisting of a combinatorial factor given by a ratio of two products of factorials. This means a total of α vertices. Notice that the procedure described here produces not one but a ﬁnite α β J collection of graphs associated to a given monomial λ aκ . we can introduce the analog. α β J The way one assigns graphs to monomials of the form λ aκ is by the following rules. These combinatorial factors are the symmetry factors of graphs. • The power of a−1 is then determined by the resulting number of internal edges obtained by pairing all the half edges to form a graph. • To each factor of λ one associates a vertex of valence equal to the λ degree of the given monomial P (x) = k! xk . a power of λ and a power of 2a in the denominator. Here.8). connecting two vertices. which are associated to the vertices and the internal and external edges of the graph. which corresponds to the number of diﬀerent pairings that produce equivalent graphs. An external edge is an unpaired half edge attached to a vertex of the graph. one thinks of graphs as being constructed out of a set of vertices and a set of half edges. as in general in perturbative quantum ﬁeld theory. The graphs we consider will not necessarily be connected. • To each factor J one associates an external edge. hence it is an edge in the usual graph theoretic sense. a power of J. in this 1dimensional toy model. The combinatorial factor can then also be described in terms of symmetries of the graphs. each occurring with a given multiplicity. The function Z(J) . The graphs are introduced as a visual way to keep track of the power counting in these terms. depending on all the diﬀerent possible pairings between the half edges.4 Feynman Motives λ d of the exponential exp( k! ( dJ )k ) and the term of order β + kα in J in the 2 Taylor expansion of the other exponential exp( J ) in (1. An internal edge of the graph consists of a pair of half edges. This collection of graphs can in turn be subdivided into isomorphism types. To see more precisely how these factors can be computed. It is sometimes convenient to adopt the convention that a graph (or connected component of a graph) consisting of a single line should be thought of as consisting of an internal and two external edges.
9) where Z = R e− 2 ax 1 +P (x) dx and the Green functions are xN R N! R GN = e− 2 ax 1 2 +P (x) dx 1 2 e− 2 ax +P (x) dx . (1. These then correspond to the sum over all the possible topologically distinct graphs obtained via these pairings. N ! n! (k!)n The meaning of this factor in terms of symmetries of graphs can be explained.9).13) Notice also how. k )n ∞ (λx − 1 ax2 dx 2 n=0 (k!)n n! e and (1. by identifying (N + kn − 1)!! with the number of all the possible pairings of half edges. k! permutations of the half edges attached to each valence k vertex and n! permutations of the n valence k vertices along with their star of half edges. in terms of graphs.4) can be thought of as a generating function for the Green functions ∞ Z(J) = N =0 JN N! 2 xN e− 2 ax R 1 2 ∞ +P (x) dx = Z · N =0 J N GN .12) As we have seen in (1. from which one factors out N ! permutations of the external edges. each divided by its own symmetry factor. leaving all the diﬀerent pairings of half edges. (1.11). the terms of the asymptotic series in the numerator of (1. one obtains GN = k n ∞ xN (λx ) − 1 ax2 2 dx n=0 N ! (k!)n n! e . R 1 2 (1.Perturbative quantum ﬁeld theory and Feynman diagrams 5 of (1.3). these are (N + kn − 1)!! for even N + kn. when computing the terms of the asymptotic series using either the Taylor series of the exponentials of (1. #Aut(Γ) a#Eint (Γ) Γ∈graphs (1. with the odd ones vanishing by symmetry. with P (x) = formally exchanging the sum with the integral.8) or by ﬁrst using the .11) Using (1. Thus. Taking into account the other coeﬃcients that appear in (1. one obtains the factor (N + kn − 1)!! .10) λ k k! x Upon expanding out the interaction term exp(P (x)).10) become of the form λ#V (Γ) J #Eext (Γ) .9) and (1. we then see that one way of computing the coeﬃcient of a term α β J in λ aκ in the asymptotic expansion of Z(J) is to count all the pairings (the Wick contractions) that occur in the integration xN xkn e− 2 ax dx.
one replaces the expression 1 x2 + P (x) 2 of the onedimensional toy model we saw in the previous section with a nonlinear functional. deﬁned on a conﬁguration space of classical ﬁelds. e2 (det A)1/2 (1. One can again compute the asymptotic series for the integral e− 2 x RN 1 t Ax+P (x)+Jx dx1 · · · dxN . One replaces an integral of the form e− 2 x RN 1 t Ax+Jx dx1 · · · dxN = (2π)N/2 1 JA−1 J t . 2n n! Passing from the 1dimensional case to a ﬁnite dimensional case in many variables is notationally more complicated but conceptually very similar. aimed at giving a selfcontained introduction to mathematicians. one is implicitly using the combinatorial identity (2n − 1)!! = (2n)! . A more detailed presentation. The conceptually more diﬃcult step is to adapt this computational procedure for ﬁnite dimensional integral to a recipe that is used to make sense of “analogous” computations of functional integrals in quantum ﬁeld theory.14) is obtained by diagonalizing the matrix and reducing it back to the 1dimensional case. The real number J is here an N vector. such as P (x) = 4! ( i=1 x4 ). with Jx the inner product.12). .2 From Lagrangian to eﬀective action In the case of a scalar ﬁeld theory.6 Feynman Motives expansion in Green functions and then the terms (1. can be found in the book [Connes and Marcolli (2008)]. Here we give only a very brief account of the basics of perturbative quantum ﬁeld theory. where now instead of attaching a factor a−1 to the internal edges one ﬁnds factors (A−1 )ij for edges corresponding to a Wick contraction pairing an xi and an xj . One can use the same method of i labeling the terms in the asymptotic series by graphs. The form of (1. the Lagrangian density. where the interaction term here will be a polynomial in the coordinates N λ xi of x.14) where the positive real number a > 0 of the 1dimensional case is now replaced by an N × N real matrix A with At = A and positive eigenvalues. 1.
say φ ∈ C ∞ (RD . when one talks about a scalar ﬁeld theory one means the choice of the data of the Lagrangian density and the λ spacetime dimension D. .16) The subscript L here stands for the Lorentzian signature of the metric and we’ll drop it when we pass to the Euclidean version. In the following we follow the convention of taking units where = 1 so that we do not have to write explicitly the powers of in the terms of the expansions. it is both suﬃciently simple and suﬃciently generic with respect to the renormalization properties (e. This classical action is written as the sum of two terms SL (φ) = Sf ree. To the Lagrangian density one associates a classical action functional SL (φ) = RD L(φ)dD x.L (φ) = − RD P (φ)dD x. and the Lagrangian density is given by an expression of the form L(φ) = 1 m2 2 (∂φ)2 − φ − P(φ). Thus. The interaction term P(φ) in the Lagrangian is a polynomial in the ﬁeld φ of degree deg P ≥ 3. 2 2 (1. unlike the more physical φ4 in dimension D = 4 .g. . We will give explicit examples using the special case of the φ3 theory in dimension D = 6: while this is not a physically signiﬁcant example because of the unstable equilibrium point of the potential at φ = 0. −1.15) where (∂φ)2 = g µν ∂µ φ∂ν φ for g µν the Lorentzian metric of signature (1. R). We can assume for simplicity that P(φ) = k! φk . .Perturbative quantum ﬁeld theory and Feynman diagrams 7 In the scalar case the classical ﬁelds are (smooth) functions on a spacetime manifold. where the free ﬁeld part is Sf ree. An observable of a scalar ﬁeld theory is a functional on the conﬁguration space of the classical ﬁelds. (1. −1) on RD and a summation over repeated indices understood.L (φ) = RD 1 m2 2 D (∂φ)2 − φ d x 2 2 and the interaction part is given by Sint.L (φ). . nonsuperrenormalizable).L (φ) + Sint.17) where = h/2π is Planck’s constant. . −1. The probability amplitude associated to the classical action is the expression ei SL (φ) . (1.
in general the integral is illdeﬁned mathematically. . there are in fact formidable obstacles in trying to make rigorous sense of the functional integral involved. . (1.18) as a shorthand for a corresponding asymptotic expansion.19) GN. A closer similarity between (1. .21) . . .18) and (1.L (x1 . as long as one regards the expression (1. The expectation value of an observable is deﬁned to be the functional integral O(φ)eiSL (φ) D[φ] . . xN ) = . . . (1.20) where J is a linear functional (a distribution) on the space of classical ﬁelds and J. This is.8 Feynman Motives which we write as O(φ). . after factoring out a minus sign. which turns the probability amplitude into the Euclidean version eiSL (φ) → e−S(φ) . obtained by analogy to the ﬁnite dimensional case we have seen previously. Despite the successes of constructive quantum ﬁeld theory in several important cases. with the Euclidean action 1 m2 2 (∂φ)2 + φ + P(φ) dD x. in itself. φ = J(φ) is the pairing of the space of ﬁelds and its dual.18) eiSL (φ) D[φ] where the integration is supposed to take place on the conﬁguration space of all classical ﬁelds.4) appears when one passes to Euclidean signature by a Wick rotation to imaginary time t → it.22) 2 2 RD Thus. not an obstacle to doing quantum ﬁeld theory. 1. . If J = J(x) is itself a smooth function then J. (1. In particular. Although the notation of (1. deﬁned here as φ(x1 ) · · · φ(xN ) eiSL (φ) D[φ] . (1. .φ D[φ]. φ = RD J(x)φ(x)dD x.20) and (1. (1.19) is suggestive of what the computation of expectation values should be. This has the eﬀect of switching the signature of the metric to (1.23) e−S(φ) D[φ] S(φ) = (1. one has the N point Green functions. xN ) = eiSL (φ) D[φ] for which the generating function is given again by a functional integral with source term O(φ) = eiSL (φ)+ J. . 1). in the Euclidean version we are computing functional integrals of the form φ(x1 ) · · · φ(xN ) e−S(φ) D[φ] GN (x1 .
one uses an analog of the asymptotic expansion (1. .Perturbative quantum ﬁeld theory and Feynman diagrams 9 for which the generating function resembles (1.4) in the form Z[J] = satisfying Z[J] = Z[0] for Z[0] = e − RD 2 1 m2 2 (∂φ) + 2 e − RD 2 1 m2 2 (∂φ) + 2 φ2 +P(φ)+J(x)φ(x) dD x D[φ]. . and the latter in terms of sums over graphs as in (1. These are provided by the Feynman rules of the theory. . pN ) is the Green function in momentum space. xN )ei(p1 x1 +···+pN xN ) ··· . (1. p1 .3 Feynman rules By analogy to what we saw in the 1dimensional model. . . . . (1.13). external and internal edges of the graph. . .12). where one writes the Green functions (1.23) in terms of an asymptotic series whose terms are parameterized by graphs. .26) In order to make sense of this functional integral. . .25) φ2 +P(φ) dD x D[φ]. . i.24) ∞ N =0 1 N! J(x1 ) · · · J(xN ) GN (x1 . xN ) dD x1 · · · dD xN . . pN ) = Γ V (Γ. J and a−1 to the vertices. where one expands out the exponential of the interaction term Sint (φ) = RD P(x) dD x of the Euclidean action and one follows the same formal rules about integration by parts as in the ﬁnite dimensional case to label the terms of the expansion by graphs. D (2π) (2π)D (1.6). . the Fourier transform d D pN d D p1 GN (p1 . . one also writes the Green functions (1. . .e. pN ) . pN ) = GN (x1 . . . What is needed in order to write the contribution of a given graph to the asymptotic series is to specify the rules that associate the analogs of the powers of λ. .27) where GN (p1 . 1. (1.28) The reason for writing the contributions of Feynman integrals in momentum space is that in physics one does not only think of the Feynman graphs as computational devices that do the bookkeeping of terms in integration . . . . GN (p1 .11) in terms of integrals of the form (1. #Aut(Γ) (1.
and among them the subset of all those with a ﬁxed number of loops. . where λv is the coupling constant of the monomial in the Lagrangian of order equal to the valence of v and a conservation law for all the momenta that ﬂow through that vertex. . • Each vertex v ∈ V (Γ) contributes a factor of λv (2π)D .j pj ). .e = 0 for v ∈ ∂(e). . In the case of vertices with both internal and external edges (1.30) written after chosing an orientation of the edges of the graph. pN . . . . • Each internal edge e ∈ Eint (Γ) contributes a momentum variable ke ∈ RD so that dD k1 dD kn ··· .31) where the incidence matrix of the graph Γ is the #V (Γ)×#E(Γ)matrix with +1 for v = t(e) (1. . In fact.10 Feynman Motives by parts of polynomials under a Gaussian measure. where certain particles with assigned momenta (external edges) interact (vertices) by creation and annihilation of virtual particles (internal edges). The momenta ﬂowing through the graph then represent the physical process. . pN ) are constructed according to the Feynman rules as follows.30) is equivalently written in the form n N v. kn ) is constructed according to the following procedure. . . . p) := δ( + j=1 v. D (2π) (2π)D (1. but one can think of a diagram as representing a (part of) a physical process. . pN . kn ) δv (k) := δ( s(e)=v ke − t(e)=v ke ). . k1 .i ki i=1 δv (k. . The latter speciﬁes the order in the perturbative expansion one is looking at. it is clear from this point of view that what has physical meaning is not so much an individual graph but the collection of all graphs with given external edges and assigned external momenta. . . / .29) for n = #Eint (Γ). p1 . V (Γ. . . (1. . pN ) = IΓ (p1 . The term IΓ (p1 . k1 . . p1 .32) −1 for v = s(e) v. . . (1. . The terms V (Γ.
. . The study of parametric Feynman integral we discuss in greater detail in Chapter 3 can be done also for the Lorentzian case. We work here for convenience always with Euclidean signature in the Feynman integrals. pN ). . . pN ) = C δ( n i=1 v.j pj ) q1 (k1 ) · · · qn (kn ) dD k1 dD kn ··· . k1 . . . which in the case of a scalar ﬁeld in Euclidean signature is given by 2 qe (ke ) = ke + m2 . . (1. . . . . . with qe (pe ) = p2 + m2 . .Perturbative quantum ﬁeld theory and Feynman diagrams 11 • Each internal edge contributes an inverse propagator. . when summed over the oriented external edges. . pN ) U (Γ. The external momenta are assigned so that e they satisfy the conservation law e pe = 0. kn ) is then a product λv (2π)D δv (kei . where qe is a quadratic form. . . . as shown for instance in [Bjorken and Drell (1965)]. (2π)D (2π)D (1. pN ) = e∈Eext (Γ) qe (pe )−1 . .34) with linear relations among the momentum variables kei and pej imposed by the conservation laws δv (kei .33) • Each external edge e ∈ Eext (Γ) contributes a propagator qe (pe )−1 . pej ) v∈V (Γ) ei ∈Eint (Γ) qei (kei )−1 ej ∈Eext (Γ) qej (pej )−1 . p) is given by U (Γ. . so that our propagators contain quadratic terms of the form (1. . (1. a −1 term of the form qe . .37) with C = v∈V (Γ) λv (2π)D . .36) while the factor U (Γ. (1. • The integrand IΓ (p1 . pN ) = ε(p1 . p1 .i ki + N j=1 v. . pej ) at the vertices of the graph. p1 . that is. (1. pN . . . .33). . . We can then write the Feynman integral associated to a Feynman graph Γ of the given theory in the form V (Γ. p1 .35) where the factor ε(p) is the product of the inverse propagators of the external edges ε(p1 .
(1.3. one also has the following multiplicative form of the symmetry factors of graphs. . The basic property that makes these simpliﬁcations possible is the multiplicative form (1. pN ) is multiplicative on connected components of the graph Γ. Moreover. The factorials come from the symmetries of the graph Γ that permute topologically equivalent components. Proof. there are nj connected components of Γ all isomorphic to the same graph Γj ). . Thus.27). if the graph Γ has diﬀerent connected components. kn ). . A graph with no external edges is commonly referred to as a vacuum bubble. . the symmetry factor splits multiplicatively on components according to the formula #Aut(Γ) = j (nj )! j #Aut(Γj )nj .1. . All symmetries of Γ are obtained by composing this type of symmetries with symmetries of each component. these correspond to the terms with J = 0 in the asymptotic series. and then the expansion (1. p1 . connected graphs There are some useful simpliﬁcations that can be done in the combinatorics of graphs that appear in the formal series (1.12 Feynman Motives 1. One then has a ﬁrst useful observation on the combinatorics of the graphs that appear in the asymptotic expansion of the Green functions. .34) of the Feynman integrand IΓ (p1 .4 Simplifying graphs: vacuum bubbles. Lemma 1. Lemma 1. k1 . This implies the following property.37) of the Feynman integral. . .36). (1. This follows immediately from the form (1.25) into Green functions. when one writes the expansion (1.27) do not contain any vacuum bubbles. Lemma 1.4.4.4.38) Proof. pN .27) of the latter into Feynman integrals of graphs. Proof. . The graphs of (1. The Feynman integral V (Γ.2.35) with (1. . For a graph Γ that is a union of connected components Γj with multiplicities nj (i.34) and (1. no linear relations arise between momentum variables of the edges of diﬀerent components (as these have no common vertices) and the corresponding integrals split as a product. As we have seen in the ﬁnite dimensional toy model. . In fact. .e. the expansion .
. Lemma 1. p1 .4. .40) where now the Green functions GN. .35). only a smaller class of graphs will be involved. GN. . with multiplicities nj and with Nj external edges. Suppose that Γ is a disjoint union of connected components Γ = ∪j Γj . One can then pass from multiconnected to connected graphs by rewriting the functional integral Z[J] in an equivalent form in terms of W [J] = log Z[J] Z[0] . . .41) are computed as in (1.41) where GN. .c (x1 . . . p1 . so that j nj Nj = N . p1 . xN ) will also have an expansion on graphs of the form (1. and the multiplicative properties of Lemma 1. . . pN ) = Γ connected V (Γ. pN ) = #Aut(Γ) V (Γj . . We only sketch brieﬂy why the result holds. .4.42) and the V (Γ.2 we get V (Γ. xN )ei(p1 x1 +···+pN xN ) Proof.8) in the ﬁnite dimensional case.Perturbative quantum ﬁeld theory and Feynman diagrams 13 of the functional integral Z[J] would count the contribution of all graphs including components that are vacuum bubbles as in the case of (1. . . Then by Lemma 1.40) have an expansion GN. pN ) is the Fourier transform d D p1 d D pN ··· (2π)D (2π)D (1. (1. xN ) dD x1 · · · dD xN .c (p1 . The expansion of Z[0] on the other hand only has contributions from the vacuum bubble graphs.4. . p1 . . .27). .4.c (p1 . . pN ) . . . . .1 and Lemma 1. .c (p1 . . . (1. pNj )nj . The connected Green functions GN.2 then imply that the expansion for Z[J]/Z[0] only has contributions from graphs which do not contain any connected components that are vacuum bubbles. . . . pN ) = GN. .c (x1 . . however. where. More detailed expositions can be found in standard Quantum Field Theory textbooks (for example in [LeBellac (1991)]). . . . . .c (x1 . . . pN ) in (1. xN ) of (1. . (nj )!#Aut(Γj )nj j .4. .1 and 1. .c (x1 .4. . #Aut(Γ) (1.39) One can again write a formal asymptotic series for W [J] as ∞ W [J] = N =0 1 N! J(x1 ) · · · J(xN ) GN. .
i.5 Oneparticleirreducible graphs Further simpliﬁcations of the combinatorics of graphs can be obtained by passing to the 1PI eﬀective action.5. . and it cannot be disconnected by removal of a single vertex. the graph obtained by removing from Γ the star of the vertex v.14 Feynman Motives Thus. xNj ) = exp N 1 N! J(x1 ) · · · J(xN ) GN. The terminology more commonly in use in the physics literature is the following. Here what one means by removing a vertex from a graph is the following. it has at least 3 vertices. the graph Γ v is the graph with vertex set V (Γ) {v} and edge set E(Γ) {e : v ∈ ∂(e)}. 1. Given a graph Γ and a vertex v ∈ V (Γ).c (x1 . . etc. • For k ≥ 3. or higher loop versions like the 2PI eﬀective action. . and it cannot be disconnected by removal of any set of k − 1 vertices. . Deﬁnition 1. . it has at least k + 1 vertices.1. 1vertexconnected and 1edgeconnected simply mean connected. . . xN ) i dD xi nj = N j nj Nj =N j 1 nj J(x1 ) · · · J(xNj )GNj . We ﬁrst recall the following notions of connectivity of graphs. while for k ≥ 2 the condition of being kvertexconnected is stronger than that of being kedgeconnected. . Thus. • A graph is 2vertexconnected if it has no looping edges. . xN ) . We discuss here brieﬂy only the 1PI eﬀective action. though we will later need to return to discussing higher connectivity conditions on Feynman graphs. a graph is kvertexconnected if it has no looping edges and no multiple edges. . . The notion of kconnectivity of graphs is given as follows: • A graph is kedgeconnected if it cannot be disconnected by removal of any set of k or fewer edges.e. where vertex removal is deﬁned as below. we can write Z[J] 1 = Z[0] N! N J(x1 ) · · · J(xN ) GN (x1 .c (x1 . . .
a solution of the variational equation δ ( φ. . J − W [J]) J=J(φ) . even though a tree consisting of just n edges attached to a single valence n vertex cannot be disconnected by removal of a single edge (such edges are not internal though in this case). A graph that is disconnected by the removal of any one internal edge is a tree with at least one internal edge. Find an edge that disconnects the graph. after replacing the vertices by 1PI graphs with the number of external edges equal to the valence of the vertex. Γ1 .5. so that the semiclassical calculations (i. involving only graphs that are trees) done with the eﬀective action recover . namely Seﬀ (φ) = ( φ. this must still be true in the graph obtained after collapsing all the Γi . This suggests that there should be a way to further simplify the combinatorics of graphs in the asymptotic expansion of the functional integrals. Proof. which is deﬁned as the Legendre transform of the functional W [J].2. Notice that trees are all considered not to be 1PI. . It then suﬃces to show that the graph obtained from Γ by shrinking each Γi to a vertex is a tree.5.Perturbative quantum ﬁeld theory and Feynman diagrams 15 Deﬁnition 1. This is done by passing to the 1PI eﬀective action. Γn . Look at each component and again repeat the operations ﬁnding edges that disconnect them further until one is left with a collection of 1PI graphs.3. Suppose the graph is not 1PI.e. δJ The asymptotic expansion of the eﬀective action collects the contributions of all the 1PI graphs. J − W [J]) = 0. For k ≥ 2 a (k + 1)edgeconnected graph is also called kparticleirreducible (kPI).43) evaluated at a stationary J. . For k = 1. a 2edgeconnected graph that is not a tree is called oneparticleirreducible (1PI) graph. Lemma 1. If the connected graph Γ is 1PI the tree consists of a single vertex with the number of edges attached equal to the number of external edges of the graph. Since each of the internal edges that remain in this graph was an edge whose removal increased the number of connected components. These cannot be disconnected by removal of a single (internal) edge. . and getting back from these the contributions of all connected graphs. by counting separately the contributions of trees and that of 1PI graphs. that is. (1. Any connected graph can be obtained from a tree.
which belong to the two diﬀerent components after the edge removal. Since . Then. we recall the following observations from [Aluﬃ and Marcolli (2009a)]. and their relation to the 1PI condition.5. so that Γ is not 2vertexconnected. the vertices v and v2 belong to diﬀerent connected components. 2vertexconnectivity implies 2edgeconnectivity. To explain more precisely the diﬀerence between edge and vertex connectivity in Deﬁnition 1. Assume then that there exists v = v1 ∈ V (Γ1 ). Assume that Γ is not 1PI. (2) We need to show that 2vertexconnectivity is equivalent to all splittings Γ being 1PI.4. We have the following relation between 2vertexconnectivity and 2edgeconnectivity (1PI). for a graph Γ with at least 3 vertices and no looping edges. (1) If Γ is 2vertexconnected then it is also 2edgeconnected (1PI). at least one of the two components contains at least two vertices. Lemma 1. (2) Γ is 2vertexconnected if and only if all the graphs Γ obtained as splittings of Γ at any v ∈ V (Γ) are 2edgeconnected (1PI). e has two distinct endpoints v1 and v2 .5. after the removal of the vertex v1 from Γ. Suppose ﬁrst that Γ is not 2vertexconnected. Since Γ has no looping edges. Since Γ has at least 3 vertices. Then there exists an edge e such that Γ e has two connected components Γ1 and Γ2 .16 Feynman Motives the full quantum corrections to the classical action given by all the connected graphs that appear in the expansion of W [J]. (1) We have to show that.5. as in the example in the ﬁgure. Let Γ be a graph with at least 3 vertices and no looping edges.1 above. Deﬁnition 1. E = E1 ∪ E2 and inserting a new edge e to whose end vertices v1 and v2 the edges in the two sets E1 and E2 are respectively attached.5. A graph Γ is a splitting of Γ at a vertex v ∈ V (Γ) if it is obtained by partitioning the set E ⊂ E(Γ) of edges adjacent to v into two disjoint nonempty subsets. We do not prove here how one gets the asymptotic expansion for the eﬀective action and we refer the interested reader to [Connes and Marcolli (2008)] and [LeBellac (1991)]. Proof.
So e must be the edge added in the splitting of Γ at the vertex v. . and Γ v is contained in Γ e. Since we are assuming that Γ has no multiple or looping edges.5. Let v1 and v2 denote the remaining two endpoints. it follows that removing v from Γ would also disconnect the graph. then Γ would not be 1PI (and hence not 2vertex connected. If w is adjacent to v. by hypothesis. None of the other edges adjacent to v1 or v2 is a looping edge. or v. We know it is 1PI because of the previous lemma. there exist two edges e1 and e2 such that the removal of both edges is needed to disconnect the graph.Perturbative quantum ﬁeld theory and Feynman diagrams 17 Γ has at least 3 vertices and no looping edges. Similarly. Proof. If the vertices v1 and v2 belong to diﬀerent connected components after removing e1 and e2 . the failure of 2vertexconnectivity means that there exists a vertex v whose removal disconnects the graph. Thus Γ is not 2vertexconnected.6. Thus. then removing / v and v1 leaves v2 and w in diﬀerent connected components. Then the splitting Γ of Γ at v obtained by inserting an edge e such that the endpoints v1 and v2 are connected by edges. Since the graph has at least 4 vertices and no multiple or looping edges. If e already belonged to Γ. vi ∈ ∂ei . so that Γ is not 3vertexconnected (in fact not even 2vertexconnected). respectively. assume that there exists a splitting Γ of Γ at a vertex v that is not 1PI. Let Γ be a graph with at least 4 vertices. to the vertices in V1 and V2 is not 1PI. v2 }. Assume that Γ is 3vertexconnected but not 2PI. with the other endpoint w ∈ {v. with no looping edges and no multiple edges. by (1)). then v must be in a diﬀerent component. Let V ⊂ V (Γ) be the set of vertices other than v that are endpoints of the edges adjacent to v. there exists at least one other edge attached to either v1 . and a vertex v2 = v1 adjacent to v2 . There exists an edge e of Γ whose removal disconnects the graph. v1 . v1 = v2 . We argue by contradiction. then the removal of the vertex v disconnects the graph. Then 3vertexconnectivity implies 3edgeconnectivity. If v1 and v2 belong to the same connected component. Suppose ﬁrst that they have a common endpoint v. This set is a union V = V1 ∪ V2 where the vertices in the two subsets Vi are contained in at least two diﬀerent connected components of Γ v. the two edges have at most one endpoint in common. Since v1 and v2 are in Γ v. Since Γ e is disconnected. Conversely. v1 and v2 are in distinct connected components of Γ e. Lemma 1. Let v1 and v2 be the endpoints of e. therefore there exist at least another vertex v1 = v2 adjacent to v1 . as removal of e would disconnect it. v2 .
the 2PI eﬀective action is often considered in quantum ﬁeld theory in relation to nonequilibrium phenomena. . At least one pair {vi .18 Feynman Motives if w is adjacent to (say) v1 .1. namely the renormalization problem: how to extract in a consistent and physically signiﬁcant manner ﬁnite values from the divergent integrals (1. if one tries to assign to such integrals a numerical value. However. then the removal of the two vertices v1 and v2 leave v and w in two diﬀerent connected components. one soon realizes that most of them are in fact divergent. The problem of renormalization consists of three main aspects: • Regularization • Subtraction • Renormalization .6 The problem of renormalization So far we have treated the integrals U (Γ. [Rammer (2007)]. Let v1 and w1 be the endpoints of e1 and v2 and w2 be the endpoints of e2 . It also cannot happen that v2 and w2 belong to the same component. Jimbo. p1 . . suppose that e1 and e2 have no endpoint in common. Oshima (1976)]).37) as purely formal expressions. In particular. else the removal of e1 alone would disconnect the graph. say. wi } belongs to two separate components after the removal of the two edges. §10. .g. Next. see e. Miwa. 1.5. Suppose then that v1 and w1 are in diﬀerent components. This was historically one of the main problems in the development of perturbative quantum ﬁeld theory. Then the removal of v1 and w2 leaves v2 and w1 in two diﬀerent components so that the graph is not 3vertexconnected. v2 belongs to the same component as v1 while w2 belongs to a diﬀerent component (which may or may not be the same as that of w1 ). pN ) of (1. Hence Γ is not 3vertexconnected. Conditions of 3connectivity (3vertexconnected or 3edgeconnected) arise in a more subtle manner in the theory of Feynman integrals. though not all four points can belong to diﬀerent connected components. In the following we restrict our attention to Feynman integrals of graphs that are at least 1PI. in the analysis of Laundau singularities (see for instance [Sato.37) that appear in the asymptotic expansion of the functional integrals of quantum ﬁeld theory. else the graph would not be 1PI. We can then assume that. .
but has ﬁnite values for other values of the regularization parameters. of a formal series of contributions coming from all the divergent graphs of the theory. Renormalization means being able to perform the subtraction procedure by modifying the parameters in the Lagrangian (which become themselves functions of the regularization parameter). 2 2 k! (1. This is what makes it possible. To understand more clearly the last point. other monomials δλi ki (ki )! φ . then the Lagrangian can be modiﬁed by changing the coeﬃcients to L(φ) = 1 + δZ m2 + δm2 2 λ + δλ k (∂φ)2 + φ + φ .Perturbative quantum ﬁeld theory and Feynman diagrams 19 Regularization consists of a procedure that replaces the divergent integrals (1. that were not initially present. Notice that a theory is still renormalizable if. if one introduces a regularization of the divergent Feynman integrals in terms of a complex parameter z (as we discuss below) or in terms of a cutoﬀ Λ. In fact. Thus. a theory is no longer renormalizable if an inﬁnite number of additional terms is needed to compensate for the divergences. it is important to stress the fact that the parameters that appear in the Lagrangian. They consist. nor are they the same as the actual masses and physical quantities that one can measure in experiments.e. to correct for divergent graphs by readjusting the parameters in the Lagrangian. in a renormalizable theory. one can just think of this as being the eﬀect of having chosen an arbitrary value = 0 for the coefﬁcients of these terms in the initial Lagrangian. in addition to the modiﬁcations of the coeﬃcients of the terms initially present in the Lagrangian. in fact. δm2 and δλ depend on the regularization parameter. However. i. This is not all there is yet. the parameters in the Lagrangian can be modiﬁed without aﬀecting the physics one observes. Subtraction then consists of removing the divergent part of the regularized integrals by a uniform procedure (such as removing the polar part of a Laurent series in the main example we use below).37) by functions of some regularization parameters. In fact. to compensate for divergent graphs one needs to also add a ﬁnite number of other terms. such as masses and coupling constants of the interaction terms. in such a way that the resulting function has a pole or a divergence for particular values or limits of the additional parameters that correspond to the original divergent integral.44) where the functions δZ. . are not physical observables.
(1. First recall that the Gamma function is deﬁned by the integral ∞ Γ(t + 1) = 0 st e−s ds.48) Another related way to reformulate the expression (1. The ﬁrst useful operation on Feynman integrals is the introduction of Schwinger parameters.47) This is a special case of the more general useful identity k q1 1 Rn + e−(s1 q1 +···+sn qn ) sk1 −1 · · · skn −1 ds1 · · · dsn . These are based on the very simple identity 1 = q ∞ 0 e−sq ds. (1. n 1 (1. namely Γ(n+1) = n! for nonnegative integers.37) as 1 = q1 · · · qn 1 1 = kn Γ(k1 ) · · · Γ(kn ) · · · qn Rn + e−(s1 q1 +···+sn qn ) ds1 · · · dsn . As we are going to see in more detail below.49) .46) is the Feynman trick 1 = ab 1 0 1 dt.46) This allows the reformulation of integrals where quadratic forms qi in the momentum variables appear in the denominator in terms of Gaussian integrals where quadratic forms appear in the exponent. one writes the denominator of a Feynman integral of the form (1. Schwinger and Feynman parameters We digress momentarily to recall some useful formulae involving Gamma functions. In terms of Schwinger parameters. hence it extends the factorial.20 Feynman Motives 1. (ta + (1 − t)b)2 (1.7 Gamma functions. Here the basic example. which are extensively used in Feynman integral computations and are the basis of both the dimensional regularization procedure we describe below and the parametric representation of Feynman integrals that we discuss later and which is the basis for the relation between Feynman integrals and periods of algebraic varieties.37) of Feynman integrals is based on the Feynman parameters. analogous to (1. (1. The function Γ(t) deﬁned in this way extends to a meromorphic function with poles at all the nonpositive integers. a typical way of dealing with divergences in Feynman integrals is to ﬁrst identify them with poles of some meromorphic function and typically a product of Gamma functions.45) It satisﬁes Γ(t+1) = t Γ(t).
much less is needed.47) is obtained by considering the general formula (1. . (1. one can use the right hand side of (1. p1 . In fact. . Due to the very special form of the Feynman integrals U (Γ.52) 1. 2 (1. .1]n Thus. using the method of Schwinger parameters described brieﬂy in §1.51) [0. (1. Clearly. since the right hand side of (1.37) this gives 1 Γ(k1 + · · · + kn ) = k1 kn Γ(k1 ) · · · Γ(kn ) q1 · · · qn 1 = (n − 1)! q1 · · · qn δ(1 − i ti ) dt1 · · · dtn .53) continues to make sense for D ∈ C. tn ) ∈ Rn  + i ti = 1}.8 Dimensional Regularization and Minimal Subtraction Dimensional regularization (DimReg) is based on a formal procedure aimed at making sense of integrals in “complexiﬁed dimension” D−z. one can then reformulate the Feynman integrals in terms of Gaussian integrals. It would seem at ﬁrst that one needs to develop a notion of geometry in “complexiﬁed dimension” along with measure spaces and a suitable theory of integration. . e−λt dD t = π D/2 λ−D/2 .53) as the deﬁnition of the left hand side and set e−λt dz t := π z/2 λ−z/2 . it suﬃces to have a good deﬁnition for the Gaussian integrals in D dimensions. . instead of integral dimension D ∈ N. with z ∈ C∗ .7.54) . (t1 q1 + · · · + tn qn )n [0. .50) In the particular case of the denominators of (1. the integration is performed on the ndimensional topological simplex σn = {t = (t1 . In fact. 2 ∀z ∈ C.53) in the case where D is no longer a positive integer but a complex number.1]n (1. (t1 q1 + · · · + tn qn )n (1. pN ). in order to deﬁne dimensional regularization property.48) and performing a change of variables si = Sti with S = s1 + · · · + sn so as to obtain tk1 −1 · · · tkn −1 δ(1 − i ti ) n 1 dt1 · · · dtn .Perturbative quantum ﬁeld theory and Feynman diagrams 21 The more general expression analogous to (1. . .
. i. taking (1 − T)U (Γ)z=0 does not suﬃce as a renormalization method. As we see below (see also [Connes and Marcolli (2008)] for a more detailed explanation). The method of minimal subtraction consists of removing the divergent part of the Laurent series expansion at z = 0 of a meromorphic function obtained from the regularization of the Feynman integrals. .e. due to the role of subdivergences. With this notation. It is not an easy result to show that the dimensionally regularized integrals give meromorphic functions in the variable z. This is a convergent power series in the variable z ∈ C∗ . pN )) = IΓ (p1 . See a detailed discussion of this point in Chapter 1 of [Connes and Marcolli (2008)]. We will not enter into details here and talk loosely about (1. It will play an important role later on. then (1 − T)U (Γ) would be its minimal subtraction. . smaller subgraphs γ of the Feynman graph Γ which themselves already contribute divergent Feynman integrals U (γ). . with a Laurent series expansion at z = 0. hence the renormalization group ﬂow. . . pN . (1. [Collins (1986)] the simplest example of dimensional regularization. . if U (Γ) = U (Γ)(z) is shorthand for the meromorphic function obtained from the dimensional regularization of the Feynman integral associated to the graph Γ (suppressing the explicit dependence on the external momenta). . k ) µz dD−z k1 · · · dD−z k . We report here from [Connes and Marcolli (2008)]. k1 . . which then has a ﬁnite value at z = 0. for the selfenergy graph of the theory T with L (φ) = φ3 and D = 6.55) as a meromorphic function of z depending on the additional parameter µ. .1. . The correct procedure that uses dimensional regularization and minimal subtraction to extract ﬁnite values from divergent Feynman integrals is given by the Bogolyubov recursion described below in §5. . which one writes formally as z Uµ (Γ(p1 .22 Feynman Motives One then obtains well deﬁned Feynman integrals in complexiﬁed dimension D − z. One denotes by T the operator of projection of a Laurent series onto its polar part. as it sets the dependence on the energy scale of the renormalized values of the Feynman integrals. .55) The variable µ has the physical units of a mass and appears in these integrals for dimensional reasons.
It seems then that one could simply cure these divergences by removing the polar part of the Laurent series obtained by dimensional regularization. 2 + m2 (p + k)2 + m2 k s>0. In this case. due to the eﬀect of the delta functions imposing the conservation law at the two vertices. This follows 2 2 6 directly from the application of the Feynman rules: the variables assigned to the two internal edges are −k and p + k. One obtains in this way a Gaussian in q = k + xp and using the prescription (1. Slightly more complicated examples with nested divergent graphs . the method of Schwinger parameters described in §1.7 consists of replacing the integral above by 2 2 2 2 1 1 = e−s(k +m )−t((p+k) +m ) ds dt. t>0 One can then diagonalize the quadratic form in the exponential to get −Q(k) = −λ ((k + xp)2 + ((x − x2 )p2 + m2 )) with s = (1 − x)λ and t = x λ. which makes sense for D ∈ C∗ and shows the presence of a pole at D = 6. where it is usually interesting to consider this scalar ﬁeld theory. with p the external momentum. It is easy to see that this integral is divergent.Perturbative quantum ﬁeld theory and Feynman diagrams 23 p+k p k p This corresponds to the divergent Feynman integral (neglecting the constant multiplicative factor containing the coupling constant λ and powers of 2π) is given by 1 1 dD k k 2 + m2 (p + k)2 + m2 for the theory with Lagrangian L(φ) = 1 (∂φ)2 + m φ2 + λ φ3 .54) one then gets 1 0 0 1 ∞ 0 1 ∞ 2 e−(λ(x−x 2 )p2 +λ m2 ) e−λ q dD q λ dλ dx λ−D/2 λ dλ dx 2 = π D/2 0 e−(λ(x−x 2 )p2 +λ m2 ) = π D/2 Γ(2 − D/2) 0 ((x − x2 )p2 + m2 )D/2−2 dx. for instance in dimension D = 4 or D = 6.
In such cases. Deﬁnition 1. In fact. the theory will end up depending on inﬁnitely many parameters. 6 While one can check that this works for the example given here above. . with polynomial interaction terms P(φ) deﬁnes a renormalizable quantum ﬁeld theory if all the divergences arising from the corresponding Feynman integrals U (Γ) can be corrected by repeatedly altering the bare constants of the existing terms in the Lagrangian or by adding a ﬁnite number of new polynomial terms to L(φ). However.8. We conclude this introductory section by recalling more formally the main distinction we already mentioned between renormalizable and nonrenormalizable theories. by redoing the Feynman integral computation for the modiﬁed Lagrangian L(φ) = 1 (∂φ)2 (1 − δZ) + 2 m2 − δm2 2 φ2 − g + δg 3 φ . that is. the main point is that one would like to cancel the divergence by correcting the coeﬃcients of the Lagrangian by functions of z. since inﬁnitely many new terms will have to be added to the Lagrangian to correct for the divergences of increasingly complicated graphs.1 below. one can still follow the same renormalization procedure. A quantum ﬁeld theory with Lagrangian L(φ) = 1 m2 2 2 2 (∂φ) + 2 φ + P(φ).24 Feynman Motives (see [Connes and Marcolli (2008)] [Collins (1986)]) show why a renormalization procedure is indeed needed at this point.1. This will be explained in more detail in §5. by correcting the Laurent series U (Γ) by a linear combination of other Laurent series associated to the subgraphs. one needs to account for the way the divergences of the subgraphs have already been renormalized. the counterterms. it does not seem to work any longer for more complicated examples where the graph Γ contains divergent subgraphs. In the case of a nonrenormalizable theory.
for which the more complicated theory of mixed motives is needed. 2. This will be a simple introduction.Chapter 2 Motives and periods After reviewing the main tools from quantum ﬁeld theory that we need.1 The idea of motives Grothendieck introduced the idea of motives as a universal cohomology theory for algebraic varieties. these do not suﬃce for the quantum ﬁeld theoretic applications. We ﬁrst sketch the construction and main properties of pure motives. a universal Euler characteristic given by the class in the Grothendieck ring of varieties. at least conjecturally. we recall the role of Galois symmetries in categories of motives. we give here a very brief and sketchy account of the other subject whose relation to quantum ﬁeld theory will then be the main topic of the rest of the book. mainly arise in quantum ﬁeld theory. like the e 25 . We are even sketchier when it comes to describing the world of mixed motives. We describe brieﬂy the role of periods. There are diﬀerent notions of cohomology that can be deﬁned for algebraic varieties: Betti cohomology. As will become clearer in the following chapter. Finally. There are relations between them. and focus only on the minimal amount of information that is necessary in order to explain how the connection to quantum ﬁeld theory arises. ´tale cohomology. the theory of motives of algebraic varieties. which will be the type of motives that. which will be the main point of contact with Feynman integral computations. were we will inevitably avoid the deeper and more diﬃcult aspects of the theory. with special attention to the case of multiple zeta values. algebraic de Rham cohomology. We also show how one can estimate the complexity of a motive through an associated invariant. but we give some concrete examples of mixed Tate motives.
A more recent reference that gives a broad survey of the current state of the ﬁeld is [Andr´ (2004)]. or “noncompact”. e The theory of motives comes in two ﬂavors. the motivic Galois group. that already exist at the level of this universal cohomology. this will be by no account an accurate introduction to motives and we refer the readers interested in a more serious and detailed account of the current state of research in this area to the recent book [Andr´ (2004)]. with a corresponding Galois group of symmetries. The idea of a universal cohomology that lies behind all of these diﬀerent realizations asserts that properties which occur in all the reasonable cohomologies and relations between them should come from properties that are motivic. one obtains the theory of pure motives. meaning varieties that may be singular. An extensive series of papers on motives is e available in the proceedings of the 1991 Seattle conference. The ﬁeld is in fact still a very active and fast developing area of algebraic and arithmetic geometry.e. although problems like explicitly identifying the motivic Galois groups of various signiﬁcant subcategories of the category of pure motives. For our purposes we will try to follow the shortest path through this intricate theory that will allow us to draw the connection to quantum ﬁeld theory and discuss the main problems currently being investigated. This part of the theory of motives is at this point on solid ground. and where the morphisms are no longer just given by morphisms of varieties but by more general correspondences given by algebraic cycles in the product. In the ﬁrst case. which is still an excellent short introduction to the subject. Serre (1994)]. one can obtain in this way an abelian category of pure motives and in fact a Tannakian category. collected in the two volumes [Jannsen. which is built by ﬁrst abstracting the main properties that a “good” cohomology theory of such algebraic varieties should satisfy (Weil cohomologies) and then constructing a category whose objects are smooth projective varieties together with the data of a projector and a Tate twist. Thus. While the general theory of pure motives is a beautiful and . or identifying when a correspondence is realized by an algebraic cycle. Under a suitable choice of an equivalence relation on the set of algebraic cycles. was [Manin (1968)]. where one only considers motives of smooth projective varieties. Kleiman. Setting up a theory of motives is no easy task. i. For a long time the only available written reference giving a detailed account of the theory of motives. roughly corresponding to the distinction between working with smooth projective varieties or with everything else.26 Feynman Motives period isomorphism between de Rham and Betti cohomologies. remain extremely challenging. along with some interesting algebrogeometric applications.
with Poincar´ duality being replaced by the pairing of cohomologies e and cohomologies with compact support. is based on triangulated categories. and for very particular subcategories of the triangulated category of mixed motives. Only in very special cases. . This is the case for the subcategory of mixed motives that is most directly of interest to quantum ﬁeld theory. the main problem of establishing Grothendieck’s standard conjectures remains at this time open. and with a prominent role played by Mayer–Vietoris type long exact sequences. in the form which is presently available after the work of Voevodsky. so that they cannot be described. where these types of cohomology exact sequences manifest themselves in the form of distinguished triangles. using only pure motives. namely the category of mixed Tate motives over a number ﬁeld. When one is studying the motivic nature of certain classes of varieties. as will be the case here with certain hypersurfaces associated to Feynman graphs. namely motives associated to varieties that are no longer necessarily smooth and projective. with the same inclusionexclusion property of an Euler characteristic.Motives and periods 27 well established part of arithmetic algebraic geometry. The theory of mixed motives. but one can sometime get useful information about the motives by looking at their classes in the Grothendieck ring. As it happens in topology. the reason why it is mixed motives instead of the simpler pure motives that arise in the quantum ﬁeld theoretic context lies in the fact that certain algebraic varieties naturally associated to Feynman graphs are typically singular. it is often too compicated to work directly in the triangulated category of mixed motives. also in the context of algebraic varieties passing to the “noncompact” case profoundly alters the main properties of the cohomology. at the motivic level. The ﬁrst issue in this case is that of identifying the right set of properties that a good cohomology theory is expected to have and that should then be also associated to a good category of motives. This consists of isomorphism classes of varieties up to equivalence relations induced by very rough “cutandpaste” operations on algebraic varieties. The situation is far more involved when it comes to mixed motives. is it possible to extract from the triangulated structure a heart giving rise to an abelian category. As we are going to see in the next chapter.
2.3) which is functorial with respect to pullbacks and pushforwards and satisfying j n i γX×Y (Z × W ) = γX (Z) ⊗ γY (W ). (2.1) H i (X1 ) ⊗K Hj (X2 ). . This means that. such a variety X deﬁnes a smooth complex manifold X(C). Z → K being the standard if H denotes a smooth hyX the inclusion. • The weak Lefschetz isomorphisms hold: perplane section of X. • Poincar´ duality holds: H 2 dim(X) (X) = K and there is a none degenerate bilinear pairing ·. and normalized by γpt : inclusion. (2. . which is embedded in some suﬃciently large projective space PN .2) with the isomorphism induced by the projection maps πi : X1 ×K X2 → Xi . at each archimedean place of the number ﬁeld. . • Let Zi (X) denote the abelian group of algebraic cycles in X (smooth connected subvarieties) of codimension i. · : H i (X) × H 2 dim(X)−i (X) −→ K. then ι∗ : i ≤ dim(X) − 2 and injec . for i = 1.2 Pure motives The main properties of a good cohomology theory of algebraic varieties are summarized by the concept of Weil cohomology. 2 dim(X)}. • H i (X) = 0 outside of the range i ∈ {0. We assume we are working with smooth projective varieties over a ﬁeld K. • The K¨nneth formula holds: u H n (X1 ×K X2 ) = i+j=n (2. i with i + j = n. . which for our purposes we can assume is Q or a ﬁnite algebraic extension (a number ﬁeld).28 Feynman Motives 2. with ι : H → H i (X) → H i (H) is an isomorphism for tive for i = dim(X) − 1. • H · is a contravariant functor from the category VK of smooth projective varieties to the category GrVectK of graded ﬁnite dimensional Kvector spaces. Then there exists a cycle map i γX : Zi (X) → H 2i (X). The axioms of a Weil cohomology are the following.
An abelian category C is a category satisfying the following properties. the trivial abelian group. 0) = 0. is closely related to this question. such as the period isomorphism between Betti and de Rham cohomology. Then Ldim(X)−i : H i (X) → 2 dim(X)−i H (X) is an isomorphism for all i ≤ dim(X). K) and Betti (singular) cohomol∗ ∗ ogy HB (X) = Hsing (X(C). • There is a zeroobject 0 ∈ Obj(C) with HomC (0. The fact that in general algebraic cycles are very complicated objects and one does not know much about how to characterize them homologically constitutes the main diﬃculty in the theory of pure motives. It has additional objects and morphisms that make it into a “linearization” of the category of varieties. Y ) is an abelian group (i. Grothendieck’s idea that gave birth to the whole theory of motives was that there should be a cohomology theory for (smooth projective) algebraic varieties that should map via realization functors to any cohomology satisfying the axioms recalled above. It is in general extremely diﬃcult to determine whether a given class in a ∗ Weil cohomology H ∗ (X) lies in the image of the cycle map γX . a category that is good enough to behave as a category of modules over a ring. The composition of morphisms is compatible with the abelian group structure of the Homsets. Q) are both examples of Weil cohomologies. motivated this search for an underlying universal theory. The existence of isomorphisms between the diﬀerent known Weil cohomologies. Y ∈ Obj(C) the set of morphisms HomC (X. is that of an abelian category. namely a category that is good enough to do homological algebra.e. Deﬁnition 2. as cohomologies typically do. . One modiﬁes instead the maps between varieties so as to obtain a category that still contains the usual category of smooth projective varieties. ∗ Algebraic de Rham cohomology HdR (X. The Hodge conjecture. The approach to constructing such a universal cohomology theory is to remain suﬃciently close to the geometry of the varieties themselves. • For all X. see [Gelfand and Manin (1994)]. instead of passing to associated vector spaces or modules over a ring. that is.Motives and periods 29 • The hard Lefschetz isomorphisms hold: let L : H i (X) → H i+2 (X) 1 be given by L : x → x ∪ γX (H).1. as the construction of the category of pure motives that we now recall will show more clearly. for example.2. C is preadditive). The precise notion for a category that behaves like a category of modules over a ring.
we regard the graph G(f ) of a morphism f : X → Y as a correspondence from Y to X. (2.8) f1 f2 for Z ⊂ X1 × X2 and Z ⊂ X2 × X3 . which admits realization functors to vector spaces determined by the various possible Weil cohomologies. K = Coker(f ) ∈ Obj(C). k i j c (2.6) Since one is aiming at constructing a universal cohomology theory. To obtain a category of pure motives one then aims at constructing an abelian category of smooth projective varieties and correspondences. and with K = Ker(f ) ∈ Obj(C). (2. and I = Coker(k) = Ker(c) ∈ Obj(C). This means that.5) This is a closed connected subvariety of the product X × Y of codimension codim(G(f )) = dim(Y ).4) with j ◦ i = f . the generalization of morphisms f : X → Y is given by correspondences of the form Z = j nj Zj .7) where Corri (Y. • All morphisms have kernels and cokernels in the category and every morphism f : X → Y has a canonical decomposition K→X→I→Y →K. A morphism f : X → Y of varieties can be regarded as a correspondence by considering its graph G(f ) = {(x. The property that an abelian category “behaves like a category of modules over a ring” is made precise by the Freyd–Mitchell theorem. y) ∈ X × Y  y = f (x)} ⊂ X × Y. which shows that indeed a small abelian category can be faithfully embedded as a subcategory of a category of modules over a ring. Thus. The usual composition of morphisms X1 → X2 → X3 is replaced by a composition of correspondences given by the ﬁber product Z ◦ Z = Z ×X2 Z. X). to obtain the composition Z ◦ Z one does the following operations: . (2.30 Feynman Motives • There are ﬁnite products and coproducts in C. which should be a contravariant functor. X) is the abelian group of correspondences from Y to X given by algebraic cycles in Zi (X × Y ). that are ﬁnite linear combinations with nj ∈ Z of irreducible smooth algebraic subvarieties Zj ⊂ X × Y . G(f ) ∈ Corrdim(Y ) (Y. (2.
nj Zj and Because we are working with algebraic varieties as opposed to smooth manifolds. at least in some cases. −1 −1 Z ◦ Z = (π31 )∗ (π32 (Z ) ∩ π21 (Z)) ⊂ X1 × X3 . Although this second relation depends on the choice of a Weil cohomology. #(Z1 ∩ Z) = #(Z2 ∩ Z). We only mention here numerical and homological equivalence. one needs to use some more sophisticated tools like excess intersection to deal with intersection products in nontransverse situations.9) j • Extend by linearity to correspondences of the form Z = Z = r nr Zr . for arbitrary subvarieties Z of complementary dimension. This is compatible with the composition of correspondences deﬁned using the ﬁber product. ﬁnite dimensional. • Push forward the result along the remaining projection π31 : X1 × X2 × X3 → X1 × X3 . and use an equivalence relation on the set of algebraic cycles that preserves the intersection product and allows for diﬀerent choices of representatives within the same class. thus making the latter independent of the choice of the cohomology. Z ∼hom 0. i if its image under the cycle map γX×Y : Zi (X × Y ) → H 2i (X × Y ) is i trivial. Thus. We also .Motives and periods 31 −1 −1 • Take the preimages π32 (Z ) and π21 (Z) in X1 × X2 × X3 . with πij : X1 × X2 × X3 → Xi × Xj the projections. γX×Y (Z) = 0. Reducing the size of the groups of algebraic cycles by an equivalence relation is also desirable to obtain groups that are. as one does in algebraic geometry to deﬁne the product in the Chow ring. The homological equivalence requires the choice of a Weil cohomology H ∗ and prescribes that a cycle Z ∈ Zi (X × Y ) is homologically trivial. (2. We refer the reader to [Andr´ (2004)] for a discussion of rational equivalence and the corresponding e properties. There is more than one possible choice of an equivalence relation on algebraic cycles. Imposing the numerical equivalence relation means identifying two cycles Z1 ∼num Z2 if their intersection products agree. and the properties of the resulting category of motives change according to which equivalence relation one considers. it is part of Grothendieck’s standard conjectures that numerical and homological equivalence agree. −1 −1 • Intersect π32 (Z ) ∩ π21 (Z) in X1 × X2 × X3 . One denotes by Corri (X ×Y ) ∼ the abelian group generated by the algebraic cycles of codimension i in X × Y modulo the equivalence relation ∼. one in general cannot simply deform them to be in general position −1 −1 so that the intersection π32 (Z ) ∩ π21 (Z) is transverse.
The category Meﬀ (K) has a tensor product that comes form the prodnum uct of varieties. with a bifunctor ⊗ : C × C → C and an object 1 ∈ Obj(C) for which there are functorial isomorphisms aX. Y ∈ Obj(C). It is in general very diﬃcult to give an explicit description in terms of algebraic cycles of a projector that isolates a given component of the cohomology of a variety. for example the piece that corresponds to a certain degree only. ∼ ∼ One can then deﬁne a category of motives as follows.10) The reason for introducing projectors p2 = p is in order to be able to “cut out” pieces of the cohomology of a variety X. Recall ﬁrst that a category C where the set HomC (X. the property of being idempotent also depends on the ∼ choice of the equivalence relation. Deﬁnition 2. These satisfy consistency conditions expressed in the form of a triangle.2. Y ) := Corrdim(X) (Y × X) ⊗Z Q. Y )p. The morphisms are num given by HomMeff (K) ((X. One e . p). A tensor category over K is a Klinear category C.Z : X ⊗ (Y ⊗ Z) → (X ⊗ Y ) ⊗ Z cX. ∼ num (2. The semisimplicity property num means that there exists a set S of objects such that for all X ∈ S one has Hom(X. Y ) is a Kvector space.11) (2. q)) = qC0 (X.Y : X ⊗ Y → Y ⊗ X lX : X ⊗ 1 → X and rX : 1 ⊗ X → X. This makes it into a tensor category. Y ) = 0 for all X = Y ∈ S (simple objects). Notice also that. a pentagon.32 Feynman Motives use the notation C0 (X. The category Meﬀ (K) of eﬀective numerical pure monum tives has as objects pairs (X. and a hexagon diagram. X) = Q and Hom(X. X) is an idempotent. X). and such that every object X of the category has a decomposition X = ⊕i Xi with summands Xi ∈ S. A desirable property for a tensor category is a duality that makes it into a rigid tensor category. An important recent result [Jannsen (1992)] shows that the category Meﬀ (K) is a semisimple abelian category. see [Andr´ (2004)] for details. (Y. for any X.Y. p). where X is a smooth projective variety over K and p ∈ C0 (X. is called Klinear. The set of morphisms here are Qvector spaces instead of abelian groups (the category is Qlinear). since p ∈ C0 (X.2. p2 = p.
with the identiﬁcation realized by a natural isomorphism. This means that one writes the motive of P1 as 1⊕L. One deﬁnes the Tate motive Q(1) to be the formal inverse of the Lefschetz motive L. then the transpose Z † is in Corrdim(X) (X×Y )⊗Z Q. the Tate motives Q(n). . which corresponds to the cohomology H 0 (P1 ). with 1 = (Spec(K). A Klinear tensor category C is rigid if there is a duality ∨ : C → C op such that.Y Deﬁnition 2. −1). id. 0) being the motive of a point. As an object in the category Mnum (K) (see Deﬁnition 2. (1 ⊗ ) ◦ (δ ⊗ 1) = 1X ∨ . Similarly. to have a num num duality one needs to include among morphisms also cycles of other codimensions. Y ) = num Corrdim(X) (Y ×X)⊗Z Q. num num but the latter is not C0 (Y. 1) = K. G(Y )). and the corresponding cohomology ring H ∗ (Pn . there are an evaluation : X ⊗ X ∨ → 1 (the unit object for tensor product) and a unit δ : 1 → X ∨ ⊗ X satisfying ( ⊗ 1) ◦ (1 ⊗ δ) = 1X . X) = Corrdim(Y ) (X ×Y )⊗Z Q. Namely. has the eﬀect of shifting the codimensions of cycles in the deﬁnitions of morphisms in the category of motives in the following way.3. which means that cY. with Q(n) = Q(1)⊗n and Q(0) = 1. One also assumes that HomC (1. although this can be replaced by a signed version. Introducing the formal inverse Q(1) = L−1 = (Spec(K).2. The category of numerical eﬀective pure motives does not have a duality. One can see that. Z) = Z[u]/un+1 with the generator u in degree 2. the Lefschetz motive is given by the triple L = (Spec(K). to have such a duality.X = c−1 . id. Y ) = HomC (X. id. one writes the motive of Pn as 1 ⊕ L ⊕ L2 ⊕ · · · ⊕ Ln . for all X ∈ Obj(C) the functor · ⊗ X ∨ is left adjoint to · ⊗ X and X ∨ ⊗ · is right adjoint to X ⊗ ·. This can be achieved by introducing new objects.4 below). one would need to take the transpose of a correspondence Z ⊂ X × Y . 1). if Z ∈ C0 (X.2. Recall here that given two functors F : C → C and G : C → C satisfying HomC (F (X). Thus. which corresponds to the decomposition of Pn into cells A0 ∪ A1 ∪ · · · ∪ An . but this would have the eﬀect of changing the codimension. Moreover. one says that F is leftadjoint to G and G is rightadjoint to F . X. for all X ∈ Obj(C) and Y ∈ Obj(C ). This is the motive whose cohomological realization is H 2 (P1 ). where 1 is the motive of a point.Motives and periods 33 also generally assumes that commutativity holds.
12) (2. one considers a diﬀerent set of properties that a universal cohomology theory should satisfy in this case. p. 2. where the varieties involved may be noncompact or singular.3 Mixed motives and triangulated categories Having just quickly mentioned how one constructs a good category of pure motives. p† . where dim(X)+n−m Cm−n (X.9 below. Y ) = Corrnum (Y × X) ⊗Z Q (2. p. instead of working with Weil cohomologies. see §2. m). m) with (X. we jump ahead and leave the world of smooth projective varieties to venture into the more complicated and more mysterious realm of mixed motives. The diﬃcult nature of mixed motives therefore lies in the nontrivial extensions of pure motives. p. and the duality is of the form (X. n)) = qCn−m (X. From the cohomological point of view. m) ⊗ (Y. (Y. The category Mnum (K) has objects given by triples (X. m)∨ = (X. p) ∈ Obj(Meﬀ (K)) an eﬀective numerical motive and num m ∈ Z. q.15) (2. The morphisms are given by HomMnum (K) ((X. q.14) One obtains in this way a rigid tensor category. the resulting category is in fact more than a rigid tensor category: it has enough structure to be identiﬁed uniquely as the category of ﬁnite dimensional linear representations of a group (more precisely of an aﬃne group scheme) via the Tannakian formalism. Y )p.13) The tensor product is then given by (X.9 below.34 Feynman Motives Deﬁnition 2.16) . One should think of mixed motives as objects that come endowed with ﬁltrations whose graded pieces are pure motives. As we discuss in §2.4.2. (2. p ⊗ q. n) = (X × Y. (2. n + m). p. dim(X) − m). The main properties that one expects such a cohomology theory to satisfy become of the following form: • Homotopy invariance: the projection π : X × A1 → X induces an isomorphism π ∗ : H ∗ (X) → H ∗ (X × A1 ).
• A triangle isomorphic to a distinguished triangle is distinguished (where isomorphisms of triangles are isomorphisms of the objects such that the resulting diagram commutes). with T (f ) : T (X) → T (X ) completing the diagram. there exists a morphism Z → Z such that all squares commute. • Rotations Y → Z → T (X) → T (Y ) and T −1 (Z) → X → Y → Z of a distinguished triangle X → Y → Z → T (X) are distinguished. u This replaces the notion of a Weil cohomology with “mixed Weil cohomologies”. in the form of a triangulated category of mixed motives DM(K). although after a long series of developments and deep results in the ﬁeld (see [Voevodsky (2000)]. • Given distinguished triangles X → Y → Z → T (X) and X → Y → Z → T (X ) and morphisms f : X → X and h : Y → Y such that the square commutes. Ideally. [Hanamura (1995)]) one has at least a weaker version available. in particular in the form of long exact cohomology sequences.3. Deﬁnition 2. these three triangles can h◦f f h f id (2.18) .Motives and periods 35 • Mayer–Vietoris sequence: if U. Triangulated categories still allow for several techniques from homological algebra. This is (at present) not available. one would like to have an abelian category MM(K) of mixed motives that has realization functors to all the mixed Weil cohomologies. Y ). • Given X → Y → C(f ) → T (X) and Y → Z → C(h) → T (Y ). and the composite X → Z → C(h ◦ f ) → T (X). V are an open covering of X then there is a long exact cohomology sequence · · · → H i−1 (U ∩ V ) → H i (X) → H i (U ) ⊕ H i (V ) → H i (U ∩ V ) → · · · (2. such that X → Y → C(f ) → T (X) is a distinguished triangle.1. • The K¨nneth formula still holds. • For all f ∈ HomT (X. A triangulated category T is an additive category endowed with an automorphism T : T → T (the shift functor) and a family of distinguished triangles X → Y → Z → T (X) with the following properties: • X → X → 0 → T (X) is a distinguished triangle for all X ∈ Obj(T ).17) • Poincar´ duality is replaced by the duality between cohomology H ∗ e ∗ and cohomology with compact support. [Levine (1998)]. there exists an object C(f ) ∈ Obj(T ) (the mapping cone) with a morphism C(f ) → T (X). Hc .
36
Feynman Motives
be placed on the vertices and edges of an octahedron with morphisms completing the picture so that all the faces are commutative diagrams. One typically writes X[1] for T (X) and X[n] = T n (X) for n ∈ Z. One more property of triangulated categories that follows from these axioms, and which we will need explicitly later, is the following. Lemma 2.3.2. Given a triangulated category T , a full subcategory T is a triangulated subcategory if and only if it is invariant under the shift T of T and for any distinguished triangle X → Y → Z → X[1] for T where X and Y are in T there is an isomorphism Z Z where Z is also an object in T . A full triangulated subcategory T ⊂ T is thick if it is closed under direct sums. It is sometimes possible to extract an abelian category from a triangulated category using the method of tstructures. Deﬁnition 2.3.3. A tstructure on a triangulated category T consists of two full subcategories T ≤0 and T ≥0 with the following property. Upon denoting by T ≥n = T ≥0 [−n] and T ≤n = T ≤0 [−n], one has T ≤−1 ⊂ T ≤0 , with the property that HomT (X, Y ) = 0, ∀X ∈ Obj(T ≤0 ), ∀Y ∈ Obj(T ≥1 ), and T ≥1 ⊂ T ≥0
and such that, for all Y ∈ Obj(T ), there exists a distinguished triangle X → Y → Z → X[1] with X ∈ Obj(T ≤0 ) and Z ∈ Obj(T ≥1 ). When one has a tstructure on a triangulated category, one can consider the full subcategory T 0 = T ≤0 ∩ T ≥0 . This is called the heart of the tstructure and it is known to be an abelian category, see [Be˘ ılinson, Bernstein, Deligne (1982)], [Kashiwara and Schapira (1990)]. Theoretical physicists have recently become familiar with the formalism of tstructures in triangulated categories in the context of homological mirror symmetry, see for instance [Bridgeland (2009)].
Motives and periods
37
2.4
Motivic sheaves
We also mention brieﬂy, for later use, a recent construction due to [Arapura (2008)] of a category of motivic sheaves over a base scheme S, modeled on Nori’s approach to the construction of categories of mixed motives. We will return in Chapter 7 to discuss how this ﬁts with the motives associated to Feynman diagrams. The category of motivic sheaves constructed in [Arapura (2008)] is based on Nori’s construction of categories of motives via representations of graphs made up of objects and morphisms (cf. [Brugui`res (2004)]). e According to [Arapura (2008)], one constructs a category of motivic sheaves over a scheme S, by taking as vertices of the corresponding graph objects of the form (f : X → S, Y, i, w) , (2.19)
where f : X → S is a quasiprojective morphism, Y ⊂ X is a closed subvariety, i ∈ N, and w ∈ Z. One thinks of such an object as determining a motivic version hi (X, Y )(w) of the local system given by the (Tate twisted) S i ﬁberwise cohomology of the pair HS (X, Y ; Q) = Ri f∗ j! QX Y , where j = jX Y : X Y → X is the open inclusion, i.e. the sheaf deﬁned by U → H i (f −1 (U ), f −1 (U ) ∩ Y ; Q). The edges are given by the geometric morphisms: • morphisms of varieties over S, (f1 : X1 → S, Y1 , i, w) → (f2 : X2 → S, Y2 = F (Y ), i, w), with f2 ◦ F = f1 ; • the connecting morphisms (f : X → S, Y, i + 1, w) → (f Y : Y → S, Z, i, w), • the twisted projection morphisms (f : X × P1 → S, Y × P1 ∪ X × {0}, i + 2, w + 1) → (f : X → S, Y, i, w). (2.22) The product in this category of motivic sheaves is given by the ﬁber product (X → S, Y, i, w) × (X → S, Y , i , w ) = (X ×S X → S, Y ×S X ∪ X ×S Y , i + i , w + w ). for Z ⊂ Y ⊂ X; (2.21) (2.20)
(2.23)
38
Feynman Motives
2.5
The Grothendieck ring of motives
An invariant of algebraic varieties and motives that contains a lot of information and can be used to evaluate the motivic complexity of a given algebraic variety is the class in the Grothendieck ring. This is constructed by breaking up the variety into pieces, by simple cutandpaste operations of the scissors congruence type. The operation of assigning to a variety its class in the Grothendieck ring behaves like a universal Euler characteristic. Deﬁnition 2.5.1. Let VK denote the category of algebraic varieties over a number ﬁeld K. The Grothendieck ring K0 (VK ) is the abelian group generated by isomorphism classes [X] of varieties, with the relation [X] = [Y ] + [X Y ], (2.24) for Y ⊂ X closed. It is made into a ring by the product [X × Y ] = [X][Y ]. Let M(K) denote, as above, the abelian category of (numerical) pure motives over K. As above we write the objects of M(K) in the form (X, p, m), with X a smooth projective variety over K, p = p2 ∈ End(X) a projector, and m ∈ Z the Tate twist. Let K0 (M(K)) denote the Grothendieck ring of the abelian category M(K). We’ll see that the class of a variety, or of a motive, in the corresponding Grothendieck ring, can be thought of as a “universal Euler characteristic”. To this purpose, we ﬁrst recall the notion of additive invariants of varieties. Deﬁnition 2.5.2. An additive invariant is a map χ : VK → R, with values in a commutative ring R, satisfying the following properties. ∼ • Isomorphism invariance: χ(X) = χ(Y ) if X = Y are isomorphic. • Inclusionexclusion: χ(X) = χ(Y ) + χ(X Y ), for a closed embedding Y ⊂ X. • Multiplicative property: χ(X × Y ) = χ(X)χ(Y ). The topological Euler characteristic clearly satisﬁes all of these properties and is in fact the prototypical example of such an invariant. Counting points over ﬁnite ﬁelds is another invariant that behaves in the same way and factors through the Grothendieck ring. By the deﬁnition of the relations on the Grothendieck ring one sees that assigning an additive invariant with values in R is equivalent to assigning a ring homomorphism χ : K0 (VK ) → R.
Motives and periods
39
The results of [Gillet and Soul´ (1996)] show that there exists an additive e invariant χmot : VK → K0 (M(K)). This assigns to a smooth projective variety X the class χmot (X) = [(X, id, 0)] ∈ K0 (M(K)). The value assigned to a more general variety is more complicated to describe and it is given in [Gillet and Soul´ (1996)] in terms of an object W (X) in the category e of complexes over M(K). (Technically, here the category M(K) is taken with the rational equivalence relation on algebraic cycles.) It is sometime referred to as “motivic Euler characteristic”. If L denotes the class L = [A1 ] ∈ K0 (VK ) then its image in K0 (M(K)) is the Lefschetz motive L = Q(−1) = [(Spec(K), id, −1)]. Since the Lefschetz motive is invertible in K0 (M(K)), its inverse being the Tate motive Q(1), the ring homomorphism χmot : VK → K0 (M(K)) induces a ring homomorphism (2.25) χmot : K0 (VK )[L−1 ] → K0 (M(K)). This in particular means that one can work either in the Grothendieck ring of varieties or of motives, being able to map the ﬁrst to the latter through χmot . In the following, we will usually refer to the Grothendieck ring K0 (VK ). A particularly interesting subring is the one generated by the Lefschetz motive L = [A1 ]. This is a polynomial ring Z[L] ⊂ K0 (VK ), or a ring of Laurent polynomials Z[L, L−1 ] ⊂ K0 (M(K)), which corresponds to the classes of varieties that motivically are mixed Tate motives. See the comment at the end of §2.6 below on the relation between the information given by knowing that the class in the Grothendieck ring lies in the subring Z[L] ⊂ K0 (VK ) and the question of whether the motive, as an object of the triangulated category DM(K) is a mixed Tate motive.
2.6
Tate motives
Let DM(K) be the Voevodsky triangulated category of mixed motives over a ﬁeld K [Voevodsky (2000)]. The triangulated category DMT (K) of mixed Tate motives is the full triangulated thick subcategory of DM(K) generated by the Tate objects Q(n). It is known that, over a number ﬁeld K, there is a canonical tstructure on DMT (K) and one can therefore construct an abelian category MT (K) of mixed Tate motives (see [Levine (1998)]). For the kind of applications we are going to see in the rest of this book, we will be especially interested in the question of when a given mixed motive is mixed Tate. In suﬃciently simple cases, one can try to answer this question using the properties of DMT (K) as a triangulated subcategory of
1. which replaces Poincar´ duality that holds e for Weil cohomologies in the pure motives setting. a closed embedding Y ⊂ X determines a distinguished triangle in DM(K) m(Y ) → m(X) → m(X Y ) → m(Y )[1].8 of [Voevodsky (2000)]) Homotopy invariance implies the identity of motives m(X × A1 ) = m(X)(1)[2]. one does not need the full strength of the general construction. denoted by C∗ (X) in [Voevodsky (2000)]. Although the available explicit constructions of DM(K) given in [Voevodsky (2000)]. one can try to ﬁnd a geometric decomposition. if two of the terms are known to be in DMT (K) then the third one also is. We reported above the distinguished triangles and homotopy invariance properties for the motives with compact support in Voevodsky’s terminology. in such an approach one wants to make repeated use of two properties of the triangulated category DM(K): • (Proposition 4.4 of [Voevodsky (2000)]) Over a ﬁeld of characteristic zero.26) where the notation m(X) here stands for the motive with compact c support. of the variety. since DMT (K) is a triangulated subcategory of DM(K). The distinguished triangle (2. often in the form of a stratiﬁcation. This can often be used as a way to reduce a motive to simpler building blocks for which it may be easier to answer the question of whether they are mixed Tate or not. and then derive information about the original motive via the distinguished triangles and homotopy invariance property. and (n) is the Tate twist by Q(n).1.26) implies that. but one can directly use the formal properties that the triangulated category DM(K) satisﬁes.27) . in order to understand whether the motive of a given variety is mixed Tate. • (Corollary 4. often when dealing with very speciﬁc concrete examples. (2.40 Feynman Motives the triangulated category DM(K) of mixed motives. In particular. (2. where [n] is the shift. More precisely. [Levine (1998)]. are technically very complicated and certainly beyond the scope of this book. [Hanamura (1995)]. One can obtain analogous statements for motives without compact support using the duality in DM(K) between mixed motives and mixed motives with compact support. in such a way that one has enough control on the strata to build the motive out of pieces that are mixed Tate.
namely the case of determinant hypersurfaces. which can be done in the following way. – Note that W2 is equipped with a trivial vector bundle E2 = E × W2 .6. Clearly. which will be relevant to Feynman integrals (see §3. Theorem 2.13 below). by the ﬁrst property above on the distinguished triangle of a closed embedding. if the result holds for the hypersurface complement it also holds for the hypersurface itself.2. – Let W2 ⊆ E1 be the complement E1 L1 . as that will be the natural geometric object to consider in the case of the Feynman integrals. Fix an dimensional vector space E. in order to describe 2 ˆ the locus A D we need to parameterize matrices of rank exactly equal to . – Denote by W1 the variety E {0}. and two line subbundles of E2 : the pullback of L1 (still denoted L1 ) and the linebundle L2 whose ﬁber over v2 ∈ W2 consists of the line spanned by v2 . 2 Let D ⊂ P −1 be the corresponding projective hypersurface deﬁned by the homogeneous polynomial det(x). Let D ⊂ A ishing of the determinant 2 be the hypersurface deﬁned by the van(2. The determinant hypersurface complement A ﬁnes an object in the category DMT of mixed Tate motives. Since D consists of × matrices of rank < . We have the following result on the motivic nature of the determinant hypersurfaces. – Note that W1 is equipped with a trivial vector bundle E1 = E × W1 . and with a line bundle S1 := L1 ⊆ E1 whose ﬁber over v1 ∈ W1 consists of the line spanned by v1 .6.Motives and periods 41 We give here an example from [Aluﬃ and Marcolli (2009a)]. and the fact that aﬃne and projective spaces are Tate motives. ˆ Deﬁnition 2. . It is a well known fact that. motivically. – By construction. these varieties are mixed Tate (see [Belkale and Brosnan (2003a)]).28) 2 ˆ D = {x = (xij ) ∈ A  det(x) = 0}.1. L1 and L2 span a rank2 subbundle S2 of E2 . We formulate it in terms of hypersurface complements. 2 ˆ D de ˆ Proof. but we illustrate here explicitly how one can argue that this is the case using only the two properties listed above of the triangulated category of mixed Tate motives.
At the ﬁrst step. this matrix has rank exactly k. namely the existence of distinguished triangles (2. At the next step one considers the complement of the line bundle S1 inside the trivial vector bundle E1 over W1 . Lk spanning a rankk subbundle Sk of the trivial vector bundle Ek := E × Wk . . The sequence stops at the th step. vk resp. . To see this. . . Lk to be the pullbacks of the similarly named line bundles on Wk . which gives a mixed Tate motive by the ﬁrst observation above on distinguished triangles associated to closed embeddings. . and deﬁne the line subbundles L1 . .42 Feynman Motives – Let W3 ⊆ E2 be the complement E2 S2 . . . both m(S1 ) and m(E1 ) are mixed Tate motives. and so on. Let Ek+1 = E × Wk+1 . . . . (and the remaining rows are 0). 2 Each variety Wk maps to A as follows: a point of Wk determines k vectors v1 . by construction. in the construction described above. so that E S = ∅. . for all 1 ≤ k ≤ . Lk+1 span a rank k + 1 subbundle Sk+1 of Ek+1 . . and can be mapped to the matrix whose ﬁrst k rows are v1 . . . . . .26) associated to closed embeddings Y ⊂ X and the homotopy invariance (2. vr )) → c1 v1 + · · · + cr vr . We then apply the two properties of the triangulated category of mixed motives described above. . .27) inductively to the loci constructed above. . . If Sk Ek . since we know that the bundles Sk are trivial. cr ). . . The line bundles L1 . since both are products by aﬃne spaces. We ﬁrst show that the variety W constructed as above is isomorphic 2 ˆ to the locus we want. the bundle Sk over Wk is in fact trivial. . all of which are varieties deﬁned over Q. namely A D. Again. . Therefore m(E1 S1 ) is also mixed Tate. single points obviously belong to the category of mixed Tate motives and we are taking the complement W1 of a point in an aﬃne space. . any such rank k matrix is the image of a point of Wk . An explicit trivialization is given by the map α Kk × Wk → Sk deﬁned by α : ((c1 . Conversely. The same argument shows that. (v1 . endowed with k line bundles L1 . deﬁne Wk+1 := Ek Sk . By construction. . . vk . equal to the rank of E . where S has rank . . the procedure described above produces a variety Wk . the motive m(Ek Sk ) is mixed Tate. by repeatedly using the triviality of Sk and the two properties of DMTQ recalled above. where points of Wk are parameterized by ktuples of vectors v1 . . . and the construction can continue. Let Lk+1 be the line bundle whose ﬁber over vk+1 is the line spanned by vk+1 . Also notice that. vk spanning Sk ⊆ K × Wk = Ek . . ﬁrst observe that at the kth step.
knowing that the class in the Grothendieck ring is a function of the Lefschetz motive L is a very good indication of the possible mixed Tate nature of the motive. the Tate conjecture predicts that knowing the e number Np of points of the reduction mod p for almost all p would suﬃce to determine the motive.2. The formula (2.6.3. we already know that the class [A D] will be in the mixed Tate part Z[L] of the Grothendieck ring of varieties K0 (VQ ). One ﬁnds the following expression [Aluﬃ and Marcolli (2009a)]. One can use the same geometric construction described above to ˆ compute explicitly the class of D in the Grothendieck ring of varieties.29) where L is the class of A1 .2.29) using the fact that D is the aﬃne cone over D .Motives and periods 43 ˆ The result for the projective hypersurface D is then the same. 2 ˆ Notice that. Using the same argument as in Theorem 2. In fact. This suﬃces to obtain the expression (2. (2. as in principle there may be cancellations that happen in the Grothendieck ring that do not take place at the level of the motives themselves. In the projective case. and assuming the conjecture holds. The class in the Grothendieck ring of varieties of the aﬃne determinant hypersurface is [A 2 ˆ D ] = L(2) i=1 (Li − 1) (2.6. In general. as observed for instance in [Andr´ (2009)]. since D is the aﬃne cone over D and the same two properties of the triangulated ˆ category of mixed motives used above show that if D is mixed Tate so is D . one shows inductively that the class of Wk is given by [Wk ] = (L − 1)(L − L)(L − L2 ) · · · (L − Lk−1 ) k = L(2) (L − 1)(L (2. However. computing the class in the Grothendieck ring may be easier than knowing explicitly the motive in DMQ . the class is [P 2 −1 D ] = L(2) i=2 (Li − 1). this does not imply that the motive itself will be in the subcategory DMTQ of mixed Tate motives. knowing that this .30) is then ˆ obtained from (2. by Theorem 2.6. If one only knows that the class belongs to the Tate part Z[L] of the Grothendieck ring.30) Proof. For mixed Tate motives the numbers Np are polynomial in p. Theorem 2.31) −1 − 1) · · · (L −k+1 − 1).29).
Q) = ∗ Hsing (X. The role in this of the class in the Grothendieck ring comes from the fact that the counting of points Np over ﬁnite ﬁelds has the properties of an Euler characteristic (as we saw above. directly at the level of objects in DMK . 2. (2.6. Thus.7 The algebra of periods Betti and de Rham cohomology are two of the possible Weil cohomologies for smooth projective varieties. while being transcendental. Periods are typically transcendental numbers. Q) with the dual of singular homology Hi (X. Even without these general considerations and the role of the Tate conjecture. arguably.2. an additive invariant). In fact. the same type of geometric information on stratiﬁcations that one can use to compute the class in the Grothendieck ring K0 (VK ) may already contain enough information to run an argument similar to the one used in Theorem 2. knowing that the class in the Grothendieck ring is mixed Tate would suﬃce. as we have seen in the example of the determinant hypersurfaces. ∗ where one identiﬁes Betti cohomology HB (X) = Hom(H∗ (X). The isomorphism is induced (over aﬃne pieces) by the pairing of diﬀerential forms and cycles ω⊗γ → γ ω. (2. They are related by the period isomorphism ∗ HdR (X) ⊗K C ∗ HB (X) ⊗Q C. in the best cases. Z) = ∗ ∗ Ker(∂i )/Im(∂i+1 ).32) for varieties deﬁned over a number ﬁeld K.44 Feynman Motives is the case would conversely imply that the motive is mixed Tate.33) for ω ∈ Ker(d : Ωi (X) → Ωi+1 (X)) and γ ∈ Ker(∂i : Ci (X) → Ci−1 (X)). When one is given a basis of HdR (X) and of HB (X). From the numbertheoretic point of view they are. the period isomorphism is speciﬁed by a period matrix. the most interesting class of numbers beyond the algebraic ones. they are obtained from a procedure that uses only algebraic data: an algebraic diﬀerential form on an algebraic variety. assuming the conjecture. paired with a cycle deﬁned . hence it factors through the Grothendieck ring.
36) (2. ω. f ∗ (ω). An interesting problem is understanding the structure of this algebra: what relations (over Q) exist between periods. The equivalence in (2. f∗ (γ))]. ω. D(1) .34) consisting of a smooth (aﬃne) variety X deﬁned over Q. Q). dω. ω.Motives and periods 45 by algebraic relations. . 2. γ)] → γ ω (2. equivalence classes of elements of the form [(X. ∂γ)]. Let us denote by PC the Qsubalgebra of C generated by periods. D.8 Mixed Tate motives and the logarithmic extensions An interesting class of mixed Tate motives is associated to logarithm and polylogarithm functions [Be˘ ılinson and Deligne (1994)]. (2. that is. γ)].35) ˜ ˜ where D is the normalization of D and D(1) is the normalization of ∪i. γ)] = [(D. [Bloch (1997)]. D. D. D. an algebraic diﬀerential form ω ∈ Ωdim(X) (X) (not necessarily closed). see also [Ayoub (2007)]. D) → (X . a normal crossings divisor D in X. γ) are Qlinear in ω and γ. γ)] = [(X . and a class in the relative homology γ ∈ Hdim(X) (X(C).34) is given by the following relations • Linearity: the expressions (X. ω. Kontsevich conjectured that the map P → PC ⊂ C given by [(X. D(C).37) is injective. D. Following [Kontsevich and Zagier (2001)]. (2.j (Di ∩ Dj ) where Di are the components of the divisor D. • Stokes’ formula: ˜ ˜ [(X. one can deﬁne a Qalgebra P as the algebra generated by “formal periods”. D . • Change of variable formula: given a morphism f : (X. D(C). D ) and ω ∈ Ωdim(X ) (X ) and γ ∈ Hdim(X) (X(C). ω. Q) one has [(X. This means that all possible nontrivial relations between periods come from either a change of variable formula or Stokes’ theorem.
This is the Kummer extension QGm (1) → K → QGm (0) → QGm (1)[1].39) is also referred to as the logarithmic motive.38) When. instead of working with motives over the base ﬁeld K. (2. Q(1)) of Tate motives are given by the u Kummer motives M = [Z → Gm ] with u(1) = q ∈ K∗ . and denoted by K = Log.39) where over the point s ∈ Gm one is taking the Kummer extension Ms = u [Z → Gm ] with u(1) = s. (2. one can then consider the logarithmic motives Logn . (2.43) .46 Feynman Motives The extensions Ext1 DM(K) (Q(0). When working with Qcoeﬃcients. deﬁned by Logn = Symn (K). The mixed motive Ln has period matrix 1 0 (n) MLi MLogn−1 (1) (2.42) where M (1) = M ⊗ Q(1) and L1 = Log. one works with the relative setting of motivic sheaves over a base scheme S. there is a natural way to assemble the Kummer motives into a unique extension in Ext1 DM(Gm ) (QGm (0). the Kummer extension (2. cf.41) The polylogarithms appear naturally as period matrices for extensions involving the symmetric powers Logn = Symn (K). σ∈Σn (2. so that one can include denominators in the deﬁnition of projectors. in the form [Bloch (1997)] 0 → Logn−1 (1) → Ln → Q(0) → 0. This extension has period matrix of the form 1 0 log q 2πi . [Ayoub (2007)] [Be˘ ılinson and Deligne (1994)]. In the case where S = Gm . instead of the Tate motives Q(n) one considers the Tate sheaves QS (n). QGm (1)). Let DM(Gm ) be the Voevodsky category of mixed motives (motivic sheaves) over the multiplicative group Gm .40) where the symmetric powers of an object in DMQ (Gm ) are deﬁned as Symn (X) = 1 #Σn σ(X n ). Because of the logarithm function log(s) that appears in the period matrix for this extension. These correspond to the constant sheaf with the motive Q(n) over each point s ∈ S. (2.
. . . to the distinguished triangles in DM(Gm ) of the form Logn−1 (1) → Logn → Q(0) → Logn−1 (1)[1]. .47) ← − n The analog of the period matrix (2.44) du . .46). . . (2. . .e. . u equivalently deﬁned (on the principal branch) using the power series sk Lin (s) = . . ··· . i. . −Lin (s))τ . in the particular case m = 1. . ··· . ··· logn (s) logn−1 (s) logn−2 (s) 2πi n! (2πi)2 (n−1)! (2πi)3 (n−2)! · · · (2πi)n · · · .45) The period matrices for the motives Logn correspond to the description of Logn as the extension of Q(0) by Logn−1 (1).Motives and periods 47 with MLi = (−Li1 (s). . Lemma 4. . . and the map Symm (K) → Q(0) of (2. .48) . −Li2 (s). . and Lin (s) = Lin−1 (u) 0 k and with 2πi 0 2πi log(s) (2πi)2 log2 (s) (2πi)2 log(s) = 2πi 2! . . as in [Ayoub (2007)]. 2πi logn!(s) (2πi)2 log (s) (2πi)3 log (s) · · · (2πi)n (n−1)! (n−2)! . · · · .46) n The motives Log form a projective system under the canonical maps βn : Logn+1 → Logn given by the composition of the morphisms Symn+m (K) → Symn (K) ⊗ Symm (K). (2. . .45) then becomes the inﬁnite matrix 2πi 0 0 ··· 0 ··· 2πi log(s) (2πi)2 0 ··· 0 ··· 2 3 2πi log2!(s) (2πi)2 log(s) (2πi) ··· 0 ··· . . . . n−2 ··· ··· ··· ··· 0 0 0 . . .35. Let Log∞ denote the promotive obtained as the projective limit Log∞ = lim Logn . where τ means transpose and where s (n) (2. . kn Li1 (s) = − log(1 − s). . n n−1 MLogn (1) 0 0 (2πi)3 . MLog∞ (1) = . given by the fact that Symn+m (K) is canonically a direct factor of Symn (K) ⊗ Symm (K). . (2. ··· (2.
F • ) of a ﬁnitedimensional Qvector space. n V p.q .C ) ⊂ F p V2. These comprise the analytic side of the theory of motives and they provide one of the possible realizations of categories of mixed motives. The Hodge structure (V. satisfying W grn VC = F p grW VC ⊕ F n−p+1 grW VC . A Qmixed Hodge structure consists of the data M = (V. For Hodge–Tate structures one has an associated period matrix obtained −1 as in §2 of [Goncharov (1999)].1.48 Feynman Motives This can best be expressed in terms of mixed Hodge structures. F p (V1 ⊗Q V2 ) = r+s=p A Qmixed Hodge structure M = (V. by considering V1 ⊗Q V2 with ﬁltrations Wm (V1 ⊗Q V2 ) = i+j=m Wi (V1 ) ⊗Q Wj (V2 ). W• .C .49) is equivalent to a Hodge decomposition grW VC = ⊕p+q=n V p. F • ) is mixed Tate if grW M = ⊕Q(−p). expressed in a 2p chosen basis of grW VQ . W• .C ). as the composition SHT ◦ SW of an isomorW phism SHT : ⊕p gr2p (VQ ) ⊗Q C → VC and a splitting of the weight ﬁltration. 2p . These deﬁne.8. 2p and grW M = 0. These correspond to the pure Tate motives. p+q p+q (2.q = F p grW VC ∩ F q grW VC . Tensoring of QMHS is also deﬁned. 2p−1 The condition (2. W• .49) ¯ where F • is the complex conjugate of F • with respect to VR ⊂ VC . which is an isomorphism SW : ⊕p grW (VQ ) ⊗Q C → VC . One has the following deﬁnition as in [Steenbrink (1994)].50) (2. hence another way to test complexity of various types of mixed motives.C ) ⊗C F s (V2. F • ) is pure of weight n if Wn = V and Wn−1 = 0. Deﬁnition 2. A morphism M1 → M2 of QMHS is a linear map φ : V1 → V2 such that φ(Wk V1 ) ⊂ Wk V2 and φ(F p V1. Tate twists on Qmixed Hodge structures. n n (2. F r (V1.51) An example of pure Hodge structure of weight −2n is the Tate Hodge structure with V = (2πi)n Q. by tensoring. with an increasing ﬁltration W• V (weight ﬁltration) and a decreasing ﬁltration F • VC for VC = V ⊗Q C (the Hodge ﬁltration).
with target the category of ﬁnitedimensional vector spaces over a ﬁeld extension K of k. with ﬁber functor ω : C → Vectk . Y ) → HomC (ω(X). for all X.2.45) on vectors in Qn with the ﬁrst k − 1 entries equal to zero. see [Saavedra Rivano (1972)]. [Saavedra Rivano (1972)] Let C be a neutral Tannakian category. the corresponding sequence 0 → ω(X) → ω(Y ) → ω(Z) → 0 in C is also exact.Y : ω(X ⊗ Y ) → ω(X) ⊗ ω(Y ). is the following statement that reconstructs a group from the category. A Tannakian category C is neutral if K = k. 2. ω : C → VectK . of which we are going to see an application to motives in the next section and one to quantum ﬁeld theory in a later chapter. The main result of the theory of Tannakian categories. a functor ω is additive if the map of Hom sets above is a klinear map.Motives and periods 49 In the example above.9 Categories and Galois groups We review here a type of structure on a category which is more rigid than that of an abelian category we encountered before.e.9. the mixed Hodge structure associated to the motives Logn is the one that has as the weight ﬁltrations W−2k the range of multiplication by the matrix MLogn deﬁned as in (2. Recall here that a functor ω : C → C is faithful if. namely the notion of Tannakian category. Y ∈ Obj(C). the mapping ω : HomC (X. for any exact sequence 0 → X → Y → Z → 0 in C. An additive functor ω is exact if. [Deligne and Milne (1982)]. i. ω(Y )) is injective. an exact faithful tensor functor. . Let G be the group of automorphisms (invertible natural transformations) of the ﬁber functor. If C and C are klinear categories. A functor ω : C → C between klinear tensor categories is a tensor functor if there are functorial isomorphisms τ1 : ω(1) → 1 and τX. Deﬁnition 2. A klinear rigid abelian tensor category C is a Tannakian category if it admits a ﬁber functor. where RepG is the category of ﬁnite dimensional linear representations of G.9. and which is suﬃcient to identify it uniquely as a category of ﬁnite dimensional linear representations of a group. while the Hodge ﬁltration F −k is given by the range of multiplication of MLogn on vectors of Cn with the entries from k + 1 to n equal to zero [Bloch (1997)]. Theorem 2. Then ω induces an equivalence of rigid tensor categories ω : C → RepG .1.
Q(n)) = K2n−1 (K) ⊗ Q (2. For example. Appendix B. here G is really an aﬃne group scheme in the sense that we are going to recall in §5. it is possible to form an abelian category MT (K) of mixed Tate motives. The aﬃne group scheme G is referred to as the Tannakian Galois group of C. This means that objects in this category are mixed Tate motives over K that are unramiﬁed at each ﬁnite place v ∈ V . So is the one of the Tannakian subcategory generated by the pure Tate motives Q(n). for V a set of ﬁnite places of K. see [Levine (1993)] and [Levine (1998)].2 below. M → ω(M ) = ⊕n ωn (M ) with ωn (M ) = Hom(Q(n). with ﬁber functor ω to Zgraded Qvector spaces. In fact. then it is known that the Tannakian Galois group is reductive.53) where Grw (M ) = W−2n (M )/W−2(n+1) (M ) is the graded structure asso−2n ciated to the ﬁnite increasing weight ﬁltration W . Q(n)) = 0. The resulting abelian category MT (K) of mixed Tate motives over a number ﬁeld is in fact also a Tannakian category. −2n (2. / In this case one also has a Ktheory group parameterizing extensions in the . as the heart of a tstructure inside the triangulated subcategory DMT (K) of the triangulated category DM(K) of mixed motives. If a Tannakian category C is semisimple. the mixed Tate motives that correspond to extensions in Ext1 DM(K) (Q(0). over a number ﬁeld one has an explicit description of the extensions in DM(K) between pure Tate motives in terms of algebraic Ktheory as Ext1 DM(K) (Q(0).10 Motivic Galois groups We recalled above that.52) and Ext2 DM(K) (Q(0). The construction of a Tannakian category of mixed Tate motives MT (OV ) and the explicit form of its motivic Galois group are also known for motivic sheaves over the scheme OV of V integers of a number ﬁeld K. Q(1)) are the Kummer motives we have encountered in the previous sections. over a number ﬁeld K. the latter is the simplest example of a reductive group.50 Feynman Motives In fact. 2. In fact. the Tannakian Galois group of the category of pure motives is reductive. Grw (M )). the multiplicative group Gm . For example.
Q(n)). since it provides a motivic Galois group (noncanonically) isomorphic to the Tannakian Galois group of renormalization of [Connes and Marcolli (2004)].10.Motives and periods 51 form K2n−1 (K) ⊗ Q n ≥ 2 1 ∗ ExtDM(OV ) (Q(0). Proposition 2. the corresponding Lie algebra Lie(U) is freely generated by a set of homogeneous generators in degree n identiﬁed with a basis of the dual of Ext1 (Q(0).54) and Ext2 DM(OV ) (Q(0). The aﬃne group scheme that gives the motivic Galois group in this case is of the form U Gm . Q(n))∨ . The following special case of this general result of [Deligne and Goncharov (2005)] is especially relevant to the context of renormalization. There is however no canonical identiﬁcation between Lie(U) and the free Lie algebra generated by the graded vector space ⊕Ext1 (Q(0). Q(n)) = 0. In fact. and U is a prounipotent aﬃne group scheme. Q(n)) = OV ⊗ Q n=1 0 n ≤ 0.3 of [Deligne and Goncharov (2005)]. (2. where the Lie algebra Lie(U) is (noncanonically) isomorphic to the free graded Lie algebra with one generator en in each degree n ≤ −1. which we review in Chapter 6. . the motivic Galois group of the Tannakian category MT (O[1/N ]) is of the form U Gm . as proved in Proposition 2. where the reductive group Gm comes from the pure Tate motives. that takes into account the nontrivial extensions between pure Tate motives.1. [Deligne and Goncharov (2005)] For K = Q(ζN ) the cyclotomic ﬁeld of level N = 3 or N = 4 and O its ring of integers.
as in [Aluﬃ and Marcolli (2008a)]. using the classical Cremona transformation and dual graphs. we show 53 . that do not descend to the level of isomorphism classes of varieties in the Grothendieck ring. and the question of whether this (after removing divergences) will be a period of a mixed Tate motive can then be formulated as the question of whether a certain relative cohomology of the hypersurface complement relative to the normal crossings divisor is a realization of a mixed Tate motive. in the form of a hypersurface complement. a way of formulating Feynman rules directly at the algebrogeometric level of graph hypersurface complements. how one can sometimes obtain explicit computations of the classes in the Grothendieck ring of varieties for some families of Feynman graphs. as proved in [Aluﬃ and Marcolli (2009b)]. We show. deﬁned using characteristic classes of singular varieties. and a (possibly divergent) period integral on a chain with boundary on a normal crossings divisor. by showing how to associate to individual Feynman integrals an algebraic variety. We give examples of such Feynman rules that factor through the Grothendieck ring and others. perturbative quantum ﬁeld theory and motives. We follow ﬁrst what we referred to as the “bottomup approach” in the introduction. following [Aluﬃ and Marcolli (2009a)]. We begin by describing the parametric form of Feynman integrals and the main properties of the two Kirchhoﬀ–Symanzik graph polynomials involved in this formulation and the graph hypersurfaces deﬁned by the ﬁrst polynomial. although they still satisfy an inclusionexclusion principle. These classes also satisfy a form of deletion/contraction relation. We discuss the diﬀerence between working with aﬃne or projective hypersurfaces in describing Feynman integrals. Finally. We then describe. The motivic nature of this period.Chapter 3 Feynman integrals and algebraic varieties In this chapter we begin to connect the two topics introduced above. following the results of [Aluﬃ and Marcolli (2008b)].
. . . .j pj ) q1 (k1 ) · · · qn (kn ) dD k1 · · · dD kn in the form known as the Feynman parametric representation. . The incidence matrix = ( v. .37) depends on the orientation given to the edges of the graph. .i involved in writing the conservation laws at vertices in (1. Z). Lemma 3. we show how to reformulate the Feynman integral U (Γ. given by +1 if edge ei ∈ loop lk . . reverse orientation (3. 3. a choice of a maximal set of independent loops in the graph. . pN ) = δ( n i=1 v. . Now we also make a choice of a set of generators for the ﬁrst homology group H1 (Γ.54 Feynman Motives that one can reformulate the period computation underlying the Feynman amplitudes in terms of a simpler hypersurface complement. i. .i ki + N j=1 v. same orientation ηik = −1 if edge ei ∈ loop lk . .37) we have made a choice of an orientation of the graph Γ. . using determinant hypersurfaces. .7 above.1 The parametric Feynman integrals Using the Feynman parameters introduced in §1. This holds independently of the choice of the orientation of the graph and the basis of H1 (Γ. p1 . . the circuit matrix η = (ηik ). {l1 . . / There is a relation between the circuit matrix and the incidence matrix of the graph.1. We then deﬁne another matrix associated to the graph Γ. since the matrix v. l } with = b1 (Γ) the ﬁrst Betti number.37) in the form of an integration over the topological simplex σn as in (1.i ) and the circuit matrix η = (ηik ) of a graph Γ satisfy the relation η = 0. . (We neglect here an overall multiplicative factor in the coupling constants and powers of 2π. ranging over the chosen basis of loops.51). in terms of the Feynman parameters t = (t1 . which is given as follows. In writing the integral (1. We describe the advantages and the diﬃculties of this viewpoint. with i ∈ E(Γ) and k = 1. Z). but with a more complicated normal crossings divisor as the second term in the relative cohomology whose motivic nature one wishes to establish.) The ﬁrst step is to rewrite the denominator q1 · · · qn of (1. whose mixed Tate nature is established.1) 0 if edge ei ∈ loop lk .1.e. tn ) ∈ σn .
(3.2. as above. .1.i ηik =− v. Lemma 3. 2 t = (t1 . .i ηik = 0 for only two edges in the loop lk . .3) where A denotes the aﬃne line over a ﬁeld (here mostly C or R or Q).2 of [Nakanishi (1971)]. . The result follows from the observation that. one has v. Z). tn ). The Kirchhoﬀ–Symanzik matrix MΓ (t) of the graph Γ is the × matrix given by n (MΓ (t))kr = i=1 ti ηik ηir .4) Notice that.j ηjk so that one obtains v. both incident to a vertex v. the graph polynomial is independent of these choices. (3.3. The Kirchhoﬀ–Symanzik polynomial ΨΓ (t) is independent of the choice of edge orientation and of the choice of generators for H1 (Γ.2) Equivalently. tn ) → MΓ (t) = (MΓ (t))kr (3. Deﬁnition 3.1. since v. . with diﬀerent signs. independently of the choices of orientations. Z). We then deﬁne the Kirchhoﬀ matrix of the graph. . . see §2. it can be written as MΓ (t) = η † Λ(t)η. where † is the transpose and Λ(t) is the diagonal matrix with entries (t1 . also known as the Symanzik matrix. while the construction of the matrix MΓ (t) depends on the choice of an orientation on the graph Γ and of a basis of H1 (Γ. We think of MΓ as a function MΓ : A n → A . Deﬁnition 3.4. for two given edges ei and ej in a loop lk .i ηik i = 0. . . The Kirchhoﬀ–Symanzik polynomial ΨΓ (t) of the graph Γ is deﬁned as ΨΓ (t) = det(MΓ (t)).1.Feynman integrals and algebraic varieties 55 Proof.
We view it as a function ΨΓ : An → A. After rewriting the denominator of the integrand in (1. Z) changes MΓ (t) → AMΓ (t)A−1 . (3. and the integration in these variables. We deﬁne the aﬃne graph ˆ hypersurface XΓ to be the locus of zeros of the graph polynomial ˆ XΓ = {t ∈ An  ΨΓ (t) = 0}. We set ki = ui + r=1 ηir xr . . we want to replace in the Feynman integral U (Γ. A change of basis for H1 (Γ.i ui i=1 N + j=1 v.j pj = 0. .8) one ﬁnds the conservation condition n v. (3. pN ) the variables ki associated to the internal edges.8) that is. by variables xr associated to the independent loops in the graph and an integration only over these variables. . where A ∈ GL (Z) is the matrix that gives the change of basis. The determinant is once again unchanged. . i=1 ∀r = 1. (3. The change of sign in the corresponding row of η † leaves the determinant of MΓ (t) = η † Λ(t)η unaﬀected. (3.5) The polynomial ΨΓ is by construction a homogeneous polynomial of degree = b1 (Γ).7) and the constraint (3. (3.7) with the constraint n ti ui ηir = 0.6) ˆ ˆ of which XΓ is the aﬃne cone XΓ = C(XΓ ).37) in terms of an integration over σn using the Feynman parameters.37) give n v. p1 . (3. we require that the column vector Λ(t)u is orthogonal to the rows of the circuit matrix η. The momentum conservation conditions in the delta function in the numerator of (1. XΓ = {t ∈ Pn−1  ΨΓ (t) = 0}. using the linear constraints at the vertices.10) .1.j pj = 0.5. . Using the change of variables (3.9) Lemma 3. . . hence we can view it as deﬁning a hypersurface in projective space Pn−1 = (An {0})/Gm . A change of orientation in a given edge results in a change of sign in one of the columns of η = ηik . .i ki i=1 N + j=1 v. .56 Feynman Motives Proof.
First we give a convenient reformulation of the graph polynomial.1.7). In the case of graphs with several connected components. Proposition 3.1. The polynomial ΨΓ = det(MΓ (t)) can equivalently be described as the polynomial (see [Itzykson and Zuber (2006)].1. one deﬁnes the Kirchhoﬀ–Symanzik polynomial as in (3. as follows.10) constitute the Krichhoﬀ laws of circuits applied to the ﬂow of momentum through the Feynman graph. In particular they determine uniquely the ui = ui (p) as functions of the external momenta. The two equations (3.8) and (3.Feynman integrals and algebraic varieties 57 Proof.6. by which we mean here collections of a spanning tree in each connected component. (3.32) ΨΓ (t) = S e∈S te . In fact. This can be seen to be equivalent to the formulation (3.1.11) in terms of spanning trees of the graph Γ (see [Nakanishi (1971)] §1. . This polynomial is therefore multiplicative over connected components.9 below. This will be useful in Theorem 3. (3.4) is given by ΨΓ (t) = T ⊂Γ e∈E(T ) / te .3). each spanning tree has #V (Γ) − 1 edges and is the complement of a set S as above. We see the explicit form of the solution in Proposition 3.1. such that the removal of all the edges in S leaves a connected graph.11). The Kirchhoﬀ–Symanzik polynomial ΨΓ (t) of (3. The graph polynomial ΨΓ (t) has a more explicit combinatorial description in terms of the graph Γ. with the sum over spanning forests. This follows immediately from the orthogonality relation between the incidence matrix and circuit matrix of Lemma 3.7 below. Proof.11) where the sum is over all the spanning trees T of the graph Γ and for each spanning tree the product is over all edges of Γ that are not in that spanning tree. §623 and [Nakanishi (1971)] §1.12) where S ranges over all the subsets S ⊂ Eint (Γ) of = b1 (Γ) internal edges of Γ. We then have the following description of the resulting term i ti u2 i after the change of variables (3.
i v . p) .4) and PΓ (t.10). §623. We set VΓ (t.v such that − v Ai.1.v Pv Pv Ai. (3. The function VΓ (t.14) Proof. the collections of b1 (Γ) + 1 edges that divide the graph Γ into exactly two connected components Γ1 ∪Γ2 . with n (DΓ (t))v. p) of the form PΓ (t. as functions of the external momenta p = (pj ).15) In the massless case (m = 0).8.v ti . p) = p† RΓ (t)p + m2 . We have the following result.1. p)m=0 = PΓ (t.v ∈V (Γ) Pv (DΓ (t)−1 )v. ΨΓ (t. with the polynomial PΓ (t. . (3.7.13).v = i.13) and Pv = e∈Eext (Γ).16) of which the denominator is the graph polynomial (3.v Ai. p) = C⊂Γ sC e∈C te . with N = #Eext (Γ) with p† RΓ (t)p = v. i.i ti (3. hence −1 2 i ti u i = i v. we will see below that this is a ratio of two homogeneous polynomials in t. The term i ti u2 is of the form i ti u2 = p† RΓ (t)p. there is a matrix Ai. The result is a consequence of the fact that the ui . We give a quick sketch of the argument and we refer the reader to [Nakanishi (1971)] and [Itzykson and Zuber (2006)]. i i where RΓ (t) is an N × N matrix.v = i=1 −1 v.v so that one obtains the matrix DΓ (t) as in (3. p) is a homogeneous polynomial of degree b1 (Γ) + 1.t(e)=v pe . VΓ (t. p) satisﬁes (3.17) where the sum is over the cutsets C ⊂ Γ. are determined by the Kirchhoﬀ law (3.58 Feynman Motives Proposition 3.e. §623 for more details. Proposition 3.v Pv . p) (3.16) in the massless case. see [Itzykson and Zuber (2006)]. Thus. (3.v Pv = ti ui . The constraints on the ui given by the Kirchhoﬀ laws also show that Ai.
.9.Feynman integrals and algebraic varieties 59 The coeﬃcient sC is a function of the external momenta attached to the vertices in either one of the two components. see [Bjorken and Drell (1964)] §8 and [Bjorken and Drell (1965)] §18.the Feynman integral U (Γ. This is related to the polynomial ΨΓ (t) by ΨΓ (t) = ( e te ) det(DΓ (t)) = T e∈T / te . We can now rewrite the Feynman integral in its parametric form as follows. p1 . for m = 0.1. .15) becomes p† RΓ (t)p = v. Up to a multiplicative constant Cn. . Theorem 3. . .n dD x1 · · · dD x =C n ( i=0 ti qi )n det(MΓ (t))−D/2 ( i=0 ti (u2 + m2 ))−n+D i /2 . as the sum of the incoming external momenta. 2 2 sC = v∈V (Γ1 ) Pv = v∈V (Γ2 ) Pv . p1 . Proof.v Pv = 1 ΨΓ (t) sC C⊂Γ e∈C te . Thus. We give a brief sketch of the argument and refer the reader to the more detailed treatment in [Nakanishi (1971)]. (3.19) σn where ωn is the volume form on the simplex σn .13) has determinant det(DΓ (t)) = T e∈T t−1 . . (3.14). . (3. . p) /2 . the function (3. .v ∈V (Γ) Pv (DΓ (t)−1 )v. pN ) can be equivalently written in the form U (Γ. e where T ranges over the spanning trees of the graph Γ. The matrix DΓ (t) of (3. We ﬁrst show that we have n . pN ) = Γ(n − D ) 2 (4π)D /2 ΨΓ (t)D/2 V ωn n−D Γ (t.18) where the Pv are deﬁned as in (3. Proof.20) .
In the massless case.n . p). . .60 Feynman Motives where ui = ui (p) as above. the nature as a period) of the remaining integral I(Γ.n a−n+D /2 k=1 λk −D/2 . p)−n+D /2 ωn .e. using the expression (3. + m2 ) + k. pN ) = Γ(n − D ) 2 (4π)D /2 PΓ (t. after the change of variables (3. . We observe here that these hypersurfaces are in general singular and with singular locus of positive dimension. . the left hand side of (3. p1 .2 The graph hypersurfaces We introduced above the projective hypersurfaces XΓ ⊂ Pn−1 . p) as U (Γ. and the corresponding aﬃne hypersurface XΓ ⊂ An . p)−n+D /2 ωn . see §3. = dD x1 · · · dD x .20) reads ( n 2 i=0 ti (ui dD x1 · · · dD x . (1 + k x2 )n k We then write det MΓ (t) = ΨΓ (t) and we use the expression of Proposition 3. .r (MΓ )kr xk xr )n The integral can then be reduced by a further change of variables that diagonalizes the matrix MΓ to an integral of the form d D y1 · · · d D y =C 2 (a + k λk yk )n with C . In fact. ΨΓ (t)−n+D( +1)/2 (3. one writes equivalently the integral U (Γ. p) as in (3. with n = ˆ #Eint (Γ). .21) σn Up to a divergent Gammafactor.15).22) 3.7).5 below for more details. . with VΓ (t. The fact that we are dealing with singular hypersurfaces in projective spaces . ΨΓ (t)−n+D( +1)/2 (3. one is interested in understanding the motivic nature (i.7 to express the term ( i ti (u2 + m2 ))−n+D /2 in terms of i 2 ti (ui + m2 ) = i i 2 ti ui + m2 = VΓ (t. . pN ) = σn PΓ (t. p) in terms of the polynomials PΓ (t. p) and ΨΓ (t). p1 .1.16) of VΓ (t.
This is one of the main themes that we are going to develop in the rest of the book. In terms of graph hypersurfaces.23) can be regarded (modulo the problem of divergences) as a period obtained by integrating an algebraic diﬀerential form deﬁned on the hypersurface complement Pn−1 XΓ over a domain given by the topological simplex σn . we will necessarily be dealing with mixed motives.19) in the stable range where n ≤ D /2. It is important to stress the fact that identifying the motivic nature of the relative cohomology above would only suﬃce up to the important issue . We only make some general considerations about this case in §3. when we consider motives associated to these geometric objects.26) A possible strategy to showing that the Feynman integrals evaluate to numbers that are periods of mixed Tate motives would be to show that the motive (3. p1 . the motivic nature of this period should be detected by the mixed motive m(Pn−1 X Γ .24) We will not discuss the unstable range of small D. though it is often of signiﬁcant physical interest. the algebraic diﬀerential form that is integrated in the parametric representation of the Feynman integral only has singularities along the graph hypersurface XΓ .23) for n = #Eint (Γ) and = b1 (Γ). . the Feynman integral with the condition (3. This means that we look at the parametric Feynman integrals (3. . Σn (Σn ∩ XΓ )). Thus. for which the integral (3. In this range. . while the polynomial PΓ (t.Feynman integrals and algebraic varieties 61 implies that. p) caused by the additional dependence on the external momenta.25) whose realization is the relative cohomology H n−1 (Pn−1 X Γ . the special case of the log divergent graphs. We consider only the case where the spacetime dimension D of the quantum ﬁeld theory is suﬃciently large. . p) only appears in the numerator. (3. because of additional diﬃculties in treating the hypersurfaces deﬁned by the second graph polynomial PΓ (t.3 below.25) is mixed Tate. in particular. (3. whose boundary ∂σn is contained in the normal crossings divisor Σn given by the algebraic simplex (the union of the coordinate hyperplanes). pN ) = σn ωn . These are the graphs with n = D /2. ΨΓ (t)D/2 (3. This range covers. Σn (Σn ∩ XΓ )). (3.22) reduces to the simpler form I(Γ.
ΨΓ (t) > 0.14 below. . pk ). We now discuss brieﬂy the diﬀerence between working with the aﬃne or the projective hypersurfaces. While it is natural to work projectively when one has equations deﬁned by homogeneous polynomials. with pj = (pj. Notice that the intersections XΓ ∩σn can only occur on the boundary ∂σn . Esnault.27) −n+D( +1)/2 σn ΨΓ (t) with a period of the motive (3. one also needs to know that the motive remains mixed Tate after the blowups. the Feynman rules are multiplicative over disjoint unions of graphs. which in the projective case is replaced by the more complicated join operation. with the result of separating the domain of integration and the hypersurface and therefore regularizing the integral by replacing the original integration with one performed in the blown up variety. .1 . . In fact. This regularization procedure adds a further complication to the problem of identifying the motivic nature of (3. We will return to this issue in more detail in §3. These intersections are usually nonempty. before one can identify the integration PΓ (t. pj. pk ). As we saw in §1. as can easily be seen by the expression (3. The regularization method proposed in [Bloch.28) for Γ = Γ1 ∪ · · · ∪ Γk a disjoint union of graphs with Nj = #Eext (Γj ) external edges. there are various reasons why the aﬃne context appears to be more natural in this case. Divergences occur where the graph hypersurface XΓ meets the domain of integration σn . .Nj ). we have XΓ ∩ σn ⊂ XΓ ∩ Σn .25). we argue as in [Aluﬃ and Marcolli (2008b)] that the multiplicative property of the Feynman rules is more naturally reﬂected by a corresponding multiplicative property of the aﬃne hypersurface complements. . .62 Feynman Motives of divergences. which allows one to determine the . namely U (Γ. Namely. after we discuss the approach of [Aluﬃ and Marcolli (2009a)]. p1 ) · · · U (Γk . and the graph polynomial is then also strictly positive. Kreimer (2006)] consists of performing a number of blowups of (strata of) the locus XΓ ∩ Σn . in the interior of σn all the coordinates te are strictly positive real numbers. p) = U (Γ1 . (3. with external momenta p = (p1 . so that some method of regularization needs to be introduced to remove the corresponding divergences in the integral. Thus. p)−n+D /2 ωn (3.3. .25). . In particular. The other “multiplicative” property that the Feynman rules satisfy is the one we discussed in the ﬁrst chapter.11).
p)p=0 = U (e.2. as in (1. More precisely. Then the aﬃne hypersurface complements satisfy ˆ ˆ ˆ (3. Deﬁnition 3. (3. pN ) = e∈Eext (Γ) qe (pe )−1 .Feynman integrals and algebraic varieties 63 Feynman integrals of non1PI graphs using the integrals for 1PI graphs.33) An XΓ = (An1 XΓ1 ) × · · · × (Ank XΓk ). . . p) = V (Γv . In the particular case of a massive theory m = 0 where one sets all the external momenta equal to zero. . with the property that. pN ). . pe )pe =0 U (Γv . . upon representing an arbitrary (ﬁnite) graph as a tree T with vertices v ∈ V (T ) replaced by 1PI graphs Γv with a number of external edges #Eext (Γv ) = val(v) equal to the valence of the vertex.30) v∈V (T ) where in this case the edge propagator U (e. . . In fact.v∈∂(e) where. following [Aluﬃ and Marcolli (2008b)]. One can then.e∈Eint (T ). with ε(p1 . this becomes an actual product #Eint (T ) U (Γ. . we have the following result from [Aluﬃ and Marcolli (2008b)]. for a disjoint union Γ = Γ1 ∪ · · · ∪ Γk this satisﬁes the multiplicative property U (Γ) = U (Γ1 ) · · · U (Γk ) (3. one obtains the expression for the Feynman integral of the resulting graph Γ as V (Γ. deﬁne an abstract Feynman rule in the following way. is the fact that the aﬃne hypersurface complements behave multiplicatively as abstract Feynman rules are expected to do. while the projective hypersurface complements do not. . . where ni = #Eint (Γi ). pe )pe =0 = m−2 is just a constant depending on the mass parameter m = 0.29) v∈V (T ). pv )qe (pv )−1 δ((pv )e −(pv )e ). .1. p1 . An abstract Feynman rule is an assignment of an element U (Γ) in a commutative ring R for each ﬁnite graph Γ. . . . . . pN ) U (Γ. p1 .2.35) we write V (Γ. (3. which is directly relevant to the quantum ﬁeld theory setting. Lemma 3.2.31) and for a non1PI graph obtained by inserting 1PI graphs Γv at the vertices of a tree T it satisﬁes U (Γ) = U (L)#Eint (T ) U (Γv ).32) v∈V (T ) A ﬁrst diﬀerence one encounters between the projective graph hypersurˆ faces XΓ and the aﬃne XΓ . . Let Γ = Γ1 ∪ · · · ∪ Γk be a disjoint union of ﬁnite graphs. (3. pv )pv =0 . pN ) = ε(p1 .
1 : · · · : t2. Instead of getting directly a product of the complements one gets a torus bundle.35) would not be deﬁned everywhere.1 : · · · : t2.1 .1 : · · · : t2. This means that the aﬃne hypersurface XΓ is a union k ˆ XΓ = i=1 ˆ (An1 × · · · × Ani−1 × XΓi × Ani+1 × · · · × Ank ).1 : · · · : ut1.n1 : t2.34) is then induced by the isomorphism in (3. Equivalently..1 : · · · : t1.33). Under the assumption that neither Γi is a forest. The hypothesis that neither graph is a forest ensures that there is a regular map Pn1 +n2 −1 given by (t1. To see this.1 : · · · : t1.j=1.. suppose given a disjoint union of two graphs Γ = Γ1 ∪ Γ2 and assume. .j )i=1. then there is a regular map ˆ An XΓ → Pn−1 XΓ .. where t = (te ) = (ti. (3.n1 ). .. (t2. . respectively. ti. notice that if Γ is not a forest. .ni ). For instance.. (t2. the map (3.34) .33).n2 )).. respectively over XΓ1 and XΓ2 with vertices Pn2 −1 and Pn1 −1 . one sees easily that if Γ = Γ1 ∪ · · · ∪ Γk is a disjoint union. The situation with the projective hypersurfaces is more complicated..1 : · · · : vt2. This is no longer the case if Γ is a forest. Then the projective hypersurface complement Pn1 +n2 −1 XΓ is a Gm bundle over the product (Pn1 −1 XΓ1 ) × (Pn2 −1 XΓ2 ).n2 ) → ((t1.34) is surjective and the ﬁber over a point ((t1.35) while if either Γi happened to be a forest. that neither graph is a forest (the assumption is not necessary in the aﬃne case).1 : · · · : t1. and the map (3. with (u : v) ∈ P1 .ni are the edge variables of the internal ˆ edges of Γ. XΓ → (Pn1 −1 XΓ1 ) × (Pn2 −1 X Γ2 ) (3. the corresponding XΓi would be empty.64 Feynman Motives Proof..n1 ). then one has k ΨΓ (t) = i=1 ΨΓ1 (ti. Using the deﬁnition of the graph polynomials in terms of the determinant ΨΓ (t) = det MΓ (t) of the Kirchhoﬀ matrix of the graph. observe that in this case the projective hypersurface XΓ is a union of cones C n2 (XΓ1 ) and C n1 (XΓ2 ) in Pn1 +n2 −1 . uv = 0. moreover.n2 ).n1 : vt2. and (3.n2 )) is a copy of the multiplicative group Gm given by all the points of the form (ut1.k. which implies that the hypersurface complement is then given by the product (3.
For simplicity. A 1PI graph Γ satisﬁes the generic condition on the external momenta if. some of the divergences of the integral. the situation is more complicated because of the explicit dependence on the additional parameters p. the second graph polynomial PΓ (t. (3. i.1. p) and ΨΓ .37) ˆ ˆ or else on the complement of the union of hypersurfaces YΓ (p) ∪ XΓ . §18. is typically not the range of physical interest for theories where D = 4 and the number of loops and internal edges violates the above inequality. In such cases. can occur in the interior of the domain. Away from this stable range. This. p) = 0}. not only on its boundary as in the case of XΓ .3. t). and D. p) and ΨΓ (t) have no common factor. however. the polynomials PΓ (t. since PΓ (·. for a ﬁxed value of the external momenta p. again depending on the values of n. . This is certainly the case within what we referred to as the “stable range”. Deﬁnition 3. In particular.36) or in projective space. p) is homogeneous of degree b1 (Γ) + 1. are sometimes referred to as the Landau varieties. for p in a dense open set in the space of external momenta. see [Bjorken and Drell (1965)]. The ˆ hypersurfaces YΓ (p) deﬁned by the vanishing of PΓ (p.e. p) also appears in the denominator. We only discuss here the case where −n + D /2 ≥ 0. though some of the same arguments extend to the other case as in [Marcolli (2008)]. and the diﬀerential form of the parametric Feynman integral is then deﬁned on the complement of the hypersurface (family of hypersurfaces) ˆ YΓ (p) = {t ∈ An  PΓ (t. for D (considered as a variable) suﬃciently large so that n ≤ D /2. The general type of arguments described above using the graph hypersurˆ faces XΓ should then be adapted to the case of the Landau varieties.Feynman integrals and algebraic varieties 65 3. coming from the intersection of the hypersurface YΓ (p) with the domain of integration σn . it is also convenient to make the following assumption on the polynomials PΓ (t. on the complement of YΓ (p) = {t ∈ Pn−1  PΓ (t.3 Landau varieties For simplicity we restrict here everywhere to the case where the diﬀerential form in the parametric Feynman integral has singularities only along the ˆ graph hypersurface XΓ deﬁned by the vanishing of the graph polynomial ΨΓ (t). p) = 0} (3. however. so that we avoid cancellations of common factors.
for a degree one polynomial L(t).66 Feynman Motives Recall that PΓ (t.40) by LT (t) = p2 e∈Tv1 . the simplest possible conﬁguration of external momenta is the one where one puts all the external momenta to zero. These cut sets consist of the complement of a spanning tree T and an edge of Tv1 . we obtain in this case PΓ (p. Upon writing the polynomial PΓ (t.38). one has PΓ (t. one would have PΓ (p. t) = ΨΓ (t) · L(t).v2 ⊂ T is the unique path in T without backtrackings connecting the vertices v1 and v2 .e te e∈T c te . t) = p2 T ( e ∈Tv1 .e = sC for the cutset C = T c ∪ {e }. . e2 } ⊂ Eext (Γ). We then have.40).8.40) where Tv1 . p) = T e ∈T sT. in this case.41) we see that. We use as in (3.38) where sT.v2 te . In terms of spanning trees. (3. Let vi be the unique vertex attached to the external edge ei of the chosen pair. and this would give (LT (t) − L(t)) T e∈T / te ≡ 0. p) in the form (3. Consider for simplicity the case where the polynomial ΨΓ is irreducible.39) Thus.1. p) is deﬁned in terms of external momenta and cutsets through the functions Pv and sC as in Proposition 3.40). Pv1 = p = −Pv2 . The parameterizing space of the external momenta is the hyperplane in the aﬃne space AD·#Eext (Γ) obtained by imposing the conservation law pe = 0.18) 2 2 sC = v∈V (Γ1 ) Pv = v∈V (Γ2 ) Pv to get sC = p2 for all the nonzero terms in this (3. (3. These are all the terms that correspond to cut sets C such that the vertices v1 and v2 belong to diﬀerent components. if the polynomial ΨΓ (t) divides (3.v2 te ) e∈T / te . except for a pair pe1 = p = −pe2 associated to a choice of a pair of external edges {e1 . (3. e∈Eext (Γ) (3.v2 . If we denote the linear functions of (3.
we ﬁrst recall some basic facts about diﬀerential forms on aﬃne and projective spaces and their relation.43) is the operator of contraction with the Euler vector E= i ti ∂ . one may be able to ﬁnd vertices and momenta as above such that the degree one polynomials LT (t) are all equal to the same L(t). vk−1 ). vk−1 ) = ω(E. p) ˆ as computed over the aﬃne hypersurface complement An XΓ .1.42) The relation between the volume form dt1 ∧ · · · ∧ dtn and the homogeneous form Ω of degree n of (3. 3. One then sees. or an ˆ aﬃne hypersurface XΓ ⊂ An . . the validity of the condition of Deﬁnition 3. with n = #Eint (Γ). assuming we are in the stable range of suﬃciently high D where −n + D /2 ≥ 0. and Graev (1980)]. the aﬃne cone over XΓ .Feynman integrals and algebraic varieties 67 for all t. . see also [Gelfand.4 Integrals in aﬃne and projective spaces As we have seen above. Gindikin.44) (3. where ∆ : Ω → Ω ﬁeld k k−1 (3. In fact.1 can be checked algorithmically to be satisﬁed for large classes of 1PI graphs. . . for a graph that is not 1PI.3. In order then to reformulate in projective space Pn−1 integrals originally deﬁned in aﬃne space An . following mostly [Dimca (1992)]. one can regard the Feynman integral U (Γ. the homogeneous graph polynomial ΨΓ deﬁnes either a projective hypersurface XΓ ⊂ Pn−1 . . ∂ti (3. Generally. though a complete combinatorial characterization of the graphs satisfying these properties does not appear to be known. .42) is given by Ω = ∆(ωn ). v1 . . . for example. that the 1PI condition on the graph Γ is necessary in order to have the condition of Deﬁnition 3. Thus.45) ∆(ω)(v1 .3. but one can also reformulate it in terms of the projective hypersurface complement. for = b1 (Γ) the number of loops. . (3. The projective analog of the volume form ωn = dt1 ∧ · · · ∧ dtn is given by the form n Ω= i=1 (−1)i+1 ti dt1 ∧ · · · ∧ dti ∧ · · · ∧ dtn .
180 and [Dolgachev (1982)]) ˆ Given ω ∈ Ωk (D(f )). Let ω ∈ Ωk deg(f ) be a closed kform.44). Proposition 3. we can always write a form ω ∈ Ωk (D(f )) as η ω = m . Let Rm denote the subset of homogeneous polynomials of degree m. With the notation introduced above. .46) m f We then have the following characterization of the pullback along π : ˆ D(f ) → D(f ) of forms on D(f ). tn ] be the ring of polynomials of An .47) ∆ ∆ ∆ ∆ In particular. . with η ∈ Ωk deg(f ) .45) with the Euler vector ﬁeld E of (3. fm with η ∈ Ωk deg(f ) . and by D(f ) = An Xf and D(f ) = Pn−1 Xf the hypersurface complements. . since the sequence 0 → Ωn → Ωn−1 → · · · → Ω1 → Ω0 → 0 is exact at all but the last term.1. Proposition 3. Similarly. We then obtain the following result formulating integrals in aﬃne spaces in terms of integrals of pullbacks of forms from projective spaces.2. m (3. . one can write π ∗ (ω) = ∆(η) . Let π : An {0} → Pn−1 be the standard projection t = ˆ (t1 . fm with h ∈ Rm deg(f )−n (3. (3. .4. let Ωk denote the Rmodule of kforms on An and let Ωk denote the subset of m kforms that are homogeneous of degree m. . tn ) → π(t) = (t1 : · · · : tn ). Thus.42).68 Feynman Motives Let R = C[t1 . where ∆ is the contraction (3. We denote by Xf = {t ∈ An  f (t) = 0} n−1 the aﬃne hypersurface and by Xf = {π(t) ∈ P  f (t) = 0} the projective ˆ ˆ hypersurface. any (n − 1)form on D(f ) ⊂ Pn−1 can be written as hΩ . Let . the pullback π ∗ (ω) ∈ Ωk (D(f )) is characterized by the properties of being invariant under the Gm action on An {0} and of satisfying ∆(π ∗ (ω)) = 0. (see [Dimca (1992)]. . and consider the form ω/f m on An . .4. Let f be a homogeneous polynomial function f : An → A of degree deg(f ).48) and with Ω = ∆(dt1 ∧ · · · ∧ dtn ) the (n − 1)form (3. p. which is hom mogeneous of degree m deg(f ). homogeneous of degree n.
df ω = f dω − m df ∧ ω. . For simplicity.53) By Stokes’ theorem we have ∆(ω) = fm d σ ∂σ ∆(ω) fm . dω = 0.54) (3.51) Since the form ω is closed. In the range −n + D( + 1)/2 ≥ 0 we consider the Feynman integral ˆ U (Γ. Then the integration of ω/f m over σ satisﬁes m deg(f ) σ ω = fm ∂σ ∆(ω) + fm df ∧ σ ∆(ω) . f m+1 (3. f m+1 (3. fm f m+1 (3. and we have ∆(df ∧ ω) = deg(f ) f ω − df ∧ ∆(ω). f m+1 ∂σ σ σ which gives (3. p) as deﬁned on the aﬃne hypersurface complement An XΓ and we use the result above to reformulate the parametric Feynman integrals in terms of integrals of forms that are pullbacks to An {0} of forms on a hypersurface complement in Pn−1 . we remove here the divergent Γfactor from the parametric Feynman integral and we concentrate on the residue given by the integration over the simplex σ as in (3.52) (3.50) where.54) this gives ∆(ω) = m deg(f ) fm ω − fm df ∧ ∆(ω) . [Dolgachev (1982)]) d ∆(ω) fm =− ∆(df ω) . Recall that we have ([Dimca (1992)].55) below.49) Proof. for a form ω that is homogeneous of degree m deg(f ). we have d ∆(ω) fm =− ∆(dω) ∆(df ∧ ω) +m . Thus. fm f m+1 (3. Using (3. we obtain from the above d ∆(ω) fm = m deg(f ) ω df ∧ ∆(ω) − .49).Feynman integrals and algebraic varieties 69 σ ⊂ An {0} be a kdimensional domain with boundary ∂σ = ∅.
p)−n+D /2 ωn with ωn = dt1 ∧ · · · ∧ dtn the volume form on An . for a generic choice of the external momenta the polynomials PΓ and ΨΓ have no common factor.57) π ∗ (η) = m . p)−n+D σ dt1 ∧ · · · −n+D( +1)/2 ΨΓ ∆(ω) + fm /2 ∧ dtn = σ ω fm = = C(n. ) π ∗ (η) + ∂σ σ df ∧ π ∗ (η) f . p)a ∆(ωn ) + Ψm Γ df ∧ σ with a = −n + D /2 and m = −n + D( + 1)/2. we obtain PΓ (t. The coeﬃcient C(n. D. )−1 1 ( m deg(f ) df ∧ σ ∂σ ∆(ω) ) f m+1 PΓ (t. hence it is given by (3. and assuming that −n + D /2 ≥ 0. ) is given by C(n. D. First notice that. p)a ∆(ωn ) Ψm+1 Γ . .56) where π : An {0} → Pn−1 is the projection and η is the form on the hypersurface complement D(f ) in Pn−1 with ∆(ω) (3. The coeﬃcient C(n.3. ∂σ PΓ (t. p) = σ VΓ −n+D /2 D/2 ΨΓ ωn . ) = m deg(f ). where f = ΨΓ and m = −n+D( +1)/2. and ω as above. D. hence it is the pullback of a form η on D(f ) ⊂ Pn−1 . is contained in a fundamental domain of the action of the multiplicative group C∗ on Cn {0}.e.56) is given by C(n. Under the generic condition on the external momenta. Moreover. f on An . the form ∆(ω)/f m is Gm invariant on An {0}. it also satisﬁes ∆(α) = 0.4.1.58) Proof. given by the simplex σ = σn = { i ti = 1. p). and ω = PΓ (t. ) in (3.3. p)/ΨΓ (t. p) = PΓ (t.58). D. i. (3.55) with VΓ (t. (3. the parametric Feynman integral U (Γ.4. with f .2 above. D. can be computed as U (Γ. ) = (−n + D( + 1)/2) . with m and f as above. D. Consider on An the form given by ∆(ω)/f m .70 Feynman Motives Proposition 3. We assume the condition of Deﬁnition 3. Also notice that the domain of integration σ ⊂ An . Applying the result of Proposition 3. m. (3. p) = 1 C(n. since the polynomial ΨΓ is homogeneous of degree and PΓ is homogeneous of degree +1. ti ≥ 0}. since it is of the form α = ∆(ω)/f m .
Thus.5. if T is a spanning tree containing the edge e. where Γe is the graph obtained by removing the edge e of Γ. it makes it especially useful to adopt methods from singularity theory to study these hypersurfaces. For example. notice that.9 below. for e ∈ E(Γ). Lemma 3. given any two C(XΓe ) and C(XΓe ). . . while if T is a spanning tree of Γ which does not contain e. then T e is no longer a spanning tree of Γe . where Γe is the graph obtained from Γ by removing the edge e.1. Moreover. 0. we are going to see in the following sections how characteristic classes of singular varieties can be employed in this context. the hypersurfaces XΓ tend to have singularity loci of low codimension.Feynman integrals and algebraic varieties 71 3. the Euler formula mΨΓ (t) = ∂ e te ∂te ΨΓ (t) implies that ∩e Z(∂e ΨΓ ) ⊂ XΓ . Proof. In fact. The variables te appear in the polynomial ΨΓ (t) only with degree zero or one. so the corresponding terms disappear in passing from ΨΓ to ΨΓe . the singular locus of XΓ is just given by the equations ∂e ΨΓ = 0. . The resulting polynomial is therefore of the form ΨΓe . The singular locus of XΓ is given by the intersection of cones over the hypersurfaces XΓe . Ongoing work [Bergbauer and Rej (2009)] gives an analysis . Let Γ be a graph with deg ΨΓ > 2. one can see in terms of spanning trees that. To see that these cones do not intersect transversely. . These provide a way of measuring how singular a hypersurface is. Thus. In fact. . The cones C(XΓe ) do not intersect transversely. As we observe again later. the zero locus Z(ΨΓe ) ⊂ Pn−1 is a cone C(XΓe ) over the graph hypersurface XΓe ⊂ Pn−2 with vertex at the coordinate point ve = (0. in the case where deg ΨΓ > 2. as we discuss more in detail in §3. . where Z(∂e ΨΓ ) is the zero locus of the te derivative. First observe that. 0. then T is still a spanning tree of Γe and the corresponding monomial mT of ΨΓe is the same as the monomial mT in ΨΓ without the variable te . 0) with te = 1. This can easily be seen by the following observation.5 Nonisolated singularities The graph hypersurfaces XΓ ⊂ Pn−1 deﬁned by the vanishing of the polynomial ΨΓ (t) = det(MΓ (t)) usually have nonisolated singularities. 1. the vertex of one cone is contained in the graph hypersurface spanning the other cone. this has important consequences from the motivic viewpoint. with ΨΓ a polynomial of degree m. hence the polynomial ∂e ΨΓ consists of only those monomials of ΨΓ that contain the variable te . where one sets te = 1. since XΓ is deﬁned by a homogeneous equation ΨΓ (t) = 0. .
We follow [Aluﬃ and Marcolli (2008a)].1. . (3. . This was pointed out in [Bloch (2007)] and we illustrate the principle here in the particular completely explicit case of the “banana graphs” computed in [Aluﬃ and Marcolli (2008a)].72 Feynman Motives of the singular locus of the graph hypersurfaces using a formula for the Kirchhoﬀ polynomials ΨΓ under insertion of subgraphs at vertices. let L be the hyperplane deﬁned by the equation L = {(t1 : · · · : tn ) ∈ Pn−1  t1 + · · · + tn = 0}. (3. as shown in the recent results of [Bloch (2008)]. Also. t2 · · · tn ). one can still use the image of the graph hypersurface under the Cremona transformation to derive useful information on its motivic nature. t1 t3 · · · tn . . which for planar graphs relates the hypersurface of a graph with that of its dual.6.61) The Cremona transformation is a priori deﬁned only away from the algebraic simplex of coordinate axes Σn := {(t1 : · · · : tn ) ∈ Pn−1  i ti = 0} ⊂ Pn−1 . Then G(C) is a subvariety of Pn−1 × Pn−1 with projections G(C) ?? ??π2 ?? ? C Pn−1 _ _ _ _ _ _/ Pn−1 π1 (3. (3. The standard Cremona transformation of Pn−1 is the map C : (t1 : · · · : tn ) → 1 1 : ··· : t1 tn .60) We introduce the following notation. Deﬁnition 3. 3. Even for graphs that are nonplanar.62) (3. Let Sn ⊂ Pn−1 be deﬁned by the ideal ISn = (t1 · · · tn−1 .63) . .6 Cremona transformation and dual graphs A useful tool to study the algebraic geometry of the projective graph hypersurfaces XΓ is the Cremona transformation.59) Let G(C) denote the closure of the graph of C. t1 · · · tn−2 tn .
as we now show. Proof.2. An example where it is easier to visualize the eﬀect of the Cremona transformation is the case where n = 3.7.61) generated by the partial derivatives of the equation of the algebraic simplex (i. so that the resulting image is again a triangle of three lines in P2 . π1 : G(C) → Pn−1 is the blowup along Sn .64) Using (3. its locus of indeterminacies being only the singularity subscheme Sn of the algebraic simplex Σn .1 below. Let C.64). . These three points. Lemma 3. are blown up in G(C) and each replaced by a line.6.Feynman integrals and algebraic varieties 73 though. Properties (3) and (4) of the proposition above only play a role implicitly in the proof of Proposition 3. Σn cuts out a divisor with simple normal crossings on L. The Cremona transformation (the horizontal rational map in the diagram (3. the singular locus of Σn ). L intersects every component of Sn transversely. the zerodimensional strata of Σ3 .60). it is in fact well deﬁned also on the general point of Σn . (1) (2) (3) (4) The indeterminacy locus of C is Sn . one can view the algebraic simplex Σ3 in the projective plane P2 as a union of three lines. The proper transforms of the 1dimensional components of Σ3 are then blown down by the other map π2 of (3. (3. The components of the map deﬁning C given in (3. The properties (3) and (4) then follow. so we leave here the details to the reader. In this case. and L be as above.59) can be rewritten as (t1 : · · · : tn ) → (t2 · · · tn : t1 t3 · · · tn : · · · : t1 · · · tn−1 ). each two intersecting in a point.e. the map π1 : G(C) → Pn may be identiﬁed with the blowup of Pn along the subscheme Sn deﬁned by the ideal ISn of (3. G(C).60)) is an isomorphism of the complements of the two triangles. Sn .
as the following ﬁgure shows. but also on the choice of the embedding ι : Γ → S 2 . the dual graph has #E(Γ∨ ) = #E(Γ). since in fact this depends not only on Γ itself. w ∈ V (Γ∨ ) for each edge of Γ that is in the boundary between the regions of S 2 ι(Γ) containing the points v and w. Thus. ι2 of the same graph Γ. for which the corresponding dual graphs Γ∨ and 1 Γ∨ are topologically inequivalent. (3.74 Feynman Motives The dual graph Γ∨ of a planar graph Γ is deﬁned by choosing an embedding ι : Γ → S 2 in the sphere and then taking as vertices v ∈ V (Γ∨ ) a point in each connected component of S 2 ι(Γ). (3. One can give examples of diﬀerent embeddings ι1 .66) Notice that it is somewhat misleading to talk about the dual graph Γ∨ . #V (Γ∨ ) = b0 (S 2 ι(Γ)).65) and edges E(Γ∨ ) given by adding an edge between two vertices v. 2 .
. with dual graph Γ∨ . . . tn ). . . . This follows from the combinatorial identity ΨΓ (t1 . . so that deg ΨΓ + deg ΨΓ∨ = #E(Γ). . . . tn ) = ( te ) ΨΓ∨ (t−1 . C(XΓ ∩ (Pn−1 Σn )) = XΓ∨ ∩ (Pn−1 Σn ). .67) 1 e∈E(Γ) hence the graph hypersurfaces are related by the Cremona transformation C of (3. . Then the graph polynomials satisfy −1 ΨΓ (t1 . . . sn ) e from which (3. . (3. . Written in the coordinates (s1 : · · · : sn ) of the target Pn−1 of the Cremona transformation. .59). . . . . 1 The third equality uses the fact that #E(Γ) = #E(Γ∨ ) and #V (Γ∨ ) = b0 (S 2 Γ). Suppose given a planar graph Γ with #E(Γ) = n. (3. tn ). tn ) = ( e∈E(Γ∨ ) s−1 )ΨΓ∨ (s1 .68) Proof.6. . .3.67) gives ΨΓ (t1 .68) follows. tn ) = T ⊂Γ e∈E(T ) te / =( =( e∈E(Γ) te ) e∈E(Γ) te ) T ⊂Γ T ⊂Γ∨ −1 e∈E(T ) te −1 e∈E(T ) te / −1 = ( e∈E(Γ) te )ΨΓ∨ (t−1 . Lemma 3. . and the fact that there is a bijection between complements of spanning tree T in Γ and spanning trees T in Γ∨ obtained by shrinking the edges of T in Γ and taking the dual graph of the resulting connected graph. the identity (3. .Feynman integrals and algebraic varieties 75 dual graph dual graph It was shown in [Bloch (2007)] that the Cremona transformation relates the graph hypersurfaces of Γ and its dual Γ∨ in the following way (see also [Aluﬃ and Marcolli (2008a)]).
The class in the Grothendieck ring of the hypersurface XΓn of the nth banana graph is explicitly given by the formula [XΓn ] = Ln − 1 (L − 1)n − (−1)n − − n (L − 1)n−2 . one observes easily . The banana graph Γn has graph polynomial 1 1 ΨΓ (t) = t1 · · · tn ( + · · · + ).76 Feynman Motives 3. The formula (3. In fact.69) is proved using the method of Cremona transformation and dual graphs described above. We gave in [Aluﬃ and Marcolli (2008a)] an explicit formula for the class in the Grothendieck ring of the graph hypersurfaces for the banana graphs. These are known as the banana graphs: for n ≥ 2 the graph Γn in this family has two vertices of valence n and n parallel edges between them.69) In particular. Proposition 3. the class [XΓn ] manifestly lies in the mixed Tate part Z[L] of the Grothendieck ring.1. t1 tn Thus. this formula shows that. the parametric integral in this case is (t1 · · · tn )( 2 −1)(n−1)−1 ωn σn D ΨΓ (t)( 2 −1)n D . L−1 L (3. This is given by the following result. Proof. We sketch the proof here and refer the reader to [Aluﬃ and Marcolli (2008a)] for more details. in this example. up to an overall multiplicative factor proportional to the sum of external momenta at a vertex.7 Classes in the Grothendieck ring An example of an inﬁnite family of Feynman graphs for which the classes in the Grothendieck ring K0 (V) can be computed explicitly was considered in [Aluﬃ and Marcolli (2008a)].7.
which is of the form χ(XΓn ) = n + (−1)n . and therefore the associated graph hypersurface XΓ∨ = L n is just a hyperplane in Pn−1 . A diﬀerent computation of the Euler characteristic χ(XΓn ). is also given in [Aluﬃ and Marcolli (2008a)] along with the more complicated explicit computation of the Chern–Schwartz–MacPherson characteristic class of the hypersurfaces XΓn . for n ≥ 3. one can still consider a dual hypersurface XΓ obtained as the image of XΓ under the Cremona transformation and still have an ∨ isomorphism of XΓ and XΓ away from the algebraic simplex Σn . since the class in the Grothendieck ring is a universal Euler characteristic. the scheme of singularities of Σn . the use of the Cremona transformation method can lead to interesting results even in the case where the graphs are not necessarily ∨ planar. as one can see from the fact that the Euler characteristic is a ring homomorphism from K0 (V) to Z and that in (3. In particular. . where one uses the Cremona transformation to identify [XΓn ] = [Sn ] + [L (L ∩ Σn )]. More generally. in the case of the banana graphs. The latter has class [Sn ] = [Σn ] − nTn−2 . the dual graph Γ∨ is simply a polyn gon with n sides. One then checks that the complement in this hyperplane of the algebraic simplex has class [L (L ∩ Σn )] = [L] − [L ∩ Σn ] = Tn−1 − (−1)n−1 T+1 where T = [Gm ] = [A1 ] − [A0 ] is the class of the multiplicative group. In fact. Moreover. XΓ (XΓ ∩ Σn ) C ∨ XΓ ∨ (XΓ ∩ Σn ). based on the use of characteristic classes of singular varieties.69) L = [A1 ] has Euler characteristic χ(A1 ) = 1.Feynman integrals and algebraic varieties 77 that. This then gives [XΓn ] = [XΓn ∩ Σn ] + [XΓn (XΓn ∩ Σn )]. this calculation also yields as a consequence the value for the topological Euler characteristic of the (complex) variety XΓn (C). one ﬁnds that XΓn ∩ Σn = Sn .
although the individual hypersurfaces XΓ of Feynman graphs are not always mixed Tate motives. [van Suijlekom (2006)] and also [Kreimer and van Suijlekom (2009)] in the setting of Dyson–Schwinger equations. is always in Z[L]. This approach via families of graphs also ﬁts in well with the treatment of gauge theories and Slavonov–Taylor identities within the context of the Connes–Kreimer approach to renormalization via Hopf algebras.8 Motivic Feynman rules We have discussed above a deﬁnition of “abstract Feynman rules” based on the multiplicative properties (3. 3. individual Feynman graphs do not represent observable physical processes and only sums over graphs (usually with ﬁxed external edges and external momenta) can be physically meaningful. rather than over a ﬁxed number of vertices. the result of [Bloch (2008)] suggests that a more appropriate formulation of the conjecture on Feynman integrals and motives should perhaps be given directly in terms that involve the full expansion of perturbative quantum ﬁeld theory.32) that express the expectation value of Feynman integrals associated to arbitrary Feynman graphs of a . Although in the more physical setting it would be natural to sum over graphs with a given number of loops and with given external momenta. This result ﬁts in well with the general fact that. where one knows by the general result of [Belkale and Brosnan (2003a)] that nonmixedTate contributions will eventually appear. #Aut(Γ) where the sum is over all graphs with a ﬁxed number of vertices. This shows that there will be very interesting cancellations between the classes [XΓ ] at least for a suﬃciently large number of loops. if one sums the classes over graphs with a ﬁxed number of vertices one obtains a class that is always in the Tate part Z[L] of the Grothendieck ring. rather than in terms of individual graphs. as in [van Suijlekom (2007)].31) and (3. in quantum ﬁeld theory.78 Feynman Motives ∨ although in nonplanar cases XΓ is no longer identiﬁed with the graph hypersurface of a dual graph XΓ∨ . More precisely. with sums over graphs. it is shown in [Bloch (2008)] that the class SN = #V (Γ)=N [XΓ ] N! . By applying this method to the complete graph ΓN on N vertices. Bloch proved in [Bloch (2008)] a very interesting result showing that.
If Γ is a forest with n = #E(Γ).32) reducing connected graphs to 1PI graphs and inverse propagators. One can ﬁrst observe. 1 2 ˆ Lemma 3.2. then An XΓ = An . We will specify the ring more precisely below as a reﬁnement of the Grothendieck ring of varieties. For a connected graph Γ that is obtained by inserting 1PI graphs Γv at the vertices of a ﬁnite tree T .31).31) and (3. in the sense that it satisﬁes the multiplicative properties (3. (3. Also notice that if Γ1 and Γ2 are graphs joined at a single vertex. Lemma 3. with the same formal properties as the original Feynman integrals with respect to the combinatorics of graphs. One can similarly see that the second property of Feynman rules. As we discuss later on in the book.70) v v∈V (T ) ˆ Proof. This more abstract deﬁnition of Feynman rules allows one to make sense of algebrogeometric and motivic U (Γ). the hypersurface complement satisﬁes ˆ ˆ An XΓ = A#E(T ) × (A#Eint (Γv ) XΓ ).1.2.8. Thus. namely the ring of immersed conical varieties introduced in [Aluﬃ and Marcolli (2008b)].2 for the disjoint union still gives the desired result. More precisely.2.Feynman integrals and algebraic varieties 79 given scalar quantum ﬁeld theory ﬁrst in terms of connected graphs and then in terms of 1PI graphs. that the hyˆ persurface complement An XΓ where the parametric Feynman integral is computed behaves itself like a Feynman integral. (An1 XΓ ) × · · · × (An#V (T ) XΓ v1 v#V (T ) . when we think of the Cartesian ˆ ˆ product (An1 XΓ1 ) × (An2 XΓ2 ) as deﬁning the product operation in a suitable commutative ring.32). is also satisﬁed by the hypersurface complements.2 shows that the hypersurface complement An XΓ satisﬁes the multiplicative property (3. these two observations combine to give ˆ ˆ ) × A#E(T ) .2 above that for a disjoint union of graphs Γ = Γ1 ∪ Γ2 the hypersurface complement splits multiplicatively as ˆ ˆ ˆ An XΓ = (An1 XΓ ) × (An2 XΓ ). then the same argument given in Lemma 3. this will be closely related to deﬁning Feynman rules in terms of characters of the Connes–Kreimer Hopf algebra of Feynman graphs and to the abstract form of the Birkhoﬀ factorization that gives the renormalization procedure for Feynman integrals. in the case of a connected graphs Γ obtained by inserting 1PI graphs at vertices of a tree. the multiplicative property (3. we have seen in Lemma 3. as in [Aluﬃ and Marcolli (2008b)].
2.32) as an immediate consequence of Lemma 3. by identifying varieties not up to isomorphism. The relation between the classes in the Grothendieck ring of the aﬃne and projective hypersurface complements is given by the formula ˆ [An XΓ ] = (L − 1)[Pn−1 XΓ ].2. .1. Deﬁnition 3. By imposing equivalence under isomorphisms one falls back on the usual Grothendieck ring K0 (V).32). This then satisﬁes both (3. The ring F is then the receptacle of the universal algebrogeometric Feynman rule given by ˆ U(Γ) = [An XΓ ] ∈ F. and one imposes the inclusion–exclusion relation [X] = [X Y ] + [Y ] for closed subvarieties Y ⊂ X.80 Feynman Motives ˆ In the above. For example.8. Thus. but only up to linear changes of coordinates in an ambient aﬃne space.31) and (3. for a given commutative ring R. The ring of immersed conical varieties FK is generated by classes [V ] of equivalence under linear coordinate changes of varieties V ⊂ AN deﬁned by homogeneous ideals (hence the name “conical”) immersed in some arbitrarily large aﬃne space. as in [Aluﬃ and Marcolli (2008b)] in terms of a ring homomorphism I : F → R by setting ˆ U(Γ) := I([An ]) − I([XΓ ]) and by taking as value of the inverse propagator U(L) = I([A1 ]).71) This satisﬁes (3. one can consider the hypersurface complements An XΓ modulo certain equivalence relations. (3. one can deﬁne an Rvalued algebrogeometric Feynman rule.8. with the inclusionexclusion and product relations [V ∪ W ] = [V ] + [W ] − [V ∩ W ] [V ] · [W ] = [V × W ]. one can deﬁne a Feynman rule with values in the Grothendieck ring of varieties ˆ ˆ U(Γ) := [An XΓ ] = [An ] − [XΓ ] ∈ K0 (V).31) and (3. ˆ which expresses the fact that XΓ is the aﬃne cone over XΓ . One can also impose a weaker equivalence relation.2 and Lemma 3. This leads to the following reﬁnement of the Grothendieck ring of varieties ([Aluﬃ and Marcolli (2008b)]). if these varieties are considered up to isomorphism.
Feynman integrals and algebraic varieties
81
A Feynman rule deﬁned in this way is motivic if the homomorphism I : F → R factors through the Grothendieck ring K0 (VK ). In this case the inverse propagator is the Lefschetz motive (or the propagator is the Tate motive). The reason for working with FK instead is that it allowed us in [Aluﬃ and Marcolli (2008b)] to construct invariants of the graph hypersurfaces that behave like algebrogeometric Feynman rules and that measure to some extent how singular these varieties are, and which do not factor through the Grothendieck ring, since they contain speciﬁc information on how the ˆ XΓ are embedded in the ambient aﬃne space A#Eint (Γ) .
3.9
Characteristic classes and Feynman rules
In the case of compact smooth varieties, the Chern classes of the tangent bundle can be written as a class c(V ) = c(T V ) ∩ [V ] in homology whose degree of the zero dimensional component satisﬁes the Poincar´–Hopf theoe rem c(T V ) ∩ [V ] = χ(V ), which gives the topological Euler characteristic of the smooth variety. Chern classes for singular varieties that generalize this property were introduced independently, following two diﬀerent approaches, in [Schwartz (1965)] and [MacPherson (1974)]. The two deﬁnitions were later proved to be equivalent. The approach followed by Marie H´l`ne Schwartz generalized the deﬁee nition of Chern classes as the homology classes of the loci where a family of k + 1 vector ﬁelds become linearly dependent (for the lowest degree case one reads the Poincar´–Hopf theorem as saying that the Euler characterise tic measures where a single vector ﬁeld has zeros). In the case of singular varieties a generalization is obtained, provided that one assigns some radial conditions on the vector ﬁelds with respect to a stratiﬁcation with good properties. The approach followed by Robert MacPherson is instead based on functoriality. The functor F of constructible functions assigns to an algebraic variety V the Zmodule F(V ) spanned by the characteristic functions 1W of subvarieties W ⊂ V and to a proper morphism f : V → V the map f∗ : F(V ) → F(V ) deﬁned by f∗ (1W )(p) = χ(W ∩ f −1 (p)), with χ the Euler characteristic. A conjecture of Grothendieck–Deligne predicted the existence of a unique natural transformation c∗ between the functor F and the homol
82
Feynman Motives
ogy (or Chow group) functor, which in the smooth case agrees with c∗ (1V ) = c(T V ) ∩ [V ]. MacPherson constructed this natural transformation in terms of data of Mather classes and local Euler obstructions and deﬁned in this way what is usually referred to now as the Chern–Schwartz–MacPherson (CSM) characteristic classes of (singular) algebraic varieties cSM (X) = c∗ (1X ). It is convenient here to work inside an ambient space, for instance an ambient projective space, so that one can regard c∗ (1X ) as taking values in the homology (or Chow group) of the ambient PN , so that the characteristic class cSM (X) = c∗ (1X ) can be computed for X an arbitrary locally closed subset of PN , without requiring a compactness hypothesis. This is important in order to have an inclusionexclusion relation for these classes, as explained below. The results of [Aluﬃ (2006)] show that, in fact, it is possible to compute these classes without having to use Mather classes and Euler obstructions that are usually very diﬃcult to compute. Most notably, these characteristic classes satisfy an inclusionexclusion formula for a closed Y ⊂ X, cSM (X) = cSM (Y ) + cSM (X Y ), but are not invariant under isomorphism, hence they are naturally deﬁned on classes in FK but not on K0 (VK ). For a smooth locally closed X in some ambient projective space PN , the class cSM (X) can be obtained by a computation using Chern classes of sheaves of diﬀerentials with logarithmic poles in a resolution of the closure ¯ X of X. Then inclusionexclusion gives a way to compute any embedded cSM without recourse to MacPherson’s Euler obstructions or ChernMather classes. The CSM classes give good information on the singularities of a variety: for example, in the case of hypersurfaces with isolated singularities, they can be expressed in terms of Milnor numbers. To construct a Feynman rule out of these Chern classes, one uses the ˆ following procedure, [Aluﬃ and Marcolli (2008b)]. Given a variety X ⊂ AN , N one can view it as a locally closed locus in P , hence one can apply to its characteristic function 1X the natural transformation c∗ that gives an ˆ element in the Chow group A(PN ) or in the homology H∗ (PN ). This gives as a result a class of the form c∗ (1X ) = a0 [P0 ] + a1 [P1 ] + · · · + aN [PN ]. ˆ One then deﬁnes an associated polynomial given by GX (T ) := a0 + a1 T + · · · + aN T N . ˆ
Feynman integrals and algebraic varieties
83
ˆ It is in fact independent of N as it stops in degree equal to dim X. It is by construction invariant under linear changes of coordinates. It also satisﬁes an inclusionexclusion property coming from the fact that the classes cSM satisfy inclusionexclusion, namely GX∪Y (T ) = GX (T ) + GY (T ) − GX∩Y (T ) ˆ ˆ ˆ ˆ ˆ ˆ It is a more delicate result to show that it is multiplicative, namely that the following holds. Theorem 3.9.1. The polynomials GX (T ) satisfy ˆ GX×Y (T ) = GX (T ) · GY (T ). ˆ ˆ ˆ ˆ Proof. The proof of this fact is obtained in [Aluﬃ and Marcolli (2008b)] using an explicit formula for the CSM classes of joins in projective spaces, where the join J(X, Y ) ⊂ Pm+n−1 of two X ⊂ Pm−1 and Y ⊂ Pn−1 is deﬁned as the set of (sx1 : · · · : sxm : ty1 : · · · : tyn ), with (s : t) ∈ P1 , and is related to the product in aﬃne spaces by the property that the ˆ ˆ product X × Y of the aﬃne cones over X and Y is the aﬃne cone over J(X, Y ). One has c∗ (1J(X,Y ) ) = ((f (H) + H m )(g(H) + H n ) − H m+n ) ∩ [Pm+n−1 ], where c∗ (1X ) = H n f (H) ∩ [Pn+m−1 ] and c∗ (1Y ) = H m g(H) ∩ [Pn+m−1 ], as functions of the hyperplane class H in Pn+m−1 . We refer the reader to [Aluﬃ and Marcolli (2008b)] for more details. The resulting multiplicative property of the polynomials GX (T ) shows ˆ that one has a ring homomorphism ICSM : F → Z[T ] deﬁned by ˆ ICSM ([X]) = G ˆ (T )
X
and an associated Feynman rule ˆ UCSM (Γ) = CΓ (T ) = ICSM ([An ]) − ICSM ([XΓ ]). This is not motivic, i.e. it does not factor through the Grothendieck ring K0 (VK ), as can be seen by the example given in [Aluﬃ and Marcolli (2008b)] of two graphs (see the ﬁgure below) that have diﬀerent UCSM (Γ), CΓ1 (T ) = T (T + 1)2 CΓ2 (T ) = T (T 2 + T + 1) but the same hypersurface complement class in the Grothendieck ring, ˆ [An XΓ ] = [A3 ] − [A2 ] ∈ K0 (V).
i
84
Feynman Motives
There is an interesting positivity property that the CSM classes of the graph hypersurfaces XΓ appear to satisfy, namely the coeﬃcients of all the powers H k in the expression of the CMS class of XΓ as a polynomial in H are positive. This was observed in [Aluﬃ and Marcolli (2008a)] on the basis of numerical computations of these classes, done using the program of [Aluﬃ (2003)], for sample graphs. As the graphs that are accessible to computer calculations of the corresponding CSM classes are necessarily combinatorially simple (and all planar), it is at present not known whether the observed positivity phenomenon holds more generally. This property may be related to other positivity phenomena observed for CSM classes of similarly combinatorial objects such as Schubert varieties [Aluﬃ and Mihalcea (2006)]. We see in the last chapter that, in the case of quantum ﬁeld theories on noncommutative spacetimes and the corresponding modiﬁcation of the graph polynomials, positivity fails in nonplanar cases.
3.10
Deletioncontraction relation
We report here brieﬂy the recent results of [Aluﬃ and Marcolli (2009b)] on deletion–contraction relations for motivic Feynman rules. We have seen above that one can construct a polynomial invariant CΓ (T ) of graphs that is an algebrogeometric Feynman rule in the sense of [Aluﬃ and Marcolli (2008b)], which does not factor through the Grothendieck ring of varieties but satisﬁes an inclusion–exclusion principle. The polynomial CΓ (T ) is deﬁned in terms of CSM classes of the graph hypersurface complement. Several well known examples of polynomial invariants of graphs are examples of Tutte–Grothendieck invariants and are obtained as specializations of the Tutte polynomial of graphs, and they all have the property that they satisfy deletion–contraction relations, which make it possible to compute the invariant indictively from simpler graphs. These relations express the invariant of a given graph in terms of the invariants of the graphs obtained by deleting or contracting edges.
However. y) + TΓ/e (x. y) = xTΓ • If Γ has no edges then TΓ (x. and a “looping edge” if e starts and ends at the same vertex. Since such relations make it possible to control the invariant for more complicated graphs in terms of combinatorially simpler ones.12 below). Among them one has well known invariants such as the chromatic polynomial and the Jones polynomial. y). it was shown in [Aluﬃ and Marcolli (2009b)] that the polynomial CΓ (T ) is not a specialization of the Tutte polynomial. contains a term expressed as the class of an intersection of hypersurfaces which does not appear to be easily controllable in terms of combinatorial information alone.72) • If e ∈ E(Γ) is a looping edge then TΓ (x. a form of deletion–contraction relation does hold for the motivic Feynman ˆ rule [An XΓ ] ∈ K0 (V). y) . Here we call an edge e a “bridge” if the removal of e disconnects the graph Γ. y) = 1. y). y) = yTΓ/e (x. y] of ﬁnite graphs Γ is completely determined by the following properties: • If e ∈ E(Γ) is neither a looping edge nor a bridge the deletion– contraction relation holds: TΓ (x. We start by recalling the case of the Tutte polynomial of graphs and the form of the deletion–contraction relation in that well known classical case. y) = TΓ e (x. A ﬁrst observation of [Aluﬃ and Marcolli (2009b)] is that the Tutte polynomial can be regarded as a Feynman rule. such as replacing an edge by multiple parallel copies. unlike the relation satisﬁed by Tutte– Grothendieck invariants. In particular. the validity or failure of such relations can provide useful information that can help to identify where the motivic complexity of the graph hypersurfaces increases beyond the mixed Tate case (see §3. (3.Feynman integrals and algebraic varieties 85 The results of [Aluﬃ and Marcolli (2009b)] investigate to what extent motivic and algebrogeometric Feynman rules satisfy deletion–contraction relations. However. Tutte–Grothendieck invariants are specializations of the Tutte polynomial. • If e ∈ E(Γ) is a bridge then TΓ (x. this form of deletion– contraction relations suﬃces to obtain explicit recursive relation for certain operations on graphs. The Tutte polynomial TΓ ∈ C[x. The latter. in the sense that it has the e (x.
by assigning U (Γ) = TΓ (x. The Tutte polynomial invariant deﬁnes an abstract Feynman rule with values in the polynomial ring C[x.73) Proof. with Γv 1PI graphs inserted at the vertices of the tree T . y) = γ⊂Γ (x − 1)#V (Γ)−b0 (Γ)−(#V (γ)−b0 (γ)) (y − 1)#E(γ)−#V (γ)+b0 (γ) . y) = x#Eint (T ) TΓ ∪e∈Eint (T ) e (x. Then one obtains an abstract Feynman rule with values in R = C[x. y) = γ⊂Γ (x − 1)b0 (γ)−b0 (Γ) (y − 1)b1 (γ) . This is sometime referred to as the “sum over states” formula for the Tutte polynomial. since for Γ = Γ1 ∪Γ2 we can identify subgraphs γ ⊂ Γ with V (γ) = V (Γ) and E(γ) ⊂ E(Γ) with all possible pairs of subgraphs (γ1 . (3. y). γ2 ) with V (γi ) = V (Γi ) and E(γi ) ⊂ E(Γi ). y].1. with b0 (γ) = b0 (γ1 ) + b0 (γ2 ). we get TΓ (x. hence the property of the Tutte polynomial for the removal of bridges gives TΓ (x. The property for connected and 1PI graphs follows from the fact that. The multiplicative property under disjoint unions of graphs is clear from the closed expression (3.10.γ2 ) (x − 1)b0 (γ1 )+b0 (γ2 )−b0 (Γ) (y − 1)b1 (γ1 )+b1 (γ2 ) = TΓ1 (x. Thus. y). This can be written equivalently as TΓ (x. Proposition 3. the internal edges of the tree are all bridges in the resulting graph. and #E(γ) = #E(γ1 ) + #E(γ2 ).74). with inverse propagator U (L) = x. #V (Γ) = #V (Γ1 ) + #V (Γ2 ).86 Feynman Motives right multiplicative properties over disjoint unions and over decompositions of connected graphs in terms of trees and 1PI graphs. One can ﬁrst observe that the properties listed above for the Tutte polynomial determine the closed form TΓ (x. y) = γ=(γ1 . when writing a connected graph in the form Γ = ∪v∈V (T ) Γv .74) where the sum is over subgraphs γ ⊂ Γ with vertex set V (γ) = V (Γ) and edge set E(γ) ⊂ E(Γ). . y] of the form (3. y) TΓ2 (x. y). (3.73).
This implies that the universal algebrogeometric Feynman rule with values in the Grothendieck ring F of immersed conical varieties satisﬁes U(Γ) = ([A1 ] − 1)U(Γ/e) ˆ if e is a looping edge of Γ and U(Γ) = [An XΓ ] ∈ F. Proposition 3.5 and §2. One can then observe. in terms of graph hypersurfaces.76) then follows since the image of the class [A ] is the inverse propagator (T +1). Proof. For example. In the case where e ∈ E(Γ) is a looping edge.2. the relations for bridges and looping edges imply that one has to identify x(T ) = T + 1 and y(T ) = T . this is not compatible with the behavior of the invariant CΓ (T ) on more complicated graphs. However. (See §3.9 above and Proposition 2. for the triangle graph one has CΓ (T ) = T (T + 1)2 while the specialization TΓ (x(T ). if CΓ (T ) has to be a specialization of the Tutte polynomial. The property 1 (3.75) In fact.10. (3.2 of [Aluﬃ and Marcolli (2008b)]. the polynomial CΓ (T ) satisﬁes the relation CΓ (T ) = (T + 1)CΓ e (T ). The polynomial invariant CΓ (T ) is not a specialization of the Tutte polynomial.Feynman integrals and algebraic varieties 87 It follows that Tutte–Grothendieck invariants such as the chromatic and Jones polynomials can also be regarded as Feynman rules. (3.76) In fact. to taking a cone on the graph hypersurface and intersecting it with the hyperplane deﬁned by the coordinate of the looping edge. (T + 1) is the inverse propagator of the algebrogeometric Feynman rule U (Γ) = CΓ (T ) and the property of abstract Feynman rules for 1PI graphs connected by a bridge gives (3. y(T )) = (T + 1)2 + (T + 1) + T . . y(T )). we have CΓ (T ) = T CΓ/e (T ). if e ∈ E(Γ) is a bridge. however. First notice that.75). that the algebro–geometric Feynman rule ˆ CΓ (T ) = ICSM ([An XΓ ]) that we discussed in the previous section is not a specialization of the Tutte polynomial. We show that one cannot ﬁnd functions x = x(T ) and y = y(T ) such that CΓ (T ) = TΓ (x(T ).) This implies that. adding a looping edge to a graph corresponds.
77) also holds. . . . (1) If e ∈ E(Γ) is neither a bridge nor a looping edge then the following deletion–contraction relation holds: [An XΓ ] = L · [An−1 (XΓ e ∩ XΓ/e )] − [An−1 = L · [An−1 XΓ e ]. Let Γ be a graph with n > 1 edges.81) Proof. (3. tn ) = tn F (t1 . . (3.83) ˆ Given a projective hypersurface Y ⊂ P we denote by Y the aﬃne ¯ cone in AN .3. .80) (5) Under the same hypotheses. . . en−1 . . N −1 (3. An answer is given in [Aluﬃ and Marcolli (2009b)] for the motivic case. (3. . . (3.82) F := ∂tn The Kirchhoﬀ polynomial ΨΓ satisﬁes ΨΓ (t1 . . As above we consider the Kirchhoﬀ polynomial ΨΓ and the aﬃne and projective graph ˆ hypersurfaces XΓ ⊂ An and XΓ ⊂ Pn−1 . tn−1 ). .88 Feynman Motives One can then investigate what kind of deletion–contraction relations hold in the case of motivic and algebro–geometric Feynman rules.78) (3) If e is a looping edge in Γ. and we use the notation Y for the projective cone in PN . .77) (2) If the edge e is a bridge in Γ. for the Feynman rule U(Γ) = [An ˆ XΓ ] ∈ K0 (V). . Theorem 3. Then the polynomials for the deletion Γ e and contraction Γ/e of the edge e = en are both nontrivial and given by ∂ΨΓ = ΨΓ e and G := ΨΓ tn =0 = ΨΓ/e . the Euler characteristics satisfy χ(XΓ ) = n + χ(XΓ e ∩ XΓ/e ) − χ(XΓ e ). e = en . then [An XΓ ] = (L − 1) · [An−1 XΓ e] = (L − 1) · [An−1 XΓ/e ]. with (t1 : . : tn ) the corresponding variables in Pn−1 . tn−1 ) + G(t1 .10. then [An XΓ ] = L · [An−1 XΓ e] XΓ/e ]. We assume deg ΨΓ = > 0. . . . . then the projective version of (3.79) (4) If Γ contains at least two loops. that is. (3. . We consider an edge e that is not a bridge or looping edge. in the form [Pn−1 XΓ ] = L · [Pn−2 (XΓ e ∩ XΓ/e )] − [Pn−2 XΓ e ]. (1) and (4): Let Γ be a graph with n ≥ 2 edges e1 . (3.
if deg ψ = 1). then we also have [Pn−1 XΓ ] = L · [Pn−2 1 (XΓ e ∩ XΓ/e )] − [Pn−2 XΓ e ]. the projection restricts to a regular map XΓ (XΓ ∩ X Γ e) → Pn−2 . In fact. (3. assume deg F > 0. the ideal of XΓ ∩ X Γ e is (ψ. and hence p ∈ X (XΓ ∩ X Γ e ). G). (3. Thus. ΨΓ (p) = F (p) = 0.85) If Γ has at least two loops.86) where L = [A ] is the Lefschetz motive and XΓ/e is the hypersurface G = 0. (3. this is a polynomial of degree exactly 1 in t. Indeed. then XΓ e = X Γ e = ∅ and the statement is trivial. The line from p to q is parametrized by (q1 : · · · : qn−1 : t).84) where XΓ e is the hypersurface in Pn−2 deﬁned by the polynomial F . as needed. If F is constant (that is. Theorem 3. In this case. F ) = (tn F + G. F ) = (F. We show that this line meets XΓ (XΓ ∩ X Γ e ) transversely at one point. (3. The intersection with XΓ is determined by the equation tF (q1 : · · · : qn−1 ) + G(q1 : · · · : qn−1 ) = 0. Since F (q) = 0. This means that XΓ ∩ X Γ e = XΓ e ∩ X Γ/e . in the Grothendieck ring K0 (V) we have identities [Pn−1 XΓ ] = [Pn−1 (XΓ ∩ X Γ e )] − [Pn−2 XΓ e ].87) . and determines a reduced point. the projection Pn−1 Pn−2 from p = (0 : · · · : 0 : 1) acts as (t1 : · · · : tn ) → (t1 : · · · : tn−1 ). Therefore. The image is clearly contained in Pn−2 this map induces an isomorphism XΓ (XΓ ∩ X Γ e) ∼ XΓ and the statement is that XΓ . −→ Pn−2 e Let q = (q1 : · · · : qn−1 ). so that deg ΨΓ > 1.3. e. hence p ∈ XΓ ∩ X Γ e . in a more general context that includes the case under consideration. that the projection from the point (0 : · · · : 0 : 1) induces an isomorphism XΓ (XΓ ∩ X Γ e) −→ Pn−2 ∼ XΓ e. It then follows that.Feynman integrals and algebraic varieties 89 It is proved in [Aluﬃ and Marcolli (2009b)].
The ﬁbers of the projection Pn−1 (X Γ e ∩ X Γ/e ) → Pn−2 1 XΓ e with center p are then all isomorphic to A . then ΨΓ does not depend on the variable te and F ≡ 0. if deg XΓ > 1.84) the identity χ(Pn−1 (XΓ ∪ X Γ e )) = 0. (2) If e is a bridge. so that G ≡ 0. The equation for XΓ/e is obtained by dividing ΨΓ through by te . The equation for XΓ/e is the same. then deg F > 0. The equation for XΓ e is ΨΓ = 0 again. In this case. which have the property that the problem of describing the intersection XΓ e ∩ XΓ/e can be bypassed and the class of more complicated graphs can be computed inductively only in terms of combinatorial data. One such operation replaces a chosen edge e in a graph Γ with m parallel edges connecting the same two vertices ∂(e). .80). (3) If e is a looping edge.90 Feynman Motives If deg XΓ > 1. One has from (3.77) follows from the formula for the projective case. hence X Γ e = ∅. the classes in the Grothendieck ring satisfy [AN [AN XΓ ∩ X Γ (XΓ e e] = (L − 1) · [PN −1 (XΓ ∩ X Γ (XΓ e e )] ∩ XΓ/e )] = (L − 1) · [PN −1 ∩ XΓ/e )]. hence XΓ ∩ X Γ e and XΓ e ∩ XΓ/e contain the origin. but viewed in one fewer variables. the statement about the Euler characteristics follows from (3. while it can be checked directly for the remaning case with deg XΓ = 1. (5) Under the hypothesis that deg XΓ > 1. then ΨΓ is divisible by te . and one has XΓ e = XΓ/e .77) follows by observing that. Then (3. then F is not constant. This general result on deletioncontraction relations for motivic Feynman rules was applied in [Aluﬃ and Marcolli (2009b)] to analyze some operations on graphs. In the case where deg XΓ = 1 one has χ(XΓ ) = n − 1. It then follows that X Γ e ∩ X Γ/e contains the point p = (0 : · · · : 0 : 1). Another operation consists of replacing an edge with a chain of triangles (referred to as a “lemon graph”). The aﬃne version (3. and it follows that [Pn−1 (X Γ e ∩ X Γ/e )] = L · [Pn−2 (XΓ e ∩ XΓ/e )].
Feynman integrals and algebraic varieties
91
The latter operation can easily be generalized to chains of polygons as explained in [Aluﬃ and Marcolli (2009b)]. Assume that e is an edge of Γ, and denote by Γme the graph obtained from Γ by replacing e by m parallel edges. (Thus, Γ0e = Γ e, and Γe = Γ.) Theorem 3.10.4. Let e be an edge of a graph Γ. Let T = L − 1 = [Gm ] ∈ K0 (V). (1) If e is a looping edge, then U(Γme )
m≥0
sm = eTs U(Γ m!
e).
(3.88)
(2) If e is a bridge, then U(Γme )
m≥0
sm = m!
T·
eTs − e−s + s eTs + 1 U(Γ T+1
e).
(3.89)
(3) If e is not a bridge nor a looping edge, then U(Γme )
m≥0
eTs − e−s sm = U(Γ) m! T+1 + eTs + Te−s U(Γ e) T+1 eTs − e−s + s eTs − U(Γ/e). T+1 (3.90)
The result is proved in [Aluﬃ and Marcolli (2009b)] by ﬁrst deriving the relation in the case of doubling an edge and then repeating the construction. We do not report here the details of the proof, and we refer the readers to the paper for more details, but it is worth pointing out that the main step that makes it possible to obtain in this case a completely explicit formula in terms of U(Γ), U(Γ e) and U(Γ/e) is a cancellation that occurs in the calculation of the classes in the operation of doubling an edge. In fact, one ﬁrst expresses the class U(Γ2e ) as ˆ ˆ U(Γ2e ) = L · [An (XΓ ∩ XΓ )] − U(Γ),
o
where n = #E(Γ) and Γo denotes the graph obtained by attaching a looping edge named e to Γ/e. In the notation we used earlier in this section, the equation for Γo is given by the polynomial te G. Using inclusion–exclusion for classes in the Grothendieck ring, it is then shown in [Aluﬃ and Marcolli (2009b)] that one has ˆ ˆ ˆ ˆ ˆ [XΓ ∩ XΓ ] = [XΓ/e ] + (L − 1) · [XΓ e ∩ XΓ/e ].
o
92
Feynman Motives
Thus, one obtains U(Γ2e ) = (L − 2) · U(Γ) + (L − 1) · U(Γ e) + L · U(Γ/e).
This expression no longer contains classes of intersections of graph hypersurfaces, and can be iterated to obtain the general formula for U(Γme ). It is shown in [Aluﬃ and Marcolli (2009b)] that the recursive relation obtained in this way for the operation of multiplying edges is very similar in form to the recursive relation that the Tutte polynomial satisﬁes under the same operation. In fact, they both are solutions to a universal recursive relation with diﬀerent initial conditions. One also derives from the same universal recursion a conjectural form for the recursion relation, under the operation of multiplying edges Γ → Γme , for the polynomial ˆ invariant CΓ (T ) = ICSM ([An XΓ ]). The operation of replacing an edge with a chain of triangles, or “lemon graph” depicted in the ﬁgure produces, similarly, a recursive formula for the motivic Feynman rule U(ΓΛ ), where ΓΛ is the graph obtained from Γ m m by replacing an edge e by a lemon graph Λm .
e !
Proposition 3.10.5. Let e be an edge of a graph Γ, and assume that e is neither a bridge nor a looping edge. Let ΓΛ be the “lemonade graph” m obtained by building an mlemon fanning out from e. Then U(ΓΛ )sm = m
m≥0
1 1 − T(T + 1)s − T(T + 1)2 s2 e) + (T + 1)2 s U(Γ/e) .
· (1 − (T + 1)s) U(Γ) + (T + 1)Ts U(Γ
This result is proven in [Aluﬃ and Marcolli (2009b)] by ﬁrst obtaining a recursive formula for the classes U(Λm ) of the lemon graphs. This is a succession of operations where one ﬁrst doubles an edge and then splits the added edge by inserting a vertex. The ﬁrst operation is covered by the expression derived before for the class U(Γ2e ) while splitting an edge with a vertex corresponds to taking a cone over the corresponding graph
Feynman integrals and algebraic varieties
93
hypersurface. Thus, both operations are well understood at the level of classes in the Grothendieck ring and they give the recursive relation U(Λm+1 ) = T(T + 1)U(Λm ) + T(T + 1)2 U(Λm−1 ), for m ≥ 1, which gives then
m
U(Λm ) = (T + 1)m+1
i=0
m − i m−i T . i
It is interesting to observe that the sequence am = U(Λm ) is a divisibility sequence, in the sense that U(Λm−1 ) divides U(Λn−1 ) if m divides n. This is the property satisﬁed, for instance, by the sequence of Fibonacci numbers. (See [Aluﬃ and Marcolli (2009b)], Corollary 5.12 for more details.) Returning to the classes U(ΓΛ ), one then shows that these also satisfy m a relation of the form U(ΓΛ ) = fm (T)U(Γ) + gm (T)U(Γ m where the fm , gm , hm satisfy the with diﬀerent seeds given by f0 (T) = 1 , g (T) = 0 , 0 h (T) = 0 ,
0
e) + hm (T)U(Γ/e),
same recursion that gives U(Λm ), but f1 (T) = T2 − 1 g1 (T) = T(T + 1) h1 (T) = (T + 1)2 .
For more details we refer the reader to [Aluﬃ and Marcolli (2009b)]. 3.11 Feynman integrals and periods
Numerical calculations performed in [Broadhurst and Kreimer (1997)] gave very strong evidence for a relation between residues of Feynman integrals and periods of mixed Tate motives. In fact, the numerical evidence indicates that the values computed in the log divergent case are Qlinear combinations of multiple zeta values, which, as we recalled in the previous chapter, can be realized as periods of mixed Tate motives. Multiple zeta values are real numbers obtained by summing convergent series of the form 1 ζ(n1 , . . . , nr ) = (3.91) n n , k1 1 · · · kr r
0<k1 <k2 <···<kr
where the ni are positive integers with nr ≥ 2. There are ways to realize multiple zeta values as periods of mixed Tate motives; see [Goncharov
94
Feynman Motives
(2001)], [Goncharov and Manin (2004)], [Terasoma (2002)]. In [Goncharov and Manin (2004)], for instance, as well as in [Brown (2006)], multiple zeta values are explicitly realized as periods on the moduli spaces of curves. It is the occurrence of multiple zeta values in Feynman integral computations, along with the fact that these numbers arise as periods of mixed Tate motives, that gave rise to the idea of a deeper relation between perturbative quantum ﬁeld theory and motives. Among the examples of Feynman graphs computed in [Broadhurst and Kreimer (1997)], the wheel with n spokes graphs give simple zeta values ζ(2n − 3). It is also shown in [Broadhurst and Kreimer (1997)] that the nonplanar graph given by the complete bipartite graph K3,4 evaluates to the double zeta value ζ(5, 3). Recent results of [Brown (2009a)] for multiloop Feynman diagrams provide a new method for evaluating the case of primitive divergent graphs (i.e. those that do not contain any smaller divergent subgraph, and are therefore primitive elements in the Connes–Kreimer Hopf algebra). This shows the occurrence of values at roots of unity of multiple polylogarithms, in addition to the multiple zeta values observed in [Broadhurst and Kreimer (1997)]. The multiple polylogarithm function, introduced in [Goncharov (2001)] is deﬁned as Lin1 ,...,nr (x1 , . . . , xr ) =
0<k1 <k2 <···<kr
xk1 · · · xkr r 1 n n . k1 1 · · · kr r
(3.92)
It is absolutely convergent for xi  < 1 and it extends analytically to an open domain in Cr as a multivalued holomorphic function. As shown in [Goncharov (2001)], the multiple polylogarithms are also associated to objects in the category of mixed Tate motives. 3.12 The mixed Tate mystery
The results recalled in the previous section suggest that, in its strongest possible form, one could formulate a conjecture as follows. Conjecture 3.12.1. Are all residues of Feynman graphs of perturbative scalar ﬁeld theories (possibly after a suitable regularization and renormalization procedure is used) periods of mixed Tate motives? In this form the conjecture is somewhat vaguely formulated, but it does not matter since it is unlikely to hold true in such generality. One can
Feynman integrals and algebraic varieties 95 however reﬁne the question and formulate it more precisely in the following form. and the renormalization procedure are the residues of Feynman graphs periods of mixed Tate motives? As we have seen above. one is working with a relative cohomology.1. In fact.12. Question 3. it was later disproved in [Belkale and Brosnan (2003a)]. Kontsevich formulated the conjecture that the graph hypersurfaces XΓ themselves may always be mixed Tate motives. one is in fact considering only a certain relative cohomology group. However. only a part of the cohomology of the complement of the hypersurface XΓ is involved in the period computation of the Feynman integral.94) σn can be seen as a pairing of a diﬀerential form on the hypersurface complement with a chain σn with boundary ∂σn ⊂ Σn contained in the normal crossings divisor. and we have seen that the question of whether the integral above (i. ignoring divergences. one can write the Feynman integrals in a form that looks exactly like a period of an algebraic variety obtained as a hypersurface complement. In fact. As we have seen earlier in this chapter. the residue of . The fact that the classes [XΓ ] can be arbitrarily far from the mixed Tate part Z[L] of the Grothendieck ring K0 (V) seems to indicate at ﬁrst that Conjecture 3. the scalar ﬁeld theory. Under what conditions on the graphs. (3. modulo the issue of divergences which needs to be handled separately. Thus. the evaluation of the integral PΓ (p. namely H n−1 (Pn−1 X Γ .1 cannot hold. In the stable range for D. Although numerically this conjecture was at ﬁrst veriﬁed up to a large number of loops in [Stembridge (1998)].2. the XΓ generate the Grothendieck ring of varieties. that is not necessarily the case. or the union of the two.93) where n = #Eint (Γ) and Σn = {t ∈ Pn−1  i ti = 0} denotes the normal crossings divisor given by the union of the coordinate hyperplanes. this hypersurface is the graph variety XΓ . which would imply Conjecture 3. while in the unstable range it is the Landau variety YΓ .e. using the Feynman parameters.12. They proved that the varieties XΓ can be arbitrarily complicated as motives: indeed. Σn (Σn ∩ XΓ )).12. t)−n+D /2 ωn ΨΓ (t)−n+( +1)D/2 (3. if one ignores momentarily the issue of divergences.
so that the situation is further complicated. consisting of so called “zigzag graphs”. Thus. a weaker statement than the relative cohomology (3. if something like Conjecture 3.95).12. it is due to a very subtle interplay between the geometry of the graph hypersurface XΓ and the locus of intersection XΓ ∩ Σn . Kreimer (2006)]. at present. Esnault.95) This could still be the case even though the motive of XΓ itself need not be mixed Tate. However.2.95) is shown to give a mixed Tate motive and for which the corresponding Feynman integrals evaluate to multiple zeta values.93) is a realization of a mixed Tate motive m(Pn−1 XΓ . by possibly performing blowups of this locus. (3. the main problem remains that as the graphs increase in combinatorial complexity. when one tries to eliminate divergences.12. one should also take into account that the possible occurrence of nonmixed Tate periods in the Feynman amplitudes for more complicated graphs may cancel out in sums over . Notice that this locus is also the one that is responsible for the divergences of the integral (3. according to the formulation of Question 3.1 holds. This identiﬁes in a signiﬁcant class of examples a mixed Tate piece of the cohomology. identiﬁes speciﬁc families of graphs for which (3. Σn XΓ ∩ Σn ). this observation makes it clear that.12. Some new work of Francis Brown.94).95) is in fact mixed Tate. but not a necessary one. since the period only captures a piece of the relative cohomology (3. Very nice recent results of [Doryn (2008)] compute the middle cohomology of the graph hypersurfaces for a larger class of graphs than those originally considered in the work of [Bloch. which became available in the preprint [Brown (2009b)] while this book was going to print. In fact. it becomes extemely diﬃcult to control the motivic nature of (3. not only is the conjecture in its stronger form 3. but not much is understood on speciﬁc conditions that would ensure that (3. Even without introducing immediately the further diﬃculties related to eliminating divergences.1 still open. so that the question on the motivic nature of this relative cohomology provides a suﬃcient condition for a mixed Tate period. As we mentioned in relation to the recent result of [Bloch (2008)] on sums of graphs with ﬁxed number of vertices.93).93) being mixed Tate may still suﬃce to prove that the residue of the Feynman integral is a period of a mixed Tate motive.96 Feynman Motives the Feynman graph after dropping the divergent Gamma factor in the parametric Feynman integral) evaluates to a period of a mixed Tate motive can be addressed in terms of the question of whether the relative cohomology (3.
One should also keep in mind that. σ ).Feynman integrals and algebraic varieties 97 graphs with ﬁxed external structure or ﬁxed number of vertices and leave only a mixed Tate contribution. 2 ˆ D = {x = (xij ) ∈ A  det(xij ) = 0}. In the case of parametric Feynman integrals one can proceed in the following way. A period σ ω associated to the data (X. one can hope to obtain in this way some suﬃcient conditions that will ensure that the periods associated to Feynman integrals are mixed Tate.13 From graph hypersurfaces to determinant hypersurfaces An attempt to provide speciﬁc conditions.12. D. and an integration domain σ with boundary ∂σ ⊂ D can be computed equivalently after a change of variables obtained by mapping it via a morphism f of varieties to another set of data (X . As long as no information is lost in the period computation. As shown in [Aluﬃ and Marcolli (2009a)]. The matrix MΓ (t) associated to a Feynman graph Γ determines a linear map of aﬃne spaces Υ : An → A . D . 3. ω . for any 3edgeconnected graph with at least 3 vertices and no looping edges. which admit a closed 2cell embedding of face width at least 3. the map Υ . σ) of a variety X. a divisor D. that would ensure that certain Feynman integrals evaluate to periods of mixed Tate motives was developed in [Aluﬃ and Marcolli (2009a)].2 above. in the case of divergent Feynman integrals. whose motivic nature is easier to control. ω. a diﬀerential form ω on X. 2 Υ(t)kr = i ti ηik ηir (3. after subtraction of inﬁnities. as stated in Question 3. One can use the properties of periods. handling divergences in diﬀerent ways might aﬀect the nature of the resulting periods. in particular the change of variable formula.96) such that the aﬃne graph hypersurface is obtained as the preimage ˆ ˆ XΓ = Υ−1 (D ) under this map of the determinant hypersurface. One can give explicit combinatorial conditions on the graph that ensure that the map Υ is an embedding. with ω = f ∗ (ω ) and σ = f∗ (σ). to recast the computation of a given integral σ ω computing a period in a diﬀerent geometric ambient variety.
one can start with the following observation on the properties of the matrix MΓ (t) that deﬁnes the map Υ of (3. • The variable te appears in η † Λη if and only if e is part of at least one loop. • If there are two loops i . One can then give conditions based on properties of the graph that . where the tk correspond to the edges common to the ith and jth loop. as needed. . Since we are assuming τi to be injective.2. The matrix MΓ (t) = η † Λη deﬁning the map Υ has the following properties. added to the variables corresponding to the other edges forming the loop i .13. i) in the matrix η † Λη. then te only appears in the diagonal entry (i. • For i = j. Lemma 3. . j) and (j.13. Let (t1 . . . and the sign is +1 if the orientations of the edges both agree or both disagree with the loop orientations. . for a speciﬁc edge e. Lemma 3. the entry is the sum of the variables tk corresponding to the edges in the ith loop (all taken with sign +). and not having any other edge in common. To see why this is the case. the corresponding entry is the sum of ±tk . tn ) = (c1 . cn ) be in the kernel of τ . viewed as a map of the variables corresponding only to the edges that belong to the ith loop in the chosen bases of the ﬁrst homology of the graph Γ.1. i). Similarly. and −1 otherwise. j) entry in the target matrix is a combination of edges in the ith loop. Proof.96). we denote by Υi the composition of the map Υ with the projection to the ith row of the matrix η † Λη. that tuple is the zerotuple. one has the following. . • If e belongs to a single loop i . .98 Feynman Motives is injective. the map τi must send to zero the tuple of cj ’s corresponding to the edges in the ith loop. then the ±te appears by itself at the entries (i. • For i = j. with te the corresponding variable. . In the following. Since each (i. If Υi is injective for i ranging over a set of loops such that every edge of Γ is part of a loop in that set. j containing e. it follows that every cj is zero. Since every edge is in some loop for which τi is injective. then Υ is itself injective.
It is known that an embedding of a connected graph is minimal genus if and only if it is a 2cell embedding ([Mohar and Thomassen (2001)]. tiv are the variables corresponding to the edges of a loop i . . We discuss below conditions on the existence of closed 2cell embeddings. In this situation. . More explicitly. The map Υi is injective if the following conditions are satisﬁed: • For every edge e of the ith loop. Γ is planar if and only if it may be embedded in a sphere. up to rearranging the entries in the corresponding row of η † Λη and neglecting other entries. the advantage of having a closed 2cell embedding for a graph Γ is that the faces of such an embedding determine a choice of loops . . ±tiv ) if i has no edge not in common with any other loop. if and only if its genus is 0. and (ti1 . . . For our purposes. Proof. . and the possible last remaining variable appears summed together with the other variables.5. .13.13. [Mohar and Thomassen (2001)]. and • The ith loop has at most one edge not in common with any other loop. ±ti1 .4. . all but at most one edge variable appear by themselves as an entry of the ith row. . the map Υi is given by (ti1 . cf. if ti1 . Thus. §5.4. The minimum genus of an orientable surface in which Γ may be embedded is the genus of Γ. . An embedding of Γ in S is a closed 2cell embedding if the closure of every face is a disk. In either case the map Υi is injective. ±ti1 . . . ±tiv−1 ) if i has a single edge tv not in common with any other loop. tiv ) → (ti1 + · · · + tiv . Every (ﬁnite) graph Γ may be embedded in a compact orientable surface of ﬁnite genus. . Deﬁnition 3. .2. . . . .1 and Theorem 3. . .4). Lemma 3. or regions determined by the embedding). An embedding of a graph Γ in an orientable surface S is a 2cell embedding if the complement of Γ in S is homeomorphic to a union of open 2cells (the faces. tiv ) → (ti1 + · · · + tiv .Feynman integrals and algebraic varieties 99 ensure that the components Υi are injective. as claimed. This is done in [Aluﬃ and Marcolli (2009a)] as follows. there is another loop having only e in common with the ith loop.3. Proposition 3.
Any f −1 loops. The fact that the closure of each face is a 2disk guarantees that the boundary is nullhomotopic.13. where f is the number of faces of the embedding. so there will be no other relations.13. Z) determine a basis of H1 (Γ. Z) given by 2g + f − 1 loops. these f −1 loops.13. Indeed. generate H1 (Γ). deﬁned with respect to a choice of basis for H1 (Γ) as in Lemma 3.13. Lemma 3. however. and then add the edges on the boundary of each face with sign determined by the orientations.13. Thus. then Υ is injective. as one can see from the Euler characteristic formula b0 (S) − b1 (S) + b2 (S) = 2 − 2g = χ(S) = v − e + f = b0 (Γ) − b1 (Γ) + f = 1 − + f.1). together with 2g generators of the homology of S. One refers to the chosen one among the f faces as the “external face” and the remaining f − 1 faces as the “internal faces”. by taking the boundaries of the 2cells of the embedding together with a basis of generators for the homology of the Riemann surface in which the graph is embedded. given a closed 2cell embedding ι : Γ → S. A closed 2cell embedding ι : Γ → S of a connected graph Γ on a surface of (minimal) genus g. This produces a number of loops equal to the number f of faces. will be independent in H1 (Γ). Z) constructed as in Lemma 3. Then the f − 1 maps Υi . Assume that Γ is closed2cell embedded in a surface. We then have the following result. assume that • any two of the f faces have at most one edge in common. .5. up to sign. Lemma 3.5.6. With notation as above.5 to compute the map Υ and the maps Υi of (3. Proof. Orient (arbitrarily) the edges of Γ and the faces. If further • every edge of Γ is in the boundary of two of the f faces. The homology group H1 (Γ) has rank 2g + f − 1. It is clear that these f loops are not independent: the sum of any f − 1 of them must equal the remaining one. together with the choice of a face of the embedding and a basis for the homology H1 (S. are all injective.100 Feynman Motives of Γ. we can use a basis of H1 (Γ.
Then. Theorem 3. Let ι : Γ → S be a given embedding of a graph Γ on a Riemann surface S. Conjecture 5. Recall the notions of connectivity of graphs that we discussed in the ﬁrst chapter in relation to the 1PI condition. (1) Let Γ be a graph with at least 3 vertices and with no looping edges.5. the map Υ is injective.8.Feynman integrals and algebraic varieties 101 Proof. The face width f w(Γ.2. see Proposition 5. The injectivity of the f − 1 maps Υi follows from Lemma 3. of face width at least two (see [Mohar and Thomassen (2001)]. if any two of the faces have at most one edge in common. namely. For a graph Γ with at least 3 vertices and with no looping edges. When S is a sphere.13.7. hence ι : Γ → S is a planar embedding.3.3. ι) ≥ 2. this accounts for all but at most one of the edges in . so that every edge is accounted for. Since has at most one edge in common with the external region. the condition that an embedding ι : Γ → S is a closed 2cell embedding is equivalent to the properties that Γ is 2vertexconnected and that the embedding has face width f w(Γ. ι) = ∞. It is not known whether every 2vertexconnected graph Γ admits a closed 2cell embedding. By Lemma 3. At least one of them is internal. ι) is the largest number k ∈ N such that every noncontractible simple closed curve in S intersects Γ at least k times.5. If is a loop determined by an internal face. as shown in Lemma 3.11 of [Mohar and Thomassen (2001)].13. We also recall the notion of face width for an embedding of a graph in a Riemann surface. The stated condition guarantees that the edge appears in the loops corresponding to the faces separated by that edge. .16). the variables corresponding to edges in common between and any other internal loop will appear as (±) individual entries on the row corresponding to . which is closed2cell embedded in an orientable surface S. The “strong orientable embedding conjecture” states that this is the case. Deﬁnition 3. the map Υ is injective if every edge is in one of the f − 1 loops and the f − 1 maps Υi are injective.5. The following conditions imply the injectivity of Υ. that every 2vertexconnected graph Γ admits a closed 2cell embedding in some orientable surface S. one sets f w(Γ.13. the injectivity of Υi follows.13. Deﬁnition 1.1. Finally.13.
2edgeconnected is just the usual 1PI condition. Then that edge must bound the same face on both of its sides. with at least 3 vertices and no looping edges. F2 are (path)connected. it follows that removing the edge splits Γ into two connected components. Assume that two faces F1 . (1) It suﬃces to show that.13. However. (2) The previous result shows that the second condition stated in Lemma 3. Υ are all injective. Since the embedding has face width ≥ 3. Since F1 . Then the maps Υi . hence Γ is not 2edgeconnected.102 Feynman Motives (2) Let Γ be a 3edgeconnected graph. the fact that Γ has at least 3 vertices and no looping edges and it admits a closed 2cell embedding implies that Γ is 2vertexconnected. In fact. The closure of the face is a cell. Assume an edge is not in the boundary of two faces. while 3edgeconnected or 2PI is the next strengthening of this condition (the 2PI eﬀective action is often considered in quantum ﬁeld theory). Further.6 holds. That is.13. the second condition of Lemma 3. so that only the ﬁrst condition remains to be checked. ι) ≥ 3.6 is automatically satisﬁed. Therefore Γ is split into two connected components by removing the two edges. Since γ splits the cell into two connected components. it is shown there that the injectivity property of the map Υ extends from the classes of graphs described in the previous theorem to other graphs obtained from these by simple combinatorial operations such as attaching a looping edge at a vertex of a given graph or subdividing edges by adding intermediate valence two vertices. and the face width condition is also the next strengthening of face width 2. The combinatorial conditions of Theorem 3. F2 have more than one edge in common. there are paths γi in Fi connecting corresponding sides of the edges. With suitable care.6 is automatically satisﬁed. we show that every edge of Γ is in the boundary of two faces. and in particular it splits it into two connected components. it can be arranged that γ1 ∪ γ2 is a closed path γ meeting Γ in 2 points.8 are fairly natural from a physical viewpoint. hence Γ cannot be 3edgeconnected. admitting a closed2cell embedding ι : Γ → S with face width f w(Γ. which a well known combinatorial conjecture on graphs [Mohar .13.5. A more geometric description of these combinatorial properties in terms of wheel neighborhoods at the vertices of the graph is discussed in [Aluﬃ and Marcolli (2009a)]. and this gives a contradiction. Proof.5. so the only thing left to check is that the ﬁrst condition stated in Lemma 3. γ must be nullhomotopic in the surface. by assumption.13. under these conditions on the graph Γ. hence in particular it is 1PI by Lemma 1. Let γ be a path from one side of the edge to the other.
it is then easy to show (see [Aluﬃ and Marcolli (2009a)]) that adding looping edges does not aﬀect the injectivity of the map Υ. Υ(∂σn ) ⊂ ˆ ΣΓ .2 and 2. One can in fact rewrite the Feynman integral (as usual up to a divergent Γfactor) in the form U (Γ) = Υ(σn ) 2 PΓ (x. the diﬃculty has been moved from understanding the motivic nature of the hypersurface complement to .6. and with ωΓ (x) ˆ the image of the volume form. in this approach. In fact. det(x)−n+( +1)D/2 for a polynomial PΓ (x. One then sees that. as we showed in Lemma 1. as we have already seen in detail in Theorems 2.5. ΣΓ ˆ ˆ (ΣΓ ∩ D )) (3. The question on the motivic nature of the resulting period can then be reformulated (again modulo divergences) in this case as the question of whether the motive m(A 2 ˆ ˆ D . p). Let then ΣΓ be a normal crossings divisor in 2 A . it is well known that the determinant hypersurface D is a mixed Tate motive. When the map Υ is injective. Motivically.3 above.) The condition that the graph has no looping edges is only a technical device for the proof. We recall below how to crossings divisor in A such that Υ(∂σn ) ⊂ Σ ˆ construct the divisor ΣΓ . Notice also that if the map Υ is 2 injective then one has a well deﬁned map Pn−1 → P −1 . ΣΓ (D ∩ ΣΓ )). (The latter condition is a bit more than 1PI: for graphs with at least two vertices and no looping edges it is equivalent to all the splittings of the graph at vertices also being 1PI. p)−n+D /2 ωΓ (x) .97) is mixed Tate. The advantage of moving the period computation via the map Υ = ˆ ΥΓ from the hypersurface complement An XΓ to the complement of the 2 ˆ determinant hypersurface A D is that. unlike what happens with the ˆ graph hypersurfaces.5. when the map Υ : An → A is injective.6. p) on A that restricts to PΓ (t. it is possible to rephrase the computation of the parametric Feynman integral as a period of the complement of the 2 ˆ determinant hypersurface D ⊂ A . which contains the boundary of the domain of integration.Feynman integrals and algebraic varieties 103 and Thomassen (2001)] expects should simply follow for graphs that are 2vertexconnected. the complexity of Feynman integrals of the graph Γ as a period is 2 ˆ ˆ ˆ ˆ ˆ D . where ΣΓ is a normal controlled by the motive m(A 2 ˆ Γ . which is otherwise 2 not everywhere deﬁned.
More explicitly. There exists a normal crossings divisor Σ which is a union of N = f linear spaces. then knowing that A D is always mixed Tate. ˆ Proposition 3. ﬁrst of all. as a question on when the moˆ ˆ ˆ tive of ΣΓ (ΣΓ ∩ D ) is mixed Tate. This requires.97) whose realization is the relative cohomology would also be mixed Tate.g can be described by the N = f equations 2 xij = 0 xi1 + · · · + xi. On each face.f −1 = 0 1≤i<j ≤f −1 1 ≤ i ≤ f − 1. Thus. 2 ˆ ˆ ˆ ˆ If the motive of ΣΓ (ΣΓ ∩ D ) is mixed Tate.104 Feynman Motives having some control on the other term of the relative cohomology. the minimal genus of an orientable surface in which the graph can be embedded).2 can be reformulated.g . Question 3. the components of ˆ the divisor Σ . for all graphs Γ with loops and genus g closed 2cell embedding. with = 2g +f −1. 2 ˆ Σ . This divisor is given by a union of linear spaces.g for all graphs Γ with a ﬁxed number of loops and a ﬁxed genus (that is.g ⊂A .12.98) 2 := L1 ∪ · · · ∪ LN .99) where f = − 2g + 1 is the number of faces of the embedding of the graph Γ on a surface of genus g. A ﬁrst observation in [Aluﬃ and Marcolli (2009a)] is that one can use ˆ the same normal crossings divisor Σ . Thus. such that. Such an embedding has f faces. Proof. we can make the following convenient choice of these orientations. the fact that mixed Tate motives form a triangulated subcategory of the triangulated category of mixed motives would show that the motive (3.g is the algebraic simplex Σn in An . namely ˆ the normal crossings divisor ΣΓ and the way it intersects the determinant hypersurface.13. We have chosen a closed 2cell embedding of Γ into an orientable surface of genus g. under the combinatorial conditions given above for the embedding. (3. Then each edgevariable in common between . ˆ the preimage under Υ = ΥΓ of the union ΣΓ of a subset of components of ˆ Σ .9. We can arrange MΓ (t) so that the ﬁrst f −1 rows correspond to the f − 1 loops determined by the ‘internal’ faces of the embedding. The polynomial det MΓ (t) does not depend on the choice of orientation for the loops of Γ. a better ˆ description of the divisor ΣΓ that contains the boundary ∂(Υ(σn )). (3. we choose the positive orientation (counterclockwise with respect to an outgoing normal vector).
99) produce a list of all the edge variables. the intersection of a linear subspace with the determinant is not mixed Tate.10. with a positive sign. . Therefore. we have to understand in what sense the intersections ∩i∈I Li are special. and we can arrange so that the minor is the upperleft part of the × ambient matrix.100) will be mixed Tate. possibly with redundancies.100) (∩i∈I Li ) D is mixed Tate.13. The ˆ ˆ components of Σ .e. the intersection with a general P2 is a nonsingular cubic curve. N } determines a choice of linear subspaces V1 . coming from the zero entries of the matrix MΓ (t)). . It follows that the pullbacks of the equations (3. In fact.g that form the divisor ΣΓ are selected by eliminating ˆ . D3 is a degree3 hypersurface in P8 . We have in fact seen that the injectivity of an ( − 2g) × ( − 2g) minor of the matrix MΓ suﬃces to control the injectivity of the map Υ. v ) ∈ A  ∀i. The following characterization. These entries are both in the ( − 2g) × ( − 2g) upperleft minor. which is given in [Aluﬃ and Marcolli (2009a)]. with singularities in codimension > 1. . working projectively. j will appear with a minus sign in the entries (i.Feynman integrals and algebraic varieties 105 two faces i. . Thus. In fact. On the diagonal of the (f − 1) × (f − 1) submatrix we ﬁnd all edges making up each face. . Let E be a ﬁxed dimensional vector space. therefore a curve of genus = 1. vi ∈ Vi }. in order to understand under what conditions the locus (3. . i) of MΓ (t). using inclusionexclusion. . Every I ⊆ {1. . A second observation of [Aluﬃ and Marcolli (2009a)] is then that. 2 (3. Then the hyperplanes in An associated to the coordinates ti can be obtained by pulling back linear spaces along this minor. it suﬃces to show that arbitrary intersections of the ˆ components Li of Σ . Notice that the intersection ∩i∈I Li is a linear subspace of codimension 2 #I in A . The aﬃne version is a cone over this. V of E with the property that ∩k∈I Lk = {(v1 . . leads to a reformulation of the problem in terms of certain “manifolds of frames”. j) and (j. In general. . Lemma 3. this would then imply that also the locus ˆ ˆ ˆ ΣΓ (D ∩ ΣΓ ) is mixed Tate. For example.101) .g have the property that ˆ (3. by repeatedly using inclusion–exclusion. . the intersection of a general ˆ A3 with D3 is a cone over a genus1 curve. .g that contain the image of the graph hypersurface those components of Σ (i.
.g ∩ D )) (3. v ) for which exactly one row vi belongs to a ﬁxed hyperplane of E. . . . Further.13. i = 1. . Thus. e ) mentioned in the statement..102). for a ﬁxed . . the manifold of frames is mixed Tate for all choices of the subspaces Vi . shows that if. . . . . and more precisely to one of the hyperplanes x1 + · · · + x −2g = 0. .102) The statement follows by choosing Vi to be the intersection of the hyperplanes corresponding to the Lk in the ith row. x ) of the dual space to E. . . to obtain the basis (e1 .11. . . V ) is deﬁned as the locus 2 2 ˆ F(V1 . e ) of E such that each space Vi is the span of a subset (of cardinality ≥ i − 1) of the vectors ej .9) that the components Lk of Σ . . among those listed in (3. . together with the use of inclusion–exclusion arguments.g consist of matrices for which either the (i.104) . .. . .106 Feynman Motives 2 Here. . (3. x2 . Further still. x2 = 0. . The previous characterization of the locus (3. . . ˆ ˆ ˆ A suﬃcient condition that would ensure that the locus ΣΓ (ΣΓ ∩ D ) is mixed Tate is then described in terms of manifolds of frames. where each vi is constrained to belong to the given subspace Vi . simply choose the basis dual to the basis (x1 + · · · + x −2g . Since there are at most − 2g − i + 1 hyperplanes Lk in the ith row. . Σ . Recall (Proposition 3. . (3. there exists a basis (e1 . j) entry xij equals 0. the manifold of frames F(V1 . .13. Vr ) ⊂ V1 × · · · × Vr denotes the locus of rtuples of linearly independent vectors in a given vector space V . this suﬃces to guarantee that the motive 2 ˆ ˆ ˆ ˆ m(A D . each Lk consists of tuples (v1 .100). for 1 ≤ i < j ≤ − 2g. we denote an × matrix in A by its rowvectors vi ∈ E. −2g =0 for 1 ≤ i ≤ − 2g. . ˆ Proof. x −2g = 0. . . . F(V1 .g (Σ . v ) ∈ A  vk ∈ Vk } ∩ (A D ). V ) := {(v1 . . . dim Vi ≥ − ( − 2g − i + 1) = 2g + i − 1 ≥ i − 1 . dim Vi ≥ i − 1. .103) More generally. . Finally. or xi1 + · · · + xi. For a given ambient space V = A and an assigned collection of linear subspaces Vi . . . Deﬁnition 3. .
one can verify explicitly that the manifolds of frames F(V1 . are all periods of mixed Tate motives. a . one can show that the motive deﬁnes an object in the category of mixed Tate motives as a subtriangulated category of the category of mixed motives. V2 ) and F(V1 . In the case of two and three loops. We want to parameterize all pairs (v1 . as in the case of the determinant hypersurface discussed above. the residues of Feynman graphs Γ with loops and genus g. This exhausts all the possible ways of obtaining linearly independent vectors with the ﬁrst one in V1 and the second one in V2 and the manifold of frames F(V1 . The tautological line bundle O12 (−1) over P(V1 ∩V2 ) sits inside V12 . from which. For given subspaces V1 . V2 ) is a mixed Tate motive. v2 = 2. Consider the projective space P(V1 ∩ V2 ). V3 ) are mixed Tate. v2 ) of vectors such that v1 ∈ V1 . V2 )] = Ld1 +d2 − Ld1 − Ld2 − Ld12 +1 + Ld12 + L. The desired pairs of vectors (v1 . V2 . Proposition 3. being obtained by taking products and complements of aﬃne spaces. This locus can be decomposed into the two (possibly empty) pieces (1) v1 ∈ V1 (V1 ∩ V2 ). They are of the following form. and v2 ∈ V2 v1 . v2 ) are obtained by choosing a point p ∈ P(V1 ∩ V2 ). V2 ) is the union of these two loci. the result then implies that. Proof.Feynman integrals and algebraic varieties 107 is also mixed Tate. and the trivial bundles V12 ⊆ V2 with ﬁber V1 ∩ V2 ⊆ V2 . and dim v1 . Its class in the Grothendieck ring is of the form (Ld1 − Ld12 )(Ld2 − 1). The locus deﬁned by the second case can be described equivalently by the following procedure.13. hence inside V2 . The ﬁrst case describes the locus (V1 (V1 ∩ V2 )) × (V2 {0}) which is clearly mixed Tate. the manifold of frames F(V1 . v2 ∈ V2 . for which the injectivity condition on Υ holds. One can also compute explicitly the corresponding classes in the Grothendieck ring. with di = dim(Vi ) and dij = dim(Vi ∩ Vj ). V2 . as a function of the Lefschetz motive L. whose class in the Grothendieck ring is [F(V1 . When this is the case. modulo divergences. and v2 ∈ V2 {0}. (2) v1 ∈ (V1 ∩ V2 ) {0}. This is done in [Aluﬃ and Marcolli (2009a)] by exhibiting an explicit stratiﬁcation.12.
as in §6. dij = dim(Vi ∩ Vj ). V3 ) is a mixed Tate motive.2 of [Aluﬃ and Marcolli (2009a)].6. using the same two properties of the existence of distinguished triangles in the category of mixed motives associated to closed embeddings and homotopy invariance. the class of F(V1 . The case of three subspaces already requires a more delicate analysis of the contribution of the various strata. The following ﬁve loci deﬁne such a stratiﬁcation of V3 {0}: (1) S123 := (V1 ∩ V2 ∩ V3 ) {0}. This time one can proceed in the following way to obtain a stratiﬁcation.108 Feynman Motives vector v1 = 0 in the ﬁber of O12 (−1) over p. . Then the manifold of frames F(V1 . where di = dim(V1 ). we want to construct a stratiﬁcation of V3 such that the dimensions of the spaces π(V1 ). V3 )] = (Ld1 − 1)(Ld2 − 1)(Ld3 − 1) −(L − 1)((Ld1 − L)(Ld23 − 1) + (Ld2 − L)(Ld13 − 1) + (Ld3 − L)(Ld12 − 1) +(L − 1)2 (Ld1 +d2 +d3 −D − Ld123 +1 ) + (L − 1)3 . the class Fα := [F(π(V1 ). One can look for a stratiﬁcation {Sα } of V3 which is ﬁner than the one induced by the subspace arrangement V1 ∩ V3 . dijk = dim(Vi ∩ Vj ∩ Vk ). V2 ∩ V3 and such that. Proof. is then given by the expression (Ld12 − 1)(Ld2 − L). as in Theorems 2. The class in the Grothendieck ring. with class in the Grothendieck ring given by [F(V1 . V2 . and D = Dijk = dim(Vi + Vj + Vk ). π(V2 ))]. extends in fact to the locally trivial case of vector bundles or projective bundles. which we reproduce here below. Thus. Let V1 . V2 ) is the sum of these two classes. which is formulated for products. V2 .6.13. see [Voevodsky (2000)]. with π : V → V := V / v3 the projection. V3 be assigned subspaces of a vector space V . depends only on α and not on the chosen vector v3 ∈ Sα . V2 )] = (Ld1 +d2 − Ld1 − Ld2 +d12 + Ld12 ) + (Ld2 +d12 − Ld12 +1 − Ld2 + L) = Ld1 +d2 − Ld1 − Ld2 − Ld12 +1 + Ld12 + L. In other words. [F(V1 .2 and 2. This again deﬁnes a mixed Tate locus. and a vector v2 in the ﬁber of V2 O12 (−1) over p. The homotopy invariance property.13. Theorem 3. π(V2 ) and π(V1 ∩ V2 ) are constant along the strata.3. for v3 in Sα . in this case. V2 .
(3. For the table of dimensions obtained above the corresponding classes [Fα ] are as follows. S(12)3 := ((V1 + V2 ) ∩ V3 ) ((V1 ∪ V2 ) ∩ V3 ). [Sα ] d123 sα 2 1 1 0 0 S123 S13 S23 S(12)3 S3 L −1 Ld13 − Ld123 Ld23 − Ld123 d1 +d2 +d3 −D−d12 L − Ld13 − Ld23 + Ld123 d3 d1 +d2 +d3 −D−d12 L −L .Feynman integrals and algebraic varieties 109 (2) (3) (4) (5) S13 := (V1 ∩ V3 ) (V1 ∩ V2 ∩ V3 ). S3 := V3 ((V1 + V2 ) ∩ V3 ). V3 )] = α Lsα [Fα ][Sα ]. [Fα ] S123 S13 S23 S(12)3 S3 L −L − Ld2 −1 − Ld12 + Ld12 −1 + L Ld1 +d2 −1 − Ld1 −1 − Ld2 − Ld12 +1 + Ld12 + L Ld1 +d2 −1 − Ld1 − Ld2 −1 − Ld12 +1 + Ld12 + L Ld1 +d2 − Ld1 − Ld2 − Ld12 +2 + Ld12 +1 + L Ld1 +d2 − Ld1 − Ld2 − Ld12 +1 + Ld12 + L d1 +d2 −2 d1 −1 One then obtains the following values for the class of the stratum [Sα ] and the number sα . V2 . i = 1. This is also independent of the choice of v3 and only dependent on α by the properties of the stratiﬁcation. The dimensions are indeed constant along the strata and given by dim π(V1 ) S123 S13 S23 S(12)3 S3 d1 − 1 d1 − 1 d1 d1 d1 dim π(V2 ) d2 − 1 d2 d2 − 1 d2 d2 dim(π(V1 ) ∩ π(V2 )) d12 − 1 d12 d12 d12 + 1 d12 This information can then be translated into explicit classes [Fα ] in the Grothendieck ring associated to the strata Sα so that the class of the frame manifold is of the form [F(V1 .105) where the number sα is the number of subspaces Vi . S23 := (V2 ∩ V3 ) (V1 ∩ V2 ∩ V3 ). 2 containing the vector v3 ∈ Sα .
V2 . an argument similar to the one of Theorems 2. One can use these explicit computations of classes of manifolds of frames for two and three subspaces to obtain speciﬁc information about the motives 2 ˆ ˆ ˆ ˆ m(A D . . t5 t2 t1 1 3 t6 2 t3 t4 This graph has matrix MΓ (t) given by t 1 + t 2 + t5 −t1 −t2 . −t1 t 1 + t3 + t 4 −t3 −t2 −t3 t 2 + t 3 + t6 Labeling entries of the matrix as xij . Although the stratiﬁcation is used there only to compute the class in the Grothendieck group. .6. ΣΓ (ΣΓ ∩ D )) for speciﬁc Feynman graphs with up to three loops. . V3 )] also show that F(V1 . V3 ) deﬁne objects in the triangulated category of mixed Tate motives.3 can be used to show that the same stratiﬁcations used to compute [F(V1 . . V2 )] and [F(V1 . For example. t6 as pullbacks of the following: t1 = −x12 t2 = −x13 t = −x 3 23 t4 = x21 + x22 + x23 t5 = x11 + x12 + x13 t6 = x31 + x32 + x33 ˆ Thus.110 Feynman Motives Using the formula (3.105) then gives the stated result. V2 ) and F(V1 . we report here brieﬂy the case of the wheelwiththreespokes graph (the 1skeleton of a tetrahedron) described in [Aluﬃ and Marcolli (2009a)].2 and 2. V2 .6. . we can obtain t1 . we can consider as ΣΓ the normal crossings divisor deﬁned by the equation x12 x13 x23 (x11 + x12 + x13 )(x21 + x22 + x23 )(x31 + x32 + x33 ) = 0.
. V ) will continue to be mixed Tate for large and for arbitrarily complex subspace arrangements. given . ΣΓ ˆ ˆ (ΣΓ ∩ D3 )) is mixed Tate for the case of the wheel with three spokes. V3 )]. but the combinatorics of the possible subspace arrangements quickly becomes very diﬃcult to control. were computed explicitly in [Aluﬃ and Marcolli (2009a)]. V2 . . together with the fact that the hypersurface complement itself is a mixed Tate motive. which shows that the resulting class is then of the form ˆ [ΣΓ ˆ ˆ (D3 ∩ ΣΓ )] = L(6L4 − 3L3 + 2L2 + 2L − 1)(L − 1)3 . Vr ) with more than three subspaces. The situation may in fact be somewhat similar to that of “Murphy’s law in algebraic geometry” described in [Vakil (2006)]. and the distinguished triangle corresponding to the long exact cohomology sequence. Then this fact. . as explained in [Aluﬃ and Marcolli (2009a)]. By arguing as in Theorem 2. for each of the 64 possibilities. but we only give the ﬁnal result. one can improve the result from a statement about the class in the Grothendieck ring being a polynomial function of the Lefschetz ˆ ˆ ˆ motive L to one about the motive m(ΣΓ (D3 ∩ ΣΓ )) itself being an object in the triangulated subcategory DMTQ of mixed Tate motives inside the triangulated category DMQ of mixed motives. We do not reproduce them all here. The result for other graphs with three loops can be derived from the same analysis used for the wheel with three spokes.Feynman integrals and algebraic varieties 111 In order to compute the class in the Grothendieck ring of the intersection ˆ ˆ ΣΓ ∩ (A9 D3 ). where a suﬃciently general case may be as bad as possible. suﬃce to show that the mixed motive m(A9 ˆ ˆ D3 . hence there are 2 = 64 such intersections. .6. Each of them determines a choice of three subspaces V1 . . in general it seems unlikely that the frame manifolds F(V1 . . . by restricting only to certain components of the divisor. For the cases of manifolds of frames F(V1 . and using the same stratiﬁcation of the ˆ ˆ complement A9 D3 induced by the divisor ΣΓ as in the computation of the class above. . it would seem at ﬁrst that one should be able to establish an inductive argument that would take care of the cases of more subspaces. In fact. V3 corresponding to the linearly independent vectors given by the rows of the matrix (xij ) in A9 .2. The connection can probably be made more precise. V2 . This 6 divisor has 6 components. All of the corresponding classes [F(V1 . one can use inclusionexclusion and compute the classes ˆ for all the intersections of subsets of components of the divisor ΣΓ . while speciﬁc cases one can explicitly construct are much better behaved.
the source of other possible divergences in the parametric Feynman integral is the locus of intersection of the graph hypersurface XΓ with the domain of integration σn . In terms of the geometry. even when in the parametric form one removes the divergent Gamma factor and only looks at the residue. In other words. or (to . A reformulation of the problem of understanding when frame manifolds are mixed Tate is given in [Aluﬃ and Marcolli (2009a)] in terms of intersections of unions of Schubert cells in ﬂag varieties. while the result of [Belkale and Brosnan (2003a)] is nonconstructive. Thus.14 Handling divergences All the considerations above on the motivic nature of the relative cohomology involved in the computation of the parametric Feynman integral. the issue of divergences is not dealt with in this approach. in the sense that it shows that the graph hypersurfaces generate the Grothendieck ring of motives but it does not provide an explicit construction of matroids on which to test the mixed Tate property. by a regularization procedure. Notice that the poles of the integrand that fall inside the integration domain σn occur necessarily along the boundary ∂σn . Another observation from the point of view e of matroids is that. one knows very well that most Feynman integrals are divergent. since in the interior the graph polynomial ΨΓ takes only strictly positive real values.112 Feynman Motives that both the result of [Belkale and Brosnan (2003a)] on the general motivic properties of the graph hypersurfaces XΓ and Murphy’s law result of [Vakil (2006)] are based on the same universality result for matroid representations of [Mn¨v (1988)]. This version of the problem suggests a possible connection to Kazhdan–Lusztig theory [Kazhdan and Lusztig (1980)]. the intersections XΓ ∩ ∂σn . by constructing explicit matroids. assume that the integral is convergent and therefore directly deﬁnes a period. However. one needs to modify the integrals suitably in such a way as to eliminate. after removing the divergent Gamma factor. it 2 ˆ ˆ may be that the loci ΣΓ ∩(A D ) considered here may provide a possible way to make that general result more explicit. 3. along the lines of [Gelfand and Serganova (1987)]. and that some regularization and renormalization procedure is needed to handle these divergences. either in the hypersurface complement Pn−1 XΓ or in the complement of 2 the determinant hypersurface P −1 D .
in the context of log canonical thresholds. We mention here three possible approaches. [Bogner and Weinzierl (2007b)]. The result was more recently extended algorithmically to the nonlogdivergent case [Bogner and Weinzierl (2007a)].g . which is a sum of Tate twisted copies of the motive of the base. Kreimer (2006)]. but when the locus is singular. They prove that this Lfunction has a Laurent series expansion where all the coeﬃcients are periods.104) is mixed Tate. Yet another method was proposed in [Marcolli (2008)]. Another method. rewrite the Feynman integral in the form of a local Igusa Lfunction I(s) = σ f (t)s ω. Similarly. We discuss the third one in more detail in the following chapter. since one has an explicit formula for the motive of a projective bundle in the Voevodsky category. when the polynomial PΓ (t. To make sure that this operation does not change the motivic properties of the ˆ ˆ resulting period integrals. p) is not present and only the denominator ΨΓ (t)D/2 appears in the parametric Feynman integral. in this case. for f = ΨΓ : An → A1 . the issue of eliminating divergences becomes similar to the techniques used.g ∩ D maintain the mixed Tate properties if we know that (3. used in [Bloch. . one needs to ensure that blowups along Σ . as is typically the case here. then the required analysis is more delicate. One method was developed in [Belkale and Brosnan (2003b)] in the logarithmically divergent case where n = D /2. for f = ΨΓ . For blowups performed over a smooth locus one has projective bundles. for instance. working in the setting of the determinant hypersurface 2 ˆ complement A D one can perform blowups to separate the locus of ˆ ˆ intersection of the determinant hypersurface D with the divisor Σ . that is. Esnault. one can form a tubular neighborhood D (X) = ∪s∈∆∗ Xs . There are diﬀerent possible ways to achieve such a regularization procedure.Feynman integrals and algebraic varieties 113 work in algebrogeometric terms) the intersections XΓ ∩ Σn which contains the former. using dimensional regularization. consists of eliminating the divergences by separating Σn and XΓ performing a series of blowups. As we have already mentioned. for which one knows that the mixed Tate property is maintained. based on deformations instead of resolutions. In this setting. By considering the graph hypersurface XΓ as the special ﬁber X0 of a family Xs of varieties deﬁned by the level sets f −1 (s). one can.
From this viewpoint. One then uses a procedure of extraction of ﬁnite values to eliminate the polar parts of these Laurent series in a way that is consistent over graphs. as well as the theory of limiting mixed Hodge structures for a degeneration of a family of algebraic varieties. The use of deformations X to regularize the integral. where one works with Milnor ﬁbers of singularites. while it is precisely because of the fact that they are highly singular that the XΓ continue to be mixed Tate for a fairly large number of loops and even though one knows that eventually one will run into graphs for which XΓ is no longer mixed Tate. they would typically not be mixed Tate even for a small number of loops. and a circle bundle π : ∂D (X) → X . It is not yet clear how to use this type of methods to obtain motivic information on Feynman integrals after subtraction of divergences. One can then regularize the Feynman integral by integrating “around the singularities” in the ﬁber π −1 (σ ∩ X ). In fact. the fact that they continue to be singular in low codimension stil makes it possible to have a signiﬁcant part of the cohomology that will still be a realization of a mixed Tate motive. which we describe more in detail in the following chapter. or the deformation parameter in the example here above) in which the resulting function has a Laurent series expansion around the pole that corresponds to the divergent integral originally considered. The regularized integral has a Laurent series expansion in the parameter . . In fact. one should take into account the fact that the presence of singularities in the hypersurface XΓ increases the likelihood that the part of the cohomology involved in the period computation can be mixed Tate. that is. if these were smooth hypersurfaces. the generic ﬁber in a family used as deformation can be a general hypersurface that will not be motivically mixed Tate any longer. while the special singular ﬁber may be mixed Tate. However. a regularization procedure for Feynman integrals replaces a divergent integral with a function of some regularization parameters (such as the complexiﬁed dimension of DimReg. Motivically. but many techniques of singularity theory are designed to cover also this more general case.2. The singularities of the hypersurfaces XΓ are nonisolated.114 Feynman Motives for ∆∗ a punctured disk of radius . is natural from the point of view of singularity theory. a renormalization procedure. In general. as we discuss more at length in Chapter 4. it is not clear though how to perform the regularization and subtraction of divergences in a way that does not alter the motivic properties. there are notions such as a motivic tubular neighborhood [Levine (2005)] that may be useful in this context.
This was done in [Kapranov (2000)]. Later. suggested that the zeta function itself may be lifted at the motivic level. where the motivic zeta function is deﬁned as ZX (t) = n≥0 [sn (X)] tn (3.106) Notice that this can be written equivalently in terms of symmetric products as ZX (t) = n≥0 #sn (X)(Fq ) tn . . if a variety. is motivically mixed Tate. factors through the Grothendieck ring. where sn (X) denotes the nth symmetric power of X. Kapranov proved that.15 Motivic zeta functions and motivic Feynman rules As we mentioned brieﬂy in describing the Grothendieck ring as a universal Euler characteristic. as such. it was proved in [Larsen and Lunts (2003)] that in general this is not true in the case of algebraic surfaces.107) where [sn (X)] are the classes in the Grothendieck ring and the zeta function can be seen as an element of K0 (VK )[[t]]. for a ﬁnite ﬁeld K = Fq . the zeta function ZX.µ (T ) ∈ R[[T ]] is a rational function of T . the counting of points of a variety over ﬁnite ﬁelds is an example of an additive invariant that.R (t) = n≥0 µ(sn (X)) tn ∈ R[[t]]. hence it descends from the Grothendieck ring. One obtains other kinds of zeta functions by applying other “motivic measures” µ : K0 (VK ) → R. then the zeta function is a rational function. One recovers the zeta function of the variety. For example. say deﬁned over Q. by applying a “counting of points” homomorphism K0 (VK ) → Z to the motivic zeta function. The behavior of the number of points over ﬁnite ﬁelds is one of the properties of a variety that reveal its motivic nature. which give ZX. The fact that the counting of the number of points behaves like an Euler characteristic.Feynman integrals and algebraic varieties 115 3. given a motivic measure µ : K0 (M) → R. For example. in the sense that. when X is the motive of a curve. The information on the number of points over ﬁnite ﬁelds is conveniently packaged in the form of a zeta function ZX (t) = exp n #X(Fqn ) n t n (3. one has ZP1 (t) = (1 − t)−1 (1 − Lt)−1 . then the number of points of the reductions mod p is polynomial in p.
for example. via the map ΨΓ . When one considers motivic Feynman rules. instead of looking only at the contribution of individual Feynman graphs. Γ for s a complex variable. For instance. the number of vertices minus the number of connected components. In quantum ﬁeld theory it is customary to consider the full partition function of the theory. other suitable gradings δ(Γ) are given by the number n = #Eint (Γ) of internal edges. [Larsen and Lunts (2003)]. regarded as a zeta function ZΓ (T ) := n≥0 logn ΨΓ n T = ΨT .77. arranged in an asymptotic series by loop number. where it is also observed e that the motivic zeta function should play a useful role in the case of motives associated to Feynman graphs. As shown in [Connes and Marcolli (2008)] p. where the motive LogΓ is the pullback of the logarithmic motive (2.116 Feynman Motives It is well known that the motivic zeta function contains all the information on the behavior of the number of points of reductions mod p. Another reason is the formulation of parametric Feynman integrals as local Igusa Lfunctions given in [Belkale and Brosnan (2003b)]. or by #Eint (Γ) − b1 (Γ) = #V (Γ) − b0 (Γ).Γ (T ) := n≥0 sn (LogΓ ) T n . Γ n! may be related to a motivic zeta function ZLog. the observation made at the end of [Marcolli (2008)] which suggests that the function Ψs . indeed. that appears in the dimensional regularization of parametric Feynman integrals. There is a motivic Igusa Lfunction constructed by [Denef and Loeser (1998)]. these all deﬁne gradings on the Connes–Kreimer Hopf algebra of Feynman graphs. these partition functions appear to be interesting analogs of the motivic zeta functions considered in [Kapranov (2000)]. Kapranov’s motivic zeta functions may be especially suitable tools to investigate motivic properties of Feynman integrals.39). There are other reasons why one can expect that. One is. one can . A recent survey of these ideas is given in [Andr´ (2009)]. hence potentially the information on the mixed Tate nature of a variety. Besides the loop number = b1 (Γ). which may provide the right tool for a motivic formulation of the dimensionally regularized parametric Feynman integrals.
108) where δ(Γ) is any one of the gradings described above and where U (Γ) = ˆ [An XΓ ] ∈ K0 (VK ). #Aut(Γ) (3. it is possible that interesting cancellations of a similar nature may occur in suitable evaluations of the zeta functions (3.108). #Aut(Γ) In the case of the sums over graphs SN = #V (Γ)=N [XΓ ] N! #Aut(Γ) considered in [Bloch (2008)].Feynman integrals and algebraic varieties 117 consider a partition function given by the formal series Z(t) = N ≥0 δ(Γ)=N U (Γ) tN . . one ﬁnds that cancellations between the classes occur in the Grothendieck ring K0 (VK ) and the resulting class SN is always in the Tate subring Z[L] even if the individual terms [XΓ ] may be nonmixed Tate. this gives a zeta function with values in R[[t]] of the form ZR (t) = N ≥0 δ(Γ)=N µ(U (Γ)) N t . Given a motivic measure µ : K0 (VK ) → R. Similarly.
Lyashko. GuseinZade. see [Arnold. on the similarities between the parametric form of the Feynman integrals and the oscillatory integrals used in singularity theory to describe integrals of holomorphic forms over vanishing cycles of a singularity and to relate these to mixed Hodge structures.u (φ) αu (log α)k . (4. Varchenko (1988)]. based on the use of Leray cocycles. It is well known that. following [Marcolli (2008)].2) 119 . as mentioned in §3. 4.1 Oscillatory integrals An oscillatory integral is an expression of the form I(α) = Rn eiαf (x) φ(x) dx1 · · · dxn . (4. then (4.Chapter 4 Feynman integrals and Gelfand–Leray forms We focus in this chapter. we show that dimensionally regularized parametric Feynman integrals can be related to the Mellin transform of a Gelfand–Leray form.II of [Arnold.1) where f : Rn → R and φ : Rn → R are smooth functions and α ∈ R∗ is a + real parameter. whose Fourier transform is an oscillatory integral of the sort considered in the context of singularity theory.1) has an asymptotic development for α → ∞ given by a series n−1 I(α) ∼ eiαf (x0 ) u k=0 ak. aimed at eliminating the divergences coming from the intersections XΓ ∩σn . Goryunov. We show that one can deﬁne in this way a regularization procedure for the parametric Feynman integrals. Vasilev (1998)] and Vol. In particular.14 above. if the phase f (x) is an analytic function in a neighborhood of a critical point x0 .
7) The oscillatory integral. is then written as I(α) = R eiαt Xt (R) φ(x)ωf (x. expressed in terms of Gelfand–Leray forms. hence one writes it with the notation ωf (x. deﬁned as follows. (4. It is also well known that the integral (4.1. (4. t) to Xt is uniquely deﬁned. satisfying α df ∧ = α. (4.4) df As above. §2. More generally.6. which we denote by α/df .3) Notice that there is an ambiguity given by diﬀerent possible choices of an (n − 1)form satisfying (4. Then the Gelfand–Leray form ωf (x. t). the restriction of the form (α/df )(x.1. df f− The Gelfand–Leray function is the associated function J(t) := Xt (4. t) above is the Gelfand–Leray form of the volume form α = dx1 ∧ · · · ∧ dxn . Vol.8) . t) = dx1 ∧ · · · ∧ dxn .1) can be reformulated in terms of onedimensional integrals using the Gelfand–Leray forms. The form ωf (x.120 Feynman Motives where u runs over a ﬁnite set of arithmetic progressions of negative rational numbers depending only on the phase f (x). Let Xt be the level hypersurfaces Xt = {x  f (x) = t}. t) dt. and the ak.6) φ(x)ωf (x.1.II of [Arnold. we say that α admits a Gelfand–Leray form with respect to f if there is a diﬀerential form.3). df (4.5) The Gelfand–Leray form is given by the Poincar´ residue e α α = Res =0 . (4. GuseinZade. given a form α. Deﬁnition 4. t) of f is the unique (n − 1)form on the level hypersurface Xt with the property that df ∧ ωf (x. cf. Varchenko (1988)]. t) = dx1 ∧ · · · ∧ dxn . but their restrictions to Xt all agree so that the Gelfand–Leray form on Xt is well deﬁned.u are distributions supported on the critical points of the phase.
see §2.Feynman integrals and Gelfand–Leray forms 121 where Xt (R) ⊂ Rn is the level set Xt (R) = {x ∈ Rn : f (x) = t} and ωf (x. The Leray coboundary L(σ) of a kchain σ in X is a (k + 1)chain in D(f ) obtained by considering a tubular neighborhood of X in An . The Leray coboundary L(σ) is a cycle if σ is a cycle. with deﬁning polynomial equation f = 0 (where f will be the graph polynomial ΨΓ ) and the hypersurface complement D(f ) ⊂ An . df ∧ df Its diﬀerential gives α α dα = d df ∧ = −df ∧ d .6 and §2. in the following way. For more details.2. Let σ be a kchain in X = {t ∈ An f (t) = } and let L(σ ) be its Leray coboundary in D(f − ).2.2 Leray regularization of Feynman integrals We consider here the case of complex hypersurfaces X ⊂ An = Cn . Deﬁnition 4. we can assume here that the integration is over the values t ∈ R such that the level set Xt is a smooth hypersurface. one has 1 α df ∧ = α. One considers the preimage of σ under the projection map as a chain in D(f ). df (4.1.2. Then. and if one changes σ by a boundary then L(σ) also changes by a boundary. for a form α ∈ Ωk that admits a Gelfand–Leray form.II [Arnold.10) Proof. the boundary of its tubular neighborhood is a circle bundle over X. df df . Varchenko (1988)]. t) is the Gelfand–Leray form of f . 4. If X is a smooth hypersurface. Notice that.9) 2πi L(σ( )) f− σ( ) where d d α= σ( ) σ( ) dα − df ∂σ( ) α .7. Vol. First let us show that if α has a Gelfand–Leray form then dα also does. up to throwing away a set of measure zero. (4. We have a form α/df such that α = α. Lemma 4. such that the restriction of f to the interior of the domain of integration σn ⊂ An takes strictly positive real values. GuseinZade.
so one gets (4. The last term can be made arbitrarily small. To obtain (4. .10) notice that 1 d α 1 α df ∧ = df ∧ . Then we proceed to prove the ﬁrst statement.9). the form dα α = −d df df is a Gelfand–Leray form for dα. s− ∈ C.9) d dα α α= − . One can write 1 2πi df ∧ L(σ( )) α f− = 1 2πi α γ σ(s) ds . 2πi d L(σ( )) f− 2πi L(σ( )) (f − )2 One then uses d α f− = α dα − f− (f − )2 α f− α f− α df to rewrite the above as 1 2πi L(σ( = 1 2πi )) dα − f− df ∧ dα df d L(σ( )) L(σ( )) f− dα df − 1 2πi L(∂σ( )) 1 = 2πi df ∧ L(σ( )) f− 1 − 2πi df ∧ L(∂σ( )) f− .122 Feynman Motives Thus. df ∧ df and α/df is a Gelfand–Leray form with the property that α df ∧ = α. df This then gives by (4. where dα/df is a Gelfand–Leray form such that dα = dα. d σ( ) σ( ) df ∂σ( ) df This completes the proof. This can α σ( ) where γ ∼ S 1 is the boundary of a small disk centered at = then be written as ds ds 1 1 1 α + α − = 2πi γ σ( ) s − 2πi γ σ(s) s − 2πi γ ds s− .
11) ∗ ∗ where Xs = {tf (t) = s} and ∆ ⊂ C is a small punctured disk of radius > 0. The Leray regularized Feynman integral is obtained from (3.4. We consider forms ∆(ω) . possibly with boundary.14 above. and ω = PΓ (t. As mentioned brieﬂy in §3.56) by replacing the part π ∗ (ηm ) (4. one considers a neighborhood D (X) of a hypersurface X in a projective space Pn−1 .2.Feynman integrals and Gelfand–Leray forms 123 The formulation given in Proposition 3. similarly to what happens in the case of DimReg. (4.14) fm as in Proposition 3. We know that the divergences of the parametric Feynman integral appear at the intersections XΓ ∩ σn .12) It is a circle bundle over the generic ﬁber X . with projection π : ∂D (X) → X . (4. π ∗ (ηm ) = Deﬁnition 4.3.3. The boundary ∂D (X) is given by ∂D (X) = ∪s∈∂∆∗ Xs . This contains the locus σ∩X which is of interest to us in terms of divergences in the Feynman integral.13) This satisﬁes L (∂σ) = ∂L (σ). in the following way. .15) π ∗ (ηm ) + df ∧ f ∂σn ∩D (XΓ ) σn ∩D (XΓ ) of (3. (4. p)−n+D /2 ωn .4. Given a domain of integration σ. written in the form of Proposition 3. where X = {f = 0} with order of vanishing m = −n + D( + 1)/2.16) f− f− L (∂σn ) L (σn ) where f = ΨΓ and m = −n + D( + 1)/2. given by the level sets D (X) = ∪s∈∆∗ Xs . we consider the intersection σ∩D (X). (4.3. so we can concentrate only on the part of the integral that is supported near this locus.3 of the parametric Feynman integrals suggests a regularization procedure based on Leray coboundaries. (4. We let L (σ) denote the set L (σ) = π −1 (σ ∩ X ). We can then regularize the Feynman integral.4.56) with the integral π ∗ (ηm−1 ) π ∗ (ηm ) + df ∧ . which has the eﬀect of replacing a divergent integral with a meromorphic function of a regularization parameter . with ωn the standard volume form.
By Lemma 4.4.18) df ∂σ∩X σ∩X Theorem 4. and the fact that dπ ∗ (η) = 0. Proof.56) can be equivalently written in the form U(Γ) = 1 C(n.124 Feynman Motives Thus. ) π ∗ (ηm ) + ∂σn ∩D (XΓ )c σn ∩D (XΓ )c df ∧ π ∗ (ηm ) f + ∂σn ∩XΓ.2. We now study the dependence on the parameter > 0 of the Leray regularized Feynman integral (4. it extends to a holomorphic function for ∈ ∆∗ ⊂ C. we can write d I = A (ηm ) − A (ηm−1 ). let us write A (η) = σ∩X π ∗ (η). To prove the diﬀerentiability of I . Lemma 4. (4.5.2. with a pole of order at most m at = 0.16). The Leray regularization of the Feynman integral (3. The function I of (4.57). Thus.2 applied to (4.4.20) d df ∂σ∩X where X = {f = } and π ∗ (η)/df is the Gelfand–Leray form of π ∗ (η). ) × S 1 .14) and Proposition 3.3 and Lemma 4. (4. with m = −n + D( + 1)/2. d . Proof.18) is inﬁnitely diﬀerentiable in . π ∗ (ηm−1 ) + df π ∗ (ηm ) . Moreover. D. is the level set of f = ΨΓ .2. D.19) with π ∗ (η) as in (3. Here XΓ. the Leray regularization introduced here consists in replacing the integral over σn ∩ D (XΓ ) with an integral over L (σn ) (σn ∩ XΓ. we obtain d π ∗ (η) A (η) = − .4. which avoids the locus σn ∩ XΓ where the divergence can occur by going around it along a circle of small radius > 0. i.16). σn ∩XΓ. of an integral of the form π ∗ (ηm−1 ) I := + π ∗ (ηm ).2. The result follows directly from Proposition 3. ) 2πi C(n.3 and D (XΓ )c denoting the complement of D (XΓ ).2 above. a small punctured disk. (4.e. (4.17) with π ∗ (ηm ) = ∆(ω)/f m as in (4.
Feynman integrals and Gelfand–Leray forms 125 Thus. Assuming then that dk π ∗ (η) A =− δ k−1 d k df ∂σ∩X we obtain again by a direct application of Lemma 4. In fact.17) and in (4. Finally. Varchenko (1988)]. GuseinZade. to see that I has a pole of order at most m at = 0.18) depends holomorphically on the parameter by the general argument on holomorphic dependence on parameters given in Part III.2. which we denote by δ(α) = α d df Υ(α) = . to check the diﬀerentiability in the variable to all orders of I is equivalent to checking that of A . (4. . §10. df df (4. It can then be seen that the expression (4.21) df where α/df is a Gelfand–Leray form for α.2 of Vol.3 is. the equivalent expression given in the second line of (4. the nform Υ(α) also has a Gelfand–Leray form. notice that the form π ∗ (ηm ) of (4.2.II of [Arnold.22) We then prove that. We then deﬁne Υ : Ωn → Ωn by setting α Υ(α) = d .2. applying Lemma 4. (4.14) is given by ∆(ω)/f m and has a pole of order at most m at X. only deﬁned for > 0. In turn.16) used in Deﬁnition 4. for k ≥ 2. Notice then that. This proves diﬀerentiability to all orders.18) is clearly deﬁned for any complex ∈ ∆∗ in a punctured disk around = 0 of suﬃciently small radius. while the expression (4. a priori.23) df .2 gives d =− ∂σ∩X π ∗ (η) df .2 dk+1 A =− d k+1 =− ∂σ∩X d ∂σ∩X δ k−1 π ∗ (η) d f df df π (η) df ∗ δk . we ﬁrst see that d2 d π ∗ (η) A =− 2 d d ∂σ∩X df which. dk π ∗ (η) A =− δ k−1 d k df ∂σ∩X This follows by induction.
Algebrogeometric notions of tubular neighborhood were also elaborated upon in Grothendieck’s “Esquisse d’un Programme”. possibly involving a motivic version of the notion of “tubular neighborhood”. To be able to do so one would have to ﬁrst derive an algebraic version of the procedure described here. [Levine (2005)]. but cannot be used directly to derive information at the motivic level. . with Σ a normal crossings divisor containing the boundary ∂σn of the domain of integration. applied to the locus Σ ∩ XΓ . such as Σ = Σn the union of the coordinate hyperplanes. see e.g.126 Feynman Motives Notice that the regularization procedure described here is natural from the point of view of singularity theory and oscillatory integrals. The role of such notion in the algebrogeometric theory of Feynman integrals awaits further investigation.
In this chapter we give a very brief account of the main steps in the Connes–Kreimer theory. We restrict to the essential skeleton of what we need in order to continue with our narrative in the following chapters. i. it is well known in quantum ﬁeld theory that the process of removing divergences from Feynman integrals via renormalization cannot be achieved without taking into account the nested structure of divergences. the presence of subgraphs inside a given Feynman graph that already contribute divergences to the Feynman integral.Chapter 5 Connes–Kreimer theory in a nutshell So far we focused on individual Feynman graphs and associated algebrogeometric constructions. so we do not duplicate them here. However. For brevity. and the BPHZ procedure becomes an extraction of ﬁnite values via the Birkhoﬀ factorization of loops in a prounipotent Lie group of characters of the Hopf algebra of Feynman graphs. where the coproduct encodes the information on the subdivergences. these are already described in full detail in the original papers of Connes and Kreimer and also in the ﬁrst chapter of the book [Connes and Marcolli (2008)]. A process of subtracting divergences that takes into account this hierarchy of nested subdivergences was developed in [Bogolyubov and Parasiuk (1957)] and later completed by [Hepp (1966)] and [Zimmermann (1969)].e. such as how to deal with external momenta and we also do not reproduce full proofs of the main results. In fact. A very elegant geometric formulation of the BPHZ renormalization procedure was given in [Connes and Kreimer (2000)] and [Connes and Kreimer (2001)] in terms of a Hopf algebra of Feynman graphs. 127 . we skip several important technical points.
2) only terms C(γ) for strictly smaller graphs γ ⊂ Γ are involved. (5. They are deﬁned as the ¯ polar part of the Laurent series R(Γ).55) by the expression ¯ R(Γ) = U (Γ) + γ∈V(Γ) C(γ)U (Γ/γ). The expression (5.128 Feynman Motives 5.e.2) below. (Again. The coeﬀcients C(γ).2 Step 2: Counterterms These are the expressions by which the Lagrangian needs to be modiﬁed to cancel the divergence produced by the graph Γ. µ and the external momenta p for simplicity of notation. The residue Res(Γ) of a connected Feynman graph Γ is the Feynman graph consisting of a single vertex and as many half edges attached to it as the number of external edges of Γ.1. The sum in (5. ¯ C(Γ) = −T(R(Γ)). themselves Laurent series.1) Here we suppress the dependence on z.1. Here the residue is deﬁned as follows.1) is over the set V(Γ) of (not necessarily connected) proper nonempty subgraphs γ ⊂ Γ with the property that the residue Res(γi ) of each component γi of the graph γ has valence equal to one of the monomials in the Lagrangian L(φ). are deﬁned recursively in (5. We discuss here brieﬂy a property of the operator T that will be useful later. where in the right hand side of (5.1. i. the graph obtained by contracting all the internal edges of Γ to a single vertex.1. 5. Deﬁnition 5.5 below that this property corresponds to the fact .1 Step 1: Preparation One replaces the Feynman integral U (Γ) of (1. We’ll see in §5.2) Here T denotes the operator of projection of Laurent series onto their polar part. (5. for more details the reader is invited to look at Chapter 1 on [Connes and Marcolli (2008)]). 5. depending on the extra parameter µ. Notice how the counterterms C(Γ) are deﬁned here by a recursion.1 The Bogolyubov recursion The main steps of the Bogolyubov–Parasiuk–Hepp–Zimmermann renormalization procedure (BPHZ) are summarized as follows.1) is to be understood as a sum of Laurent series in z.
If f1 = ∞ k=−N1 (5. On the other hand. (5.2. (5. 5. Proof.Connes–Kreimer theory in a nutshell 129 that the diﬀerential ﬁeld K of convergent Laurent series with the operator T has the structure of a Rota–Baxter algebra. Lemma 5.1.1.3) ak z k and f2 = −1 ∞ r r=−N2 br z then T(f1 f2 ) = m=−(N1 +N2 ) k+r=m ak br z m . which has the eﬀect of removing the terms that had been counted twice in the previous sum and leaving exactly the terms that constitute the polar part T(f1 f2 ). This satisﬁes the identity T(f1 )T(f2 ) = T(f1 T(f2 )) + T(T(f1 )f2 ) − T(f1 f2 ).3). the operator that projects a Laurent series onto its polar part. Let K be the ﬁeld of convergent Laurent series (germs of meromorphic functions at the origin).5) . we also have −1 −1 T(f1 )T(f2 ) = ( k=−N1 ak z k )( r=−N2 br z r ). One then ﬁnds the identity (5. as above. not of U (Γ) itself but of its preparation. since all the terms that appear in T(f1 f2 ) are accounted for with multiplicity one by summing −1 ∞ T(T(f1 )f2 ) = T(( k=−N1 ak z k )( r=−N2 −1 br z r )) and ∞ T(f1 T(f2 )) = T(( k=−N1 ak z k )( r=−N2 br z r )) and then subtracting T(f1 )T(f2 ).4) γ∈V(Γ) which we also write equivalently as R(Γ) = U (Γ) + γ∈V(Γ) C(γ)U (Γ/γ) − T U (Γ) + γ∈V(Γ) C(γ)U (Γ/γ) .3 Step 3: Renormalized values One can then extract a ﬁnite value from the integral U (Γ) by removing the polar part. Let T be. Namely. one sets ¯ R(Γ) = R(Γ) + C(Γ) = U (Γ) + C(Γ) + C(γ)U (Γ/γ).
. the following statement is the ﬁnal result of the combined [Bogolyubov and Parasiuk (1957)]. In fact. because the Hopf algebras that appear in applications to the theory of renormalization in quantum ﬁeld theory are usually of this kind.2 Hopf algebras and aﬃne group schemes In the following we work with algebras deﬁned over a ﬁeld K of characteristic zero. A very nice conceptual understanding of the BPHZ renormalization procedure with the DimReg+MS regularization was obtained by Connes and Kreimer [Connes and Kreimer (2000)]. if one wants to be able to remove the polar part by correcting the Lagrangian by adding suitable counterterms. the coeﬃcient of pole of the corrected Lau¯ rent series R(Γ) provided by the Bogolyubov preparation is always a local expression in the variables of the external momenta.3. The coeﬃcient of pole of R(Γ) is local.2).1.130 Feynman Motives that is.1. based on a reformulation of the BPHZ procedure in geometric terms. so we limit ourselves to a more sketchy overview here and refer the reader to that more detailed treatment. K will be either Q or C. The main result of BPHZ is that. hence one that can be corrected either by altering the coeﬃcients of the already existing terms in the Lagrangian or by a adding a ﬁnite number of new polynomial terms to the Lagrangian. 5. A commutative Hopf algebra over K is a commutative Kalgebra H endowed with an algebra homomorphism ∆ : H → H ⊗ H. A detailed account of the Connes–Kreimer theory was given in Chapter 1 of [Connes and Marcolli (2008)]. these themselves have to be local functions as the terms already present in the Lagrangian are. the minimal subtraction of the polar part not of the original Laurent ¯ series U (Γ) but of the corrected one R(Γ). ¯ Theorem 5.2. Namely. unlike what happens typically with the Laurent series U (Γ) alone. because of the recursive nature of the deﬁnition of the counterterms C(Γ) in (5. We also restrict our attention here to commutative Hopf algebras. [Hepp (1966)] and [Zimmermann (1969)]. Here local means that it can be expressed as a local function in the momentum variables (involving only polynomials and derivatives). Deﬁnition 5. for our main application. In fact. The BPHZ procedure also goes under the name of Bogolyubov recursion. [Connes and Kreimer (2001)].
antipode S. Elements of G(A) are homomorphisms of Kalgebras φ : H → A. As we recall brieﬂy here (see also [Waterhouse (1979)]) this is a functor from algebras to groups naturally associated to a commutative Hopf algebra. The property that this is a unit with respect to the multiplication in G(A) come from the compatibility relation between counit and coproduct in the ε . is that of aﬃne group scheme. by setting G(A) = HomAK (H. ∀x. and the compatibility of comultiplication ∆. the antipode.2. A useful notion we refer to often in the following. ∆(x) . and multiplication µ. A). which also already appeared in §2. We need to show that the set G(A) is a group. an algebra homomorphism ε : H → K. Proof. (5.9 and §2. (5. φ(1) = 1 . (id ⊗ ε)∆ = id = (ε ⊗ id)∆ as morphisms H → H.Connes–Kreimer theory in a nutshell 131 the coproduct. Proposition 5. φ(x y) = φ(x)φ(y).2. and an algebra antihomomorphism S : H → H. Notice that in general the coproduct will not be cocommutative. A commutative Hopf algebra H over a ﬁeld K of characteristic zero determines a covariant functor G from the category AK of commutative algebras over K to the category of groups. for A ∈ Obj(AK ). y ∈ H . µ(id ⊗ S)∆ = µ(S ⊗ id)∆ = 1 ε as morphisms H → H. There is a unit element in G(A) which is determined by the counit ε of H by 1G(A) : H → K → A.7) It is associative because of the coassociativity of ∆.6) There is a product operation on G(A) deﬁned by the coproduct ∆ of H φ1 ∗ φ2 (x) = φ1 ⊗ φ2 . These satisfy the coassociativity (∆ ⊗ id)∆ = (id ⊗ ∆)∆ as morphisms H → H ⊗K H ⊗K H.10. in the context of the Tannakian formalism and motivic Galois groups. the compatibility of the coproduct ∆ with the counit ε. the counit.
φ ∗ φ−1 = φ−1 ∗ φ = 1. So the above shows that a commutative Hopf algebra determines an aﬃne group scheme. there is an inverse φ−1 = φ◦S. counit ε(t) = 0 and antipode S(t) = −t. The map φ → φ−1 is an antihomomorphism. with coproduct ∆(t) = t ⊗ t. t−1 ]. H0 = K. where by algebraic groups we mean here subgroups of the aﬃne group scheme GLn . We discuss more the example of GLn as an aﬃne group scheme in Proposition 5. An aﬃne group scheme is a representable covariant functor from the category of commutative algebras over the ﬁeld K to the category of groups. . counit ε(t) = 1 and antipode S(t) = t−1 . to be the inverse in the group G(A) follows from the remaining axiom of the Hopf algebra structure that gives the compatibility between the counit. respectively. Ga and Gm . The main examples in the following will be prounipotent aﬃne group schemes. These prounipotent cases arise from Hopf algebras that are graded. for some commutative algebra H. as the Connes–Kreimer Hopf algebra is. which in turn inherits the structure of Hopf algebra from the group structure of G(A). we refer the reader to [Demazure and Grothendieck (1970)] and [Waterhouse (1979)]. The fact that it satisﬁes the correct properties with respect to the multiplication and unit. • The multiplicative group Gm corresponds to the Hopf algebra H = K[t.2 below. comultiplication. H = ⊕n≥0 Hn and connected. which correspond to the Hopf algebras • The additive group Ga corresponds to the Hopf algebra H = K[t] with coproduct ∆(t) = t ⊗ 1 + 1 ⊗ t. because S is an antihomomorphism of the Hopf 2 1 algebra H. Any such representable functor will be of the form (5. deﬁned by precomposing a morphism φ : H → A with the antipode S : H → H. These are projective limits of unipotent algebraic groups.4. The simplest examples of aﬃne group schemes are the additive and the multiplicative group. For more detailed information on aﬃne group schemes. Similarly.6).132 Feynman Motives axioms of Hopf algebra. and antipode. (φ1 ∗ φ2 )−1 = φ−1 ∗ φ−1 .
This grading corresponds to the order in the perturbative expansion.8) The sum is over the class V(Γ) of proper subgraphs of Γ with the property that the quotient graph Γ/γ is still a 1PI Feynman graph of the same theory. and. with deg(Γ1 · · · Γn ) = i deg(Γi ). The antipode is deﬁned inductively by S(X) = −X − S(X )X . The class of subgraphs V(Γ) can then be charcterized by requiring that the components of γ ⊂ Γ are 1PI Feynman graphs with the property that the residue Res(Γ) has valence corresponding to one of the monomials in the Lagrangian of the theory. in the sense that it involves only those graphs that are Feynman graphs for the speciﬁed Lagrangian L(φ). In the case of several connected components. deg(1) = 0. the quotient graph Γ/γ is understood as the graph obtained by shrinking each component of γ ⊂ Γ to a vertex. The most interesting part of the structure is the coproduct.3 The Connes–Kreimer Hopf algebra The Hopf algebra of Feynman graphs introduced in [Connes and Kreimer (2000)] depends on the choice of the physical theory. It is graded by loop number. This in fact ensures that the corresponding vertex in the quotient graph Γ/γ has an admissible valence for it to be a Feynman graph. . As an algebra HCK it is the free commutative algebra with generators the 1PI Feynman graphs Γ of the speciﬁed physical theory. as we see below. The condition that Γ/γ is a 1PI Feynman graph can be formulated in terms of residue graphs Res(γ) of the sugraphs γ ⊂ Γ. which is given on generators by ∆(Γ) = Γ ⊗ 1 + 1 ⊗ Γ + γ∈V(Γ) γ ⊗ Γ/γ.Connes–Kreimer theory in a nutshell 133 5. or by the number of internal lines. the coproduct formula also involves only those subgraphs whose residue graph corresponds to one of the monomials in the Lagrangian. The valence of this vertex is then equal to the number of external edges. The fact that Γ/γ is 1PI then follows from the fact that Γ and all the components of γ are 1PI. (5. The subgraphs in V(Γ) are not necessarily connected. For a given graph Γ the residue Res(Γ) is the graph that is obtained by keeping all the external edges of Γ and shrinking all internal edges to a single vertex.
It is referred to as the core Hopf algebra and it is simply deﬁned by considering as generators all 1PI graphs without restriction on the valence of vertices. The aﬃne group scheme dual to the Connes–Kreimer Hopf algebra is referred to as the group of diﬀeographisms. This additional structure is described in detail in [Connes and Kreimer (2000)] as well as in Chapter 1 of [Connes and Marcolli (2008)]. the coproduct is taken over all the subgraphs whose connected components are 1PI. Generalizations to vector valued ﬁelds do not present any problem. was considered in [Bloch and Kreimer (2008)]. where all the X and X have lower degrees. The results were recently generalized to gauge theories in [van Suijlekom (2007)]. [van Suijlekom (2006)]. σ) of a 1PI graph and a distribution on the space of test functions of the external momenta. This continuous version of the the Hopf algebra of renormalization is necessary when taking into account the external momenta of the graphs in the extraction of subdivergences. see also [Kreimer (2009)]. Correspondingly. while extending the same setting to gauge theories and chiral theories is more subtle. It was proved in [Connes and Kreimer (2000)] that this group acts by local diﬀeomorphisms on the coupling constants of the theory. [van Suijlekom (2008)]. which is useful in algebrogeometric applications. A way to describe the prescription of [Breitenlohner and Maison . Finally. The algebrogeometric Feynman rules discussed above and introduced in [Aluﬃ and Marcolli (2008b)] deﬁne ring homomorphisms from the core Hopf algebra to the Grothendieck ring of varieties. see [Jegerlehner (2001)]. or to the ring of immersed conical varieties deﬁned in [Aluﬃ and Marcolli (2008b)]. without restriction on the residue graphs. C) is a prounipotent complex Lie group G(C). the Connes–Kreimer construction of the Hopf algebra is given for scalar quantum ﬁeld theories. A further important remark about the Connes–Kreimer Hopf algebra: the version we recalled brieﬂy here is only the combinatorial part of a larger Hopf algebra. Its set of complex points G(C) = Hom(HCK .134 Feynman Motives where X is an element with coproduct ∆(X) = X ⊗ 1 + 1 ⊗ X + X ⊗ X . This would correspond in physical terms to arbitrarily large powers φk in the Lagrangian. A variant of the Connes–Kreimer Hopf algebra. whose generators are pairs (Γ. The case of chiral theories runs into a well known problem of reconciling the use of dimensional regularization with the chirality operator γ5 . by showing that the Ward identities deﬁne Hopf ideals.
In this setting the operation of dimensional regularization is then expressed as a cup product of spectral triples. with values in the polynomial algebra Q = C[z −1 ]. such that φ = (φ− ◦ S) φ+ . using periods of supermanifolds.4 Birkhoﬀ factorization We give an algebraic formulation of Birkhoﬀ factorization in the group scheme G of a commutative Hopf algebra in terms of the Hopf algebra H and its characters. i.Connes–Kreimer theory in a nutshell 135 (1977)] to express the chirality operator γ5 within DimReg in terms of noncommutative geometry was given in the unpublished [Connes and Marcolli (2005b)]. Another interesting recent result is a categoriﬁcation of the Connes–Kreimer Hopf algebra obtained in [Kremnizer and Szczesny (2008)].9) where S is the antipode in the Hopf algebra and is the product in G dual to the coproduct in the Hopf algebra.10) is the counit of the Hopf where − : C[z −1 ] → C is the augmentation and algebra H. based on zeta function regularization instead of DimReg and adapted to curved backgrounds was developed in [Agarwala (2009)]. and φ− ∈ Hom(H. Recently. a version of the Connes–Kreimer renormalization as Birkhoﬀ factorization.9). is not unique. O). To make it unique one also requires the normalization condition − ◦ φ− = . though the most appropriate form of the Connes–Kreimer renormalization procedure in that context remains to be properly discussed. . Suppose given an algebra homomorphism φ ∈ Hom(H. with O the complex algebra of convergent power series.1. The factorization (5. Q). An extension of the parametric Feynman integrals to theories with bosonic and fermionic ﬁelds. (5.e. Deﬁnition 5. germs of meromorphic functions at z = 0. if it exists. where K is the ﬁeld of convergent Laurent series. 5. K). was given in [Marcolli and Rej (2008)]. see also Chapter 1 of [Connes and Marcolli (2008)].4. where the dual Hopf algebra is identiﬁed with the Hall algebra of a ﬁnitary abelian category. (5. Then φ has a Birkhoﬀ factorization if there exist algebra homomorphisms φ+ ∈ Hom(H.
One of the main results of the Connes–Kreimer theory is a recursive formula for the Birkhoﬀ factorization in the case of graded connected commutative Hopf algebras. t]/(det(xij ) − t) with the coproduct ∆(xij ) = k = xik ⊗ xkj . Proposition 5. .2. so that the recursive formula for the factorization of φ gives the Birkhoﬀ factorization of loops written in the more traditional form γ(z) = γ− (z)−1 γ+ (z). Then there exist elements φ ∈ Hom(H. (5. Proof.9). K) in terms of loops γ : ∆∗ → GLn . the existence of a factorization (5.11) shows that the resulting bundle is trivial.9). The aﬃne group scheme dual to the Hopf algebra H = C[xij . To see that there exist elements of φ ∈ Hom(H. if we use the function γ(z) to deﬁne the transition function of a holomorphic vector bundle over the 2sphere P1 (C). The following example illustrates a case where there are elements in Hom(H. C) of algebra homomorphisms from H to C. as above. K) that do not admit a factorization of the form (5. K) can be equally interpreted as loops deﬁned on an inﬁnitesimal punctured disk ∆∗ around the origin z = 0 in C∗ and with values in the prounipotent complex Lie group G(C) = Hom(H. Consider the commutative Hopf algebra H C[xij . we can give an equivalent formulation in terms of loops in a complex Lie group. so it suﬃces to take one such nontrivial vector bundle E and a transition function γ(z) relating the trivializations of E on the two contractible sets ∆ and P1 (C) ∆ to see that this loop cannot admit a factorization of the form (5. Then.136 Feynman Motives The characters φ ∈ Hom(H. We review this result in the next section. K). K) that do not admit a factorization of the form (5. one knows that there are nontrivial holomorphic vector bundles on P1 (C). However. it suﬃces to notice that.11) with the normalization condition γ− (∞) = 1.4. We can use the equivalent description of characters φ ∈ Hom(H. K) that do not admit a Birkhoﬀ factorization in the form stated above. t]/(det(xij ) − t) with the given coproduct is the group GLn . Not all commutative Hopf algebras admit a Birkhoﬀ factorization for arbitrary elements φ ∈ Hom(H.11).
Given a commutative unital Rota–Baxter algebra (A. for which the Rota–Baxter identity is the integration by parts formula.5. Such an operator is called a Rota–Baxter operator of weight λ. We consider. using the formalism of Rota–Baxter algebras.2. R) of weight λ = −1. Notice that here the condition that H is a commutative graded Hopf algebra does not mean that it is “gradedcommutative”: the commutators are not graded commutators but ordinary commutators. we can formulate their result in a slightly more general form. To this purpose we recall some preliminary notions on Rota–Baxter algebras and operators. the smallest unital algebra containing R(A) as an ideal) and A+ = (1 − R)(A).1. This is an immediate consequence of Lemma 5. one obtains two commutative unital algebras A± by taking as A− the image R(A) with a unit adjoined (i. for later use.1. (5. Lemma 5.e.5 Factorization and RotaBaxter algebras Let H be a graded connected commutative Hopf algebra H = ⊕n≥0 Hn .5. where the Hn are ﬁnite dimensional.2. It behaves like the case of a graded–commutative Hopf algebra where the graded pieces are nonzero only in even degrees. Let K denote the ﬁeld of convergent Laurent series. elements φ ∈ Hom(H. for an algebra deﬁned over the ﬁeld of complex numbers. Kreimer (2004)]. K) and we ask whether all such elements admit a Birkhoﬀ factorization as in (5. Deﬁnition 5. T) of the ﬁeld of convergent Laurent series with the projection on the polar part is a Rota–Baxter algebra of weight λ = −1. in the form of an explicit recursive formula. The connectedness condition means that H0 = C. A commutative Rota–Baxter algebra of weight λ is a commutative algebra A endowed with a linear operator R satisfying the Rota–Baxter identity R(x)R(y) = R(xR(y)) + R(R(x)y) + λR(xy). as above. The prototype of a Rota–Baxter operator is integration. The pair (K. In fact.5. Guo. .12) Lemma 5.9). for the class of Hopf algebras mentioned above. as was done for instance in [EbrahimiFard.Connes–Kreimer theory in a nutshell 137 5.3. Proof. The Connes–Kreimer formula [Connes and Kreimer (2000)] gives a positive answer.
then any homomorphism of unital commutative algebras φ ∈ Hom(H. The normalization condition is also formulated in this case as − ◦ φ− = .14) of lower (5. R) of weight λ = −1. Then xy = (a − R(a))(b − R(b)) = ab − aR(b) − bR(a) + R(a)R(b) = ab − aR(b) − bR(a) + R(aR(b) + R(a)b + λab).5. K) we can similarly consider the same problems for algebra homomorphisms from the Hopf algebra H with values in any Rota–Baxter algebra (A. where − : A− → C is the augmentation map and is the counit of the Hopf algebra H.13) where the homomorphisms φ± ∈ Hom(H. that if H is a graded connected commutative Hopf algebra and (A. we look for algebra homomorphisms φ± ∈ Hom(H. The product R(a)R(b) = R(aR(b) + R(a)b + λab) is also in the range of R. A) admits a Birkhoﬀ factorization (5. ∀X ∈ H viewed as an equality of elements in A. In fact. for λ = −1. b ∈ A. Guo. with X and X φ− (X )φ(X ) . where here we use the fact that A± ⊂ A so that the equality above is understood as the equality φ(X) = (φ− ◦ S)(X) φ+ (X).138 Feynman Motives Proof. (5. A). Similarly. one obtains xy = (ab − aR(b) − bR(a)) − R(ab − aR(b) − bR(a)) ∈ A+ . Given φ ∈ Hom(H. satisfying the factorization φ = (φ− ◦ S) φ+ . and reformulated in the RotaBaxter context in [EbrahimiFard. given x.15) X ⊗ X .4. Kreimer (2004)].1 formulates the Birkhoﬀ factorization for algebra homomorphisms φ ∈ Hom(H. R) is a unital Rota–Baxter algebra of weight λ = −1. φ+ (X) = (1 − R) φ(X) + .13) as in (5. Just as Deﬁnition 5. A± ) with the algebras A± obtained as in Lemma 5. y ∈ A+ we have x = a − R(a) and y = b − R(b) for some a. and φ− (X )φ(X ) . b ∈ A. for x.3. Thus. hence in A− . These are indeed algebras.9). y ∈ A− we have x = R(a) and y = R(b) for some a. A± ) are deﬁned inductively by the formulae φ− (X) = −R φ(X) + where ∆(X) = X ⊗ 1 + 1 ⊗ X + degree. It was proved in [Connes and Kreimer (2000)]. (5.
given above for any algebra homomorphism from a graded connected commutative Hopf algebra H to a commutative Rota–Baxter algebra (A.5) are exactly the recursive formulae for the Birkhoﬀ factorization of characters of the Connes– Kreimer Hopf algebra with values in a Rota–Baxter algebra. (5. Thus. following [Aluﬃ and Marcolli (2008b)]. (5. A variant of such motivic Feynman rules is obtained by setting ˆ [An XΓ ] .2). φ± (XY ) = φ± (X)φ± (Y ). Then the C(Γ) = U− (Γ) are the counterterms and the R(Γ) = U+ (Γ)z=0 are the renormalized values.17) This is just the same as the formulae (5. the results of [Connes and Kreimer (2000)] and [Connes and Kreimer (2001)] show that.16) U+ (Γ) = (1 − T) U (Γ) + γ⊂Γ U− (γ)U (Γ/γ) . in terms of Birkhoﬀ factorization. where one inverts the Lefschetz motive. or equivalently. with e values in the Grothendieck ring of motives K0 (M).Connes–Kreimer theory in a nutshell 139 The RotaBaxter property of the operator R with weight −1 is needed to show that the φ± deﬁned recursively in this way are algebra homomorphisms. For . (5. Dividing by Ln has ˆ the eﬀect of normalizing the “motivic Feynman rule” [An XΓ ] by the value it would have if Γ were a forest on the same number of edges. and the homomorphism U is the usual Feynman amplitude regularized using DimReg. the formulae of the BPHZ renormalization procedure (5.14). that one can deˆ ﬁne motivic Feynman rules in terms of the class [An XΓ ] of the graph hypersurface complement in the Grothendieck ring of varieties K0 (V). R) of weight −1. applied to the case where H = HCK is the Connes–Kreimer Hopf algebra and where the RotaBaxter algebra is given by Laurent series in the DimReg variable z and R = T is the projection onto the polar part. (5. namely U− (Γ) = −T U (Γ) + γ⊂Γ U− (γ)U (Γ/γ) .4).18) U (Γ) = n L with values in the ring K0 (V)[L−1 ]. via the additive invariant of [Gillet and Soul´ (1996)].14) and (5.6 Motivic Feynman rules and RotaBaxter structure We have seen. (5. 5.
In fact. gives some information on the birational geometry of the graph hypersurface. We work here with the Grothendieck ring K0 (VC ) for varieties over the ﬁeld of complex numbers. The class of a graph hypersurface XΓ in the ring Z[SB] of stable birational equivalence classes detects some of the properties of the graph Γ. if ˆ a graph is not 1PI then in the Grothendieck ring one has [An XΓ ] = n−1 ˆ L · [A XΓ ]. m ≥ 0. the renormalized value of the Feynman integral. Thus. In the ring K0 (V)[L−1 ] or K0 (M) one can still consider the Rota–Baxter structure given by the operator R of projection onto the polar part in the variable L. [An1 Xγ ] and [An2 XΓ/γ ] that are contained in the ideal (L) ⊂ K0 (V0 ) contribute cancellations to the Ln in the denominator. For example. where Γ is the graph obtained by removing a disconnectˆ ing edge of Γ. It is possible that this invariant and the Birkhoﬀ factorization of U (Γ) may help to detect the presence of nonmixedTate strata in the graph hypersurface XΓ coming from the contributions of hypersurfaces of smaller graphs γ ⊂ Γ or quotient graphs Γ/γ. whenever the graph is not 1PI. we ﬁrst recall the following useful result of [Larsen and Lunts (2003)]. More interestingly. We then see that the renormalized value is in this case U+ (Γ)L=0 = (1 − R) U (Γ) + γ⊂Γ U− (γ)U (Γ/γ) L=0 ∈ Z[SB].19) ˆ ˆ ˆ Notice that the parts of [An XΓ ]. [XΓ ]SB = 1. Using then the fact that . Recall that two (irreducible) varieties X and Y are stably birationally equivalent if X × Pn and Y × Pm are birationally equivalent for some n. In fact. [An XΓ ]SB = 0 ∈ Z[SB]. It is proved in [Larsen and Lunts (2003)] that the quotient of the Grothendieck ring K0 (VC ) by the ideal generated by L = [A1 ] is isomorphic as a ring to Z[SB]. The renormalized Feynman rule U+ (Γ) = (1 − R) U (Γ) + γ⊂Γ U− (γ)U (Γ/γ) then gives a class in the Grothendieck ring of varieties K0 (V). (5. the ring of the multiplicative monoid SB of stable birational equivalence classes of varieties in VC . it was shown in [Aluﬃ and Marcolli (2008b)] that the class is trivial.140 Feynman Motives the original Feynman integrals this would measure the amount of linear dependence between the edge momentum variables created by the presence of the interaction vertices. in this case.
or in the ring of immersed conical varieties. . X It would be interesting to see if there are other natural Rota–Baxter structures on the Grothendieck ring of varieties. that can be used in deﬁning a renormalized version of motivic and algebrogeometric Feynman rules. one obtains ˆ Γ ]SB = −1 + [XΓ ]SB = 0.Connes–Kreimer theory in a nutshell 141 [An [An ˆ XΓ ] = Ln − 1 − (L − 1)[XΓ ] in the Grothendieck ring.
hence the treatment given here will be shorter and will only aim at summarizing the main steps involved in the main result of [Connes and Marcolli (2004)]. hence they can be classiﬁed in terms of ﬁnite dimensional linear representations of an associated aﬃne group scheme. The purpose of this approach is to parameterize the data of divergences in perturbative scalar quantum ﬁeld theories in terms of suitable gauge equivalence classes of diﬀerential systems with irregular singularities. 6. A very detailed account of this topic is available in [Connes and Marcolli (2008)].Chapter 6 The Riemann–Hilbert correspondence This chapter is dedicated to the Riemann–Hilbert correspondence of renormalization. The interested reader is encouraged to look at [Connes and Marcolli (2006b)] and at the ﬁrst chapter of [Connes and Marcolli (2008)] for more extensive details. This is a way of formulating what is known in physics as the ’t Hooft–Gross relations. The Riemann–Hilbert correspondence is the equivalence of categories between the analytic data of diﬀerential systems and the representation theoretic data. which means that one shows that these diﬀerential systems form a Tannakian category. These are then classiﬁed in terms of the Riemann–Hilbert correspondence. and to [Connes and Marcolli (2005a)] for a shorter survey. These express the fact that counterterms 143 . as derived in [Connes and Marcolli (2004)].1 From divergences to iterated integrals The ﬁrst step in reformulating the divergences of renormalization in terms of diﬀerential systems consists of writing the negative piece γ− (z) of the Birkhoﬀ factorization as an iterated integral depending on a single element β in the Lie algebra Lie(G) of the aﬃne group scheme dual to the Connes– Kreimer Hopf algebra.
in the Birkhoﬀ factorization. there is a dependence on a mass (energy) scale µ. and the action θt is induced by the grading of the Hopf algebra by θu (X) = un X. The renormalization group analysis of [Connes and Kreimer (2001)] shows that γµ (z) satisﬁes the scaling property γet µ (z) = θtz (γµ (z)) ∂ γ− (z) = 0. so as to give the time ordered exponential above. Thus. inherited from the same dependence of the dimensionally regularized Feynman integrals Uµ (Γ). discussed in (1.55) above.144 Feynman Motives only depend on the beta function of the theory. with deg(X) = n. that is. This negative part γ− (z) can then be written as a time ordered exponential γ− (z) = T e− z where dn (β) = s1 ≥s2 ≥···≥sn ≥0 1 ∞ 0 θ−t (β)dt =1+ dn (β) . and where β ∈ Lie(G) is the beta function. and X ∈ H. where one knows by reasons of dimensional analysis (see [Collins (1986)]) that the negative part is independent of µ.2) . the inﬁnitesimal generator of renormalization group ﬂow. This result follows from the analysis of the renormalization group given in [Connes and Kreimer (2000)] [Connes and Kreimer (2001)]. after the recursive formula for the coeﬃcients dn is explicitly solved. with generator the grading operator Y (X) = nX. The loop γµ (z) collects all the unrenormalized values Uµ (Γ) of the Feynman integrals in the form of an algebra homomorphism φµ : H → C({z}) from the Connes–Kreimer Hopf algebra to the ﬁeld of germs of meromorphic functions. The starting point is the observation that. the inﬁnitesimal generator of the renormalization group ﬂow.1) in addition to the property that its negative part is independent of µ. zn n=1 ∞ θ−s1 (β) · · · θ−sn (β)ds1 · · · dsn . given on generators by φµ (Γ) = Uµ (Γ).+ (z). (6. for u ∈ Gm . ∂µ (6. that is. we have γµ (z) = γ− (z)−1 γµ.
and an aﬃne group scheme G. where D(f ω ) = ω and D(f ω ) = ω . suppose given the diﬀerential ﬁeld (K. This gauge equivalence is the same thing as the requirement that solutions have the same negative piece of the Birkhoﬀ factorization. Thus γµ (z) is speciﬁed by β up to an equivalence given by the regular term γreg (z). singular at z = 0 ∈ ∆∗ . where K = C({z}) is the ﬁeld of convergent Laurent series and δ(f ) = f . for a ﬂat Lie(G(C))valued connection ω.2 From iterated integrals to diﬀerential systems Having written the counterterms as iterated integrals has the advantage that it makes it possible to reformulate these data in terms of diﬀerential systems. The fact that. D(f h) = Dh+h−1 Df h. ω ω ω = Dh + h−1 ωh ⇔ f− = f− . as we recall below. δ). 6. In [Connes and Marcolli (2004)] these diﬀerential systems are considered up to a gauge equivalence relation by a regular gauge transformation. for h ∈ G(C{z}). counterterms are independent of the energy scale corresponds to the fact that the diﬀerential systems are given by connections satisfying an equisingularity condition. In particular.+ (z) = T e− z 1 −z log µ 0 θ−t (β)dt θz log µ (γreg (z)). ∈ π1 (∆∗ ). and one can consider diﬀerential equations of the form D(f ) = ω. More precisely. b An iterated integral (or timeordered exponential) g(b) = T e a α(t)dt is the unique solution of a diﬀerential equation dg(t) = g(t)α(t)dt with initial condition g(a) = 1. The existence of solutions is ensured by the condition of trivial monodromy on ∆∗ M (ω)( ) = T e 1 0 ∗ ω = 1. In fact.The Riemann–Hilbert correspondence 145 The Birkhoﬀ factorization is then written in [Connes and Marcolli (2004)] in terms of iterated integrals as γµ. with C{z} the ring of convergent power series (germs of holomorphic functions). by dimensional analysis. One then has a logarithmic derivative G(K) f → D(f ) = f −1 δ(f ) ∈ Lie(G(K)). in [Connes and Marcolli (2004)] a geometric . The equivalence corresponds to “having the same negative part of the Birkhoﬀ factorization”. one can use the fact that iterated integrals are uniquely solutions of certain diﬀerential equations.
It is obtained by working with all ﬂat equisingular vector bundles. equation (6. σ2 of B with σ1 (0) = σ2 (0) have “the same type of singularities”. while the second property expresses geometrically the independence of the counterterms from the mass scale. then the restrictions along diﬀerent sections σ1 . 6. uY (g)) ∀u ∈ Gm . that is.2).3 Flat equisingular connections and vector bundles The setting described above still depends on the choice of a given physical theory through the group G of diﬀeographisms. The multiplicative group acts by u(b. regular at zero. where z ∈ ∆ is the complexiﬁed dimension of DimReg and the ﬁber µz ∈ Gm over z corresponds to the changing mass scale. u(v)) = uY ( (z.1) for the solutions of the corresponding diﬀerential systems. namely ∗ ∗ σ1 (γ) ∼ σ2 (γ).2) correspond to the properties that the ﬂat connection on P ∗ is equisingular. u)). without loss of information. The aﬃne group scheme G is uniquely determined by its Tannakian category of ﬁnite dimensional linear representations RepG . one can replace singular connections on the principal Gbundle P = B × G with singular connections on associated vector bundles E = B × V obtained by ﬁnite dimensional linear representations of G in GL(V ). (6. The ﬁrst property of equisingular connections corresponds to the scaling property (6. dual to the Connes–Kreimer Hopf algebra of the theory. • If γ is a solution in G(C({z})) of the equation Dγ = . Among them one will identify a subcategory of those that are obtained .1) and (6. Thus. g) = (u(b). that is. −1 where f1 ∼ f2 means that f1 f2 ∈ G(C{z}).146 Feynman Motives reformulation of the two conditions (6. The two conditions (6.1). This provides a natural larger ambient category where the diﬀerent diﬀeographism groups G of diﬀerent physical theories can be compared.2) is given in terms of properties of connections on a principal Gbundle P = B × G over a ﬁbration Gb → B → ∆. it satisﬁes: • Under the action of u ∈ Gm the connection transforms like (z.
namely linear maps intertwining the connections and compatible with the ﬁltrations. The reason why (6. under the relation U(A) = Hom(HU . We recall here brieﬂy the main steps of the construction of this category. as usual. which also arises when dealing with categories of Hodge structures.3) Notice that. (6. thus reconstructing the data associated to divergences of a particular physical theory. 6. A). The objects Obj(E) are pairs Θ = (V.The Riemann–Hilbert correspondence 147 from representations of a given G. The vector space has a ﬁltration W −n (V ) = ⊕m≥n Vm induced by the grading and a Gm action also coming from the grading. up to the equivalence relation of W equivalence. Here the condition that the connections are equisingular means that they are Gm invariant and that restrictions of solutions to sections of B with the same σ(0) are W equivalent. so the ﬁrst guess about the choice of morphisms.3) is the correct notion here is that one needs to bypass a well known obstacle. It is proved in [Connes and Marcolli (2004)] (see also Chapter 1 of [Connes and Marcolli (2008)]) that ﬂat equisingular vector bundles form a Tannakian category. the diﬃcult part in deﬁning a good categorical setting is to have the correct notion of morphisms. This is deﬁned by T ◦ 1 = 2 ◦ T for some T ∈ Aut(E) which is compatible with the W ﬁltration and trivial on Gr−n (V ). where V is a ﬁnite dimensional Zgraded vector space. [ ]). namely the fact that ﬁltered spaces do not form an abelian category. . singular over the preimage in B of the point 0 ∈ ∆. Θ ) are linear maps T : V → V that are compatible with the grading. out of which one forms a bundle E = B × V . and such that on E ⊕ E the following connections are W equivalent: 0 0 W −equiv T 0 − T . has to be corrected to account for this problem. The morphisms HomE (Θ. The class [ ] is an equivalence class of equisingular connections on the vector bundle E.4 The “cosmic Galois group” The proof given in [Connes and Marcolli (2004)] and [Connes and Marcolli (2008)] of the fact that the resulting category E is Tannakian directly exhibits an equivalence of categories with RepU∗ for an aﬃne group scheme U∗ = U Gm where U is dual. which are compatiW ble with the ﬁltration and trivial on the induced graded spaces Gr−n (V ).
where γU is the universal singular frame deﬁned as γU (z. k1 (k1 + k2 ) · · · (k1 + k2 + · · · + kn ) n≥0. which also appear in the local index formula of [Connes and Moscovici (1995)]. e−2 . GraciaBond´ Patras ia. where the element e = is the generator of a 1parameter subgroup ∗ of U which lifts the renormalization group at the level of the group U∗ . as it maps to the correct counterterms under the map of the generators e−n to the graded components of the beta function. for any given object Θ = (V. has a natural explanation in terms of the combinatorics of Dynkin idempotents and Dynkin operators for free Lie algebras. Notice that the inﬁnite sum is well deﬁned since the Lie algebra is pronilpotent. [ ]) in E there exists unique ρ ∈ RepU∗ such that Dρ(γU ) W −equiv . v) = T e− z ∞ n=1 e−n 1 v 0 uY (e) du u . (2007)]. · · · ). The result of [Connes and Marcolli (2004)] reported here above gives a positive answer to Cartier’s . that would act on all the coupling constants of scalar quantum ﬁeld theories.148 Feynman Motives to the Hopf algebra HU = U (L)∨ dual (as Hopf algebra) to the universal enveloping algebra of the free graded Lie algebra L = F(e−1 . closely related to symmetries of multiple zeta values. the universal singular frame can be regarded as a universal source for the counterterms for all given physical theories. e−3 . He referred to it as the “cosmic Galois group”. generalizing the renormalization group action.kj >0 The occurrence of these coeﬃcients.98 of [Connes and Marcolli (2008)] that the universal singular frame is given. as shown in [EbrahimiFard. Cartier conjectured the existence of a group. v) = v j kj z −n . Thus. It is shown in Proposition 1. in terms of the generators of the Lie algebra. The group homomorphism U → G that realizes the ﬁnite dimensional linear representations of G with equisingular connections as a subcategory of E is obtained by mapping the generators e−n → βn to the nth graded piece of the beta function of the theory. seen as an element β = n βn in the Lie algebra Lie(G). The key step of the proof is showing that. in the form e−k1 · · · e−kn γU (−z.
the Tannakian Galois group U∗ = U Gm obtained as above acts on the coupling constants via its map to the Connes–Kreimer group of diﬀeographisms given by the β function.The Riemann–Hilbert correspondence 149 conjecture. which in turn acts on the coupling constants by the result of [Connes and Kreimer (2000)]. . In fact. The relation between the aﬃne group scheme U∗ and symmetries of multiple zeta values can be seen via the fact that the same group appears as a motivic Galois group for a category of mixed Tate motives via the result of [Deligne and Goncharov (2005)].
151 . 7. The second method is based on the use of techniques from noncommutative geometry and was ﬁrst discussed in Chapter 1 of [Connes and Marcolli (2008)] and in the unpublished manuscript [Connes and Marcolli (2005b)]. as we have seen in the ﬁrst chapter of this book. and which makes contact with the motivic notions considered here.1 The motivic geometry of DimReg Let us consider again the mixed Tate motives (motivic sheaves) deﬁned by the logarithmic extension Log of (2.39) and its symmetric powers Logn of (2. We describe in this chapter two diﬀerent possible approaches to a geometric description of dimensional regularization. that would justify the formal procedures used in dimensional regularization. typically no attempt is made to make sense of an underlying geometry in the complexiﬁed dimension. This ﬁrst method was discussed in [Marcolli (2008)]. is deﬁned as a purely formal procedure based on analytic continuation to a complex variable of the integral of a Gaussian in dimension D. in the way we have seen through the previous chapters.Chapter 7 The geometry of DimReg Usually. through the language of mixed Hodge structures.40). This means that. as objects in the triangulated category DMQ (Gm ) of motivic sheaves over Gm . the procedure of dimensional regularization of divergent Feynman integrals. while this provides an eﬃcient recipe for analytically continuing Feynman integrals to a complexiﬁed dimension. which was not reported in [Connes and Marcolli (2008)]. The ﬁrst is based on motivic notions and it applies to individual motives (motivic sheaves) associated to given Feynman graphs. We concentrate here especially on the later part of [Connes and Marcolli (2005b)]. as a function of D.
Lemma 2. Let us consider then the corresponding grading operator. For iterated integrals we use the notation as in [Goncharov (2001)] b a ds ds ds ◦ ◦ ··· ◦ = s s s a≤s1 ≤···≤sn ≤b ds1 dsn ∧ ··· ∧ .4 above).b (n) the domain Λa. (7. the period matrix (2.2) We also denote by Λa. The result follows from the basic identity (cf.10) expressing the powers of the logarithm in terms of iterated integrals. [Goncharov (2001)]. [Goncharov (2001)].s (∞) = ∪n Λ1. (7. s1 sn (7. where the kth entry corresponds to the kth graded piece of the weight ﬁltration. s1 sn (7.3) Lemma 7.5) Proof.1) is obtained as rational multiples of the pairing sT = Λ1.b (n) = {(s1 .4) with Λ1. one can associate to the graph .g. k! (7.1.s (n) and the form η(T ) := n dsn n ds1 ∧ ··· ∧ T . e.1) has in fact an interpretation in terms of periods.1. the H 0 piece corresponds to the ﬁrst column of the matrix MLogn . . .48) for the promotive Log∞ of (2. in terms of the motivic sheaves deﬁned as in [Arapura (2008)] (see the discussion in §2. The expression (7. (7. This follows from a well known result (cf. that multiplies the kth entry by T k .b (n) Observe then that.6) Λa.10) b log a ds1 dsn ∧ ··· ∧ = s1 sn n! n .s (∞) η(T ).47) deﬁnes the mixed Hodge structure associated to the motive. Lemma 2. . .48) to the H 0 part in the MHS realization: Q· k logk (s) k T =: Q · sT . In this Hodge realization.152 Feynman Motives As we have seen. One can then associate to the h0 piece of the Log∞ motive the following formal expression that corresponds in the period matrix (2. sn )  a ≤ s1 ≤ · · · ≤ sn ≤ b}.1) The formal expression (7.
(7. as in the construction of the logarithmic specialization system given in [Ayoub (2007)]. Proof. The category of Feynman motivic sheaves. respectively. Σ) and (X . for the product of motivic sheaves deﬁned by (2. One then has. Σ ). n − 1).11) where f : σ → S and f : σ → S are the restrictions of the maps X → S and X → S. This gives a motive LogΓ := Ψ∗ (Log) ∈ DM(UΓ ). is the subcategory of the Arapura category of motivic sheaves over Gm spanned by the objects of the form (ΨΓ : An ˆ ˆ XΓ → Gm . Deﬁnition 7. In terms of periods.9) We can then consider the pullback of the logarithmic motive Log ∈ DM(Gm ) by this morphism. n − 1. (7.2. by considering the morphism ΨΓ : A#EΓ ˆ XΓ → Gm (7. and where Σn = {t ∈ An  i ti = 0} is the union of the coordinate hyperplanes. deﬁning relative homology cycles for (X. (7.23).12) .1. Σn ˆ ˆ (Σn ∩ XΓ ).10) ˆ where UΓ = A#EΓ XΓ . Σn ˆ ˆ (Σn ∩ XΓ ). the period pairing ∗ ∗ πX (ω) ∧ πX (η) = σ×S σ σ ω ∧ f ∗ f∗ (η).7) deﬁned by the graph polynomial ΨΓ (s) = det(MΓ (s)).1.3. Lemma 7. (7. for the ﬁber product (2. one has the formula ∗ ∗ πX (ω) ∧ πX (η) = X×S X X ∗ ω ∧ (πX )∗ πX (η) = X ω ∧ f ∗ f∗ (η). First recall that. The above correspond to the local systems n−1 HGm (An ˆ ˆ X Γ . Γ (7.8) with n = #Eint (Γ) and where Γ ranges over the Feynman graphs of the ˆ given scalar ﬁeld theory.The geometry of DimReg 153 hypersurfaces XΓ and the divisors Σn with ∂σn ⊂ Σn objects in the Arapura category of motivic sheaves over Gm . Q(n − 1)). Suppose given σ ⊂ X and σ ⊂ X . when integrating a diﬀerential form over a ﬁber product. for a ﬁxed scalar quantum ﬁeld theory.23) one has the following.
12) to the integration on σ ×S σ and obtain (7. Proposition 7. of the Feynman motive (7. 0.14) by the (pro)motive ∞ (XLog → Gm .11). (7.15) where Λ∞ is such that the domain of integration Λ1.4.1 deﬁnes a cycle. in the category of motivic sheaves enlarged to include projective limits. Σ ×S X ∪ X ×S Σ ). The dimensionally regularized Feynman integrals are periods on the product. (7. One can then take the product of a Feynman motive ˆ ˆ (ΨΓ : An XΓ → Gm . The product is then given by a ﬁber product as in (2. n − 1. the procedure of dimensional regularization can then be thought of as follows.1.23). so that σ ×S σ deﬁnes a relative homology class in (X ×S X . Σ) and [η] ∈ HS (X .14) by the logarithmic promotive Log∞ of (7.t (∞) of the period computation of Lemma 7. One has ∂(σ ×S σ ) = ∂σ ×S σ ∪ σ ×S ∂σ ⊂ Σ ×S X ∪ X ×S Σ . Σ ).13) X II X II f tt f tt II II tt II tt $ ztt S Suppose then given σ ⊂ X such that ∂σ ⊂ Σ and σ ⊂ X with ∂σ ⊂ Σ . n − 1). namely Ψ∗ (Log∞ ) = (An Γ An ˆ XΓ ) Gm XLog∞ / XLog∞ (7. viewed itself as a motivic sheaf XLog∞ → Gm over Gm . we then apply the formula (7. In these terms. .15). Λ∞ . · · Given elements [ω] ∈ HS (X.154 Feynman Motives which corresponds to the diagram X ×S X II v II vv II v v πX II vv πX I$ v zv (7. Consider again the logarithmic (pro)motive.1. 0).16) / Gm ˆ XΓ ΨΓ We then have the following interpretation of the dimensionally regularized Feynman integrals. Σn (Σn ∩ XΓ ).
11).2 The noncommutative geometry of DimReg We now discuss a diﬀerent approach to the geometry of DimReg.1. with the two projections πX : Ψ∗ (Log∞ ) → An Γ ˆ XΓ πL : Ψ∗ (Log∞ ) → XLog∞ Γ ∗ ∗ and the form πX (α) ∧ πL (η(T )). Consider the product (7.1 gives Ψ∗ Γ Λ1. The period computation of Lemma 7. by (7.1) with the complex DimReg variable z. H. with A an associative involutive algebra represented as an algebra of bounded operators on a Hilbert space H. Γ This is the dimensionally regularized integral. It makes sense. where α is the form involved in the parametric Feynman integral.s (∞) η(T ) = Λ1. see [Connes (1995)]. fractals. . γ 2 = 1. Then the spectral triple is ﬁnitely summable if. such as quantum groups. D). on the Hilbert space H satisfying [a. These consist of data of the form X = (A. Assume for simplicity that the operator D is invertible. Tr(Ds ) < ∞. for some real s. and with the property that the commutators [a. ∗ ∗ πX (α) ∧ πL (η(T )) = σn ×Gm Λ1. with compact resolvent. for a wide range of examples that are not ordinary manifolds. based on [Connes and Marcolli (2005b)]. etc. the operator Ds is of trace class. using spectral triples and noncommutative geometry and we discuss its points of contact with motivic notions.4).The geometry of DimReg 155 Proof. n! (7. This structure generalizes the data of a compact Riemannian spin manifold. γ] = 0 and Dγ + γD = 0. D] are bounded operators on H. A spectral triple is even if there is a linear involution.ΨΓ (t) (∞) η(T ) = n log(ΨΓ (t))n n T = ΨΓ (t)T . however. with the (commutative) algebra of smooth functions.ΨΓ (t) (∞) σn ∗ α ∧ (πX )∗ πL (η(T )) = σn ΨT α.17) We then have.16). noncommutative tori. up to replacing the formal variable T of (7. 7. and η(T ) is the form on XLog∞ that gives the period (7. for all a ∈ A. together with a selfadjoint operator D on H. the Hilbert space of square integrable spinors and the Dirac operator. The notion of metric space in noncommutative geometry is provided by spectral triples.
156 Feynman Motives For such spectral triples there are various diﬀerent notions of dimension. [Connes and Marcolli (2008)] that there exists a (type II) spectral triple Xz with the properties that the dimension spectrum is Dim = {z} and that one recoves the DimReg prescription for the Gaussian integration in the form Tr(e−λDz ) = π z/2 λ−z/2 . Rennie. the analog of a volume form. D ⊗ 1 + γ ⊗ Dz ). Sukochev (2006a)] and [Carey. H ⊗ Hz . D) = (C ∞ (X). It is shown in [Connes and Marcolli (2005b)]. given in terms of a residue for the zeta functions. Phillips. with the spectral measure Tr(χ[a.b] (Z)) = 1 2 [a. The most sophisticated one is the dimension spectrum which is not a single number but a subset of the complex plane consisting of all poles of the family of zeta functions associated to the spectral triple. which is indeed of the form as in the cup product of spectral triples. The technical aspects of the diﬀerence between ordinary spectral triples and type II spectral triples are beyond the purpose of this book. Sukochev (2006b)]. This agrees with what is known in physics as the Breitenlohner–Maison prescription to resolve the problem of the compatibility of the chirality operator γ5 with the DimReg procedure [Breitenlohner and Maison (1977)]. Rennie. The ordinary spacetime over which the quantum ﬁeld theory is constructed can itself be modeled as a (commutative) spectral triple X = (A. D ⊗ 1 + γ ⊗ Dz . especially [Carey. These are points where one has a well deﬁned integration theory on the noncommutative space. Hz . for the type II trace. The Breitenlohner–Maison prescription consists of changing the usual Dirac operator to a product.b] dt. H. S). where Z = F Z is a selfadjoint operator aﬃliated to a type II∞ von Neumann algebra N and ρ(z) = π −1/2 (Γ(1 + z/2))1/z . Dz ) = (A ⊗ Az . DX ) / and one can take a product X × Xz given by the cup product of spectral triples (adapted to the type II case) (A. D) ∪ (Az . 2 . and we simply refer the reader to the relevant literature. Dim = {s ∈ Cζa (s) = Tr(aD−s ) have poles }. The operator Dz is of the form Dz = ρ(z)F Z1/z . L2 (X. H. Phillips.
a) da. namely an interesting formal analogy between the evanescent gauge potentials in the noncommutative geometry of DimReg and the complexes of vanishing and nearby cycles in algebraic geometry.g. r = 1 ∈ Q∗ . D). Here we concentrate on a related aspect that brings us back to motivic notions. The type II∞ factor obtained in this way is dual (in the sense of the duality introduced in [Connes (1973)]) to the type III1 factor associated to the KMS state at “critical temperature” β = 1 on the Bost–Connes system.The geometry of DimReg 157 It is shown in [Connes and Marcolli (2005b)] and [Connes and Marcolli (2008)] that an explicit example of a space Xz that can be used to perform dimensional regularization geometrically can be constructed from the ad`le e class space. by taking the crossed product of the partially deﬁned action ˆ N = L∞ (Z × R∗ ) GL1 (Q) and the trace TrN (f ) = with the operator Z(1. (H ⊗ Hz . so that one has a chirality operator γ. ρ. ˆ Z×R∗ . which may correspond. H. This setting for the geometrization of DimReg provides a way of computing the anomalous graphs of quantum ﬁeld theory in terms of index cocycles in noncommutative geometry. λ) = 0. and Chapter 3 of [Connes and Marcolli (2008)]. which is the usual γ5 in dimension four. to the ordinary spacetime over which a given quantum ﬁeld theory is considered. λ) = λ. Then consider ˜ the algebra A generated by A and γ. [Connes (1999)] and Chapters 2 and 4 of [Connes and Marcolli (2008)]). as explained in [Connes and Marcolli (2005b)] and Chapter 1 of [Connes and Marcolli (2008)].18) We consider then the spectral triple where the Hilbert space and Dirac operator are given by the product with Xz . D ⊗ 1 + γ ⊗ Dz ) f (1. the noncommutative space underlying the spectral realization of the Riemann zeta function in noncommutative geometry (see e. We assume that it is an even spectral triple. [Connes and Marcolli (2006a)]. endowed with a Z/2Z grading with deg(a) = 1 for a ∈ A and deg(γ) = −1. Z(r. a]− := Da − (−1)deg(a) aD (7. The usual commutators are then replaced by graded commutators [D. ρ. Suppose given a spectral triple (A. as above. see [Bost and Connes (1995)].
which we have already described in a very similar form in [Consani and Marcolli (2006)] in the context of the archimedean cohomology of [Consani (1998)] and its relation to noncommutative geometry. we take the point of view of Saito’s polarized Hodge–Lefschetz modules. The lack of invariance of D under the chiral gauge transformation γ gives rise to new evanescent gauge potentials coming from the presence of a nontrivial commutator ˆ [D . e To describe the analogous structure associated to the complex of gauge potentials in the case of a noncommutative space and its deformation to complexiﬁed dimension. On the algebrogeometric side. D ) obtained by taking the cup product of a spectral triple (A. We use an approach based on polarized Hodge–Lefschetz modules. We assume that the map f is smooth on X∗ = X Y and that Y is a divisor with normal crossings in X.19) as in [Connes and Marcolli (2005b)]. in the algebrogeometric case. we consider the noncommutative spaces (A . [Steenbrink (1976)].20) which disappears when one is back in the case z = 0. ¯ D = D ⊗ 1. we consider the case of a geometric degeneration of a family X of smooth algebraic varieties over a disk ∆ ⊂ C. . D ) behaves in many ways like a complex of forms with logarithmic poles associated to the family X of smooth algebraic varieties over the disk z ∈ ∆. [Guill´n and Navarro Aznar (1990)]. We shall see that the complex of gauge potentials on (A . acting by a ⊗ 1. formulated as in [Consani and Marcolli (2006)]. it is known that the special ﬁber f −1 (0) of f : X → ∆ carries a mixed Hodge structure. We use the notation D = Dz . H . with the algebra A = A as above. D) with a noncommutative space Xz in “complexiﬁed dimension” ˜ z ∈ ∆. (7. (7.158 Feynman Motives ˜ and the algebra is A. [Consani and Marcolli (2004)]. ˆ D =γ⊗D ¯ ˆ D = D + D. [Saito (1988)] and [Guill´n and Navarro Aznar e (1990)]. H. proper morphism with projective ﬁbers. γ]− = 2γ D. These evanescent gauge potentials ﬁgure prominently in the computations of anomalies in [Connes and Marcolli (2005b)]. In particular. The two geometric settings we will compare are the following. H . We now show that these evanescent gauge potentials also give rise to a complex reminiscent of vanishing cycles. Here X is a complex analytic manifold of dimension dimC X = n + 1 and f : X → ∆ is a ﬂat. On the side of noncommutative geometry.
2 ) deﬁne a bigraded Hodge–Lefschetz module if all the Li. 2) i 1 deﬁne a bigraded Lefschetz module if → Li. 2 ) is a bilinear form ψ :L⊗L→R which is compatible with the Hodge structure. u) = 2.j . Bigraded Lefschetz modules correspond bijectively to ﬁnite dimensional representations of SL(2. Let 1 .) The data (L. 1 . satisﬁes ψ( k x.j+2 .j → Li.24) = 0. σ(χ(λ).j (7.22) :L are isomorphisms for i > 0 and j > 0. dσ(u. y) + ψ(x.j 2 : Li. In the case of a geometric degeneration. w := 0 1 −1 0 . −i. 2 . 2) In these terms. and u= 01 00 = d u(s)s=0 .j and : Li. one can obtain such a structure from the action of the a Lefschetz operator on the primitive part of the cohomology (induced by wedging with the K¨hler form) and of the monodromy: the Lefschetz and a the log of the monodromy give the endomorphisms 1 . . 2 ] = 0 and 1 : Li. 1) = 1. 2. ∀x ∈ Li.26) is symmetric and positive deﬁnite on L−i. respectively.j have a real Hodge structure and 1 . C i j 1 2 ·) k y) (7. (Here C is the Weil operator. (7.j . ds 1. 1 . 2 . The data (L.j → Li+2.The geometry of DimReg 159 Let L = ⊕i. 1. j 2 (7. (7.−j → Li. with the generic ﬁber a compact K¨hler manifold. with [ 1 .j be a ﬁnite dimensional bigraded real vector space. and is such that ψ(·. A polarization on (L. χ(t)) = λi tj x. 2 be endomorphisms of L.j∈Z Li.21) The data (L.23) for λ ∈ R∗ and s ∈ R. u(s) := 1s 01 . ψ) then deﬁne a bigraded polarized Hodge–Lefschetz module. 2 are morphisms of Hodge structures. We will use the following notation: χ(λ) := λ 0 0 λ−1 . R). 1 . satisﬁes dσ(1. the representation σ = σ(L.−j . R) × SL(2. k = 1.25) (7.
We let ΩD (A) ∗ denote the complex of gauge potentials of a triple (A. w)−1 ◦ d ◦ σ(w. Let γ ˜ be the grading operator. we have ˜ A = A.r. (7. 2 ) .k and δ : Ωm. where d : Li. R) × SL(2. A]. H ∗ is identiﬁed with Ker( ). ∗ where as above Dz is the Dirac operator on the space Xz . We begin now by considering the complex Ωm. Namely. where d∗ = σ(w. for an element D a ∈ A or a ∈ [D. where = d∗ d + dd∗ .g. y) = ψ(x.23) and σ = σ(L. and HD (A) its cohomology. H . while we set (aγ) := [D2 .k → Ωm. R).k .r+1. H. A]. dy). Here we deﬁne (a) = [D2 . d). d] and ψ(dx. the cohomology H ∗ (L. We consider the graded algebra A generated by A and by γ. a] as before. 1 .r.r.k . for all a ∈ A or a ∈ [D. H. We let (A. and D = Dz = D ⊗ 1 + γ ⊗ Dz .j → Li+1. D) with the noncommutative space Xz in “complexiﬁed dimension” z.160 Feynman Motives It is also convenient (e. ψ. with diﬀerentials δ : Ωm. D) be an even ﬁnitely summable spectral triple. H.k to be the span of elements of the form k H =H⊗H (ω)D2r . d) inherits the structure of a bigraded polarized Hodge–Lefschetz module.r. 2 .r. D ) be the noncommutative space obtained as the ˜ product of (A. when working at the level of forms) to consider diﬀerential bigraded polarized Hodge–Lefschetz modules (L. We now introduce an analog of this structure for noncommutative spaces. Moreover. Let (A . D). the representation of SL(2. w).j+1 is a diﬀerential (d2 = 0) satisfying [ 1 . 1 .27) ˜ with ω ∈ Ωm (A). d] = 0 = [ 2 . Here we take Ωm. a]γ. .k → Ωm+1. for w as in (7. In this case.
when decomposing D = D ⊗ 1 + γ ⊗ Dz .r. d = δ + δ ) deﬁned as follows: 1( k (ω)D2r ) = ( k k (ω)D2r ) = −1 −1 k k k+1 (ω)D2(r−1) (7. 2 ] = 0 and are compatible with the diﬀerential. d] = 0 and [ 2 . We introduce endomorphisms 1 : Ωm.k → Ωm+2. a]. resembles very closely the case of geometric degenerations.30) The index of the resulting ﬁltration is i−k. we can endow the complex Ωm. where one also has a total diﬀerential d = δ + δ . ˆ where for simplicity we write θ for the term θ = γ D.r. namely.The geometry of DimReg 161 ˜ We consider a descending ﬁltration · · · ⊃ F p ⊃ F p+1 ⊃ · · · on Ω∗ (A) D deﬁned by setting ˜ ˜ F p Ωm (A) = ⊕t+s=m. with ω ∈ Ωt (A) and θ = γ D.29) In this way. D D D We take ˜ F m+r−k Ωm (A) = ⊕t+s=m. .t≥p Ωt.k with the conditions k ≥ 0 and k ≥ 2r + m.k → Ωm. where i = 2r+m. Notice that the decomposition of the total diﬀerential d = δ + δ . D D (7.1. [Steenbrink (1976)]). hence we then take the Ωm. We then have the following result.s (A).r−1.32) (ω θ2 ) D2(r+1) . where the second is 1dimensional with FD (C) := −(r+1) C · D2r and FD (C) = 0.r. for f : X → ∆ and z the coordinate on the base ∆ (cf. so ¯ ˆ that δ a = [D. D D (7. where δ is essentially the original de Rham diﬀerential on Ω∗ (A) and δ acts by D ˆ wedging with the diﬀerential θ = γ D.r+1. D (7.r. Lemma 7. with δ the usual de Rham diﬀerential and δ given by wedging with the form θ = f ∗ (dz/z).k = F m+r−k Ωm (A) ⊗C FD (C).28) ˜ ˆ where Ωt.k of (Ω.r.s (A) = Ωt (A)θs is the span of ω θs . The endomorphisms 1 and 2 satisfy [ 1 .2.k with a tensor product of two −r “Hodge structures”. a] and δ (a) = [D. d] = 0.31) 2( (ω)D2r ) = = √ √ (ω) ∧ θ2 D2(r+1) (7. induced by da = [D . a] for a ∈ k (A).s+r≤k Ωt (A) θs . [ 1 . The diﬀerential d = δ + δ on this complex is the same as described above.k+1 and 2 : Ωm. We have −r ˜ Ωm.
[D. 2 ] = 0. H. [D. We introduce an SL(2. We have [ 1 . 2 induce endomorphisms on the cohomology of the double complex. 2 (γb)]. (A).2. σ2 (χ(λ)) = λn−m . −[D. [D2 . we can check the result on elements of the form k (a)D2r or k (γa)D2r for a ∈ A. and on elements γbD2r . Assume that the spectral triple (A. k (a)]]D2(r−1) k+1 k (a)]D2(r−1) = 0. [D. b](γ D)2 . γb]] = [D2 . we verify that [ 2 . R) × SL(2. We have ( 1δ − δ 1) k (a)D2r = [D2 . We check compatibility with the diﬀerential. [D. where 2 [D. where we use the fact that we have set (γa) := γ (a) and we get ˆ ˆ ˆ [D2 . γ[D2 . d] = 0.33) (7.162 Feynman Motives Proof. b]] = 2[D2 .34) 2 ). . b(γ D)2 ] = −[D. σ1 (u(s)) = exp(s σ2 (u(s)) = exp(s 1 ). 1 . σ1 (w) = S1 . γ (a)]]D2(r−1) k+1 (a)]D2(r−1) = 0. since ˆ ˆ [D. b]γ D. σ2 (w) = S2 . This can be seen easily on elements of the form bD2r . (7. γ ˆ −[D. d] = 0. Deﬁnition 7. Thus. Similarly. Consider σ1 and σ2 deﬁned by σ1 (χ(λ)) = λ2r+m . b]]. b]] = −[D. which is associated to the operators 1 . 2γbD] = 2[D2 . We also have k since [D2 . In fact. γb] ˆ √ ˆ = 2 −1 bθ3 = −[D. R) representation (σ1 . D) is N summable with N = 2n. for all b ∈ ( 1δ − δ 1) k ˆ (γa)D2r = [D2 . It is immediate to verify that [ 1 . σ2 ) on the complex introduced above.2. 2 . b]γ D and ˆ ˆ [D.
k of the form ˆ S1 ( ˆ S2 ( k in the following (ω)D2r ) = k −(2r+m) ( 1 k k (ω)D2r ) (7. 2. k σ2 (χ(λ)) σ2 (u(s)) σ2 (χ(λ−1 )) σ2 (χ(λ)) 1 + s 2 (ω)D2r = k + s2 2 2 2 + ··· s2 2 k λn−m 2 n−m 2λ (ω)D2r = k 1 + λ−n+m+2 s 2 λn−m + λ−n+m+4 exp(sλ2 2) + ··· (ω)D2r = (ω)D2r .3.r. d). way.33) and (7. (7. R) it is suﬃcient ([Lang (1975)] §XI. for ω ∈ ΩD (A). deﬁned by (7. We set S1 = −1 S These satisfy the following property.k → Ωm. for k = 1. We consider involutions ˆ S1 : Ωm.2. Proposition 7.−(r+m).2) to check that it satisﬁes the relations σ(w)2 = σ(χ(−1)).35) (ω)D2r ) = n−m ( 2 √ m m ˜ ˆ1 .r−(n−m). (ω)D2r ). with ω ∈ Ωm (A).2.k−(2r+m) ˆ S2 : Ωm.The geometry of DimReg 163 2. R) on (Ω. . where we have D σ1 (χ(λ)) σ1 (u(s)) σ1 (χ(λ−1 )) σ1 (χ(λ)) 1 + s 1 k (ω)D2r = k + s2 2 2 1 + ··· λ−(2r+m) s2 2 (ω)D2r = + ··· k 1 + λ2(r+1)+m s 1 λ−(2r+m) + λ2(r+2)+m = exp(sλ2 and 1) 2 −(2r+m) 1λ (ω)D2r k (ω)D2r . In order to show that σ = σk . The data speciﬁed in Deﬁnition 7.34) is indeed a representations of SL(2.r. The ﬁrst relation is clearly satisﬁed and the second can be veriﬁed easily on elements k (ω)D2r . R) × SL(2.k → Ω2n−m.36) σ(χ(λ))σ(u(s))σ(χ(λ−1 )) = σ(u(sλ2 )).2 deﬁne a representation of SL(2. Proof. Here the operators S1 and S2 are deﬁned by powers of 1 .
For a = k (η)θs D2r and b = k+1 k (η )θs D2r . = −γ k (η) k +1 (η ) D2(r+r −1) .r.25). We also have the following result.164 Feynman Motives Notice that. b) + ψ(a. 2 (b)) = −γ k √ (η ∗ ) −1 2 (b)) k (η ) D2(r+r +1) . b) = −γ while ψ(a. Lemma 7.4. The bilinear form (7. 1 (b)) (η ∗ ) (η ) D2(r−1+r ) . Thus. by setting Li. in the case of the representation associated to the “Lefschetz” 2 . where the Hodge can be realized (on the primitive cohomology) by the (n − m)th power of the Lefschetz operator.37) satisﬁes the relation (7. the analog of the Hodge on forms. 1 (b)) = 0. b) = −γ(− −1 k (η ∗ )) k (η ) D2(r+1+r ) . which appears in σ1 (w). = 0. Deﬁne the bilinear form ψ by setting ψ( k (ω)D2r .26) for polarizations of Hodge–Lefschetz modules. We now discuss the polarization. Proof. We also have √ ψ( 2 (a). is realized by a power of 2 . ia . b) + ψ(a. This deﬁnes a structure that is analogous to the diﬀerential bigraded Hodge–Lefschetz modules. we have k ψ( 1 (a).2. analogous to the requirement (7. we also have ψ( 2 (a).37) where ω = ηθs and ω = η θs . Thus. by analogy to what happens in the classical case. This construction parallels exactly what happens in the construction of the archimedean cohomology of [Consani (1998)] for the ﬁbers at inﬁnity of arithmetic varieties. in the form presented in [Consani and Marcolli (2006)].j = ⊕k Ωm. k (ω )D2r ) := −γ k (η ∗ ) k (η )D2(r+r ) (7. We work here under an assumption of “tameness” for spectral triples deﬁned in [V´rilly and Graciaa Bond´ (1993)]. and ψ(a. integration by parts gives ψ( 1 (a).k with i = 2r + m and j = −n + m.
The geometry of DimReg 165 Lemma 7.37) has the property that a∗ . i j 1 2 b) agrees on L−i. we have j 2 (b) = k (η )θs +2j D2r+2j . Proof. the local monodromy plays an important role in determining the limiting mixed Hodge structure.−j . b := ψ(a. it agrees with the inner product on the complex of gauge potentials. C) of the Milnor −1 ﬁber Mz := BP ∩ f (z). We then have i j 1 2 (b) = k +i (η )θs +2j D2r+2j−2i . BP a small ball around P . the diﬀerence between the cohomology of the generic ﬁber and of the special ﬁber is measured by the van˜ ishing cycles. The nearby cycles span the cohomology H ∗ (Mz . deﬁned for P ∈ Y . for a = k (η)θs D2r in L−i. C). and a suﬃciently small z ∈ ∆∗ . Recall that a := −aD−2n is the integration with respect to the volume form D−2n of the spectral triple (A. In fact. Under the assumption of “tameness” this satisﬁes ab = ba and a∗ a ≥ 0. so that a∗ .2. For b = k (η )θs D2r in L−i. These span the reduced cohomology H ∗ (Mz . we obtain i j 1 2 (b)) ψ(a. Thus. The bilinear form (7. b agrees with ˜ the inner product of forms in Ω∗ (A). one usually considers all choices by passing to the universal .−j with the integral with respect to the volume form of (A. H. This gives i j 1 2 (b) = k +2r+m (η )θs +2(m−n) D−2r−2n . To eliminate the noncanonical dependence of everything upon the choice of z. D The fact that the structure described here in terms of Hodge–Lefschetz modules parallels very closely the construction of the archimedean cohomology of [Consani (1998)]. in the form presented in [Consani and Marcolli (2006)].−j . H. In the algebrogeometric setting of a degeneration over a disk. D). D).5. H. geometrically. D). = −γ k (η ∗ ) k +2r+m (η )D−2n . suggests that the formalism of DimReg via noncommutative geometry may be useful also to describe a “neighborhood” of the ﬁbers at inﬁnity of an arithmetic variety. Under the assumptions of “tameness” for (A.
Ω· (log Y ) ⊗OX OY ) X/∆ (7. The analogy described above with the structure of polarized Hodge– Lefschetz modules allows one to think of the operator Θ(a) = (−1)n+1 n n=1 ∞ n (a)D−2n .39) is the derivation that appears in [Connes and Moscovici (1995)] in the context of the local index formula for spectral triples. The identiﬁcation (cf. was used ([Steenbrink (1976)]. when working with nearby cycles. and √ with the induced action of (2π −1) 1 on the cohomology corresponding to the residue of the Gauss–Manin connection in the algebrogeometric setting. C).38) shows that. C) determ ˜∗ · · mined by (H (X . with 1 satisfying eΘ = 1 + 1 . The operator Θ of (7. with an explicit resolution of the complex. (7. L· . where F is the Hodge ﬁltration and L· is the Picard–Lefschetz ﬁltration associated to the local monodromy. [Guill´n and Navarro e ˜ Aznar (1990)]) to obtain a mixed Hodge structure on H m (X∗ . C) → Hm (Y.166 Feynman Motives ˜ ˜ ˜ cover ∆∗ = H of the punctured disk and replacing Mz by X∗ = X ×∆ ∆∗ . one can use a complex of forms with logarithmic diﬀerentials. . F ). This approach. [Steenbrink (1976)]) ˜ H m (X∗ .39) as the logarithm of the local monodromy.
g. 8. Among various techniques introduced for the study of nonisolated singularities. a common procedure consists of cutting the ambient space with linear spaces of dimension complementary to that of the singular locus of the hypersurface (cf. The parametric Feynman integrals can in turn be expressed in terms of this slicing. the graph hypersurfaces XΓ are typically singular with nonisolated singularities and a singular locus of small codimension. It is given as in (3.1 Projective Radon transform We recall the basic setting for integral transforms on projective spaces (cf.Chapter 8 Renormalization. This corresponds to cases where the singular locus is of (complex) codimension one in the hypersurface. and Hodge structures As we have seen.42) by 167 . In many cases one ends up slicing the graph hypersurfaces by planes P2 intersecting the hypersurface into a curve with isolated singular points. the restriction of the function deﬁning the hypersurface to these linear spaces deﬁnes hypersurfaces with isolated singularities. and we use the resulting integrals on the slices as a way to relate the Feynman integrals to Hodge structures for isolated singularities. Gindikin. the slicing is given by 3dimensional spaces cutting the hypersurface into a family of surfaces in P3 with isolated singularities. using projective Radon transforms. [Teissier (1975)]). §II of [Gelfand. When the singular locus is of codimension two in the hypersurface. e. We follow here the approach outlined in [Marcolli (2008)]. On any kdimensional subspace Ak ⊂ An there is a unique (up to a multiplicative constant) (k − 1)form that is invariant under the action of SLk . singularities. In this case. and Graev (1980)]). to which the usual invariants and constructions for isolated singularities can be applied.
5). (8.168 Feynman Motives the expression k Ωk = i=1 (−1)i+1 ti dt1 ∧ · · · ∧ dti ∧ · · · ∧ dtk . that is. dt ∧ Ωξ = Ωn . λ ∈ Gm .e.1. one can consider the integrand f Ωξ and deﬁne its integral over the projective space π(Πξ ) ⊂ Pn−1 as above. dt ∧ · · · ∧ ξn−k . it is convenient to also consider the following variant of the Radon transform (8. n − k}. ∀t ∈ An . . These deﬁne a kdimensional linear subspace Π = Πξ ⊂ An by the vanishing Πξ = {t ∈ An  ξi . (8.1). the (k − 1)dimensional projective Radon transform (§II of [Gelfand.2) Then the integrand f Ωk is well deﬁned on the corresponding projective space Pk−1 ⊂ Pn−1 and one deﬁnes the integral by integrating on a fundamental domain in Ak {0}.6) . .4). Let σ ⊂ An be a compact region that is contained in a fundamental domain of the action of Gm on An {0}.1) is homogeneous of degree k. This deﬁnes the integral transform. (8.3) Given a choice of a subspace Πξ . for i = 1. Then. The partial (k − 1)dimensional projective Radon transform is given by the expression n−k Fσ. i = 1. . given a function f on An with the homogeneity condition (8. on a surface that intersects each line from the origin once. . The form Ωξ is not uniquely deﬁned on An . t ) Ωξ (t). there exists a (k − 1)form Ωξ on An satisfying ξ1 . . (8. (8.2). cf. n − k.42). Suppose given a function f on An which satisﬁes the homogeneity condition f (λt) = λ−k f (t). but its restriction to Πξ is uniquely deﬁned by (8. .4) with Ωn the (n − 1)form of (3.5) For our purposes. t = 0.1. . t ) Ωξ (t). where one identiﬁes σ with its image π(σ) ⊂ P . (8. and Graev (1980)]) as n−k Fk (f )(ξ) = π(Πξ ) f (t) Ωξ (t) = Pn−1 f (t) i=1 δ( ξi . (8. Deﬁnition 8.1) The form (8. . Gindikin. i. Suppose given linearly independent dual vectors ξi ∈ (An ) .k (f )(ξ) = σ∩π(Πξ ) f (t) Ωξ (t) = σ∩π(Πξ ) f (t) i=1 n−1 δ( ξi .
12) . dt is a shorthand notation for ξ.k (f ) is the Radon transform. ΨΓ (t)D/2 (8. (8.11) We then follow the same procedure we used in the derivation of the parametric form of the Feynman integral.2) and the integrand ΨΓ (t)D/2 V ωξ (t) k−D Γ (t. ti ≥ 0.1. The parametric Feynman integral can be reformulated as U (Γ) = Γ(k − D ) 2 (4π) D/2 Fσn .p) dt1 · · · dtn . dt ∧ ωξ = ωn = dt1 ∧ · · · ∧ dtn . dt and ωξ satisﬁes ξ. p)−k+D ΨΓ (t)D/2 /2 . and with fΓ (t) = Proof. with σn the simplex i ti = 1. VΓ (t. dt .10) where ξ. p)k−D Πξ i χ+ (t) (8. applied to the integral over Πξ and we write it in the form e−VΓ (t. dt .k (fΓ )(ξ) ξ. Proposition 8.8) satisﬁes the scaling property (8.9) where χ+ (t) is the characteristic function of the domain Rn . + Given a choice of an (n − k)frame ξ. dt = ξ1 . we can then write the Feynman integrals in the form U (Γ) = (4π)− D/2 Πξ χ+ (t) e−VΓ (t. (8. (8. dt ∧ · · · ∧ ξn−k . p) /2 . /2 The function fΓ (t) of (8.Renormalization. singularities.p) ωξ ΨΓ (t)D/2 ξ. and Hodge structures 169 Let us now return to the parametric Feynman integrals we are considering.p) ωξ (t) = ΨΓ (t)D/2 Πξ D ωξ (t) Γ(k − ) δ(1 − ti ) 2 ΨΓ (t)D/2 VΓ (t.7) where ξ is an (n − k)frame in An and Fσn .2.8) We write the parametric Feynman integral as U (Γ) = (4π)− D/2 An χ+ (t) e−VΓ (t. (8.
for some m ∈ N and some η ∈ Ωk deg(f ) . p) /2 (8. .15) where Πξ is a generic linear subspace of dimension equal to the codimension of the singular locus of the hypersurface XΓ ∪ YΓ . since the form ωξ is homogeneous of degree k. so we continue to restrict only to that case. p) do not contribute to singularities of the diﬀerential form. (8. in terms of the Radon transform. (8.170 Feynman Motives is therefore Gm invariant. 8.14) = σξ ωξ (t) ΨΓ (t)D/2 VΓ (t.k (fΓ )(ξ) = Πξ δ(1 − i ti ) ΨΓ (t)D/2 V ωξ (t) k−D Γ (t. The order of the pole induces a ﬁltration.46). The minimal m such that ω can be written in the m form ω = η/f m is called the order of the pole of ω along the hypersurface X and is denoted by ordX (ω). we will then consider integrals of the form U(Γ)ξ = Fσn . In the following.13) where Fσn . with XΓ = {ΨΓ (t) = 0} and YΓ = {PΓ (t. so that the integral lives on the hypersurface complement of the graph hypersurface XΓ and the zeros of the second graph polynomial PΓ (t.1. that the spacetime dimension D is in the “stable range” where n ≤ D /2. p)k−(D+z) /2 . as in Deﬁnition 8. we can reformulate the integral (8.k (fΓ )(ξ) ξ. and σξ = σn ∩ Πξ .1.12) in projective space. So far we have always made the simplifying assumption. Here the corresponding condition would then be k ≤ D /2. dt . in dealing with the parametric Feynman integrals. p)k−D /2 as well as their dimensional regularizations U(Γ)ξ (z) = σξ ωξ (t) ΨΓ (t)(D+z)/2 VΓ (t. which is also satisﬁed in the original stable range.k (fΓ ) is the Radon transform over the simplex σn . p) = 0} as before. as Γ(k − D ) (4π)− 2 D/2 Fσn . Thus.2 The polar ﬁltration and the Milnor ﬁber Algebraic diﬀerential forms ω ∈ Ωk (D(f )) on a hypersurface complement can always be written in the form ω = η/f m as in (3. the domain σn of integration is contained in a fundamental domain for the action of Gm . Moreover.
We then obtain. where f = ΨΓ and for the index r of the ﬁltration in the range r ≤ 2k − D( + 1)/2. We are assuming that PΓ and ΨΓ have no common factor. Recall that.46) with the total diﬀerential df ω = f dω − m df ∧ ω as in (3. (8. the forms Ωξ (8. Under the generic condition on the external momenta.51). singularities. for generic r. Varying the external momenta p correspondingly varies the polynomial PΓ (t. and Hodge structures 171 called the polar ﬁltration. and in the stable range n ≤ D /2.k k−1 span subspaces Pξ of the polar ﬁltration P r ΩPn−1 of a hypersurface complement U(f ) ⊂ Pn−1 .18) We are considering here the polar ﬁltration on forms on the complement of the hypersurface XΓ deﬁned by ΨΓ = 0.Renormalization. a subspace of Pξ of the ﬁltration P r Ωk−1 . for all r ≤ Pn−1 2k − D( + 1)/2.1. where Sing(X) is the singular locus of the hypersurface X = {f = 0}. we write (8. by a result of [Deligne and Dimca (1990)]. The explict isomorphism is given by the Poincar´ residue map and can e be written in the form [ω] → [j ∗ ∆(ωξ )]. and f = ΨΓ and m = −k + D( + 1)/2.1 above. for generic external momenta. in the case of isolated singularities. Proof. with f as in Proposition 8. We also have that k ≤ D /2. the linear space Πξ cuts the singular locus Sing(X) transversely and the restriction Xξ = X ∩ Πξ has isolated singularities.k ξ and varying p. In this case. One denotes by P r ΩPn ⊂ Ωkn the subspace P of forms of order ordX (ω) ≤ k − r + 1. Proposition 8. +1)/2 (8. The polar ﬁltration P • is related to the Hodge ﬁltration F • by P r Ωm ⊃ F r Ωm . p).16) D/2 k−D /2 ΨΓ VΓ r.16) in the form PΓ ∆(α) = −k+D( fm ΨΓ where α = PΓ −k+D /2 −k+D /2 Ωξ . there is an isomorphism between the cohomology of the Milnor ﬁber Fξ of Xξ and the total cohomology of the Koszul–de Rham complex of forms (3.19) . Suppose then that k = codim Sing(X). for generic ξ. on the de Rham complex of diﬀerential forms on k the hypersurface complement.2. (8.17) ωξ . if k − r + 1 ≥ 0.2. In this case. so that also k − D( + 1)/2 < 0. or P r Ωk = 0 for k − r + 1 < 0.
§6.1.4 below. with r = dim Πξ − 1 and Fξ ⊂ Πξ the Milnor ﬁber of the hypersurface with isolated singularities Xξ = X ∩ Πξ ⊂ Πξ . (8. Then let hξ = h mod Iξ . It then follows from the identiﬁcation (8. §6). We have not discussed so far the dependence of our slicing method upon the choice of the slice. the quotient of the polynomial ring in the coordinates of the ambient projective space by the ideal of the derivatives of f .21) fm where f = ΨΓ and h = PΓ . We then have the following consequence of Proposition 8.2) that. For n ≤ D /2 and for a generic (n − k)frame ξ with n − k = dim Sing(X).16) can be written as h∆(ωξ ) .172 Feynman Motives where j : Fξ → Πξ is the inclusion of the Milnor ﬁber in the ambient space (see [Dimca (1992)]. When f has isolated singularities. Proof.2. the Feynman integrand (8.16) of (8.19) above ([Dimca (1992)].20) where f is the restriction to Πξ of the function of ΨΓ . the form (8.e. the Milnor algebra is ﬁnite dimensional. a basis for the cohomology H r (Fξ ) of the Milnor ﬁber. with r = dim Πξ − 1 is given by elements of the form ωα = tα ∆(ωξ ) .2. the form (8. Let Iξ denote the ideal of derivatives of the restriction f Πξ of f to Πξ .21) deﬁnes a class in the cohomology H r (Fξ ) with r = dim Πξ − 1. Thus. Let M (f ) be the Milnor algebra of f .22) −k+D /2 For a ﬁxed generic choice of the external momenta. i.3.1. fm with tα ∈ M (f )m deg(f )−k . this deﬁnes an element in the Milnor algebra M (f Πξ ).1. where we show that one can include the choice of slices as additional data in the Hopf algebra combinatorics. (8. with X = XΓ .2. and for a ﬁxed generic choice of the external momenta p under the generic assumption of Deﬁnition 3.14) deﬁnes a cohomology class in H r (Fξ ). This will be treated in §8. so as to consider all slices simultaneously. which lies in the homogeneous component M (f Πξ )m deg(f )−k . One denotes by M (f )m the homogeneous component of degree m of M (f ). with a . for m = −k + D( + 1)/2. in the case of isolated singularities.2. in the stable range for D. (8. By Proposition 8. Corollary 8.
and Hodge structures 173 combinatorial organizing principle. C).λ ( ) ∈ H k−1 (F . where δ( ) ∈ Hk−1 (F . .λ λ α . with k = dim Πξ − 1 and with µ the Milnor number of the isolated singularity.3 DimReg and mixed Hodge structures We assume here to be in the case of isolated singularities.λ . C). there exist classes α ηr. A holomorphic kform α = hωξ /f m determines a section of the cohomological Milnor ﬁbration by taking the classes α F ∈ H k−1 (F . as above.24) df r! λ. It is also worth pointing out that the construction described in this section can also be formulated without slicing.23) df We then have the following results ([Arnold. where the Milnor ﬁber F of Xξ is homotopically a bouquet of µ spheres S k−1 .Renormalization. C) (8. The cohomological Milnor ﬁbration has ﬁber over given by the complex vector space H k−1 (F . GuseinZade. and then ranging over all choices of external momenta to get a subspace can be performed without the slicing operation. ﬁnding the class of the numerator modulo these derivatives.25) (8. Z). the operations of taking the ideal of derivatives of f . Varchenko (1988)]. possibly after replacing the original Feynman integrals with their slices along planes Πξ of dimension complementary to that of the singular locus of the hypersurface. (8. notice that in terms of the Radon transform decomposition (8.10) only the general slices matter as the special slices (for instance those where Πξ is contained in the hypersurface XΓ ) form a set of measure zero in the space of (n − k)frames.λ ( ).24) are given by α ηr. (8.23) with a locally constant section of the homological Milnor ﬁbration has an asymptotic expansion ar.r for → 0. Vol. In particular. δ( ) = ar. Moreover. singularities.λ of (8.II §13). The asymptotic formula for the Gelfand–Leray functions implies that the function of obtained by pairing the section (8.δ ∼ log( )r . Namely. What is missing in this case is the interpretation of the resulting space as the cohomology of a Milnor ﬁber. 8.26) such that the coeﬃcients ar.
λα . but the Hodge ﬁltration is diﬀerent.λ α ηr. One denotes it by λα .174 Feynman Motives Thus. so that we can consider.λ = 0. arbitrarily large values of D ∈ N. This inherits a Hodge and a weight ﬁltration from the Milnor ﬁber cohomology with its asymptotic mixed Hodge structure. determine a subspace of the cohomology H k−1 (Fξ . C) deﬁned by F r = {[α/df ]  λα ≤ k − r − 1} k−1 (8. We show that.29) where N is the nilpotent operator given by the logarithm of the unipotent monodromy. in a neighborhood of an isolated singular point of XΓ ∩Πξ . upon varying the choice of the external momenta p and of the spacetime dimension D.28) (8. The principal part of σ(α) is then deﬁned as σmax (α)( ) := where one knows that α α ηr. (8. The asymptotic mixed Hodge structure on the ﬁbers of the cohomological Milnor ﬁbration constructed by Varchenko ([Varchenko (1980)].λα + · · · + log( )k−1 α η (k − 1)! k−1. the corresponding Feynman integrands. though the two agree on the graded pieces of the weight ﬁltration.λ . [Steenbrink (1977)]).λ ( ) log( )r . This mixed Hodge structure has the same weight ﬁltration as the limiting mixed Hodge structure constructed by Steenbrink ([Steenbrink (1976)]. [Varchenko (1981)]) has as the Hodge ﬁltration the subspaces F r ⊂ H k−1 (F .27) The order of the geometric section σ(α) is deﬁned as being the smallest λ α in the discrete set Ξ ⊂ R such that η0. We concentrate again only on the case where k − D /2 ≤ 0. r! (8.30) and as weight ﬁltration W ⊂ H (F . given by N =− 1 log T 2πi with log T = r≥1 (−1)r+1 (T − id)r /r. C) the ﬁltration associated to the nilpotent monodromy operator N . one deﬁnes the “geometric section” associated to the holomorphic kform α as λ σ(α) := r. C) of the Milnor ﬁber of XΓ ∩Πξ . for ﬁxed k. . λα α η0.λ = N r η0.
18). with k − D /2 ≤ 0. Upon varying the external momenta p in PΓ (p. where the polynomials h = hT. of the ﬁbers of the cohomological Milnor ﬁbration. i. p) as a function of the external momenta.e. fm (8.Renormalization. sliced along a linear space Πξ as in (8.w. k−1 We denote by HFeynman (F . p) becomes of the form (3. by varying arbitrarily the external momenta.1. k−1 The subspace HFeynman (F .14). t) and the spacetime dimension D ∈ N. where all external momenta are zero except for a pair of opposite momenta Pv1 = p = −Pv2 associated to a pair of external edges attached to a pair of vertices v1 .v. We write the integrand in the form αξ = with −k+D /2 h = PΓ f = ΨΓ m = −k + D( + 1)/2. through the coeﬃcients sC of (3. One can see that. C) ⊂ H k−1 (F .33). C) inherits a Hodge and a weight ﬁltration .32) as in (8. spanned by elements of the form (8. with k − D /2 ≤ 0. p)−k+D /2 for varying D. C).31) (8. when considering powers PΓ (t. one can reduce to the simplest possible case. (8. the forms αξ as above determine a subspace k−1 HFeynman (F . we obtain all polynomials of the form (8. Consider Feynman integrals.41). In such a case.33). and Hodge structures 175 Proposition 8. the polynomial PΓ (t.p are of the form −k+D /2 h(t) = i=1 LTi (t) e∈Ti / te .40). hΩξ . subject to the global conservation law (3. with h of the form (8.3. v2 .31).18). considered modulo the ideal generated by the derivatives of f = ΨΓ and localized at an isolated singular point. Consider the explicit expression of the polynomial PΓ (t. Proof.39). C) the subspace of the cohomology H k−1 (F . C) of the Milnor ﬁber spanned by the classes [αξ /df ] with αξ of the form (8. viewed as elements in the Milnor algebra M (f ).33) where the Ti are spanning trees and the LTi (t) are the linear functions of (3.31). Thus. singularities.
in the case of hypersurfaces in projective spaces. the natural associated regular singular connection is the Gauss–Manin connection on the cohomology of the Milnor ﬁber and the Picard–Fuchs equation for the vanishing cycles.176 Feynman Motives k−1 k−1 F • ∩ HFeynman and W• ∩ HFeynman from the asymptotic mixed Hodge structure of Varchenko on H k−1 (F . The irregular singular connections of [Connes and Marcolli (2004)] have values in the Lie algebra of the Connes–Kreimer group of diffeographisms and are deﬁned on a ﬁbration over a punctured disk with ﬁber the multiplicative group. It is an interesting problem to k−1 see whether the subspace HFeynman recovers the full H k−1 (F . which are naturally related to regular singular connections. with the setting of motives (especially mixed Tate motives) and mixed Hodge structures. . Then the Gauss– Manin connection on the vanishing cohomology bundle. W• ∩ HFeynman ) still give a mixed Hodge structure. based on equivalence classes of certain irregular singular connections. . how regular singular Gauss–Manin connections associated to singularities of individual graph hypersurfaces XΓ can be assembled. to give rise to an irregular singular connection of the kind used in [Connes and Marcolli (2004)]. respectively representing the complex variable z of dimensional regularization and the energy scale µ (or rather µz ) upon which the dimensionally regularized Feynman integrals depend. C).34) df be a basis for the vanishing cohomology bundle. We sketch here a relation between this regular singular connection and the irregular equisingular connections of [Connes and Marcolli (2004)].4 Regular and irregular singular connections An important and still mysterious aspect of the motivic approach to Feynman integrals and renormalization is the problem of reconciling the Riemann–Hilbert correspondence of perturbative renormalization formulated in [Connes and Marcolli (2004)] (see Chapter 1 of [Connes and Marcolli (2008)]). µ (8. C) and if k−1 k−1 (F • ∩ HFeynman . over diﬀerent graphs. written with the same notation we used above for the Gelfand–Leray form. On the other hand. following [Marcolli (2008)]. 8. In the following we let ωi i = 1. which is deﬁned by . . We explain here. . at least for some classes of graphs Γ.
Z). with P (s)ij = pij (s)... s (8. deﬁning a section [ω/df ] of the vanishing cohomology bundle. singularities.Renormalization. (8. ds with Ir (s) = sr−1 I (r−1) (s)..35) where the pij (s) are holomorphic away from s = 0 and have a pole at s = 0. Vasilev (1998)].41) (8. Lyashko. §12. The Gauss–Manin connection is regular singular and its monodromy agrees with the monodromy of the singularity (see [Arnold.. For the relation between Picard–Fuchs equations and mixed Hodge structures see §12 of [Arnold.39) −1) where the order is bounded above by the multiplicity of the critical point (see [Arnold.40) as the Picard–Fuchs equation of ω.1).39). . the function I(s) = δ(s) ω1 . (8. GuseinZade.34) by GM s ωi df = s j pij (s) ωj df . df δ(s) ωµ df (8. suppose given a holomorphic nform ω and let ω/df be the corresponding Gelfand–Leray form. . df δµ (s) ω df (8.. Given a covariantly constant section δ(s) of the vanishing homology bundle.36) is a solution of the regular singular Picard–Fuchs equation d I(s) = P (s)I(s).38) satisﬁes a regular singular order I ( ) (s) + p1 (s)I ( diﬀerential equation (s) + · · · + p (s)I(s) = 0. acts on the basis (8. Then the function I(s) = δ1 (s) ω . GuseinZade. §2.3).2. Varchenko (1988)]. δµ be a basis of the vanishing homology. Let δ1 .. Varchenko (1988)] and [Kulikov (1998)]. δi (s) ∈ Hn−1 (Fs .37) ds Similarly. (8.. One refers to (8. Goryunov. or to the equivalent system of regular singular homogeneous ﬁrst order equations d I(s) = P(s)I(s). . and Hodge structures 177 the integer cohomology lattice in each real cohomology ﬁber. . ..
this means that they satisfy the following conditions. (8. dz and γ −1 Y (γ) = b(z). We let g denote the Lie algebra g = Lie(G). which determines the behavior of pullbacks of solutions along sections of the ﬁbration Gm → B → ∆. As we recalled earlier. Counting all possible ways in which a given graph can be inserted into another at a speciﬁed vertex attaches multiplicities to each insertion that appear in the form of symmetry factors. • The connections satisfy ω(z. Let K denote the ﬁeld of germs of meromorphic functions at z = 0. The equisingularity condition. The ﬁrst condition and the ﬂatness condition imply that the connection ω(z. with Y the grading operator. have the property that their pullbacks σ ∗ (γ) ∈ G(K) along any section σ : ∆ → B with ﬁxed value σ(0) have the same negative piece of the Birkhoﬀ factorization σ ∗ (γ)− .45) . (8.44) where Γ◦v Γ denotes the graph obtained by inserting Γ into Γ at the vertex v ∈ V (Γ) and the sum is over all vertices where an insertion is possible.42) where a(z) and b(z) are elements of g(K) satisfying the ﬂatness condition db − Y (a) + [a. We also let B denote a ﬁbration over an inﬁnitesimal disk ∆∗ with ﬁber the multiplicative group Gm and we denote by P the principal Gbundle P = B ×G. u) can be written in the form ω(z. can be checked by writing the equation Df = ω in the more explicit form γ −1 dγ = a(z). b] = 0. dz (8. u) = uY (a(z)) dz + uY (b(z)) du . Γ ] = v∈V (Γ) Γ ◦v Γ − v ∈V (Γ ) Γ ◦v Γ.43) Recall that the Lie bracket in the Lie algebra Lie(G) of the Connes– Kreimer Hopf algebra is obtained by assigning (see [Connes and Kreimer (2000)]. λu) = λY ω(z. We consider Lie(G)valued ﬂat connections ω that are equisingular. • Solutions of Dγ = ω. u). for λ ∈ Gm . u (8. equisingular connections of [Connes and Marcolli (2004)] which we need here. [Connes and Kreimer (2001)]) [Γ.178 Feynman Motives We also recall some of the properties of the irregular singular.
(8. We now show how to produce a ﬂat connection of the desired form (8. Then the choice of a −∞ −∞ distribution σ ∈ Cc (SΓ ) induces distributions σγ ∈ Cc (Sγ ) and σΓ/γ ∈ −∞ Cc (SΓ/γ ). dz and (φ ◦ S) ∗ Y (φ) = b. dz and (φ ◦ S) ⊗ Y (φ). Suppose given a subgraph γ ⊂ Γ. ∆(Γ) = bΓ .48) for ∆(X) = X ⊗ 1 + 1 ⊗ X + X ⊗ X .46) where S is the antipode in H and ∗ is the product dual to the coproduct in the Hopf algebra. Let SΓ denote the manifold of planes Πξ in A#E(Γ) with dim Πξ ≤ codim Sing(XΓ ). and the regular singular Picard–Fuchs equation associated to the resulting isolated singularities of XΓ ∩ Πξ . (8. starting from the graph hypersuraces XΓ . a consistent choice of slicing Πξ .4. with X and X of lower degree.49) where SΓ. and Hodge structures 179 When one interprets elements γ ∈ G(K) as algebra homomorphisms φ ∈ Hom(H. This is similar to what happens when one enriches the discrete Hopf algebra by adding the data of the external momenta. . and the antipode is given inductively by S(X) = −X − S(X )X . (φ ◦ S) ⊗ where ∆(Γ) = Γ ⊗ 1 + 1 ⊗ Γ + γ dφ .Renormalization.m is the manifold of mdimensional planes in A#E(Γ) . on generators Γ of H. We denote −∞ by C ∞ (SΓ ) the space of test functions on SΓ and by Cc (SΓ ) its dual space of distributions. We can write SΓ as a disjoint union codim Sing(XΓ ) SΓ = m=1 SΓ.1.42). ∆(Γ) = aΓ . which accounts for the fact of having to choose a slicing Πξ . (8. K). We begin by introducing a small modiﬁcation of the Hopf algebra and coproduct.m .47) γ ⊗ Γ/γ with the sum over subdivergences. one can write the above equivalently in the form (φ ◦ S) ∗ dφ = a. This means. (8. Lemma 8. singularities. with irregular singularities.
55) The following result is then proved by the same argument used in [Connes and Marcolli (2008)]. Theorem 1. (8.54) where S = ∪Γ SΓ .2. we can realize the aﬃne Xγ as a hypersurface inside a linear subspace A#E(γ) ⊂ A#E(Γ) and similarly for the aﬃne XΓ/γ . where we simply identify the edges of γ or Γ/γ with a subset of the edges of the original graph Γ. endowed with the coproduct ∆(Γ. neglecting external edges.52) (8. We consider the commutative algebra −∞ ˜ H = Sym(Cc (S)). The coproduct (8.53) The argument for Γ/γ is analogous.180 Feynman Motives Proof.51) −∞ −∞ In turn. ˜ Lemma 8.4. σ) ⊗ 1 + 1 ⊗ (Γ. We then enrich the original Hopf algebra H by adding the datum of the slicing Πξ . σγ ) ⊗ (Γ/γ. (8. σΓ/γ ). One then has a restriction map Tγ : SΓ. seen as a hypersurface inside a linear subspace A#E(Γ/γ) ⊂ A#E(Γ) .γ = m=1 SΓ. at the level of distributions. where SΓ. σ) + γ (γ.γ ⊂ SΓ is the union of the components SΓ.27. codim Sing(Xγ ) SΓ. One sets σγ = Tγ (σΓ ) and σΓ/γ = TΓ/γ (σΓ ).γ 0 otherwise.γ → Sγ .55) is coassociative and H is a Hopf algebra. σ) = (Γ.m . (8. . This induces a map Tγ : C ∞ (Sγ ) → C ∞ (SΓ ) given by Tγ (f )(Πξ ) = f (Tγ (Πξ )) Πξ ∈ SΓ. by Tγ (σ)(f ) = σ(Tγ (f )). (8. this deﬁnes a map Tγ : Cc (SΓ ) → Cc (Sγ ). Given γ ⊂ Γ.m of SΓ with m ≤ codim Sing(Xγ ). (8.50) which is given by Tγ (Πξ ) = Πξ ∩ A#E(γ) .
. . which constructs irregular singular connections from solutions of the regular singular Picard–Fuchs equation.ξ (s) is a linear combination of the functions JΓ. We pass to the projective instead of the aﬃne formulation and we ﬁx a small neighborhood of an isolated singular point of XΓ ∩ Πξ . (8. if the kform αξ is given by PΓ Ωξ as above.ξ (z) = 0 sz JΓ. GuseinZade.ξ (z) admits an analytic continuation to a meromorphic function with poles at points z = −(λ+1) with λ ∈ ΞΓ.Renormalization.57) for the diﬀerent singular points. Any solution JΓ.ξ ( ) ( −1) (s) + · · · + p (s)JΓ.56) determines a ﬂat g(K)valued connection ω(z.ξ. and αξ /df is the Gelfand–Leray form associated to the holomorphic kform αξ .58) As in §7 of [Arnold.ξ . Varchenko (1988)]. df (8. δµ is a basis of locally constant sections of the homological Milnor ﬁbration. with a pole at z = −D if −D ∈ ΞΓ.4.ξ (s) ds.i (s) = δi (s) αξ . singularities.ξ (s) = 0.3. . We then have the following result. We look at the function FΓ.42). to simplify notation. Z). . for Πξ a linear space of dimension at most equal to the codimension of Sing(XΓ ).ξ (z) in a small neighborhood of a chosen point z = −D. It has an expansion as a Laurent series. .57) where δ1 . δi (s) ∈ Hk−1 (Fs . We consider the Mellin transform ∞ FΓ. Moreover.56) with the property that any solution JΓ. the function FΓ.ξ a discrete set in R of points related to the multiplicities of an embedded resolution of the singular point of XΓ ∩ Πξ . Proof.ξ of the regular singular Picard–Fuchs equation (8.ξ (s) at a unique isolated critical point. and Hodge structures 181 We then proceed as follows.ξ (s) + p1 (s)JΓ. we can just assume that we have a single expression JΓ. Thus. Then there exists an associated regular singular Picard–Fuchs equation JΓ. In fact. then the connection is equisingular. Suppose given a holomorphic kform αξ on Πξ . This depends on the choice of a singular point and can be localized in a small neighborhood of the singular point in XΓ ∩ Πξ . Theorem 8. introducing a cutoﬀ χξ that is supported near the singularities of XΓ ∩ Πξ amounts to adding the expressions (8. (8. u) of the form (8.
More precisely.42). This deﬁnes an algebra ˜ homomorphism φµ ∈ Hom(H. K). dt θt t=0 = Y .59) can ˜ ˜ be equivalently described as a family of G(C)valued loops γµ : ∆∗ → G(C). (8. λu) = λY ω(z. we apply σ to the coeﬃcients seen as functions of ξ. by assigning the values (8. we get γµ Y (γµ ) = uY (b(z)). for µ = et . Similarly. The dependence on µ of (8.61) ˜ where S and ∗ are the antipode of H and the product dual to the coproduct ∆ of (8.60) ˜ where θt is the oneparameter family of automorphisms of H generated by d the grading. after identifying FΓ. so that we have z ∈ ∆∗ a small neighborhood of z = 0. we deﬁne φµ (Γ. These deﬁne elements aµ . one has d −1 d γµ = θt (γ −1 γ) = uY (a(z)). the connection ω(z. Here G ˜ dual to the commutative Hopf algebra H. dz and bµ (z) := (φµ ◦ S) ∗ Y (φµ ).59) where we consider FΓ. Thus.55). u) of the form (8. u). Then one sets aµ (z) := (φµ ◦ S) ∗ d φµ . which one can use to deﬁne a connection ω(z. 2 (8. u) deﬁned in this way satisﬁes by construction the ﬁrst condition of the equisingularity property. More precisely. One can see that the connection is ﬂat since we have −1 dγµ (z) d d −1 bµ (z) − Y (aµ (z)) = Y (γµ (z)) + γµ (z) (Y (γµ (z))) dz dz dz −1 −Y (γµ (z)) d d −1 γµ (z) − γµ (z) (Y (γµ (z))) dz dz −1 = −γµ (z) d −1 −1 −1 (γµ (z))γµ (z)Y (γµ (z)) − γµ (z)Y (γµ (z))γµ (z) dz .59) on generators. bµ Ω1 (g(K)). Here µ is the mass scale. σ)(z) := µ−z b1 (Γ) σ FΓ. (8.ξ as a function of ξ to which we apply the distribution σ. γµ dz dz where we set uY = etY and then extend the resulting expression to −1 u ∈ Gm (C) = C∗ . ˜ denotes the aﬃne group scheme depending on the mass scale µ. namely ω(z. for all λ ∈ Gm .182 Feynman Motives After a change of variables on the complex coordinate z.ξ with its Laurent series expansion.ξ (− D+z ) .59) implies that γµ satisﬁes the scaling property γet µ (z) = θtz (γµ (z)). The homomorphism φ deﬁned by (8.
8. This means. it is possible to show (see [Collins (1986)] §5.1 and Proposition 1. The explicit dependence on µ in the Feynman integrals is as in (8. and Hodge structures 183 = −[a(z).58) considered here correspond to slices along a linear space Πξ of the Feynman integrals.44 of [Connes and Marcolli (2008)]). The second condition of equisingularity is the property that. the negative part satisﬁes d γµ. singularities.Renormalization. The integrals (8.+ (z). dµ By dimensional analysis on the counterterms. as shown in [Connes and Kreimer (2000)] (see also Proposition 1.− (z) = 0. . Thus. the same argument as in [Collins (1986)] §5. b(z)].− (z) = 0.44 of [Connes and Marcolli (2008)] applies to this case to show that ∂µ γµ.− (z)−1 γµ. localized by a cutoﬀ χξ near the singular points.− (z) = 0. in the Birkhoﬀ factorization γµ (z) = γµ.8. in the case of dimensional regularization and minimal subtraction.1) that the counterterms obtained by the BPHZ procedure applied to the Feynman integral Uµ (Γ)(z) do not depend on the mass parameter µ.59). which is unchanged with respect to that of the original dimensionally regularized Feynman integrals. that the Feynman integrals Uµ (Γ)(z) deﬁne a G(C)valued loop γµ (z) with the property that ∂µ γµ.
Chapter 9 Beyond scalar theories In this ﬁnal chapter we give a few examples of how the theme of Feynman integrals and motives. fol185 . 9.1 Supermanifolds We describe here an extension of the parametric Feynman integral and its motivic interpretation to the case of theories with both fermionic and bosonic variables. The second case we discuss is that of scalar quantum ﬁeld theories. and in particular the φ4 theory. based on the parametric representation of Feynman integrals and the associated algebraic varieties. In this case one can again express a parametric form of the Feynman integral in terms of periods. deﬁned in terms of Berezinians instead of determinants and fermionic integration. can be extended beyond the original setting of scalar quantum ﬁeld theories. based on the results of [Marcolli and Rej (2008)]. We concentrate here only on two simple kinds of extensions. which is a natural analog in the supergeometry context. The results reported here are from [Marcolli and Rej (2008)]. The ﬁrst extension we discuss is to theories with fermionic ﬁelds. In this case it is also known that a parametric formulation of Feynman integrals exists. where one can already see some new feature appearing. over a spacetime that is deformed to a noncommutative manifold with a Moyal product. with the determinant deﬁning the graph polynomial replaced by a Berezinian. We report a few basic facts about the geometry of supermanifolds. and we report some observations of [Aluﬃ and Marcolli (2008a)] on computations of characteristic classes for the resulting graph hypersurfaces. where we can reformulate the parametric integral as a period of a supermanifold. this time on a supermanifold.
OX ) is a complex manifold. f # ) as above. where the O grading is the Z/2Zgrading by odd/even degrees. . or odd degree part with respect to the Z/2Zgrading. A morphism F : X1 → X2 of supermanifolds Xi = (Xi . . θm ). [Manin (1988)]. . If we denote by Ei . where the xi are ordinary. scalar variables. 2 the holomorphic vector bundles on Xi such that O(Ei ) = Ei = Ni /Ni2 . §4. with the exterior powers Λ• graded by odd/even degree. for i = 1. As in ordinary geometry. while the θj are anticommuting Grassmann variables. f # ) of a morphism of the underlying complex manifolds f : X1 → X2 together with a morphism f # : A2 → f∗ A1 of sheaves of supercommutative rings with the property that at each point # # x ∈ X1 the induced morphism fx : (A2 )f (x) → (A1 )x satisﬁes fx (mf (x) ) ⊂ mx . an embedding of complex supermanifolds is a morphism F = (f. on the maximal ideals of germs of sections vanishing at the point (cf. in local coordi= O nates.1). . 2. The bosonic variables xi generate the order zero.1) An equivalent characterization of an embedding of supermanifold is given as follows. It is a split supermanifold. the quotient E = N /N 2 is locally free over OX and A is locally isomorphic to the exterior algebra Λ• X (E). i = 1. with the property that f : X1 → X2 is an embedding and f # : A2 → f∗ A1 is surjective. part and the fermionic variables θj the order one. By a complex supermanifold one understands a datum X = (X. commuting. . with the notation as above.186 Feynman Motives lowing [Manin (1988)]. A) with the following properties: A is a sheaf of supercommutative rings on X. a supermanifold is parameterized by bosonic and fermionic variables (x0 . or even degree. . θ1 . Thus. The supermanifold is split if the isomorphism A ∼ Λ• X (E) is global. (X. we deﬁne the ideal sheaf of X1 to be the kernel IX1 := Ker(f # : A2 → f∗ A1 ). A simple example is given by the projective superspaces. then an embedding F : X1 → X2 is an embedding f : X1 → X2 such . . The complex projective superspace Pnm is the supermanifold (X. . A) with X = Pn the usual complex projective space and A = Λ• (Cm ⊗C O(−1)). Ai ). where OX = A/N . xn . In particular. consists of a pair F = (f. (9. with N the ideal of nilpotents in A.
. . Then [X ] = [Y] + [X where X Y is the supermanifold X Y = (X Y. . Thus. . then a projective subvariety can be obtained by assigning a number of equations of the form Ψev/odd (x0 . Given a locally closed subset Y ⊂ X and a sheaf A on X. . . . (9. (9.Beyond scalar theories 187 that the induced morphism of vector bundles f ∗ : E2 → E1 is surjective. xn )θi1 · · · θik = 0. . θm ) the bosonic and fermionic coordinates of Pnm . .2.3. . Let SVC be the category of complex supermanifolds with morphisms deﬁned as above. if we denote by (x0 . Let F : Y → X be a closed embedding of supermanifolds. = the ideal sheaf of Y satisﬁes IY U = AU .1. θm ) = i1 <···<ik PI (x0 . θ1 . ik ). θ1 .2) where I = IX is an ideal generated by ﬁnitely many homogeneous polynomials of given Z/2parity. B) is a closed subsupermanifold of X = (X. A) is given by an open embedding U → X of the underlying complex manifolds and an isomorphism of sheaves AU ∼ B.1. A) when there exists a closed embedding Y ⊂ X and the pullback of EA under this embedding surjects to EB .6) . Let K0 (SVC ) denote the free abelian group generated by the isomorphism classes of objects X ∈ SVC subject to the following relations. A subvariety in superprojective space is a supermanifold X = (X ⊂ Pn . AX X Y Y]. see [Kashiwara and Schapira (1990)] §II.4) ). xn . we say that Y = (Y. (9. In [Marcolli and Rej (2008)] an analog of the Grothendieck ring of varieties is considered for supermanifolds. deﬁned as follows. . . .5) Here the notation AX means the following. . . . .3) where I = (i1 . there exists a unique sheaf AY with the property that AY Y = AY and AY X Y = 0. An open submanifold U = (U. . . (9. . xn ) are homogeneous polynomials in the bosonic variables. . . . Deﬁnition 9. . . In other words. and the PI (x0 . When Y ⊂ X is a closed embedding and U = X Y . (9. . . xn . B) → X = (X. . (Λ• (Cm ⊗C O(−1))/I)X ).
for a closed embedding Y ⊂ X. §3.3. A) decomposes as a sum of the corresponding classes in the Grothendieck group. A] satisﬁes [X. The supermanifold X = (X.8) where T = [A01 ] is the class of the aﬃne superspace of dimension (0. The sheaf AY on X. given a pair X = (X.3. (9. A) and the sequence of sheaves 0 → i! (AU ) → A → j∗ (AY ) → 0. associated to the open embedding U ⊂ X with complement Y = X the class [X. Namely. When Y is open it satisﬁes AY = Ker(A → j∗ (AX Y )). has a chain of subsheaves AY ⊃ A1 ⊃ · · · ⊃ Ak = 0 such that each quotient Ai /Ai+1 is coherent on Y . in the case where A = OX is the structure sheaf of X. U. the relation (9. for any coherent sheaf A over a Noetherian reduced irreducible scheme there exists a dense open set U such that AU is free. Propositions 1–3. In particular. where in this case j : X Y → X is the closed embedding of the complement. which has support Y . see [Shafarevich (1996)]. This is a consequence of the d´vissage of coherent sheaves. A] = [U. The relation (9. It is then clear that the product makes K0 (SVC ) into a ring with [X ][Y] = [X × Y]. this satisﬁes AY = i∗ (AY ) where i : Y → X is the inclusion. Thus. and in fact of supermanifolds where the vector bundle E with O(E) = E = N /N 2 is trivial. one can ﬁnd a stratiﬁcation where on each open stratum the supermanifold is split and with trivial vector bundle.7) in the Grothendieck group of ordinary varieties.1. . Proof. The Grothendieck ring K0 (SVC ) of supervarieties is a polynomial ring over the Grothendieck ring of ordinary varieties of the form K0 (SVC ) = K0 (VC )[T ]. AY Y ]. We show that all supermanifolds decompose in K0 (SVC ) as a ﬁnite combination of split supermanifolds. AU U ] + [Y. Proposition 9. Proposition 3. e §3.4) then ensures that. (9. 1).188 Feynman Motives In the case where Y is closed.4) reduces to the usual relation [X] = [Y ] + [X Y ]. [Shafarevich (1996)].2.
there are now two diﬀerent types of Lefschetz motives.... The fact that the vector bundle that constitutes the fermionic part of a supermanifold is trivial when seen in the Grothendieck group is the analog for supermanifolds of the fact that any locally trivial ﬁbration is equivalent to a product in the Grothendieck group of ordinary varieties. Suppose given a matrix M of the form M= M11 M12 M21 M22 . .m (x0 . §6 of [Manin (1988)]) ω= f1. respectively fermionic and bosonic Tate motives. det(M22 ) (9. The formal inverses of Lf and Lb correspond to two types of Tate objects for motives of supermanifolds.Beyond scalar theories 189 and the stated result then follows. . namely the bosonic one Lb = [A10 ] and the fermionic one Lf = [A01 ].9) It satisﬁes Ber(MN ) = Ber(M)Ber(N ). in the supermanifold case. which is determined by the property that. Then the Berezinian of M is the expression Ber(M) := −1 det(M11 − M12 M22 M21 ) .11) . The fact that it is indeed a polynomial ring in the class [A01 ] without further relation is a consequence of the fact that A01 is parameterized by a single Grassmann variable and that the A0n are nonisomorphic supermanifolds. . for a form ω = I I fI (x0 . (9. . θm ). where the M11 and M22 are square matrices with entries of order zero and the M12 and M21 have elements of order one. . . the integral gives (see Chapter 4. xn ) dx1 · · · dxn . .10) or equivalently by the property that (h0 (x) + h1 (x)θ) dθ = h1 (x). . (9. Grassmann variables integration is deﬁned in terms of the Berezinian integral. Notice that. . .. θ1 . . . . The analog of the determinant in supergeometry is given by the Berezinian. . . xn )θ in the local coordinates (x0 . xn . while only the bosonic part can provide nonTate contributions. . This is deﬁned in the following way. The contribution of the fermionic part of a supermanifold to the class in the Grothendieck ring is always of (fermonic) Tate type.
ξr ) and ∂Y Yβ = (yj . originating from [Rosly. The perturbative expansion for such theories corresponsingly involve graphs Γ with two diﬀerent types of edges: fermionic and bosonic edges. Schwarz. §5. with φi φ−1 j a holomorphic function on Ui ∩ Uj nowhere vanishing on the underlying Ui ∩Uj . Voronov (1988)].e.13) for the quadratic and linear forms that appear in the bosonic and fermionic propagators. i. and ignore. again just to simplify notation.12) More generally. Classes of divisors correspond to equivalence classes of line bundles and can be described in terms of integer linear combinations of (n − 1m)dimensional subvarieties Y ⊂ X . In the following we use the notation q(p) = p2 − m2 . and to fermionic edges a propagator i p+m / i = . ηs ). p2 − m2 p−m / (9. / involving γmatrices expressing fermions in spinorial form. q (p) = i(/ + m) / p (9.2 Parametric Feynman integrals and supermanifolds Consider now the case of Feynman propagators and Feynman diagrams that come from theories with both bosonic and fermionic ﬁelds. see [Ramond (1990)]. There is also a well developed theory of divisors on supermanifolds. 9. we set m = 0. The Feynman rules assign to each bosonic edge a propagator of the usual form we have encountered so far. in cases with tensor indices. In the following. the resulting infrared divergence problem. we would have p = pµ γµ .190 Feynman Motives for x an ordinary and θ a Grassmann variable. It is well known that for such theories the form of the possible fermionic terms in the Lagrangian is severely constrained: for instance.3. It is shown in [Bernˇte˘ s ın and Le˘ ıtes (1977)] that Grassmann integration satisﬁes a change of variable formula where the Jacobian of the coordinate change is given by the α Berezinian Ber(J) with J the matrix J = ∂Xβ with Xα = (xi . . we also drop the mass terms in the propagator. A Cartier divisor on a supermanifold X of dimension (nm) is deﬁned by a collection of even meromorphic functions φi deﬁned on an open covering Ui → X . in dimension D = 4 renormalizable interaction terms can only involve at most two fermion and one boson ﬁeld. for simplicity.
. θ2f ): dθ1 · · · dθ2f f ! −α = f q1 · · · qf.15) .14). 0 0 0 0 · · · 0 qf / 0 0 0 0 · · · −/f 0 q Lemma 9.14).1. the terms of the form (q1 · · · qn )−1 .Beyond scalar theories 191 Thus. as well as a pair of Grassmann variables accounting for the numerator q i in (9. .14) where n = #E(Γ) is the total number of edges in the graph and f = #Ef (Γ) is the number of fermionic edges. We have an analog of the Feynman trick we saw in the ﬁrst chapter.16) + f − k + 1) . / / 1 t α 2 f (1 + 2 θ Qf θ) (9.17) . . . (t1 q1 + · · · tn qn + 1 θt Qf θ)n−f 2 (9. . Then the following identity holds: q1 · · · qf / / = Kn. which we encountered in the purely bosonic case. Let σn2f denote the superspace σn × A02f . q1 · · · qn (9. are now replaced by terms of the form q1 · · · qf / / . as each fermionic edge contributes an ordinary integration variable. for example as done in [Medvedev (1978)]. Proof.f q1 · · · qn with Kn. ··· .2.f = 2f (n − 1)! f k=1 (−n σn2f dt1 · · · dtn−1 dθ1 · · · dθ2f . Other formulations of parametric / Feynman integrals for theories with fermionic variables are possible. The derivation we present suﬀers from a kind of “fermion doubling problem”. Qf = (9. .12) and (9. Let Qf denote the 2f × 2f antisymmetric matrix 0 q1 0 0 · · · 0 0 / −/ 0 0 0 · · · 0 0 q1 0 0 0 q ··· 0 0 /2 q 0 0 −/2 0 · · · 0 0 . We ﬁrst show that the following identity holds for integration in the Grassmann variables θ = (θ1 . .12) and (9. where we now introduce an extra set of Grassmann variables associated to the fermionic edges. which essentially account for the denominator qi in (9. .
to obtain 1 (1 + θt Qf θ)−α = 2 ∞ k=0 −α ( q i θ2i−1 θ2i )k . / i=1 together with the fact that the degree zero variables xi = θ2i−1 θ2i commute. As we are going to see below.1). / k i=1 f The rules of Grassmann integration then imply that only the coeﬃcient of θ1 · · · θ2f remains as a result of the integration. so that we have 1 dt1 · · · dtn−1 dθ1 · · · dθ2f (t1 q1 + · · · tn qn + 1 θt Qf θ)n−f σn2f 2 = f ! −n + f q1 · · · qf / / 2f f = f ! −n + f q1 · · · qf / / 2f f = T −n+f T −f dt1 · · · dtn−1 σn σn dt1 · · · dtn−1 (t1 q1 + · · · + tn qn )n 2f (n −n + f q 1 · · · q f / / f! . of the log divergent graphs.17). for which the parametric form of the Feynman integral is simpliﬁed by the lack of the second graph polynomial (see §3. we then write T = t1 q1 + · · · tn qn . we expand using the Taylor series ∞ (1 + x) = k=0 β β k x k and the identity 1 t θ Qf θ = 2 f q i θ2i−1 θ2i . here the condition that characterizes the class of graphs with this property will no longer depend only on the number of internal edges and loops but also on a particular choice of a basis of loops. − 1)! f q1 · · · qn Consider now the case of graphs that have both bosonic and fermionic legs. We will deal here with a particular class of graph. We mimic the procedure described earlier in this chapter to obtain a parametric form for the Feynman integral.192 Feynman Motives In fact. This gives (9. For simplicity of notation. . in the case with fermionic variables. which can be considered an equivalent.
. labelling the loops consisiting of only fermionic edges. . . of which f are fermionic and b = n − f bosonic. . . n of the Feynman graph into two sets ib = 1. . .2. (9. bf = − ( f f + bb )} for the loops that contain both fermionic and bosonic edges. given a choice of a basis for the ﬁrst homology of the graph. nf . .20) above we write rf and rb . where now.21) to the expression t i p2 + i i if θ2if −1 θ2if p if . . .18) 2 2 Then the following identity holds: q1 · · · qf D / / Λ(t) d s1 · · · dD s b dD σ1 · · · dD σ f = dt1 · · · dtn .2. for the indexing sets of these Grassmann and ordinary variables.19) Proof. . σr ) of an s ordinary and a Grassman variable. We then consider a change of variables / p i b = u ib + / rf ηib rf σrf + rb ηib rb s rb . and the remaning variables {1. . . (9. indexed as above by r = 1. . . . . f f }. . / / p if = u if + / rf ηif rf σrf + rb ηif rb s rb . / (9. bb } labelling the loops consisting of only bosonic edges. . and to the loops containing both fermionic and bosonic legs a pair (/r . nb and if = 1. we assign to each purely fermionic loop a Grassmann momentum variable. / We apply the change of variables (9. . . q1 · · · qn Ber(M(t))D/2 σn (9. . we replace the circuit matrix ηir of (3. . That is.22) . In (9. . since the p i are even (ordinary) variables. to each purely bosonic loop an ordinary momentum variable. for reasons of homogeneity.1) with a matrix of the form ηif rf ηif rb ηib rf ηib rb .20) Here the loop indices r = 1. Consequently. we need to assume that the ηirf are of degree one and the ηirb are of degree zero. . . . . with n = nb + nf . . b = bb + bf . / and Grassmann variables σrf ∈ A0D with rf = 1. . are at ﬁrst divided into three sets {1. We divide the edge indices i = 1. We then introduce two sets of momentum variables: ordinary variables s rb ∈ AD0 . . respectively. Assume that there exists a choice of a basis for H1 (Γ) satisfying the condition f D n − + ( f − b ) = 0. / (9. . .21) analogous to the one used before. {1. . Suppose given a graph Γ with n edges. . . . respectively labeling the bosonic and fermionic legs. . .Beyond scalar theories 193 Theorem 9. . with rb = 1. . f = f f + bf .
the relations ti ui ηir = 0 / i for each loop variable r = rb and r = rf .22) in the form 2 ti u i + i if θ2if −1 θ2if uif / + rb . if if (9. i ti ηirf ηirf i ti ηirb ηirf . since for orderone variables σrf ηirf = −ηirf σrf . We can then rewrite (9. Notice the minus sign in front of the quadratic term in the σrf . and / s T = 2 i ti u i + if θ2if −1 θ2if uif .35) of [Bjorken and Drell (1965)]).rb of degree zero. (18. of degree zero and one. respectively. Since the ηi. σ = (σrf ). Notice that.rb ( i ti ηirb ηirb )/rb s rb − s / rf rf ( i ti ηirf ηirf )σrf σrf + rb rf ( i τ ti ηirb ηirf )/rb σrf − σrf s τb ( s /r i ti ηirf ηirb ) + rb ( if θ2if −1 θ2if ηif rb )/rb + s rf ( if θ2if −1 θ2if ηif rf )σrf . the matrix Mf (t) is antisymmetric. the Mbf and Mf b of degree one. Mb (t) = Mf (t) = Mf b (t) = Nb (θ) = Nf (θ) = i ti ηirb ηirb . since the ηirf are of order one. while the Nb and Nf are. the expression (9. the matrices Mb and Mf are of degree zero.24) θ2if −1 θ2if ηif rb . s = (/rb ).194 Feynman Motives We assume again. as in the purely bosonic case (cf.23) where τ denotes transposition. We also set Mbf (t) = Mf b (t) = Mf b (t)τ .23) is of degree zero. θ2if −1 θ2if ηif rf . Thus. We write the above in the simpler notation T + s τ Mb (t)/ − σ τ Mf (t)σ + σ τ Mf b (t)/ − s τ Mbf (t)σ + Nb (θ)/ + Nf (θ)σ. .rf are of degree one and the ηi. / = −Mf (t)τ . / s s / s (9.
2 2 4 We then use this change of variable to write −σ τ Mf (t)σ + σ τ Mf b (t)/ − s τ Mf b (t)τ σ + 1 (σ τ Nf (θ) − Nf (θ)τ σ) = s / 2 −η τ Mf (t)η − 1 (Mf b (t)/ + 1 Nf (θ))τ Mf (t)−1 (Mf b (t)/ + 1 Nf (θ)) s 2 s 2 4 (9.26) +σ τ Mf b (t)/ − s τ Mf b (t)τ σ + Nf (θ)σ s / and where the integrations in the variables dD σi = dσi1 · · · dσiD are Grassmann variables integrations satisfying the properties of the Berezinian integral (9. Recall that for Grassmann variables we have the following change of variable formula.28) with 1 1 s η = σ − Mf (t)−1 Mf b (t)/ + Nf (θ) . Suppose given an invertible antisymmetric N × N matrix A with entries of degree zero and an N vector J with entries of degree one. σ. σ.10) and (9. θ. θ. Then we have 1 1 σ τ Aσ + (J τ σ − σ τ J) = η τ Aη + J τ A−1 J.Beyond scalar theories 195 We then consider an integral of the form d D s 1 · · · d D s b d D σ1 · · · d D σ f / / = / ( i ti p2 + if θ2if −1 θ2if p if )n−f i dD s 1 · · · dD s b dD σ1 · · · dD σ f / / . R(/.27) 2 4 1 for η = σ − 2 A−1 J. t) = T + s τ Mb (t)/ + Nb (θ)/ − σ τ Mf (t)σ s / s s (9. This follows immediately since Aτ = −A and we have (9. t)n−f s where R(/. 2 2 We have 1 1 1 (Mf b (t)/ + Nf (θ))τ Mf (t)−1 (Mf b (t)/ + Nf (θ)) = s s 4 2 2 1 τ s Mbf (t)Mf (t)−1 Mf b (t)/ / s 4 (9.29) .11). (9. while the integrations in dD s i are ordinary inte/ grations.25) 1 1 1 η τ Aη = σ τ Aσ + J τ σ − σ τ J − J τ A−1 J.
We have. θ) = Nb (θ) − 1 Nf (θ)τ Mf (t)−1 Mf b (t) 4 1 C(t.35) in the form det(Mf (t))D/2 dD v 1 · · · dD v b / / −1 τ 1 (T + C − 4 Bb Ab Bb + v τ Ab v )n−f +D / / f /2 (9. / / dD η1 · · · dD η 1+ 2D f /2 U (t. We then proceed to the remaining ordinary integration in (9. θ. s )−1 Mf (t). θ) = − 16 Nf (θ)τ Mf (t)−1 Nf (θ). 16 (9. s )η 2 n−f .37) .34) where Cn. f is a combinatorial factor.2. θ. the Grassmann integration in (9.33) 1 / 1 + η τ Xf (t. as in Lemma 9. s )n−f / where we use the notation Xf (t. θ)/. θ)Ab (t)−1 Bb (t. θ. θ. s )n−f +D f /2 / (9. s s We then rewrite (9. θ)/)τ = Bb (t. (9. We then have / / 1 v τ Ab (t)/ = s τ Ab (t)/ + 1 s τ Bb (t. / / s s where 1 Ab (t) = Mb (t) − 4 Mbf (t)Mf (t)−1 Mf b (t) 1 Nf (θ)τ Mf (t)−1 Nf (θ). θ) + s τ Ab (t)/ + Bb (t.25). (9. θ)τ .25) gives. s ) := T + C(t. f f /2 det(Mf (t))D/2 . U (t. dropping a multiplicative constant. θ. we write the denominator of (9.30) Bb (t. θ.1.36) .31) Thus. θ)/ + 4 Bb (t. θ)τ / v / s 2/ + 1 Bb (t. s )D / (9.32) Thus. s ) := 2U (t.25) in the form U (t.35) We use the change of variables v = s + 2 Mb (t)−1 Nb (θ)τ . f 1 τ / 2 η Xf (t. as we have already encountered in the case of ordinary parametric Feynman integrals for ordinary variables. (9. det(Mf (t))D/2 dD s 1 · · · dD s b / / . s )η n−f = Cn.f. θ)/. θ. s 1 2 where Ab (t)τ = Ab (t) and (Bb (t.196 Feynman Motives 1 + (Nf (θ)τ Mf (t)−1 Mf b (t)/ + s τ Mbf (t)Mf (t)−1 Nf (θ)) s / 8 + We then let U (t. θ. (9.f.
θ)τ . θ)n−f +D f /2 with ˜ Xb (t. θ)/)n−f +D / v f /2 . (9. θ) = i u 2 ti + i j ui θ2j−1 θ2j + / i<j Cij (t)θ2i−1 θ2i θ2j−1 θ2j . Using a change of variables 1 ˜ λ i = λ i + C u. We know from (9. the integral gives ˜ T −n+f − Consider ﬁrst the term det(Mf (t))D/2 det(Ab (t))D/2 in (9.39) det(Mf (t))D/2 1 −1 M (t))D/2 fb 4 Mf b (t)Mf (t) = Ber(M(t))−D/2 .38) dD v 1 · · · dD v b / / (1 + v τ Xb (t. det(Ab (t))D/2 (9.Beyond scalar theories 197 Set then 1 ˜ T (t. Then. / 2 we rewrite the above as ˜ T (t. i<j . we have det(Mb (t) − where M(t) = Mb (t) 1 Mf b (t) 2 1 2 Mbf (t) Mf (t) . θ)−1 Ab (t).31). θ)A−1 (t)Bb (t.39) above.41) D f b ˜ We now look more closely at the remaining term T −n+f − 2 + 2 in ˜ (9.24) that we can write T (t. θ) − Bb (t. θ) + C(t. θ) in the form ˜ T (t. b 4 so that we write the above as det(Mf (t))D/2 ˜ T (t. + D b 2 det(Mf (t))D/2 . θ) = i 1 / / u 2 ti − u τ C u + i 4 Cij η2i−1 η2i η2j−1 η2j .38). (9. This can be identiﬁed with a Berezinian. θ) = T (t. D D f 2 (9. θ) = T (t.42) where the ﬁrst sum is over all edges and the other two sums are over fermionic edges.40) (9.39). We set λi = θ2i−1 θ2i . In fact. and (9. up to a multiplicative constant. (9.
43).43) ˆ where Λ(t) is T f /2 times the coeﬃcient of η1 · · · η2f in the expansion ∞ k=0 −α k 1 ˜τ ˜ 2 λ Cλ k ˆ T . where we use the notation 1 ˜τ ˜ λ Cλ = 2 Cij η2i−1 η2i η2j−1 η2j . We write ˆ T (t) = i 1 u 2 ti − u τ C u / / i 4 k and we write ∞ ˜ ˆ T −α = T −α k=0 −α k 1 ˜τ ˜ 2 λ Cλ ˆ T . Ber(M(t))D/2 (9. More explicitly.198 Feynman Motives with ˜ λi = η2i−1 η2i . we can write the Feynman integral in the form q1 · · · qf D / / d s1 · · · dD s b dD σ1 · · · dD σ q1 · · · qn κ Σn2f f = Λ(t)η1 · · · η2f dt1 · · · dtn dη1 · · · dη2f . this term is of the form Λ(t) = Ci1 i2 (t) · · · Cif −1 if (t). i<j Thus. After imposing n − f + D ( f − 2 2 integration of the resulting term = 0 and performing the Grassmann σn2f Λ(t)η1 · · · η2f dt1 · · · dtn dη1 · · · dη2f . The multiplicative constant in front of the integral on the right hand side above is given by κ= −n + f − D ( 2 f /2 b) f − b) . ˆ(t)n− f + D ( f − b ) Ber(M(t))D/2 2 2 T (9. .44) the result follows directly from (9. over indices ia with i2a−1 < i2a and for k = f /2.
46) 4 is the union of zeros and poles of Ber(M(t)). divergences arise from the intersections between the domain of integration given by the simplex σn and the variety deﬁned by the solutions of the equation Ber(M(t))D/2 = 0.31) and (9. (The factor 1/4 in (9.46) deﬁnes a hypersurface in Pn . The multiplicities are given by the powers of these determinants that appear in Λ(t)Ber(M(t))−D/2 .24) we see that the denominator of Λ(t) is given by powers of det(Mf (t)) and det(Ab (t)) = det(Mb (t) − 1 Mbf (t)Mf (t)−1 Mf b (t)). in the sense that Mf (t) is not identically zero. .B) as the graph supermanifold.47) and by poles of Λ(t). B) that is generic.3 Graph supermanifolds The result of the previous section shows that we have an analog of the period integral dt1 · · · dtn det(MΓ (t))D/2 σn given by the similar expression Λ(t) dt1 · · · dtn .27).Beyond scalar theories 199 9. the set of zeros of (9. For generic graphs. Deﬁnition 9. the set of solutions of (9. We denote by X(Γ. The equation (9. We refer to X(Γ.) Using the formulae (9. Thus.3.3. suppose given a pair (Γ. Ber(M(t))D/2 (9.47) comes from the change of variable formula (9.2. The generic condition on graphs is imposed to avoid the cases with Mf (t) ≡ 0. The support of this divisor is the same as that of the principal divisor deﬁned by Ber(M(t)). Thus. Let Γ be a graph with bosonic and fermionic edges and B a choice of a basis of H1 (Γ).1. Λ(t) (9. in this case.46) is satisﬁed by solutions of 1 det(Mb (t) − Mbf (t)Mf (t)−1 Mf b (t)) = 0 4 (9.46) Lemma 9. hence a divisor in Pn−12f of dimension (n−22f ). Proof.B) ⊂ Pn−12f the locus of zeros and poles of Ber(M(t)) = 0.45) σn Again we see that.
0) + t5 (0. . t1 t3 t 2 t 5 t4 Consider the example of the ﬁgure. t1 + t2 + t3 ) = t1 (1.B) = XΓ = {t : det(Mb (t)) = 0} ⊂ Pn−10 . It is then natural to simply assume that. B) such that there is only one loop in B containing fermionic edges. . .18). with a choice of the basis of loops satisfying the condition (9. . . we obtain in this case Mbf (t)Mf (t)−1 Mf b (t) = 0 t 1 + t2 −(t1 + t2 ) 0 . 1) + t2 (1.200 Feynman Motives In the degenerate cases of graphs such that Mf (t) ≡ 0. In the ﬁrst case. we go back to the original calculation without Grassmann variables and we therefore simply recover X(Γ.B) = Pn−12f . 5 to the edges as in the ﬁgure above. Other special cases arise when we consider graphs with only bosonic or only fermionic edges. Examples of this sort are provided by data (Γ. It is easy to see that such a choice is possible: it has two generators. 0) and Mf (t) = Thus.B) = Pf −12f . the graph supermanifold is simply given by X(Γ. Let us assign the ordinary variables ti with i = 1. we simply set X(Γ. one of them a loop made of fermionic edges and the second a loop containing both fermionic and bosonic edges. in such cases. we have det(Mb (t) − 1 −1 Mf b (t)) ≡ 0 since both Mb (t) and Mbf (t) are identically 4 Mbf (t)Mf (t) zero. 1) + t3 (0. In the case with only fermionic edges. We then have Mb (t) = t1 + t2 + t3 since only the second loop in the basis contains bosonic edges. 1) + t4 (0. while we have Mbf (t) = (t1 + t2 .
B) ⊂ P58 is the union of t1 +t2 +t3 = 0 and t1 + t2 = 0 in P5 (the latter counted with multiplicity two). especially in relation to their renormalization properties. [Seiberg and Witten (1999)]. over a noncommutative spacetime have been investigated extensively in recent years. 9. where the underlying spacetime is deformed to a Moyal type noncommutative space. Schwarz (1998)]. t1 + t2 + t3 ) −(t1 +t2 +t3 ) t1 +t2 1 = −(t1 + t2 + t3 ) + t1 + t2 + t3 ≡ 0. . These render the theory nonrenormalizable. . Thus. a . using the simple relation of Proposition 9. but which breaks the translation invariance of the theory. More recently.4 Noncommutative ﬁeld theories Scalar quantum ﬁeld theories. with the restriction of the sheaf from P58 . However. Such theories arise as eﬀective limits of string theory [Connes. coming from UV/IR (ultraviolet/infrared) mixing. perturbative renormalization can be restored in the φ4 theory in diﬀerent ways. and we refer the reader to the lectures [Grosse and Wulkenhaar (2008)] for a nice and detailed overview. Douglas. We do not give here a detailed exposition of these theories. . this is done by the addition of a harmonic oscillator term that is quite natural from the point of view of the Moyal product.2 between the Grothendieck ring of ordinary manifolds and of supermanifolds.Beyond scalar theories 201 (t1 + t2 . It is also shown in [Marcolli and Rej (2008)] that the universality result of [Belkale and Brosnan (2003a)] implies a similar result for the supermanifold case. t5 ).1. especially φ4 theories. when 4 the underlying R4 is made noncommutative by deformation to Rθ with the Moyal product. the theory suﬀers from a new type of divergences. . so that Ber(M(t)) = det(Mb (t)) det(Mf (t))−1 = (t1 + t2 + t3 )/(t1 + t2 )2 and the locus of zeros and poles X(Γ. We only recall that in dimension D = 4. In the Grosse–Wulkenhaar model [Grosse and Wulkenhaar (2005)]. t1 + t2 + t3 ) 0 1 t1 +t2 −1 t1 +t2 0 t1 + t2 t 1 + t2 + t 3 = (t1 + t2 . in this particular example we have Mbf (t)Mf (t)−1 Mf b (t) ≡ 0 for all t = (t1 .
Both models are renormalizable to all orders. one in the graph itself and one in another associated graph . the only ones that can appear as Feynman graphs of a φ4 theory are the oneloop case (with two external edges at each vertex). was introduced in [Gurau. we see that the eﬀect of making the underlying spacetime noncommutative turns the remaining two graphs into the following ribbon graphs.) " ! 2 " ! 3 The usual Kirchhoﬀ polynomial ΨΓ (t) of the Feynman graph. Among the family of “banana graphs”. Tanasa (2009)]. in these theories. which maintains translation invariance. are replaced by new polynomials involving pairs of spanning trees. What interests us most here is the fact that. [Tanasa (2007)]. only one of which is a planar graph. it is also possible to obtain a parametric representation of Feynman integrals. Magnen. the twoloop case (with one external edge at each vertex) and the threeloop case as a vacuum bubble. as well as the polynomial PΓ (t. We concentrate on some special examples among the parametric representation described in [Gurau and Rivasseu (2007)]. p). Excluding the vacuum bubble because of the presence of the polynomial PΓ (t). Rivasseu. Due to the fact that the underlying spacetime coordinates no longer commute. (Notice how the twoloop ribbon graph now has two distinct versions.202 Feynman Motives diﬀerent approach to restoring renormalization for the φ4 theory on Moyal spaces. whose motivic properties were described in [Aluﬃ and Marcolli (2008a)]. ordinary Feynman graphs are replaced in this context by ribbon graphs and the planarity or nonplanarity of the ribbon graphs plays an important role in the properties of the corresponding Feynman integrals.
˜ Proposition 9. The hypersurface corresponding to the value s = 0 has a singularity at the origin 0 ∈ An whose tangent cone is the (aﬃne) graph ˆ hypersurface XΓ .1. Then instead of a single graph hypersurface XΓ one has a oneˆ ˜ ⊂ An . the polynomials have been computed explicitly in [Gurau and Rivasseu (2007)] and they are of the form ΨΓ2 = (1 + 4s2 )(t1 + t2 + t2 t2 + t1 t2 ). ˜ ˜ In the speciﬁc examples of the banana graphs Γ2 and Γ3 of the ﬁgure above. and that it gives the part of lowest order in the variables ti when s = 0. Similarly for the two graphs Γ3 that correspond to the banana graph Γ3 one has ([Gurau and Rivasseu (2007)]) 2 ΨΓ3 (t) = (1 + 8s2 + 16s4 )(t1 t2 + t2 t3 + t1 t3 + t1 t2 t3 + t1 t2 t3 + t1 t2 t2 ) ˜ 2 3 + 16s2 (t2 + t2 t2 ) 2 1 3 (9.49) . One can see the polynomial ΨΓ2 (t) = t1 + t2 appear˜ ing as lowest order term. Unlike the commutative case.s s ∈ R+ depends upon the deformation parameter θ of the noncommutative R4 and on the parameter Ω of the harmonic oscillator term in the Grosse– θ Wulkenhaar model. where the parameter parameter family of aﬃne hypersurfaces XΓ. Proof.4.Beyond scalar theories 203 which is a dual graph in the planar case and that is obtained from an embedding of the ribbon graph on a Riemann surface in the more general case. This follows directly from the relation between the graph poly˜ nomial for the ribbon graph Γ given in [Gurau and Rivasseu (2007)] and the Kirchhoﬀ polynomial ΨΓ .48) where the parameter s = (4θΩ)−1 is a function of the deformation parameter θ ∈ R of the Moyal plane and of the parameter Ω in the harmonic oscillator term in the Grosse–Wulkenhaar action functional (see [Grosse and Wulkenhaar (2008)]). It suﬃces to see that (a constant multiple ˜ of) the Kirchhoﬀ polynomial is contained in the polynomial for Γ for all values of the parameter s. ˜ 1 2 (9. hence the corresponding graph hypersurface only makes sense as an aﬃne hypersurface. Let Γ be a ribbon graph in the noncommutative φ4 theory that corresponds in the ordinary φ4 theory to a graph Γ with n interˆ nal edges. these polynomials are no longer homogeneous. The relation of the hypersurface for the noncommutative case and the one of the original commutative case (also viewed as an aﬃne hypersurface) is given by the following statement.
49) and let XΓ3 ⊂ P3 be the hypersurface obta