SciPost Phys. Lect.

Notes 75 (2023)

Four lectures on Euler integrals

Saiei-Jaeyeong Matsubara-Heo1⋆ , Sebastian Mizera2† and Simon Telen3‡

1 Kumamoto University, Chuo Ward, Kumamoto, 860-0862, Japan

2 IAS Princeton, NJ 08540, USA
3 MPI-MiS Leipzig, 04103 Leipzig, Germany

⋆ saiei@educ.kumamoto-u.ac.jp , † smizera@ias.edu , ‡ simon.telen@mis.mpg.de

Abstract
These lecture notes provide a self-contained introduction to Euler integrals, which are
frequently encountered in applications. In particle physics, they arise as Feynman inte-
grals or string amplitudes. Our four selected topics demonstrate the diverse mathemat-
ical techniques involved in the study of Euler integrals, including polyhedral geometry,
very affine varieties, differential equations, and computational algebra.

Copyright S.-J. Matsubara-Heo et al. Received 29-06-2023

This work is licensed under the Creative Commons Accepted 06-09-2023
Check for
Attribution 4.0 International License. Published 26-10-2023 updates

Published by the SciPost Foundation. doi:10.21468/SciPostPhysLectNotes.75

Contents
Introduction 2

1 Newton polytopes and convergence 4

1.1 A little polyhedral geometry 5
1.2 Nilsson-Passare convergence 8

2 Limits and critical points 11

2.1 Computing critical points 12
2.2 Dual volumes in field theory limits 14
2.3 Saddle point approximation in high energy limits 16

3 Twisted (co)homology 19
3.1 Twisted chains 20
3.2 Twisted cochains 23
3.3 Back to Euler integrals 25

4 Differential and difference equations 29

4.1 Difference equations 30
4.2 Differential equations 33

5 Open problems 38

References 39

1
 SciPost Phys. Lect. Notes 75 (2023)

Introduction
Consider ℓ Laurent polynomials f1 , . . . , fℓ in n variables x = (x 1 , . . . , x n ) with complex coeffi-
cients. By an Euler integral, we mean any integral of the following form:
ν ν
x 1 1 · · · x n n dx 1
Z Z
dx n dx
s1 sℓ ∧ ··· ∧ = f −s x ν . (1)
Γ f1 · · · fℓ x1 xn Γ x

The right-hand side is our shorthand notation. The first example is the Euler beta function
Z1 Z∞
xν dx Γ (ν)Γ (1 − s)
B(ν, 1 − s) = = , where Γ (u) = t u−1 e−t dt , (2)
0 (1 − x)s x Γ (ν + 1 − s) 0

is the gamma function. The equality in blue will be derived below. Such integrals have been
called many different names, depending on the context in which they are studied. They were
called generalized Euler integrals by Gelfand, Kapranov and Zelevinsky [31]. This was moti-
vated by Euler’s integral representation of Gauss’ hypergeometric function. In fact, the integral
(1) represents a generalized hypergeometric function, and the name hypergeometric integral has
appeared in the literature as well [5]. When s1 = · · · = sℓ = 1 and Γ = R+ n
, our integral is a
−1
function of ν called the Mellin transform of ( f1 · · · fℓ ) [53]. This lead the authors of [9] to use
the name Euler-Mellin integrals for general s and Γ = R+ n
. Seminal contributions like [4, 30]
justify the name Aomoto-Gelfand integrals. In physics, Feynman integrals in quantum field the-
ory and string amplitudes in superstring theory take the form (1) for particular choices of f i .
We elaborate on these specific polynomials below. In Bayesian statistics, Euler integrals appear
as marginal likelihood integrals [13]. In our title, we chose to use Euler integrals as an umbrella
term for all these instances of (1).
In different sections, we will view the integral (1) as a function of different sets of param-
eters. For instance, in Section 1, we will fix Γ = R+ n
and think of (1) as a function of s and
ν. On the other hand, in Section 3, we think of the integrand as an element of a cohomology
vector space. Hence, the integral gives a linear function which sends Γ to (1). We will also
consider the case where the coefficients of f i depend on some parameters z. In this case our
integral is a function of z satisfying some interesting differential equations, see Section 4.
As mentioned above, Euler integrals appear in particle physics. The first important exam-
ple comes from quantum field theory, where Feynman integrals are used to describe particle
scattering processes. For a complete introduction to the subject, we refer to the recent book by
Weinzierl [61]. In the Lee-Pomeransky representation [43], up to a prefactor involving gamma
functions in s, ν, the Feynman integral of a graph G takes the form

xν
Z
dx
IG = , (3)
Rn (UG + FG ) x
s
+

where n is the number of internal edges of G, and UG , FG are the first and second Symanzik
polynomials associated to the graph. We illustrate this with one of our running examples.
Example 0.1. Consider the triangle diagram G in (4) with three massless internal edges. The
internal edges carry variables (x 1 , x 2 , x 3 ). The three external (open) edges attached to each
vertex carry the kinematic parameters (t 1 , t 2 , t 3 ):
t2

x3 x1 (4)
t1 x2 t3

2
 SciPost Phys. Lect. Notes 75 (2023)

The polynomial UG is the sum over spanning trees of G, with each term given by the x i ’s not
present in the tree:
UG = + + = x1 + x2 + x3 . (5)

The FG polynomial is given similarly as a sum of spanning two-forests (disjoint unions of two
trees), each weighted with minus the corresponding kinematic variable:

FG = + + = −t 1 · x 2 x 3 − t 2 · x 3 x 1 − t 3 · x 1 x 2 . (6)

The associated integral is given by

ν ν ν
x1 1 x2 2 x3 3
Z
dx 1 dx 2 dx 3
IG = . (7)
R3+ (x 1 + x 2 + x 3 − t 1 · x 2 x 3 − t 2 · x 1 x 3 − t 3 · x 1 x 2 )s x1 x2 x3

The exponents νi are typically taken to be non-negative integers and s = D/2 is half the space-
time dimension D. It is often convenient to think of (ν1 , ν2 , ν3 ) and s as generic parameters,
which is referred to as analytic and dimensional regularization respectively. ⋄

The second application of Euler integrals in physics comes from scattering amplitudes in
string theory. Instead of particles, one computes the probability of strings interacting with
each other. See [44] for a comprehensive review. This offers a nice immediate connection to
algebraic geometry: The integration is on the moduli space M0,m of genus zero curves with m
marked points. Equivalently, this is the space of configurations of m distinct points on P1 up
to its automorphisms PSL(2). We can represent these points as the columns of a 2 × m matrix
with nonzero 2 × 2 minors. Two such matrices M1 , M2 represent equivalent configurations
if there is an invertible 2 × 2 matrix T and an n × n invertible diagonal matrix D such that
T · M1 · D = M2 . We can use the action of T and D to fix 3 out of m points, leaving n = m − 3
degrees of freedom. Following [8, Eq. (1.5)], we write a point of M0,m as
 
1 1 1 1 ··· 1 0
M = , (8)
0 1 1 + x1 1 + x1 + x2 · · · 1 + x1 + · · · + x n 1

where n = m − 3 and the 2 × 2 minors f i j = M1i M2 j − M1 j M2i , i < j are nonzero. The genus
zero contribution to the m-point string amplitude is given by an Euler integral depending on
an extra parameter α′ :
Z α′ ν1 α′ νn
′ n
x1 · · · xn dx
Im = (α ) · α′ si j
. (9)
M+
Q x
0,m 1<i+1< j<m f i j

The pairs (i, j) excluded in the product in the denominator are those for which the minor f i j
is either constant or one of the x-variables. The integration is over the positive part M+ 0,m of
M0,m , which is the subset of points M satisfying f i j > 0, for all 1 ≤ i < j ≤ m. Using the
n
parameterization (8), one checks that this is R+ .
There are two physically interesting limits: α′ → 0 and α′ → ∞. The first one is called
the field theory limit in which strings become particles, and the second is the high-energy limit.
We will see in Section 2 that both of them admit an elegant geometric description.

Example 0.2 (m = 4). The moduli space M0,4 has dimension 1. The four-point string ampli-
tude (9) is I4 = α′ · B(α′ ν, −α′ ν + α′ s13 ), where B is the beta function from (2). ⋄

3
 SciPost Phys. Lect. Notes 75 (2023)

Example 0.3 (m = 5). The matrix parameterizing M0,5 is

 
1 1 1 1 0
M = , (10)
0 1 1 + x1 1 + x1 + x2 1

which has only 5 ordered minors depending on the variables (x 1 , x 2 ):

f13 = 1 + x 1 , f14 = 1 + x 1 + x 2 , f23 = x 1 , f24 = x 1 + x 2 , f34 = x 2 . (11)

The minors f23 and f34 are not included in the integrand of (9), since they would only shift
α′ ν j
the exponents of x j . The five-point string amplitude is given by

Z α′ ν1 α′ ν2
x1 x2 dx 1 dx 2
I5 = (α′ )2 · . (12)
(1 + x 1 )α′ s13 (1 + x1 + x2 )α s14 (x + x2 )α s24
′ ′
R2+ 1 x1 x2

The parameters (ν1 , ν2 , s13 , s14 , s24 ) describe momenta and angles of the 5 strings involved in
the scattering process. ⋄

Euler integrals have many other applications, including marginal likelihood integrals [13],
wave functions in cosmology [7], and correlation functions of conformal field theories [15,24].
These notes present the basics on Euler integrals from different points of view. They pro-
vide a roadmap through the literature for a reader who is new to the subject. At the same
time, we hope they serve as a helpful overview of important results for experts. Section 1
discusses convergence and meromorphic continuation, which leads us to study convex poly-
topes and polyhedral cones. Section 2 is about certain limits which are meaningful in physics
applications. This brings in algebraic equations, very affine varieties and Euler characteris-
tics. Section 3 develops the theory of (algebraic) twisted (co)homology on these very affine
varieties. Section 4 identifies difference and differential equations satisfied by Euler integrals.
Finally, Section 5 contains a list of open problems.

1 Newton polytopes and convergence

This section discusses convergence of the integral (1), viewed as a function of the exponents
s, ν. The integration contour Γ = R+
n
is fixed throughout the section. We set
Z
dx
I (s, ν) = f −s x ν . (13)
n
R+ x

n
To ensure that the integrand is finite on R+ , we make the following assumption.

Assumption 1. The coefficients of f i are real, positive numbers. That is,

X
fi = ci,α · x α , i = 1, . . . , ℓ , (14)
α ∈ supp( f i )

α α
where ci,α ∈ R+ , supp( f i ) ⊂ Zn is the support of f i (see Definition 1.1) and x α = x 1 1 · · · x n n .

Before studying convergence, we should address how to evaluate the integrand f −s x ν .

Since s and ν are complex vectors, this function may be multi-valued. For instance, we have

f i (x)si = exp(si log f i (x)) ,

4
 SciPost Phys. Lect. Notes 75 (2023)

(0,0,1)

(1,0,1) (0,1,1)

(1,0,0) (0,1,0)

(1,1,0)

Figure 1: Left: Minkowski sum of three polytopes in R2 . Right: The polytope

∆(UG + FG ) of the triangle Feynman diagram.

p
and log is only defined up to translates
p by integer multiples of 2π −1. When s is not an
integer, exp(si F ) ̸= exp(si (F + 2π −1k)) for some integer k, i.e., there are multiple branches
of f i (x)si . Assumption 1 ensures that f i takes positive values on R+ n
, so that there is precisely
one positive branch of log f i and log x j . In this section, our integrand is

f −s x ν = exp(−s1 log f1 − · · · − sℓ log fℓ + ν1 log x 1 + · · · + νn log x n ) ,

where the unique positive branches of log f i and log x j are intended.
As it turns out, statements about convergence of (13) involve convex polytopes and poly-
hedral cones. We start by introducing these objects, and then switch to convergence results
from [8, 9, 53]. In [9, 53], (13) was called an Euler-Mellin integral and weaker assumptions on
f i are used. In this text, we stick with Assumption 1 for simplicity.

1.1 A little polyhedral geometry

This section introduces properties of convex polytopes and polyhedral cones that we need later
on. We omit most proofs, and refer the reader to the standard textbook [63] for more details.
A subset P ⊂ Rn is called convex if for any p1 , p2 ∈ P, the line segment p1 p2 is contained in P.
The convex hull of A ⊂ Rn is the smallest convex subset P ⊂ Rn such that A ⊂ P. We denote this
by conv(A). A convex polytope in Rn is the convex hull of finitely many points. Since we will not
encounter any non-convex polytopes in this text, we will sometimes omit the adjective convex.
If P is a polytope and s is a nonnegative number, the s-dilation of P is the convex polytope

s · P = {s · p : p ∈ P} .

Here s · p is the usual scalar multiplication for vectors in Rn . It is easy to check that s · P is
indeed a convex polytope. The Minkowski sum of two polytopes P, Q is a new polytope

P + Q = {p + q : p ∈ P, q ∈ Q} .

This binary operation is commutative and associative. An example is shown in Figure 1 (left),
where we take the sum of three polytopes in R2 . Each is the convex hull of the points repre-
sented by black bullets. The dimension of a polytope is the dimension of the smallest affine
space containing it. Figure 1 (left) shows two polytopes of dimension two (these are also
called polygons), and two polytopes of dimension one (i.e., line segments). In the right part
of that figure, we show a three-dimensional convex polytope in R3 . The polytopes we will
encounter in this text arise as the Newton polytope of a Laurent polynomial.

5
 SciPost Phys. Lect. Notes 75 (2023)

Definition 1.1 (Newton polytope). Let f = α∈Zn cα x α ∈ C[x 1±1 , . . . , x n±1 ] be a Laurent poly-
P

nomial. The support of f is the set supp( f ) = {α ∈ Zn : cα ̸= 0}. The Newton polytope
∆( f ) ⊂ Rn is defined as the convex hull of the support, i.e., ∆( f ) = conv(supp( f )).

At the level of Laurent polynomials, Minkowski addition corresponds to multiplication.

That is, for two Laurent polynomials f , g, we have ∆( f g) = ∆( f ) + ∆(g).

Example 1.2. The polytopes in Figure 1, from left to right, are the Newton polytopes of

1 + x1 , 1 + x1 + x2 , x1 + x2 , (1 + x 1 )(1 + x 1 + x 2 )(x 1 + x 2 ) ,

and the polytope ∆(UG + FG ) for the denominator of (7). ⋄

A nonzero vector y ∈ Rn defines a face P y of a polytope P as follows:

P y = {p ∈ P : y · p = min y · q} .
q∈P

In particular, P is a face of itself: P0 = P. Every face P y of P is a polytope itself, and a face of

a face of P is a face of P itself. If dim P y = dim P − 1, then P y is called a facet of P. Faces of
dimension 0 and 1 are called vertices and edges respectively. For example, the polygon in the
middle of Figure 1 has 5 vertices, 5 facets (or edges), and one 2-dimensional face.
The faces of P divide up Rn into finitely many regions. For a given face Q ⊆ P, we set

CQ = { y ∈ Rn : Q ⊆ P y } .

For any face Q ⊆ P, CQ is a polyhedral cone. I.e., there is a finite set A ⊂ Rn such that
¨ «
X
CQ = pos(A) = c r r : c r ∈ R≥0 . (15)
r∈A

All our cones are polyhedral, so we will sometimes just refer to them as cones. The dimension
of a cone is the dimension of the smallest linear space containing it. For our cones CQ , we have
dim CQ = n − dim Q. E.g., if v ∈ P is a vertex, we have dim C v = n. If dim P = n and Q is a
facet, then CQ is a one-dimensional cone. These are called rays. When Q runs over all faces,
the cones CQ tile up Rn . The same is true for the vertices v. In symbols:
[ [
Rn = CQ = Cv . (16)
Q v

A cone C is called pointed if C ∩ (−C) = {0}. If P ⊂ Rn is full-dimensional, i.e., dim P = n,

the cone CQ is pointed for each face Q ⊂ P. If CQ is pointed and A ⊂ Rn is the minimal subset
such that (15) holds, the elements of A are called ray generators of CQ . The reason is that each
r ∈ A generates the ray CQ′ ⊂ CQ of a facet Q′ ⊃ Q. A k-dimensional cone is called simplicial if
it has a set of k ray generators. This always holds when k ≤ 2.
The collection of cones Σ P = {CQ : Q face of P} is closed under taking intersections. In
fact, one can check that CQ 1 ∩ CQ 2 = CQ 12 , where Q 12 ⊂ P is the smallest face of P containing
both Q 1 and Q 2 . This fact, together with the observation that Σ P is closed under taking faces
(we leave the definition of a face of a cone to the reader), makes Σ P into a polyhedral fan,
called the normal fan of P.

Example 1.3. A pentagon P in R2 has five vertices. These give five pointed full-dimensional
cones in its normal fan. The ray separating two neighboring cones C v1 and C v2 is the cone
C v1 v2 corresponding to the edge containing v1 and v2 . This is illustrated in Figure 2. The cone
C P = {0} is the only zero-dimensional one. The normal fan Σ P is invariant under translations
of P. I.e., Σ P = Σ P+w for w ∈ Rn . We encourage the reader to check this. ⋄

6
 SciPost Phys. Lect. Notes 75 (2023)

Figure 2: The normal fan of a pentagon has five two-dimensional cones.

Figure 3: The polar dual of P and its subdivision induced by the normal fan Σ P .

Our final construction is the polar dual P ◦ of a polytope P ⊂ Rn . This is given by

P ◦ = { y ∈ Rn : y · p ≥ −1, for all p ∈ P} .

If P is full-dimensional and it contains the origin in its interior int(P), then P ◦ is again a
polytope. Its vertices lie on the rays of the normal fan Σ P . Hence, the normal fan induces a
subdivision [
P◦ = B v , where B v = C v ∩ P ◦ . (17)
v

Here B v is the convex polytope { y ∈ C v : y · v ≥ −1}.

Example 1.4. The polar dual and its subdivision are illustrated in Figure 3. To satisfy 0∈int(P),
we translated our polytope so that (0, 0) is an interior lattice point. ⋄

The following lemma will be useful in our discussion on convergence.

Lemma 1.5. Let P be a full-dimensional polytope in Rn . We have 0 ∈ int(P) if and only if for all
vertices v of P, y · v < 0 for all y ∈ C v \ {0}.

Proof (sketch). The proof uses the facet description of int(P):

int(P) = {p ∈ Rn : rQ · p > rQ · vQ , for all facets Q ⊂ P} .

Here rQ is any ray generator of the ray CQ , and vQ is any vertex contained in Q. If and only if
P p = 0 belongs to this set. For a given vertex v of P, a
all right-hand sides rQ · vQ are negative,
vector y ∈ C v can be written as y = v∈Q cQ rQ , with cQ ≥ 0. The lemma follows.

7
 SciPost Phys. Lect. Notes 75 (2023)

1.2 Nilsson-Passare convergence

The main theorem of this section identifies a region in (s, ν)-space Cℓ+n in which the integral
(13) converges. The result for ℓ = 1 and s = 1 is due to Nilsson and Passare [53]. This was
generalized to integrals of the form (13) in [9]. The papers [9, 53] use weaker assumptions
on f . Our proof below is inspired by that of [55, Section 2] and [8, Claim 1].
Theorem 1.6. Let Re(si ) > 0 for i = 1, . . . , ℓ and suppose that ∆( f1 ) + · · · + ∆( fℓ ) has di-
mension n. The integral (13) with f i satisfying Assumption 1 converges absolutely if and only if
Re(ν) ∈ int(P(s)), where P(s) = Re(s1 ) · ∆( f1 ) + · · · + Re(sℓ ) · ∆( fℓ ).
Example 1.7. Ignoring the α′ parameter for now, the string amplitude I5 from (12) is
ν ν
x1 1 x2 2
Z
dx 1 dx 2
I5 = . (18)
R2 (1 + x 1 ) (1 + x 1 + x 2 ) (x 1 + x 2 )
s13 s14 s34 x1 x2
+

Suppose (s13 , s14 , s34 ) = (1, 1, 1). By Theorem 1.6, the integral converges if (ν1 , ν2 ) ∈ int(P),
where P is the pentagon in the middle of Figure 1. For (ν1 , ν2 ) = (1, 1), we find using

Integrate[((1+x1)(1+x1+x2)(x1+x2))^(-1), {x1,0,Infinity}, {x2,0,Infinity}] 1

in Mathematica that I5 = π2 /6. Multiplying the integrand with x1*x2, i.e., using
(ν1 , ν2 ) = (2, 2), the program prints a message saying that the integral does not converge. ⋄
R
The normalized volume of a compact set B ⊂ Rn is defined as Vol(B) = n! · B 1dx. Our
proof of Theorem 1.6 bounds the Euler integral (13) in terms of the normalized volume of
polytopes, see Equation (23) below. It uses Lemma 1.5, as well as the following lemma.
Lemma 1.8. Let C ⊂ Rn be an n-dimensional polyhedral cone and let v ∈ Rn be such that y ·v < 0
for all y ∈ C \ {0}. Then B = { y ∈ C : y · v ≥ −1} is a polytope with volume
Z
Vol(B) = exp( y · v) d y . (19)
C

If, instead, y · v ≥ 0 for some y ∈ C \ {0}, then the integral above diverges.
Proof. It suffices to show this in the case where C is simplicial, with n ray generators r1 , . . . , rn .
This is because if C is not simplicial, it can be subdivided into finitely many simplicial cones
C1 , . . . , Ck , and we would conclude
X k X k Z Z
Vol(B) = Vol(B ∩ Ci ) = exp( y · v) = exp( y · v) .
i=1 i=1 Ci C

Since y · v < 0 for all y ∈ C, we may also assume that the ray generators ri are scaled so
that ri · v = −1. This means that Vol(B) = |det(A)|, where A = (r1 , . . . , rn ) is a matrix whose
columns are the ray generators. Since C is simplicial, a point y = ( y1 , . . . , yn ) ∈ C can be
written uniquely as y = Az, where z = (z1 , . . . , zn ) are new nonnegative coordinates. Hence
Z Z Z Y n
exp( y · v) = |det(A)| exp(v ⊤ Az)dz = |det(A)| exp(zi (ri · v))dz .
n n
C R+ R+ i=1

We now perform the integration for each variable zi separately to conclude

exp(zi (ri · v)) zi =∞
Z Yn • ˜
exp( y · v) = |det(A)| .
C i=1
r i · v z i =0

The integral is finite if and only if ri · v < 0 for all rays, which is equivalent to y · v < 0 for all
y ∈ C \ {0}. In this case, it equals | det(A)| = Vol(B), as desired.

8
 SciPost Phys. Lect. Notes 75 (2023)

Proof of Theorem 1.6. We start with a change of variables x j = exp( y j ):

exp( y · ν)
Z Z
−s ν dx
I (s, ν) = f x = dy .
Rn f (exp( y))
x s
Rn +

To show absolute convergence, we need to prove that

exp( y · ν)
Z Z
exp( y · Re(ν))
dy = dy < ∞.
Rn f (exp( y)) Rn f (exp( y))
s Re(s)

p p
The equality in this display uses |r a+ −1b | = |exp(log(r)(a + −1b))| = |exp(a log(r))| = r a
for a positive real number r and real numbers a, b. Notice that this means we may assume s
and ν are real. We first consider the case where ℓ = 1. Let P be the Newton polytope ∆( f ) of
f = f1 . We have seen in (16) that its normal fan subdivides Rn into n-dimensional polyhedral
cones C v , where v runs over the vertices of P. Therefore
X XZ exp( y · ν)
I (s, ν) = I v (s, ν) = dy . (20)
−C f (exp( y))
s
v v v

Notice that we use the cones −C v instead of C v for this decomposition,Pbecause these are the
domains on which we can find easy bounds for the integrand. Let f = Pα cα · x α with cα > 0,
as in Assumption 1. With our change of variables, this becomes f (e y ) = α cα · e y·α . Since v is
a vertex of ∆ f , one of the exponents α equals v. For y ∈ −C v , y · v ≥ y ·α for α ∈ supp( f )\{v}.
This gives the following chain of inequalities:
X
c v · exp( y · v) ≤ f (exp( y)) ≤ cα · exp( y · v) . (21)
α

This leads to a chain of inequalities of integrals. Let M = α cα . Since s > 0, we have

P

Z Z
M −s · exp( y · (ν − sv)) d y ≤ I v (s, ν) ≤ c v−s · exp( y · (ν − sv)) d y .
−C v −C v

The integral appearing on the left and right of this expression can be written as
Z
exp( y · w) d y , with w = sv − ν .
Cv

By Lemma 1.8, this integral converges if and only if y · w < 0 for all y ∈ C v \ {0}. Notice that w
is a vertex of the polytope s · P − ν, and the cone Csv−ν in its normal fan equals C v . By Lemma
1.5, y · w < 0 for all y ∈ Cw \ {0} = C v \ {0} for all vertices w if and only if 0 ∈ int(s · P − ν).
This is equivalent to ν ∈ int(s · P), which proves the theorem for ℓ = 1.
When ℓ > 1, the formula (20) generalizes to a sum over the vertices of P(s):
X XZ exp( y · ν)
I (s, ν) = I v (s, ν) = dy .
v v
f
−C 1 (exp( y)) s1 · · · f (exp( y))sℓ
ℓ
v

Here P(s) = s1 · ∆( f1 ) + · · · + sℓ · ∆( fℓ ). Each vertex v of P(s) is a sum v = s1 · v1 + · · · + sℓ · vℓ of

vertices si ·vi of the summands si ·∆( f i ). Here vi is the face ∆( f i ) y ⊂ ∆( f i ) forPany interior point
of C v . With the notation for coefficients as in Assumption 1, we set Mi = α ci,α . Bounding
each of the f i via ci,vi exp( y · vi ) ≤ f i (exp( y)) ≤ Mi exp( y · vi ) as in (21), we find
ℓ
Y Z ℓ
Y Z
−s −s
Mi i · exp( y · (ν − v)) d y ≤ I v (s, ν) ≤ ci,v i · exp( y · (ν − v)) d y. (22)
i
i=1 −C v i=1 −C v

Again, the integral in these expressions converges if and only if 0 ∈ int(P(s) − ν).

9
 SciPost Phys. Lect. Notes 75 (2023)

Remark 1.9. The decomposition of the integral I (s, ν) in (20) is referred to as sector decom-
position in the physics literature [36, Section 3.4]. This is used in state-of-the-art algorithms
for evaluating Feynman integrals numerically, which use Monte Carlo sampling and tropical
geometry [11, 12].

By Lemma 1.8, if Re(si ) > 0 and ν ∈ int(P(s)), the bounds in (22) can be written as
ℓ ℓ
−Re(si ) −Re(si )
Y Y
Mi · Vol(B v−ν ) ≤ |I v (s, ν)| ≤ ci,v · Vol(B v−ν ) . (23)
i
i=1 i=1

Here B v−ν = { y ∈ C v : y · (v − ν) ≥ −1} is the portion of (P(s) − ν)◦ corresponding to the

vertex v, see Figure 3. This will be important in our discussion on field theory limits.
Theorem 1.6 identifies a region of convergence of the integral (13), which is geometrically
described by a convex polytope. As pointed out in [9, Example 2.3], the integral might con-
verge on a larger domain. E.g., the assumption Re(si ) > 0 is in general not necessary. However,
as it turns out, our domain is large enough to allow a unique meromorphic continuation of
I (s, ν) to the entire parameter space Cℓ+n . This is similar in spirit to the fact that the integral
representation of the gamma function seen in (2) only converges for Re(u) > 0. The mero-
morphic function Γ (u) is obtained by extending that integral function on R+ to a function on
C \ Z≤0 satisfying Γ (u + 1) = u Γ (u). Let us now consider the beta function.
y
Example 1.10. The coordinate change x = 1+ y brings the integral in (2) into the form (13):

1 ∞
xν yν
Z Z
dx dy
= , with s̃ = ν + 1 − s . (24)
0 (1 − x) x
s
0 (1 + y) y
s̃

Theorem 1.6 predicts convergence when Re(s̃) > 0 and ν ∈ int(P(s̃)), where

P(s̃) = {p ∈ R : p ≥ 0, −p ≥ −Re(s̃)} . (25)

To justify the blue equality in (2), we observe that when these convergence conditions hold,
Z∞ Z∞
ν−s −t yν dy
Γ (ν + 1 − s) · B(ν, 1 − s) = t e dt ν+1−s
0 0 (1 + y) y
Z ∞Z ∞ ‹ν−1  ‹−s
ty t t d y dt
= e−t .
0 0 1+ y 1+ y (1 + y)2

ty t td ydt
With the coordinate change u = 1+ y , w= 1+ y we have u + w = t and (1+ y)2
= dudw. Hence
Z ∞ Z ∞
ν−1 −u
Γ (ν + 1 − s) · B(ν, 1 − s) = u e du w−s e−w dw = Γ (ν)Γ (1 − s) .
0 0

While the integrals in these equalities only make sense in their respective convergence regions,
we may use the definition of the gamma function to extend the beta function to a meromorphic
function on C2 :
Γ (ν)Γ (1 − s)
B(ν, 1 − s) = .
Γ (ν + 1 − s)
Its poles are countably many lines in (s, ν)-space, given by ν, 1 − s ∈ Z≤0 . ⋄

The fact that the beta function extends to a meromorphic function whose poles are given
by some gamma functions is an example of a general phenomenon, proved in [9]. We include
the statement, but omit the proof. We refer the reader to [9, Theorem 2.4] for full details.

10
 SciPost Phys. Lect. Notes 75 (2023)

Figure 4: The poles of the meromorphic continuation of I5 with

(s13 , s14 , s34 ) = (1, 1, 1) are lines emanating from the boundary of the pen-
tagon P from Example 1.7.

Theorem 1.11. Suppose the Minkowski sum ∆( f1 )+· · ·+∆( fℓ ) has dimension n and the polytope
Pℓ
P(s) = i=1 Re(si ) · ∆( f i ) is given by N < ∞ inequalities:

P(s) = {p ∈ Rn : ri · p ≥ w i · Re(s) , i = 1, . . . , N } , r i ∈ R n , w i ∈ Rℓ .

Under Assumption 1, I (s, ν) from (13) admits a meromorphic continuation of the form
N
Y
Φ f (s, ν) · Γ (ri · ν − w i · s) , (26)
i=1

where Φ f (s, ν) is an entire function.

It is worth noting that only the entire factor Φ f (s, ν) in Theorem 1.11 depends on the spe-
cific positive coefficients of f . The gamma factors only depend on the polyhedral data coming
from P(s). For fixed, positive s, the poles of this meromorphic continuation are hyperplanes
emanating from the boundary of P(s). This is illustrated in Figure 4 for (18).

Example 1.12. In Example 1.10, we read off from (25) that N = 2 and r1 = 1, w1 = 0,
r2 = −1, w2 = −1. The gamma factors in Theorem 1.11 are Γ (r1 · ν − w1 · s̃) = Γ (ν) and
Γ (r2 · ν − w2 · s̃) = Γ (1 − s). The entire function Φ1+ y (s, ν) equals Γ (ν + 1 − s)−1 . ⋄

2 Limits and critical points

In this section, we continue to use the integration contour Γ = R+
n
and we work under Assump-
tion 1. We now think of the parameters s, ν to be fixed, satisfying the conditions of Theorem
1.6. The main novelty with respect to Section 1 is that we introduce a new parameter δ, of
which the integral I is now a function:
ν1 νn
x 1δ · · · x nδ dx 1
Z Z
1 dx n 1
1 dx
I (δ) = n · ∧ ··· ∧ = n f −s x ν δ
. (27)
δ δ
s1 sℓ
x1 xn x
n
R+ f 1δ · · · f ℓ δ Γ

Notice that δ−1 plays the role of the inverse string tension α′ in the string amplitude (9). We
are interested in the opposite limits limδ→∞ I (δ) and limδ→0+ I (δ). Motivated by the physics

11
 SciPost Phys. Lect. Notes 75 (2023)

application, these are called field theory limit and high energy limit respectively [8, 52]. They
are the leading terms in the series expansions of I (δ) around δ = ∞ and δ = 0.
As it turns out, both the field theory limit and the high energy limit can be expressed in
terms of complex critical points of the potential function or log-likelihood function

log L = log f −s x ν = −s1 log f1 − · · · − sℓ log fℓ + ν1 log x 1 + · · · + νn log x n .

These critical points are the complex solutions to the n rational function equations
∂ f1 ∂ fℓ
∂ (log f −s x ν ) νj ∂ xj ∂ xj
= − s1 − · · · − sℓ = 0, j = 1, . . . , n . (28)
∂ xj xj f1 fℓ
These rational functions are defined where neither x j nor f i (x) are zero. We define

X = {x ∈ Cn : x 1 · · · x n · f1 (x) · · · fℓ (x) ̸= 0 } . (29)

This is an example of a very affine variety, see [41, page 6]. The set of complex critical points
of log L is Crit(log L) = {x ∈ X : x satisfies (28)}. Since (28) consists of n equations in n
unknowns, we expect Crit(log L) to be finite. A solution x ∈ Crit(log L) is degenerate if

∂ ∂
  ‹
Hlog L = det x j xk log L(x) = 0. (30)
∂ xj ∂ xk j,k

This determinant is called the toric Hessian of log L. It is much like the usual Hessian determi-
nant, but with ∂ /∂ x j replaced by the Euler operator x j (∂ /∂ x j ). Using the toric version will
be convenient later in the section. The following result is Theorem 1 in [40].
Theorem 2.1. There is a dense open subset U ⊂ Cℓ+n such that for (s, ν) ∈ U, the number of
solutions to (28) equals the signed Euler characteristic (−1)n · χ(X ) of the very affine variety X ,
and all solutions are non-degenerate, meaning that Hlog L (x) ̸= 0 for all x ∈ Crit(log L).
Theorem 2.1 says that the number of points in Crit(log L) depends only on the topology of
the space X . We will see how to compute the Euler characteristic χ(X ) below.
In Section 2.1 we discuss how to compute Crit(log L) using numerical homotopy continu-
ation. Sections 2.2 and 2.3 explain how these critical points are used to compute field theory
and high energy limits respectively.

2.1 Computing critical points

Our goal is to solve the equations (28), assuming that (s, ν) ∈ Cℓ+n belongs to the set U from
Theorem 2.1. This ensures that there are finitely many solutions, all of them non-degenerate,
and the number of solutions we find is the Euler characteristic of X (up to sign).
We use the Julia package HomotopyContinuation.jl (v2.6.4) for all computations in
this section [14]. Our approach follows that in [59] and is illustrated by means of an example,
for which we use the integral (18). The function log L is

log L = −s13 log(1 + x 1 ) − s14 log(1 + x 1 + x 2 ) − s34 log(x 1 + x 2 ) + ν1 log x 1 + ν2 log x 2 .

Its partial derivatives give two rational function equations g1 = g2 = 0 in the unknowns x 1 , x 2 :
s13 s14 s34 ν1 s14 s34 ν
g1 = − − − + , g2 = − − + 2 . (31)
1 + x1 1 + x1 + x2 x1 + x2 x1 1 + x1 + x2 x1 + x2 x2
Importantly, we think of s and ν as parameters at this stage. We will emphasize the dependence
of g i on these parameters by writing g i (x; s, ν). The fixed complex parameters we want to solve
for are denoted by (s∗ , ν∗ ) ∈ U. Here is how to code this up in Julia:

12
 SciPost Phys. Lect. Notes 75 (2023)

Figure 5: Illustration of the monodromy method.

using HomotopyContinuation # load the package 1

n = 2; l = 3; @var ν[1:n] s[1:l] x[1:n] # declare variables and parameters 2
f = [1 + x[1]; 1 + x[1] + x[2]; x[1] + x[2]] 3
logL = - sum([s[i]*log(f[i]) for i = 1:l]) + sum([ν[j]*log(x[j]) for j = 1:n]) 4
g = differentiate(logL, x) 5
g_sys = System(g, parameters = [s; ν]) # system of equations with parameters 6
s_star = [1;1;1]; ν_star = [1;1] # choice of target parameters 7

Here we chose s∗ = (1, 1, 1) and ν∗ = (1, 1), like in Example 1.7. The strategy for solving
g1 (x; s∗ , ν∗ ) = g2 (x; s∗ , ν∗ ) = 0 consists of two steps:

1. Solve g1 (x; s̃, ν̃) = g2 (x; s̃, ν̃) = 0 for a different set of parameters (s̃, ν̃) ∈ U.

2. Deform the start parameters (s̃, ν̃) continuously into the target parameters (s∗ , ν∗ ) and,
along the way, keep track of the solutions to g1 (x; s, ν) = g2 (x; s, ν) = 0.

Both these steps require numerically tracking solution paths as we vary the parameters. This
can be phrased as numerically solving an ordinary differential equation called the Davidenko
equation. For details, we refer to the standard textbook [58].
Step 1 is done using the monodromy method [25]. We explain how this works in a nut-
shell, using Figure 5 as an illustration. In that cartoon, the solutions for a fixed point (s, ν)
are represented by the points on the blue surface lying directly above it. This surface rep-
resents the incidence space {(x, s, ν) ∈ X × Cℓ+n : g1 (x; s, ν) = g2 (x; s, ν) = 0}. Choose
a random point x̃ ∈ X , and let (s̃, ν̃) be any solution to the linear system of equations
g1 (x̃; s, ν) = g2 (x̃; s, ν) = 0. Clearly, x̃ is a point lying above (s̃, ν̃). This is called the seed
solution. Now walk a loop in (s, ν)-space while keeping track of the seed solution x̃ along the
way. When we arrive back at (s̃, ν̃), there is a good chance we picked up a new solution x̃ new
to the system g1 (x; s̃, ν̃) = g2 (x; s̃, ν̃) = 0. Now repeat this procedure to populate the solution
set. In practice, this technique is extremely effective. In Julia, all this happens via

R1 = monodromy_solve(g_sys) 1

start_pars = parameters(R1); start_sols = solutions(R1) 2

The variable start_pars stands for start parameters. It contains ℓ + n = 5 complex numbers,
the first ℓ = 3 of which give s̃, and the last n = 2 give ν̃. If all went well, the variable
start_sols contains all solutions to g1 (x; s̃, ν̃) = g2 (x; s̃, ν̃) = 0. Hence, step 1 is completed.

13
 SciPost Phys. Lect. Notes 75 (2023)

By Theorem 2.1, the number of solutions in start_solutions equals (−1)n · χ(X ). This
gives a way of computing χ(X ), which has been applied in some challenging cases [2, 59]. In
our example, the number of solutions is 2. Here is a way to verify that χ(X ) = 2. The real
part X R of the very affine variety X is the complement of an arrangement of five lines in R2 .
These lines are given by {x 1 = 0}, {x 2 = 0}, {1 + x 1 = 0}, {1 + x 1 + x 2 = 0} and {x 1 + x 2 = 0}.
By [60, Theorem 1.2.1], the signed Euler characteristic of X is the number of bounded cells of
X R . In our case, the bounded cells are two triangles.
In step 2, we use our start solutions as initial conditions for path tracking from (s̃, ν̃) to
∗ ∗
(s , ν ). That is, we use the solutions start_sols for parameters start_pars to compute
the solutions solutions(R2) for the target parameters [s_star; ν_star]:

R2 = solve(g_sys, start_sols; start_parameters = start_pars, 1

target_parameters = [s_star;ν_star]) 2

solutions(R2) 3

The last line prints an accurate numerical approximation of the two critical points:
p   p 
5−1 − 5−1
,1 , ,1 . (32)
2 2

2.2 Dual volumes in field theory limits

We switch back to the integral (27). The field theory limit of I (δ) is limδ→∞ I (δ). This has a
Pℓ
nice description in terms of our polytope P(s) = i=1 Re(si ) · ∆( f i ) from Theorem 1.6, and in
terms of the critical points Crit(log L) computed in the previous section. The statement uses
the toric Hessian determinant H− log L of minus the log-likelihood function, see (30).
Theorem 2.2. Let Re(si ) > 0 for i = 1, . . . , ℓ and suppose that ∆( f1 ) + · · · + ∆( fℓ ) has dimension
n. If f1 , . . . , fℓ satisfy Assumption 1 and ν ∈ int(P(s)), then we have
X
lim I (δ) = Vol((P(s) − ν)◦ ) = H− log L (x)−1 . (33)
δ→∞
x∈Crit(log L)

Proof. We prove the first equality. The second equality uses [8, Section 7.1, Claim 4]. See
also [59, Theorem 13]. For any fixed δ ∈ R+ , we have ν/δ ∈ int(P(s/δ)), and a vertex v of
P(s) gives a vertex (v − ν)/δ of P(s/δ) − ν/δ. We can use (23) to obtain the estimate
ℓ ℓ
−Re(si )/δ −Re(si )/δ
XY XY
Mi · Vol(B v−ν ) ≤ δ n · I (δ) ≤ ci,v · Vol(B v−ν ) . (34)
δ i δ
v i=1 v i=1

The sums are over vertices of P(s). The factor δ n in the middle comes from
−n
I (δ) = δ · I (s/δ, ν/δ), with I (s, ν) as in (13). Using the scaling property of the volume
Vol(B(v−ν)/δ ) = δ n · Vol(B v−ν ), we can cancel δ n from (34). Taking the limit δ → ∞ gives
X
lim I (δ) = Vol(B v−ν ) = Vol((P(s) − ν)◦ ) ,
δ→∞
v
as desired. For the last equality, see (17) and Figure 3.
Example 2.3. Let us verify the formula (33) for the integral representation (24) of the beta
function. The Newton polytope P(s̃) is a segment given by (25). The dual polytope (P(s̃) − ν)◦
1 1
is given by ( ν−s̃ , ν ). Its volume Vol((P(s̃) − ν)◦ ) is ν1 + s̃−ν
1 s̃
= ν(s̃−ν) . Checking that this equals
our field theory limit can be done in one line of Mathematica code:

14
 SciPost Phys. Lect. Notes 75 (2023)

Limit[1/δ*Integrate[y^(ν/δ - 1)/(1 + y)^(s/δ), {y, 0, Infinity}], δ -> Infinity] 1

Finally, we compute the sum of H− log L ( y) evaluated at the critical points of log L. We solve

d ν s̃
log L( y) = − = 0.
dy y 1+ y
ν
The unique solution is y = s̃−ν . The toric Hessian determinant of − log L is
 ‹  ‹
d d d s̃ y s̃ y
H− log L ( y) = − y y log L( y) = − y ν− = . (35)
dy dy dy 1+ y (1 + y)2
ν ν(s̃−ν)
The value at y = s̃−ν is s̃ . We have now confirmed (33). ⋄

Example 2.4. Consider the integral I (δ) from (12) with α′ = δ1 . The Newton polytope
P(s) = P(s13 , s14 , s24 ) of f1 (x 1 , x 2 ) = 1 + x 1 , f2 (x 1 , x 2 ) = 1 + x 1 + x 2 , f3 (x 1 , x 2 ) = x 1 + x 2
is two dimensional for positive s. When s13 = s14 = s24 = 1, it is as in Figure 1. We take
ν1 = ν2 = 1. The values of H− log L at the two critical points (32) are

1 p
1 p

25 + 11 5 , 25 − 11 5 .
2 2
The sum of the reciprocals of these two numbers is 5. This is the area of (P(s)−ν)◦ in the right
part of Figure 3, normalized by a factor 2! = 2 (recall our definition of Vol in the discussion
preceding Lemma 1.8). Let us illustrate the computation of the dual volume using the Julia
package OSCAR.jl (v0.12.0) [54]. It calls polymake for polytope computations [29]. The
Newton polytope P = P(s13 , s14 , s24 ) of Example 2.4 is computed as follows:

using Oscar #load the package 1

P1 = convex_hull([0 0;1 0]) #Newton polytope of f1 2

P2 = convex_hull([0 0;1 0;0 1]) #Newton polytope of f2 3

P3 = convex_hull([1 0;0 1]) #Newton polytope of f3 4

P = P1+P2+P3 #Minkowski sum 5

The dual polytope (P(s) − ν)◦ and its volume are computed by the commands polarize
and volume respectively. The vertices of the dual polytope (P − ν)◦ are (1, 1), (1, 0), (0, −1),
(−1, −1), (0, 1) as in Figure 3. The normalized volume is 5, as expected.

DP = polarize(P+[-1,-1]) #dual polytope of P-ν 1

println(vertices(DP)) #print the vertices 2

factorial(2)*volume(DP) #normalized volume - output: 5 3

The reader is encouraged to repeat this example for the Feynman integral (7). ⋄

We point out that the field theory limit α′ → 0 of the string amplitude (9) is the scattering
amplitude for a physical model called bi-adjoint scalar φ 3 theory. The expression in terms of
critical points was first discovered in [18]. The critical point equations (28) are called the
scattering equations in this context. For a connection to algebraic statistics, see [59].
As a final remark on field theory limits, note that the coordinates of the individual critical
points in Crit(log L) are algebraic functions of s, ν. They are usually not rational functions, like
in Example 2.3. For instance, eliminating x 2 from (31) gives a quadratic equation in x 1 , result-
ing in the square roots in (32) via the quadratic formula. However, the sum over Crit(log L)

15
 SciPost Phys. Lect. Notes 75 (2023)

in Theorem 2.2 is a rational function in s and ν by Galois theory. This is called the canonical
function of P(s). On the domain Re(si ) > 0, ν ∈ int(P(s)), it evaluates to Vol((P(s) − ν)◦ ).
When s with positive real part is fixed and P(s) is viewed as a polytope in n-dimensional ν-
space Rn , the canonical function defines a meromorphic top form on Rn . That form is called
the canonical form of P(s) in the theory of positive geometries [6].

2.3 Saddle point approximation in high energy limits

While the field theory limit limδ→∞ I (δ) is obtained by summing over the complex points
Crit(log L), the high energy limit limδ→0+ I (δ) is governed by a single, positive critical point.

Theorem 2.5. Let si ∈ R+ for i = 1, . . . , ℓ and suppose that ∆( f1 ) + · · · + ∆( fℓ ) has dimension

n. If f1 , . . . , fℓ satisfy Assumption 1 and ν ∈ int(P(s)), then Crit(log L) ∩ R+
n
consists of one point
{a}. Moreover, the following formula holds:
‹− 2n
2π

− 1
1
lim L(a)− δ I (δ) = H− log L (a) 2 . (36)
δ→0+ δ

We used the following conventions in (36). As above, for a positive real number r, we take
1 π 1
the branch of the logarithm for which log r ∈ R. For the square root, we set (−r) 2 = e i 2 r 2 .
Notice that, rather than requiring Re(si ) > 0, here weponly allow real values for si . The reader
can check that, using the values s13 = s14 = s34 − −1 = 1 in the example of Section 2.1
instead, there are no positive critical points.
To prove Theorem 2.5, we use two lemmas. The first is on the concavity of log L.

Lemma 2.6. Let f1 , . . . , fℓ satisfy Assumption 1 and let s1 , . . . , sℓ ∈ R+ . The function

log L(ez1 , . . . , ezn ) with L(x) = f −s x ν is strictly concave in z ∈ Rn . The toric Hessian matrix

∂ ∂
  ‹
xj xk log L(x) , (37)
∂ xj ∂ xk j,k
n
of log L is negative definite for any x ∈ R+ .

Proof. Substituting x j = exp(z j ), the toric Hessian matrix (37) is the usual Hessian matrix
Pℓ
of − i=1 si log f i (ez1 , . . . , ezn ). Each summand is strictly concave for z ∈ Rn by [17, Theorem
1.13], so the toric Hessian is indeed a negative definite matrix.

To state the next lemma, we introduce a version of the algebraic moment map µC : X −→ Cn .
This name comes from toric geometry, see Fulton’s book [28, Section 4.2]. Our moment map
is slightly different from the one used by Fulton. It is given by
ℓ ℓ
‚ Œ
X
−1 ∂ fi X
−1 ∂ f i
µC (x) = x1 si f i (x) (x), . . . , x n si f i (x) (x) . (38)
i=1
∂ x1 i=1
∂ xn

This map is closely related to our critical points. From (28), it is clear that x ∈ Crit(log L) if
and only if µC (x) = ν. The crucial properties of µC are summarized in the following Lemma.

Lemma 2.7. Let si ∈ R+ for i = 1, . . . , ℓ and suppose that ∆( f1 ) + · · · + ∆( fℓ ) has dimension n.

If f1 , . . . , fℓ satisfy Assumption 1, then µC (R+
n
) ⊂ int(P(s)). Moreover, the restriction
n
µ = (µC )|R+n = (µ1 , . . . , µn ) : R+ −→ int(P(s)) , (39)

∂ µk n
€ Š
of µC to R+
n
is a diffeomorphism, and the Jacobian matrix ∂ x j j,k=1 is positive definite.

16
 SciPost Phys. Lect. Notes 75 (2023)

Proof. This theorem follows from [28, Section 4.2] if ℓ = 1. The statement for ℓ ≥ 1 has
appeared in [8, Claim 4]. For the readers’ convenience, we provide a concise proof. Let
Ai = supp( f i ) ⊂ Zn be the set of exponents appearing in f i (x) = α∈Ai ci,α x α . We construct
P

the Cayley configuration A ⊂ Zℓ+n of A1 , . . . , Aℓ . This is given by

A = {(ei , α) : α ∈ Ai , i = 1, . . . , ℓ} ⊂ Zℓ+n , (40)

where ei is the i-th standard basis vector of Zℓ . Consider the Laurent polynomial
fˆ = s1 y1 f1 + · · · + sℓ yℓ fℓ ∈ R+ [ y1 , . . . , yℓ , x 1±1 , . . . , x n±1 ]. We define the map

∂ fˆ ∂ fˆ ∂ fˆ ∂ fˆ
 
µ̂ : Rℓ+n
+ → int(pos(A)) , ( y, x) 7−→ y1 , . . . , yℓ , x1 , . . . , xn .
∂ y1 ∂ yℓ ∂ x1 ∂ xn

Here, pos(A) is the positive hull seen in (15). The reader who is unfamiliar with moment
maps should check that the image of µ̂ indeed lies in the interior of the cone pos(A). By the
statement labeled (An ) in [28, page 83], µ̂ is a diffeomorphism.
We consider the polytope P̂(s) consisting of all points in pos(A) whose first ℓ coordinates
are (s1 , . . . , sℓ ). The preimage of P̂(s) under µ̂ is given by

X̂ + = {( y, x) ∈ Rℓ+n
+ : y1 f1 (x) = · · · = yℓ fℓ (x) = 1} .

In fact, µ̂|X̂ + : X̂ + → int( P̂(s)) is a diffeomorphism. We now relate this to the moment map
µ in (39). To identify the domains of µ and µ̂, we introduce the map κ : R+ n
→ X̂ + with
−1 −1
κ(x) = ( f1 (x) , . . . , fℓ (x) , x 1 , . . . , x n ). For the co-domains, note that ι : int(P(s)) → int( P̂(s))
with ι(v) = (s1 , . . . , sℓ , v) is an isomorphism. We obtain a diagram of diffeomorphisms

µ̂
X̂ + / int( P̂(s))
O O
κ ι
µ
n
R+ / int(P(s))

The Jacobian matrix of µ is positive definite on R+

n
if and only if the toric Jacobian matrix

ℓ
∂ µk ∂ ∂ ∂ ∂
    ‹ X   ‹
xj = −x j xk log L(x) = si x j xk log f i .
∂ xj j,k
∂ xj ∂ xk j,k i=1
∂ xj ∂ xk j,k

n
is positive definite on R+ . The positivity follows from Lemma 2.6.

Note that Lemma 2.7 implies that, under our assumptions, Crit(log L) ∩ R+ n
= µ−1 (ν) con-
sists of a single point {a}. This is the first claim in Theorem 2.5. While Theorem 2.5 uses the
fiber µ−1 (ν), Theorem 2.2 sums over the fiber µ−1 C (ν) of the complexified map µC .

Proof of Theorem 2.5. We have established that Crit(log L) ∩ R+ n

= {a} is a singleton (Lemma
2.7). Let U be a small open neighborhood of the positive critical point a. The idea of the proof
is to decompose I (δ) into two parts:
‚Z Z Œ
n 1 n 1 1 dx 1 dx
δ− 2 L(a)− δ I (δ) = δ− 2 L(a)− δ L(x) δ + L(x) δ .
n
R \U x U x
+

We will show that the integral over R+ n

\ U does not contribute to the limit δ → 0+ , and
the integral over U gives rise to a Gaussian integral. By Lemma 2.6, log L attains its unique
global maximum at x = a and log L(ez1 , . . . , ezn ) is a concave function of (z1 , . . . , zn ) ∈ Rn (see

17
 SciPost Phys. Lect. Notes 75 (2023)

Figure 6). There exists a positive number ϵ such that the inequality log L(x) − log L(a) ≤ −ϵ
n
is true for R+ \ U. We obtain the following inequality for 0 < δ < 1:
Z Z
1 dx
dx
− δ1 − δϵ
L(a) L(x) δ =e exp δ−1 (log L(x) − log L(a) + ϵ) (41)
n
R+ \U x n
R+ \U x
Z
− δϵ dx
≤e exp (log L(x) − log L(a) + ϵ) . (42)
Rn \U x
+

Here, we used the fact that log L(x) − log L(a) + ϵ is negative for any x ∈ R+n
\ U, and the
final integral in (42) is bounded because of Theorem 1.6. The inequality (42) shows that the
integral over R+ n
\ U converges to zero as δ → 0+ .
Let P be an orthogonal matrix which diagonalizes the Hessian matrix of log L:
∂ log L
 2 
⊤
P · (a) · P = D.
∂ x j∂ xk j,k

Here D is an n×n diagonal matrix, with negative diagonal entries. We perform a linear change
of coordinates y = P T (x − a). The Taylor expansion of log L(x) − log L(a) around y = 0 looks
like 12 ( y T D y + r( y)). Plugging this into our integral gives
y D y + r( y)
Z Z  T 
− δ1
1 dx dy
L(a) L(x) δ = exp .
i (P y + a)i
Q
U x P T (U−a) 2δ
The last denominator is the product of the entries of P y + a. Without loss of generality,pwe
may assume that P T (U − a) is a product of small intervals (−ε, ε)n . Replacing yi with yi / δ,
the last integral becomes
p
r( δ y)
Z  
n 1 T dy
δ2 exp y Dy + Q p .
p p
(−ε/ δ,ε/ δ)n 2 δ i ( δP y + a) i
p p p
The function r( δ y)/δ is bounded for 0 < δ < 1 and y ∈ (−ε/ δ, ε/ δ)n and it converges
to 0 when δ tends to 0. Therefore, Lebesgue’s dominance convergence theorem proves
Z Z
− 2n − δ1
1 dx 1 T dy
lim δ L(a) L(x) δ = e2 y Dy .
δ→0 +
U x Rn a1 · · · an
This leaves us with a Gaussian integral. To finish the proof, recall that
Z∞ v
t 2π
λ 2
z
e 2 dz = , for λ < 0 ,
−∞ −λ
and use the fact that H− log L (a) = (a1 · · · an )2 · i (−Dii ).
Q

Example 2.8. Let us verify the formula (36) for the integral representation (24) of the beta
function. The integral (24) is expressed by Gamma functions as in (2):
ν
s̃ − ν ν Γ ( s̃−ν
δ )Γ ( δ )
 ‹
I (δ) = B , = .
δ δ Γ ( δs̃ )
1 ν
The function L(a)− δ , where a is the unique critical point s̃−ν from (2.3), is given by
‹−s̃  − δ1
ν ν

s̃
.
s̃ − ν s̃ − ν
p 1
formula Γ (x) ∼
Stirling’s Ç 2πe−x x x− 2 (x → +∞) shows that the left-hand side of (36) is
1
s̃
given by ν(s̃−ν) . We have seen in Example 2.3 that this equals H− log L (a)− 2 . ⋄

18
 SciPost Phys. Lect. Notes 75 (2023)

Figure 6: The likelihood function of Example 2.4 attains its maximum at

(x 1 , x 2 ) ≈ (0.618, 1).

3 Twisted (co)homology
n
In this section, we abandon the concrete integration contour R+ , and we drop Assumption 1.
±1 ±1 ℓ ℓ
We fix f = ( f1 , . . . , fℓ ) ∈ C[x 1 , . . . , x n ] , s = (s1 , . . . , sℓ ) ∈ C and ν = (ν1 , . . . , νn ) ∈ Cn .
The perspective we take is that the Euler integral (1) is the result of a pairing between the
integration contour Γ and the differential n-form dx x . More generally, an n-form φ gives
Z
〈Γ , φ〉 = f −s x ν φ . (43)
Γ

This works nicely when Γ is a twisted n-cycle and φ is a twisted n-cocycle. We will introduce
these concepts, and see that the pairing (43) is a perfect pairing of finite dimensional C-vector
spaces. In particular, the integral (43) always evaluates to a (finite) complex number.
This story is reminiscent of the classical duality between singular homology and de Rham
cohomology, where one pairsRan integration contour ∆ on a complex manifold X with a differ-
ential form φ by evaluating ∆ φ [34, Chapter 0]. In our setting, X is the very affine variety
seen in (29). The material in this section is like that standard theory, but with a twist. For
instance, recall from the beginning of Section 1 that our integrand f −s x ν is multi-valued. To
make sense of the integral (43), we need to specify a branch. We will see that this information
is carried by our twisted cycle Γ . Also, a central role in de Rham’s cohomology theory is played
by Stokes’ theorem, which says that for an (n − 1)-form ψ,
Z Z
dψ = ψ. (44)
∆ ∂∆

In our setting, the twisted boundary operator ∂ω takes the choice of branch into account, and
the twisted differential ∇ω replaces the ordinary differential d to accommodate our integrals:
Z Z
f −s x ν ∇ω ψ = f −s x ν ψ . (45)
Γ ∂ω Γ

The meaning of the index ω will become clear soon. For now, it simply indicates the twist.
The theory of twisted (co)homology goes back to the seminal work of Deligne and
Grothendieck [23]. It has been investigated in the context of Euler integrals and hyperge-
ometric functions by several authors, among which we mention Aomoto, Gelfand, Iwasaki,

19
 SciPost Phys. Lect. Notes 75 (2023)

∆0,0 (ϵ) ∆1,0 (−ϵ)

[ϵ, 1 − ϵ]
0 1

∆0,π (ϵ) ∆1,π (−ϵ)

Figure 7: Five simplices in C \ {0, 1}.

Kapranov, Kita, Matsumoto and Zelevinsky. See [5, 31] and references therein. The relevance
of this theory in particle physics was first realized by Mastrolia and Mizera [46].
The section is organized as follows. We start by discussing twisted chains and cycles,
leading to a twisted version of the usual chain complex of X . Next, we switch to the dual
complex, called the twisted de Rham complex of X . We discuss properties of the (co)homology
of these complexes, ultimately leading to the perfect pairing in (43).

3.1 Twisted chains

Throughout the section, X is the very affine variety from (29). A singular k-simplex ∆ in X is
a continuous (not necessarily injective) map from the standard k-simplex to X . The C-vector
space generated by all singular k-simplices is the space of singular k-chains, denoted by Ck (X ):
M
Ck (X ) = C · ∆. (46)
∆⊂X , k-simplex

Example 3.1. Examples of 1-simplices X = C \ {0, 1} are illustrated in Figure 7. Here ϵ ∈ R

lies in (0, 1/2). There are 5 simplices in total. Four of them are semicircles, parameterized by
p
∆a,θ (ϵ) = { t 7→ a + ϵ · exp( −1(θ + tπ) } , t ∈ [0, 1] .

Here, a is 0 or 1 and θ is 0 or π. The remaining simplex is the line segment [ϵ, 1 − ϵ],
parameterized by t 7→ (1−t)ϵ+t(1−ϵ). These parameterizations fix the orientations visualized
by the arrows in Figure 7. ⋄

We need to modify this standard construction to account for multi-valuedness of f −s x ν .

The branches of f −s x ν generate the space of sections of a line bundle L−ω on X called a local
system. On an open subset U ⊂ X , these sections are

L−ω (U) = {τ : U → C : τ is holomorphic and dτ − dlog( f −s x ν ) τ = 0} . (47)

One checks that each branch τ of f −s x ν indeed satisfies the equation dτ − dlog( f −s x ν ) τ = 0.
The one-form dlog( f −s x ν ) is of crucial importance in this section, so we introduce the notation
ω = dlog( f −s x ν ). We have ω = g1 dx 1 + · · · + g n dx n , where g j are the rational functions in
(28). The symbol L−ω stresses the term −ω in the operator applied to τ in (47).
Picking a (linear combination of) branch(es) of f −s x ν on a singular k-simplex ∆ means
picking a section τ of L−ω on a sufficiently small open subset U ⊃ ∆. We formalize this
intuition by considering the direct limit

L−ω (∆) = lim L−ω (U) . (48)

−→
U⊃∆

Here the open sets U containing ∆ are ordered by inclusion, and when U ⊂ U ′ , the map
L−ω (U ′ ) → L−ω (U) is given by restriction. A reader who is unfamiliar with direct limits can
simply think of elements in L−ω (∆) as branches of f −s x ν , restricted to ∆.

20
 SciPost Phys. Lect. Notes 75 (2023)

Example 3.2. We continue Example 3.1. On each of the five simplices we define a section of
L−ω , where ω = (−s(1− x)−1 +ν x −1 ) dx is the logarithmic differential of f −s x ν = (1− x)−s x ν .
Here s and ν are fixed complex numbers. Notice that f −s x ν is the integrand of the beta function
in (2). At x = ϵ, both (1 − x) and x are positive. Let ζ = exp(−s log(1 − ϵ) + ν log ϵ) ∈ C,
where we evaluate the positive branch of the logarithm. The conditions

dτ0,0 − ωτ0,0 = 0 , and τ0,0 (ϵ) = ζ , (49)

uniquely define τ0,0 ∈ L−ω (∆0,0 (ϵ)). Constraints like τ0,0 (ϵ) = ζ are called initial conditions.
On our other simplices, we choose τa,θ ∈ L−ω (∆a,θ (ϵ)), τ− ∈ L−ω ([ϵ, 1 − ϵ]), with

τ0,π (−ϵ) = τ0,0 (−ϵ) , τ− (ϵ) = τ0,0 (ϵ) , τ1,π (1−ϵ) = τ− (1−ϵ) , τ1,0 (1+ϵ) = τ1,π (1+ϵ) .

To compute these boundary values,

p one can make use of the p parameterizations
p in Example
−s ν
3.1. For instance, τ0,0 (ϵ exp( −1tπ) is given by (1 − ϵ exp(p −1πt)) ϵ exp( −1νtπ), for
−s ν
t ∈ [0, 1]. At t = 1, this gives τp
0,0 (−ϵ) = (1 + ϵ) ϵ exp( −1πν). A similar computation for
τ0,π shows that τ0,π (ϵ) = exp( −1ν2π) ζ. In particular, τ0,0 (ϵ) ̸= τ0,π (ϵ), for non-integer ν.
We remark that having
p nontrivial sections of L−ω is the reason why we split up the circle
S0 (ϵ) = {t 7→ ϵ exp( −1t2π)}, t ∈ [0, 1] into two semicircles ∆0,0 (ϵ) and ∆0,π (ϵ). Indeed,
because of the nontrivial monodromy around x = 0, there are no nonzero holomorphic solu-
tions of dτ − ωτ = 0 on S0 (ϵ): L−ω (S0 (ϵ)) = 0. ⋄

This gives us all ingredients to define the space of twisted k-chains on X , with twist ω:
M
Ck (X , −ω) = ∆ ⊗ L−ω (∆) . (50)
∆⊂X , k-simplex

Comparing (50) with (46) motivates why this construction is sometimes called the space of
k-chains with coefficients in L−ω . In words, Ck (X , −ω) consists of finite C-linear combinations
of elements of the form ∆ ⊗ τ, where τ : ∆ → C is an element of L−ω (∆). We say that ∆ is
loaded with the branch τ. We note that Ck (X , −ω) is non-zero for any k ≥ 0.
Let us now clarify the meaning of (43) for a twisted k-chain Γ ∈ Ck (X , −ω).

Definition 3.3. Let Γ = p d p ∆ p ⊗ τ p ∈ Ck (X , −ω) be a twisted k-chain on X , with d p ∈ C.

P

Let φ = g(x) dx be a holomorphic k-form on X . We define

Z X Z X Z
〈Γ , φ〉 = f −s x ν φ = dp f −s x ν φ = dp τ p (x) φ . (51)
Γ p ∆ p ⊗τ p p ∆p

The integrals on the right are the familiar integrals of single valued k-forms on k-simplices.

We now explain how to take boundaries in this twisted setting. A k-simplex ∆ on X has
boundary ∂ ∆ = ∂ ∆0 + · · · + ∂ ∆k ∈ Ck−1 (X ). That is, if ∆ is given by the parameterization
ϕ : ∆ → X with ∆ the standard k-simplex in Rn , then ∂ ∆i is the (k − 1)-simplex in X coming
from the restriction of ϕ to the i-th boundary component of ∆. Let Γ = ∆ ⊗ τ ∈ Ck (X , −ω)
be a k-simplex ∆ on X , loaded with τ. Observe that there is a natural restriction map
ρ∆,∆′ : L−ω (∆) → L−ω (∆′ ) whenever ∆′ ⊂ ∆. The twisted boundary ∂ω (Γ ) of Γ is

∂ω (Γ ) = ∂ ∆0 ⊗ ρ∆,∂ ∆0 (τ) + · · · + ∂ ∆k ⊗ ρ∆,∂ ∆k (τ) .

Extending this C-linearly gives the twisted boundary operator

∂ω : Ck (X , −ω) 7−→ Ck−1 (X , −ω) . (52)

Morally, this boundary operator simply keeps track of the branch of the twisted chain.

21
 SciPost Phys. Lect. Notes 75 (2023)

Example 3.4. Consider again the five simplices illustrated in Figure 7, loaded with the
branches specified in Example 3.2. The twisted boundaries are

∂ω (∆0,0 (ϵ) ⊗ τ0,0 ) = {−ϵ} ⊗ τ0,0 (−ϵ) − {ϵ} ⊗ τ0,0 (ϵ) ,

∂ω (∆0,π (ϵ) ⊗ τ0,π ) = {ϵ} ⊗ τ0,π (ϵ) − {−ϵ} ⊗ τ0,π (−ϵ) ,
∂ω ([ϵ, 1 − ϵ] ⊗ τ− ) = {1 − ϵ} ⊗ τ− (1 − ϵ) − {ϵ} ⊗ τ− (ϵ) , (53)
∂ω (∆1,π (ϵ) ⊗ τ1,π ) = {1 + ϵ} ⊗ τ1,π (1 + ϵ) − {1 − ϵ} ⊗ τ1,π (1 − ϵ) ,
∂ω (∆1,0 (ϵ) ⊗ τ1,0 ) = {1 − ϵ} ⊗ τ1,0 (1 − ϵ) − {1 + ϵ} ⊗ τ1,0 (1 + ϵ) .

The orientation of the boundary components of a 1-simplex is like in standard singular ho-
mology: End point minus starting point. The restrictions of our sections to these boundary
points are simply given by their value at the point. It is instructive to reduce the number of
parameters in (53) by using our definitions and findings from p Example 3.2. For instance, we
have τ0,π (−ϵ) = τ0,0 (−ϵ), τ− (ϵ) = τ0,0 (ϵ), τ0,π (ϵ) = exp( −1ν2π) τ0,0 (ϵ), and so on. ⋄

Notice that, for a twisted k-chain ∆ ⊗ τ, we have ∂ω ∂ω (∆ ⊗ τ) = 0. This follows from

the fact that the boundary of a boundary is empty, i.e., ∂ ∂ ∆ = 0, and the fact that ∂ω simply
restricts τ to ∂ ∆. We suggest that the reader checks this carefully for a two-dimensional
simplex in X = C \ {0, 1} from our running example. In homological algebra, ∂ω ∂ω = 0 is the
key property of a boundary operator in a chain complex.

Definition 3.5. Let X be the very affine variety from (29). Let ω = dlog( f −s x ν ), and let
Ck (X , −ω) be the space (50) of twisted k-chains on X . The twisted chain complex is
∂ω ∂ω ∂ω
(C• (X , −ω), ∂ω ) : · · · −→ Ck (X , −ω) −→ Ck−1 (X , −ω) −→ · · · −→ C0 (X , −ω) −→ 0 . (54)

The homology of this complex is obtained by considering all twisted chains whose twisted
boundary is zero, modulo those that are twisted boundaries themselves.

Definition 3.6. The k-th homology vector space of (C• (X , −ω), ∂ω ) is the quotient space

{Γ ∈ Ck (X , −ω) : ∂ω (Γ ) = 0}
H k (X , −ω) = . (55)
∂ω Ck+1 (X , −ω)

Elements of H k (X , −ω) are called twisted k-cycles (or sometimes loaded k-cycles). We will
primarily be interested in the n-th homology space H n (X , −ω), because these are the cycles on
which we can integrate n-forms. While it is easy to construct cycles in the usual (non-twisted)
singular homology, this is a bit more complicated in our twisted setting. The running example
of this section illustrates a standard construction [5, Section 3.2.4].

Example 3.7. None of the five twisted chains in Example 3.4 are twisted cycles, since they
have non-zero twisted boundaries. However, we can use the expressions (53) to find a linear
combination of these chains whose twisted boundary is zero. For ease of notation, let us write
Γa,θ = ∆a,θ (ϵ) ⊗ τa,θ ∈ C1 (X , −ω), and Γ− = [ϵ, 1 − ϵ] ⊗ τ− ∈ C1 (X , −ω). Consider

Γ0,0 + Γ0,π Γ1,π + Γ1,0

Γ = p + Γ− − p ∈ C1 (X , −ω) . (56)
exp(2π −1ν) − 1 exp(−2π −1s) − 1

Note that (56) only makes sense when ν and s are non-integer. This genericity assumption will
appear in our theorems below. One checks that ∂ω (Γ ) = 0 by expanding it using (53), and
applying identities like at the end of Example 3.4. The class [Γ ] ∈ H1 (X , −ω) of Γ is non-zero.
We will show this in Example 3.22. Hence, Γ is not a boundary of a 2-chain. ⋄

22
 SciPost Phys. Lect. Notes 75 (2023)

3.2 Twisted cochains

While chains tell us where to integrate, co-chains tell us what to integrate. This is to be taken
with a grain of salt in our twisted setting. We have seen above that twisted chains also carry
some information about the integrand: They specify a branch of f −s x ν . In (43), the multival-
ued function f −s x ν is multiplied with an n-form φ = g dx, where g is a single-valued function
on X . This section explains which n-forms φ we consider. It constructs a(n algebraic) twisted
de Rham complex, dual to the twisted chain complex in Definition 3.5.
The vector spaces Ωk (X ) of the twisted de Rham complex are rather easy to describe. They
are the regular k-forms on X , which have coefficients in the coordinate ring of X :
 

 

X X
k
Ω (X ) = g j1 ,..., jk dx j1 ∧ · · · ∧ dx jk : g j1 ,..., jk ∈ a b
C· f x . (57)
1≤ j1 ≤···≤ jk ≤n ℓ
 
a∈Z 
b∈Zn

Notice that ω = dlog( f −s x ν ) ∈ Ω1 (X ). We will integrate regular n-forms φ ∈ Ωn (X ). In

particular, setting φ = dx
x ∈ Ω (X ) in (51) gives our integral (1). In analogy with the usual
n

de Rham complex, we want to regard regular k-forms φ modulo those that integrate to zero
in (51) on a twisted cycle Γ = ∆ ⊗ τ. A first step towards formalizing this is the following
important observation.

Lemma 3.8. For any ψ ∈ Ωk−1 (X ) and any twisted k-chain Γ ∈ Ck (X , −ω), we have
Z Z Z
−s ν −s ν
f −s x ν ψ .


f x (d + ω∧)ψ = d f x ψ = (58)
Γ Γ ∂ω (Γ )

Proof. The first equality is checked by expanding d( f −s x ν ψ) = d( f −s x ν ) ψ + f −s x ν dψ. The

second identity isRStokes’ theoremR (44). More precisely, if Γ = ∆ ⊗ τ is a simplex loaded with
τ, the integral is ∆ d(τ(x)ψ) = ∂ ∆ τ(x)ψ, which agrees with (58) via Definition 3.3.

Equation (58) will be our twisted version of Stokes’ theorem (45), where the twisted dif-
ferential ∇ω is given by d + ω∧. That is, for any 0 ≤ k ≤ n we define

∇ω : Ωk (X ) → Ωk+1 (X ) , with ∇ω (φ) = dφ + ω ∧ φ . (59)

A regular k-form φ is closed if its twisted differential is zero, i.e., ∇ω (φ) = 0. In particular, all
n-forms are closed, since Ωn+1 (X ) = 0. A regular k-form φ is called exact if it is the twisted
differential of some (k − 1)-form: φ = ∇ω (ψ). Here is a consequence of Lemma 3.8.

Lemma 3.9. Let Γ ∈ Ck (X , −ω) be a twisted cycle and let φ ∈ Ωk (X ) be a closed k-form, i.e.,
∂ω (Γ ) = 0 and ∇ω (φ) = 0. If Γ is a twisted boundary or φ is exact, i.e., Γ = ∂ω (Γ ′ ) for some
Γ ′ ∈ Ck+1 (X , −ω) or φ = ∇ω (ψ) for some ψ ∈ Ωk−1 (X ), we have Γ f −s x ν φ = 0.
R

Every exact k-form is closed. Indeed, using ddψ = dω = ω ∧ ω = 0, we find that

∇ω ∇ω ψ = ddψ + dω ∧ ψ − ω ∧ dψ + ω ∧ dψ + ω ∧ ω ∧ ψ = 0 , for any ψ ∈ Ωk−1 (X ) .

Here dω = 0 because ω = dlog( f −s x ν ). The property ∇ω ∇ω = 0 means that ∇ω de-

fines a flat connection on X . Lemma 3.9 says that an exact k-form φ satisfies 〈Γ , φ〉 = 0,
for any Γ ∈ H k (X , −ω). It is therefore natural to regard k-forms modulo the exact k-forms
∇ω (Ωk−1 (X )). This amounts to considering the cohomology of the following cochain com-
plex.

23
 SciPost Phys. Lect. Notes 75 (2023)

Definition 3.10. Let X be the very affine variety from (29). Let ω = dlog( f −s x ν ), and let
Ωk (X ) be the space (57) of regular k-forms. The (algebraic) twisted de Rham complex is
∇ω ∇ω ∇ω
(Ω• (X ), ∇ω ) : 0 −→ Ω0 (X ) −→ Ω1 (X ) −→ · · · −→ Ωn (X ) −→ 0 . (60)

This complex is also called the twisted cochain complex, to emphasize its duality with (54)
(see below). As said above, since exact forms integrate to zero, we pass to cohomology.

Definition 3.11. The k-th twisted cohomology vector space of (60) is the quotient space

{φ ∈ Ωk (X ) : ∇ω (φ) = 0} closed k-forms

H k (X , ω) = = .
∇ω Ω (X )
k−1 exact k-forms

We regard φ in (43) as a twisted cocycle, i.e., an element of the n-th twisted cohomology

Ωn (X )
H n (X , ω) = . (61)
∇ω Ωn−1 (X )

The twisted de Rham complex in Definition 3.10 is called algebraic because we work with
the regular k-forms Ωk (X ) in the sense of algebraic geometry. One can build an analogous
complex using holomorphic k-forms, for which the coefficients g j1 ,..., jk in (57) can be any holo-
morphic functions on X . By the Grothendieck-Deligne comparison theorem [23, Corollaire
6.3], the cohomology of this holomorphic twisted de Rham complex is isomorphic to that of
(60). Since our cocycles φ will be regarded as elements in this cohomology, it suffices to work
with the algebraic complex (60). This is also the preferred setting for doing computations,
because the regular k-forms (57) have a very concrete description.
Here is an example of how to compute relations in twisted cohomology.

Example 3.12. Consider again the Euler beta integral (2). The twisted differential is
 s ν
∇ω = d + + dx ∧ . (62)
1− x x

Applying this to 1 ∈ Ω0 (X ), we obtain the following equality in H 1 (X , ω):

dx −ν dx
• ˜ • ˜
= .
1− x s x

More generally, we shall derive in Section 4 that for a, b ∈ Z, we have the relation

(1 − s)−a (ν) b dx
 
xb dx
• ˜
= . (63)
(1 − x)a x (1 + ν − s) b−a x

Here, for a complex number γ and an integer a, we used the following notation:

γ(γ + 1) · · · (γ + a − 1)
 (a > 0) ,
(γ)a := 1 (a = 0) , ⋄
(γ − 1)−1 (γ − 2)−1 · · · (γ + a)−1 (a < 0) .


While Ω1 (X ) and ∇ω (Ω0 (X )) are infinite-dimensional C-vector spaces, Example 3.12

claims that the quotient H 1 (X , ω) is one-dimensional. Indeed, each regular 1-form can be
written as a constant multiple of [ dx
x ]. This holds, at least, when s, ν, s − ν are non-integer. We
will now proceed towards the underlying theorem (Theorem 3.14). First, we state a vanishing
result for twisted cohomology.

24
 SciPost Phys. Lect. Notes 75 (2023)

Theorem 3.13 (Vanishing theorem). Let X be the very affine variety from (29). There exists a
dense open subset U ⊂ Cℓ+n such that, for each (s, ν) ∈ U, we have
H k (X , ω) = 0 , for all k ̸= n , with ω = dlog( f −s x ν ) . (64)
A description of the open subset U ⊂ Cℓ+n follows from the proof in [3, Theorem A.1]. In
our running example, one can take U = {(s, ν) ∈ Cℓ+n : s, ν, s − ν ∈
/ Z}.
One of the consequences of this vanishing theorem is a geometric description of the di-
mension of H n (X , ω). By [5, Theorem 2.2], the topological Euler characteristic χ(X ) of the
very affine variety X is given by the alternating sum of cohomology dimensions:
n
X
χ(X ) = (−1)k dimC H k (X , ω) . (65)
k=0

Combined with (64), this immediately gives us the following result.

Theorem 3.14. Let X be the very affine variety from (29). Fix (s, ν) ∈ U, where U ⊂ Cℓ+n is as
in Theorem 3.13, and let ω = dlog( f −s x ν ). We have dimC H n (X , ω) = |χ(X )|.
The Euler characteristic χ(X ) also appeared in Theorem 2.1: It is the number of critical
points in Crit(log L) (up to a sign). Hence, for generic (s, ν), we can compute the dimension
of H n (X , ω) using the homotopy continuation techniques explained in Section 2.1.
Example 3.15. In the case of our running example, X = C∗ \{0, 1} is topologically the Riemann
sphere S 2 with three points removed. Using the inclusion-exclusion principle and χ(S 2 ) = 2,
χ(point) = 1, we get χ(X ) = −1. Hence dimC H 1 (X , ω) = 1. This confirms what we saw in
Example 3.12. A basis for H 1 (X , ω) is [ dx
x ]. ⋄
Example 3.16. Consider m-point string amplitudes, for which X = M0,m . The projection map
M0,m → M0,m−1 is obtained by dropping one of the marked points. The fiber of this map at a
given configuration of m − 1 distinct points in P1 is P1 with these m − 1 points removed. The
product property of the Euler characteristic for fibrations gives the recursion
χ(M0,m ) = χ(S 2 − {m−1 points}) · χ(M0,m−1 ) . (66)
By the same arguments as in Example 3.15, the first factor on the right-hand side is 3− m. The
endpoint of the recursion is m = 3, for which χ(M0,3 ) = χ(point) = 1. We conclude
χ(M0,m ) = (−1)m−3 (m − 3)! . (67)
We have seen this number for m = 5 in Section 2.1, where we explained that it also counts
bounded cells of hyperplane arrangements in Rm−3 . Equation (67) implies that the dimension
of H m−3 (M0,m , ω) is (m − 3)!, under the genericity assumptions of Theorem 3.14. A basis
consists of (m − 3)! regular (m − 3)-forms. In the physics literature, it is common to use the
so-called Parke-Taylor basis, see, e.g., [50, Definition 3.2] and [16, Appendix A]. ⋄

3.3 Back to Euler integrals

We have been using the shorthand notation 〈Γ , φ〉 for our integrals, see Definition 3.3.
Equipped with the tools from Sections 3.1 and 3.2, we can now formally introduce the pairing
〈·, ·〉 as a bilinear map on homology and cohomology. This was alluded to in (43).
Theorem 3.17. Let X be as in (29) and let ω = dlog( f −s x ν ). For any k, the C-bilinear map
Z
〈·, ·〉 : H k (X , −ω) × H k (X , ω) −→ C , ([Γ ], [φ]) 7−→ 〈Γ , φ〉 = f −s x ν φ , (68)
Γ

is well-defined. Moreover, the induced maps H k (X , −ω) → H k (X , ω)∨ and H k (X , ω) → H k (X , −ω)∨

(see below) are isomorphisms. In other words, the pairing (68) is perfect.

25
 SciPost Phys. Lect. Notes 75 (2023)

Proof. Well-definedness follows from Lemma 3.9. The pairing is perfect by [5, Lemma 2.9(1)],
using the Deligne-Grothendieck comparison theorem [23, Corollaire 6.3].

The pairing (68) is called the period pairing between twisted homology and cohomology.
In the theorem, V ∨ = HomC (V, C) denotes the dual vector space of a C-vector space V . The
maps H k (X , −ω) → H k (X , ω)∨ and H k (X , ω) → H k (X , −ω)∨ are given by

[Γ ] 7→ ([φ] 7→ 〈Γ , φ〉) , and [φ] 7→ ([Γ ] 7→ 〈Γ , φ〉) ,

respectively. Notice that Theorem 3.17 makes no assumptions on s and ν. We spell out three
important implications (Corollaries 3.18, 3.19 and 3.21).

Corollary 3.18. Let X , ω be as above. For any k, dimC H k (X , −ω) = dimC H k (X , ω).

Corollary 3.19. If (s, ν) lies in the open subset U from Theorem 3.13, the vanishing theorem
extends to twisted homology: H k (X , −ω) = 0 when k ̸= n, and dimC H n (X , −ω) = |χ(X )|.

This means that, when (s, ν) ∈ U , we can find a set of χ = |χ(X )| basis elements [φ1 ],. . . ,[φχ ]
for H n (X , ω), and a set of χ basis elements [Γ1 ], . . . , [Γχ ] for H n (X , −ω). In particular, for any
(a, b) ∈ Zℓ+n , there exist coefficients c1 , . . . , cχa,b ∈ C such that
a,b

 
x b dx a,b
= c1 [φ1 ] + · · · + cχa,b [φχ ] , in H n (X , ω) . (69)
fa x

Example 3.20. For the beta integral, we have seen in Example 3.12 that for φ1 = [dx/x],

a,b (1 − s)−a (ν) b

c1 = . ⋄
(1 + ν − s) b−a

Corollary 3.21. Let X , ω be as above. A regular n-form φ ∈ Ωn (X ) is zero in twisted cohomology,

i.e., [φ] = 0 in H n (X , ω), if and only if
Z
〈Γ , φ〉 = f −s x ν φ = 0 , for all [Γ ] ∈ H n (X , −ω) .
Γ

Here it suffices to let [Γ ] run over a C-basis for H n (X , −ω).

Corollary 3.21 says that relations in cohomology like (63) are equivalent to relations be-
tween Euler integrals which hold for any twisted cycle. That is, Equation (63) implies

x ν+b dx (1 − s)−a (ν) b x ν dx

Z Z
= ,
Γ (1 − x) (1 + ν − s) b−a Γ (1 − x)s x
s+a x

for any [Γ ] ∈ H1 (C \ {0, 1}, −ω). More generally, the expansion (69) in terms of a basis gives
Z ν+b Z ν Z ν
x dx a,b x x
s+a
= c1 s
a,b
φ1 + · · · + cχ s
φχ , for all [Γ ] ∈ H n (X , −ω) .
Γ f x Γ f Γ f

The integrals on the right-hand side are called a set of master integrals in physics, see [61].

Example 3.22. In Example 3.7, we promised to show that Γ from (56) is nonzero in twisted
homology. For this, let us fix values of s, ν such that 0 < ν < ν − s + 1. These are precisely
the convergence conditions derived in Example 1.10. We will come back to this later. The
twisted cycle Γ in (56) depends on ϵ, but by the Cauchy-Goursat theorem, the value of the

26
 SciPost Phys. Lect. Notes 75 (2023)

integral 〈Γ , dx
x 〉 is independent of ϵ ∈ (0, 1/2). According to the three terms in (56), we split
the integral up into three parts: 〈Γ , dx
x 〉 = I 0 (ϵ) + I − (ϵ) + I 1 (ϵ). The first summand is
Z 2π p p
1 −1θ −s ν −1θ ν
I0 (ϵ) = p (1 − ϵe ) ϵ e dθ .
e2π −1ν −1 0

Because of the assumption ν > 0, we have limϵ→0+ I0 (ϵ) = 0. Analogously, one shows
limϵ→0+ I1 (ϵ) = 0. Finally, by (2), we have
Z 1−ϵ
dx xν dx
­ ·
Γ, = lim+ I− (ϵ) = lim+ = B(ν, 1 − s) .
x ϵ→0 ϵ→0
ϵ (1 − x) x
s

Since the result is nonzero, Corollary 3.21 implies [Γ ] ̸= 0. ⋄

Our final goal in this section is to connect the Euler integrals 〈Γ , φ〉 obtained from the
pairing in Theorem 3.17 with the Euler-Mellin integrals from Section 1. For that, we first need
to reinterpret 〈Γ , φ〉 as a function of s, ν. We write ω = ω(s, ν) to emphasize the dependence
on these parameters. On the cohomology side, the natural thing to do is to fix φ ∈ Ωn (X ),
and regard it as an element in the varying cohomology vector space H n (X , ω(s, ν)). On the
homology side, we need to take into account the fact that the line bundle L−ω(s,ν) depends on
s, ν. Let Γ (s, ν) = ∆ ⊗ τ(s, ν) be the singular n-simplex ∆ loaded with τ(s, ν) ∈ L−ω(s,ν) (∆).
To see how τ varies with s, ν, note that it is a C-linear combination of the branches of

exp(−s1 log f1 − · · · − sℓ log fℓ + ν1 log x 1 + · · · + νn log x n ) ,

restricted to ∆. Such a branch is fixed after fixing the branches of the logarithms log f i , log x j ,
which are independent of s and ν. Our integral is the following function of s, ν:
Z Z
(s, ν) 7−→ f −s x ν φ = τ(s, ν)(x) φ = 〈∆ ⊗ τ(s, ν), φ〉 . (70)
∆⊗τ(s,ν) ∆

Taking derivatives of (70) in s, ν can be done under the integration sign [42, Chapter XVII,
Theorem 8.2]. This implies the following Proposition.
Proposition 3.23. The function (s, ν) 7→ 〈∆ ⊗ τ(s, ν), φ〉 from (70) is holomorphic.
It is straightforward to extend this to the case where Γ (s, ν) = i d p (s, ν) · ∆ p ⊗ τ p (s, ν),
P

where the coefficients d p (s, ν) are meromorphic functions. Similar to Definition 3.3, we set
X Z
〈Γ (s, ν), φ〉 = d p (s, ν) · τ p (s, ν)(x) φ . (71)
p ∆p

By Proposition 3.23, this is meromorphic in s, ν. Notice that we can also view this as the sum
p 〈∆ p ⊗ τ p (s, ν), d p (s, ν)φ〉, allowing meromorphic coefficients in cohomology. This will be
P

useful in Section 4.1. It turns out we have seen (71) before.

Theorem* 3.24. Let f1 , . . . , fℓ satisfy Assumption 1, and suppose that the Minkowski sum
∆( f1 ) + · · · + ∆( fℓ ) has dimension n. Let X be as in (29). There exist finite sets of meromor-
phic functions d p (s, ν), singular n-chains ∆ p on X and sections τ p (s, ν) ∈ L−ω(s,ν) (∆ p ) with the
following property. For Γ (s, ν) = p d p (s, ν) · ∆ p ⊗ τ p (s, ν), the function (s, ν) 7→ 〈Γ (s, ν), dx
P
x 〉
form (71) is the meromorphic continuation (26) from Theorem 1.11.
This statement is labeled Theorem* 3.24 (with an asterisk) because, to the best of our
knowledge, there exists no rigorous proof in the literature. Yet, it is widely accepted and used.
Our sketch of proof below makes it more than a conjecture. Providing full details is among
the proposed problems in Section 5. This requires tools beyond our present scope.

27
 SciPost Phys. Lect. Notes 75 (2023)

Sketch of proof of Theorem* 3.24. Consider the algebraic moment map in (39) with si = 1. By
Pℓ
Lemma 2.7, µ : R+ n
→ int(P) is a diffeomorphism, with P = i=1 ∆( f i ). This gives
Z Z
dy
ν dx
−s
f ( y) y = f (µ−1 (x))−s µ−1 (x)ν Jµ−1 Q ,
j µ (x) j
n y −1
R+ P

where Jµ−1 is the Jacobian determinant of µ−1 , and the last denominator is the product of
the entries of µ−1 . Like in Example 3.22, for values of s, ν where the integral on the right
converges, we replace P by a twisted cycle Γ (P), called the regularization of P, such that
Z Z ® ¸
d y dx dx
f ( y)−s y ν = f (µ−1 (x))−s µ−1 (x)ν Jµ−1 Q = Γ (P), Jµ−1 Q .
j µ (x) j j µ (x) j
y −1 −1
R
+
n
Γ (P)

The integral on the right side is the pairing of a holomorphic n-form with the twisted cy-
cle Γ (P). Here we need to use the analytic version of twisted (co)homology. The mani-
fold is X̃ = {x ∈ Cn : f1 (µ−1 (x)) · · · fℓ (µ−1 (x))µ−1 (x)1 · · · µ−1 (x)n ̸= 0}, and the twist is
ω̃ = dlog( f (µ−1 (x))−s µ−1 (x)ν ). The construction of the regularization Γ (P) is that in [5, Sec-
tions 3.2.4 and 3.2.5]. For this, when P is not smooth, one needs to replace the moment map
µ by that of a different toric variety, obtained by blowing up the toric variety of P in its singular
locus. The cycle Γ (s, ν) in the Theorem* is the pullback µ∗ (Γ (P)) of Γ (P) under µ.

Theorem* 3.24 replaces integrating over R+ n

by integrating over Γ (s, ν). The twisted cycle
Γ (s, ν) is called the regularization of R+ [5, Sections 3.2.4 and 3.2.5]. Here is an example.
n

Example 3.25. We have seen two integral formulas for B(ν, 1 − s) in Example 1.10:
1 1
xν xν yν
Z Z Z
dx dx dy
= = . (72)
0 (1 − x) x
s
0 (1 − x)s−1 x(1 − x) R+ (1 + y) y
s̃

The coordinate transformation is µ : R+ → (0, 1), with x = µ( y) = y(1 + y)−1 . This is the
moment map from Lemma 2.7 up to scaling by s̃. Its complexification µC is an isomorphism
µC
X y = C \ {0, −1} −→ C \ {0, 1} = X x .

Let ω x = dlog((1 − x)1−s x ν ) ∈ Ω1 (X ) be the regular one-form corresponding to the middle

integral of (72). We define ω y to be the pullback of ω x along µC . That is, explicitly,

ω y = µ∗ (ω x ) = dlog((1 − µC ( y))1−s µC ( y)ν ) = dlog((1 + y)−s̃ y ν ) .

The cocycle dx/(x(1− x)) pulls back to d y/ y under µ. A twisted chain Γ =∆⊗τ∈ C1 (X x , −ω x )
is naturally pulled back to a cycle µ∗ (Γ ) ∈ C1 (X y , −ω y ) via

µ∗ (Γ ) = µ−1 (∆) ⊗ (τ ◦ µ) .

As usual, this definition for simplices is extended linearly to C1 (X x , −ω x ). With this notation
in place, it is easy to check that for any Γ ∈ C1 (X x , −ω x ), we have

xν yν
Z Z
dx dx dy dy
­ · ­ ·
∗
Γ, = = = µ (Γ ), .
Γ (1 − x) x µ∗ (Γ ) (1 + y)
x(1 − x) s s̃ y y

By Example 3.22, if we pick Γ as in (56) (with s replaced by s − 1), the left integral is a
meromorphic function in (s, ν) which agrees with B(ν, 1 − s) if 0 < ν < ν − s + 1. Hence, the
regularization of R+ is µ∗ (Γ ). It depends on s̃, ν as explained above. ⋄

28
 SciPost Phys. Lect. Notes 75 (2023)

4 Differential and difference equations

It is common practice to describe a function F via the differential or difference equations it
satisfies. More precisely, one attempts to find differential operators P such that P • F = 0, or
difference/shift operators S such that S • F = 0. Here P • F reads as P applied to F , and similarly
for S • F . In this section, F is an Euler integral (1), seen as a pairing between a twisted n-
cycle [Γ ] and the twisted n-cocycle [dx/x], see Section 3. Such an integral satisfies difference
equations when seen as a meromorphic function of s and ν, as in (71).

Example 4.1. We have seen in Example 3.22 that the beta function B(ν, 1 − s) is given by

xν
Z
dx
I (s, ν) = B(ν, 1 − s) = , (73)
Γ (1 − x) x
s

where Γ is the twisted 1-cycle from (56). This agrees with the integral over (0, 1) in (2) when
Re(s) > 0 and 0 < Re(ν) < Re(ν) − Re(s) + 1. The shift operator in s, denoted σs , acts by

σs • I (s, ν) = I (s + 1, ν) .

Similarly, the action of σν is σν • I (s, ν) = I (s, ν + 1). We claim that the shift operators

S1 = 1 − σs (1 − σν ) , and S2 = ν + s σν σs , (74)

annihilate I , i.e., S1 • I = 0 and S2 • I = 0. Here coefficients that are rational functions in s, ν

simply act by multiplication. The identity S1 • I = 0 is easy to verify:

xν x · x ν dx (1 − x) · x ν dx
Z Z  Z
dx
σs (1 − σν ) • I = σs • − = .
Γ (1 − x) x Γ (1 − x) x Γ (1 − x)
s s s+1 x

The rightmost integral equals I = 1 • I . To see that S2 • I = 0, we apply Lemma 3.9:

xν ν xν
Z Z
s 
S2 • I = + dx = ∇ω (1) = 0 .
Γ (1 − x) Γ (1 − x)
s x 1− x s

Here ∇ω is as in (62). We reiterate that, in order to view I as a meromorphic function of

s, ν, it is important to keep in mind that ω = ω(s, ν) varies. Hence, so does the local system
L−ω = L−ω(s,ν) , and the twisted cycle Γ = Γ (s, ν). This was explained in Section 3.3. ⋄

The goal of Section 4.1 is to derive operators like (74) for general Euler integrals. These
operators appeared in [8, Section 3.1], [3, Section 3] and [47]. Shift operators for Feynman
integrals were studied in [10]. Section 4.2 discusses differential operators. For that, we must
view our Euler integrals as functions of a new set of parameters: The coefficients of f i .

Example 4.2. Fix two generic complex numbers s, ν ∈ C. We modify the beta integral (2) by
introducing complex valued parameters z1 , z2 for the coefficients of the denominator f :

xν
Z
dx
I (z1 , z2 ) = .
Γ (z1 + z2 x) x
s

The dependence on z = (z1 , z2 ) is subtle. For instance, the very affine variety X from (29)
depends on z and, necessarily, so does Γ . In this example, one can think about I (z1 , z2 ) as a
function on a small neighborhood of (z1 , z2 ) = (1, −1), which corresponds to our original beta
integral. In that neighborhood, one can modify Γ by keeping the 1-simplices in Example 3.1
fixed, and varying the sections in Example 3.2 with z1 , z2 . For instance, the initial condition

29
 SciPost Phys. Lect. Notes 75 (2023)

τ0,0 (ϵ; z1 , z2 ) = ζ(z1 , z2 ) from (49) is given by ζ(z1 , z2 ) = exp(−s log(z1 + z2 ϵ) + ν log ϵ), and
for the first log we use the analytic continuation of the positive branch near 1 − ϵ.
The differential operator ∂zi acts by partial derivation in zi , for i = 1, 2, and rational func-
tions in z act by multiplication. Here are two annihilating operators for I (z1 , z2 ):

P1 = z1 ∂z1 + z2 ∂z2 + s , and P2 = z2 ∂z2 + ν . (75)

The derivatives can be taken under the integration sign. We verify P1 • I = 0:

−sz1 · x ν dx −sz2 x · x ν dx
Z Z
(z1 ∂z1 + z2 ∂z2 ) • I (z) = + = −s · I (z) .
Γ (z1 + z2 x) Γ (z1 + z2 x)
s+1 x s+1 x

To see that P2 • I (z) = 0, we again need Lemma 3.9. We compute

xν ν
Z
sz2
 ‹
0 = P2 • I (z) = ∇ω(z) (1) , with ω(z) = − dx . ⋄
Γ (z1 + z2 x) x z1 + z2 x
s

The differential operators (75) form an A-hypergeometric system or GKZ system of linear par-
tial differential equations. Such systems were introduced by Gelfand, Kapranov and Zelevinsky
to study A-hypergeometric functions [30,31]. We will introduce these systems and recall their
relation to Euler integrals in Section 4.2. For a recent survey, see [56].

4.1 Difference equations

We start with difference/shift operators in (s, ν). In analogy with Example 4.1, we call these
operators σsi for i = 1, . . . , ℓ and σν j for j = 1, . . . , n. They act on the integral I = 〈Γ , φ〉 from
(71) as follows. We view (71) as the sum of pairings
X Z X
I (s, ν) = d p (s, ν) · τ p (s, ν)(x) φ = 〈 ∆ p ⊗ τ p (s, ν), d p (s, ν) φ 〉 .
p ∆p p

That is, the cocycle now depends meromorphically on s, ν. We set

X
σsi • I (s, ν) = I (s + ei , ν) = 〈 ∆ p ⊗ τ p (s + ei , ν), d p (s + ei , ν) φ 〉 (76)
p
X
= 〈 ∆ p ⊗ τ p (s, ν), d p (s + ei , ν) f i−1 φ 〉 , (77)
p
X
σν j • I (s, ν) = I (s, ν + e j ) = 〈 ∆ p ⊗ τ p (s, ν + e j ), d p (s, ν + e j ) φ 〉 (78)
p
X
= 〈 ∆ p ⊗ τ p (s, ν), d p (s + ei , ν) x j φ 〉 . (79)
p

Here e j is the j-th standard basis vector. The reader should check the passages from (76) to
(77) and (78) to (79) carefully. The expressions (77) and (79) show that the action of the
shift operators can be viewed as an action on cohomology H n (X , ω(s, ν)). For instance,

σsi • [φ(s, ν)] = [ f i−1 φ(s + ei , ν)] , σν j • [φ(s, ν)] = [x j φ(s, ν + e j )] . (80)

In [47], this action is defined on a cohomology vector space with coefficients in C(s, ν). We
will also use the inverses σs−1 , σν−1 of these shift operators. Their action is straightforward to
i j
define. The variables si and ν j act on I (s, ν) by multiplication: si • I (s, ν) = si I (s, ν), and
ν j • I (s, ν) = ν j I (s, ν). Notice that the operators si and σsi do not commute:

σsi si • I (s, ν) = (si + 1)I (s + ei , ν) ̸= si I (s + ei , ν) = si σsi • I (s, ν) .

Such commutator relations naturally lead to the following definition.

30
 SciPost Phys. Lect. Notes 75 (2023)

Definition 4.3. The ring of difference operators R = C(s, ν)〈σs±1 , . . . , σs±1 , σν±1 , . . . , σν±1 〉 is the
1 ℓ 1 n
C(s, ν)-vector space with basis σsa σνb , where (a, b) ∈ Zℓ+n . I.e., it consists of finite sums
X X
g a,b (s, ν) σsa1 · · · σsaℓ σνb1 · · · σνbn = g a,b (s, ν) σsa σνb .
1 ℓ 1 n
(a,b)∈Zℓ+n (a,b)∈Zℓ+n

The product is subject to the following relations. For any rational function g(s, ν) ∈ C(s, ν),

[σs±1 , g(s, ν)] = (g(s ± ei , ν) − g(s, ν))σsi , and [σν±1 , g(s, ν)] = (g(s, ν ± e j ) − g(s, ν))σν j .
i j

Here [A, B] = AB − BA denotes the commutator.

The ring R acts on meromorphic functions in s, ν as explained above. The annihilator
AnnR (I ) of a meromorphic function I (s, ν) consists of all shift operators annihilating I :

AnnR (I (s, ν)) = {S ∈ R : S • I (s, ν) = 0} . (81)

Clearly, if S1 , S2 ∈ AnnR (I ), then S1 + S2 ∈ AnnR (I ) as well. Also, if S1 ∈ AnnR (I ), then

S2 S1 ∈ AnnR (I ) for any S2 ∈ R. In other words, AnnR (I ) is a left ideal of R.
To describe the annihilator of our integral, we introduce the notation f i (σν ) ∈ R for the
difference operator obtained by replacing x j → σν j in f i (x). This is well-defined, since all σν j
commute. For instance, for f (x) = 1 − x from Example 4.1, we write f (σν ) = 1 − σν .
Proposition 4.4. Let J ⊂ R be the left ideal generated by the following ℓ + n operators:

1 − σsi f i (σν ) , for i = 1, . . . , ℓ, (82)

ℓ
X ∂ fi
σν−1 ν j − si · σsi (σν ) , for j = 1, . . . , n . (83)
j
i=1
∂ xj

For any cycle Γ , the annihilator of the Euler integral IΓ (s, ν) = 〈Γ , dx

x 〉 contains J.

Proof. WeP need to show that the operators (82) and (83) annihilate IΓ (s, ν). Let
f i (x) = α ci,α · x α , where ci,α ∈ C for i = 1, 2, . . . , ℓ. Hence, using (79), we find
 X  ­
α dx dx
·
f i (σν ) • IΓ (s, ν) = Γ , ci,α · x = Γ , f i (x) . (84)
α
x x

Together with (77) we find σsi f i (σν ) • IΓ (s, ν) = 〈Γ , dx

x 〉 = IΓ (s, ν), which shows that (82)
annihilates IΓ (s, ν). Notice that, for this, we do not use the fact that Γ is a cycle. To show that
the operators in (83) annihilate IΓ (s, ν) as well, we compute
∂ fi ∂ f i dx ν j − 1 dx
   
−1
(σν ) • IΓ (s, ν) = Γ , , and σν ν j • IΓ (s, ν) = Γ , .
∂ xj ∂ xj x j xj x
We combine these identities to get
*  ∂ fi

∂ x j (x)
ℓ ℓ
+
νj − 1
‚ Œ
∂ fi
 dx
X X
σν−1 ν j − si · σsi (σν ) • IΓ (s, ν) = Γ, − si .
j
i=1
∂ xj xj i=1
f i (x) x

To show that this is zero for any cycle Γ , we need to show that the n-form on the right is exact
(Lemma 3.9). That is, we need to find ψ such that it equals ∇ω (ψ). The solution is
∂ fi
 
νj − 1 Xℓ
∂ xj (x) dx

dx ĵ

 − si  = ∇ω (−1) j−1
,
xj i=1
f i (x) x x1 · · · x n

where dx ĵ is the (n − 1)-form dx 1 ∧ · · · ∧ dx j−1 ∧ dx j+1 ∧ · · · ∧ dx n .

31
 SciPost Phys. Lect. Notes 75 (2023)

Example 4.5. The operators S1 and σν−1 S2 from (74) are (82) and (83) for n = ℓ = 1 and
f = 1 − x. We can use these operators to obtain the relation (63). Observe that

s · (1 − σs + σs σν ) − (ν + sσν σs ) = s − sσs − ν ∈ J .

Since J annihilates IΓ = 〈Γ , dx/x〉 for all Γ (Proposition 4.4), this means

σs • [dx/x] = (s − ν)s−1 • [dx/x]. Applying σs−1 from the left (necessarily, because J is a
left ideal), we get
s − ν 
s−1−ν
‹
dx
• ˜ 
1−s dx
‹ • ˜
σs−1 = σs−1 = 1 mod J , i.e., σs−1 • = • .
s s−1 x 1+ν−s x

Similarly since σs−1 (1 − σs + σs σν ) ∈ J, we derive that

dx dx ν  • dx ˜
• ˜ • ˜ 
σν • = (1 − σs−1 ) • = • .
x x 1+ν−s x

Now, we use these identities to write σsa σνb •[dx/x] = [x b /(1− x)a ·dx/x] in terms of [dx/x]:

dx
• dx ˜
• dx ˜
• ˜
ν ν+1 ν


σsa σνb • = σsa σνb−1 1+ν−s • = σsa σνb−2 2+ν−s 1+ν−s • = ...
x x x

Here the second equality uses the commutation rules. Suppose b, a ∈ Z>0 are positive. After
repeating this step b times, we begin expanding σsa :

dx (ν) b dx (ν) b
• dx ˜
• ˜ • ˜
s−ν
σsa σνb • = σsa • = σsa−1 •
x (1 + ν − s) b x (ν − s)(1 + ν − s) b−1 s x
(ν) b dx
• ˜
s−ν+1

s−ν

= σsa−2 •
(ν − s − 1)(ν − s)(1 + ν − s) b−2 s+1 s
x
(ν) b 1 dx
 ‹ • ˜
= ··· = • .
(1 + ν − s) b−a (−s)(−s − 1) · · · (−s − a + 1) x

The rational function we obtain is precisely that of (63). Notice that the numerator factors
(s − ν), (s − ν + 1), . . . cancel with the denominators (ν − s), (ν − s − 1), . . . , and the minus
signs are absorbed in the denominator factors s, s + 1, . . .. The reader should check that if a, b
satisfy different sign conditions, we arrive at the same formula. ⋄

Example 4.5 illustrates the concept of contiguity relations for the Euler beta integral. In gen-
eral, the action of shift operators can be captured by contiguity matrices. Let [φ1 ], . . . , [φχ ] be
a basis for the twisted cohomology H n (X , ω(s, ν)) (for generic s, ν). There are χ = (−1)n ·χ(X )
such basis elements by Corollary 3.19. We assume the [φk ] have coefficients that are rational
functions in s, ν. There are χ × χ-matrices Csi , Cν j such that

〈Γ , φ1 〉 〈Γ , φ1 〉 〈Γ , φ1 〉 〈Γ , φ1 〉
       
. . . .
σsi •  ..  = Csi ·  ..  , σν j •  ..  = Cν j ·  ..  .
〈Γ , φχ 〉 〈Γ , φχ 〉 〈Γ , φχ 〉 〈Γ , φχ 〉

Here σsi and σν j act entry-wise on vectors, and the contiguity matrices Csi , Cν j have entries
in C(s, ν). For more on these matrices and how to compute them, see [47, Section 5].

Remark 4.6. In the context of Feynman integrals, special choices of contiguity relations are
known as dimension-shift identities, because they relate Feynman integrals evaluated in differ-
ent space-time dimensions. We refer to [61, Section 6.2] for more details.

32
 SciPost Phys. Lect. Notes 75 (2023)

Remark 4.7. Shift relations (80) give a practical way of enlarging the Nilsson-Passare domain
of convergence of Euler integrals in (s, ν), as explained
R ∞ in Section 1.2. For instance, the shift
operator σν−1 ν − s̃σs̃ annihilates the beta integral 0 y ν /(1 + y)s̃ d y/ y. Hence

y ν−1 d y yν
Z Z
dy
(ν − 1) · = s̃ · . (85)
R+ (1 + y)s̃ y R+ (1 + y)s̃+1 y

This is an equality of meromorphic functions after replacing R+ with its regularizer from The-
orem* 3.24. Suppose s̃ ∈ R+ . The convergence region from Theorem 1.6 for the left hand side
is Re(ν) ∈ (1, s̃ + 1), while that of the right hand side is Re(ν) ∈ (0, s̃ + 1). Hence, (85) is used
to evaluate the meromorphic continuation from Theorem 1.11 for Re(ν) ∈ (0, 1).

4.2 Differential equations

As illustrated in Example 4.2, Euler integrals are annihilated by differential operators in the
coefficients of f i . These are new parameters denoted by zi,α , i.e.,
X
fi = zi,α · x α , i = 1, . . . , ℓ . (86)
α ∈ Ai

Here Ai = supp( f i ) ∈ Zn is the support of f i , in the sense of Definition 1.1. We fix complex
parameters s = (s1 , . . . , sℓ ) ∈ Cℓ and ν = (ν1 , . . . , νn ) ∈ Cn . The variety X defined in (29) is
also dependent on z. For this reason, we write X z instead of X . The Euler integral (1), seen as
a function of the coefficients z = (zi,α )i,α , defines a holomorphic function on an open subset
U of coefficient space CA = CA1 × · · · × CAℓ . To define this function, we need to specify how
the twisted integration cycle Γ varies with z. For a fixed set of coefficients z ∗ = (zi,α ∗
)i,α ∈ CA,
let [Γ (z ∗ )] ∈ H n (X z ∗ ; −ω(z ∗ )) be a twisted cycle. The open set U is a sufficiently small neigh-
borhood U of z ∗ . The choice [Γ (z ∗ )] gives rise to a family of cycles [Γ (z)] ∈ H n (X z ; −ω(z))
defined for z ∈ U. For this, we fix the singular n-simplices, and vary the sections of L−ω(z) in
the only sensible way. That is, Γ (z ∗ ) = p d p · ∆ p ⊗ τ p (z ∗ ), where τ p (z ∗ ) ∈ L−ω(z ∗ ) is a branch
P

of f (x; z ∗ )−s x ν which depends holomorphically on z, and

X
Γ (z) = d p · ∆ p ⊗ τ p (z) ∈ H n (X z , −ω(z)) . (87)
p

Here d p ∈ C are constants. This was illustrated for the beta integral in Example 4.2. It is
crucial that, with this construction, the twisted boundary ∂ω(z) (Γ (z)) is zero for all z ∈ U.

Proposition 4.8. Let U ∋ z ∗ and [Γ (z)] ∈ H n (X z ; −ω) be as above. The function

Z
dx
IΓ : U −→ C , z 7−→ f (x; z)−s x ν ∈ C, (88)
Γ (z) x

is holomorphic on U.

Proof. Since τ(z) is holomorphic in z, it suffices to observe that for z ∈ U,

Z
X dx
IΓ (z) = dp · τ(z)(x) .
p ∆ x
p

The theorem follows from the definition of a holomorphic function and differentiation under
the integral sign [42, Chapter XVII, Theorem 8.2].

33
 SciPost Phys. Lect. Notes 75 (2023)

Our goal is to derive a system of differential equations satisfied by IΓ (z). This is an example
of a class of such systems, called GKZ systems (after Gelfand, Kapranov and Zelevinsky) or A-
hypergeometric systems. We introduce these in general, and then specialize to our integrals.
The proof of the main theorem in this section (Theorem 4.12) uses the theory of D-modules.
Here the ring D is the Weyl algebra, which plays an analogous role as that of the ring R of
difference operators (see Section 4.1). To state Theorem 4.12, it is unnecessary to introduce
D-modules. We refer the interested reader to [22] for a nice introduction.
Let d be a positive integer and let A ⊂ Zd be a finite subset. To each α ∈ A, we associate
a complex variable zα and a partial derivative operator ∂α = ∂∂z . For a function f (z) of
α
∂f
z = (zα )α∈A, its partial derivative ∂ z (z) with respect to zα is denoted by ∂α f (z). The toric
α
ideal IA ⊂ C[∂α , α ∈ A] is an ideal generated by all binomials
Y Y
∂αuα − ∂αvα ,
α∈A α∈A

where u = (uα )α∈A, v = (vα )α∈A ∈ NA are such that A · (u − v) = 0. That is,
X X
uα α = vα α . (89)
α∈A α∈A

Of course, there are infinitely many integer vectors u − v satisfying A· (u − v), but finitely many
suffice to generate the ideal IA. Fix a vector β ∈ Cd . The GKZ system HA(β) associated to A
and β is the following system of partial differential equations in f (z):
X
HA(β) : zα ∂α • f (z)α − f (z)β = 0 , and P(∂ ) • f (z) = 0 , for P(∂ ) ∈ IA . (90)
α∈A

Note that the first equation of (90) is an identity of vectors in Cd , and it is enough to check
that P(∂ ) f = 0 for a finite set of generators of IA. On an open subset U ⊂ CA, we define the
space of solutions SolHA(β) (U) of HA(β) as the complex vector space

SolHA(β) (U) = { f : U → C holomorphic : f satisfies (90) } .

Our Euler integral (88) satisfies the GKZ system specified by the following parameters. Set
d = ℓ + n. The Cayley configuration A ⊂ Zd of A1 , . . . , Aℓ is

A = {(ei , α) : α ∈ Ai , i = 1, . . . , ℓ} ⊂ Zd . (91)

We have seen this in (40). The vector β is −(s, ν) ∈ Cd .

Proposition 4.9. For A, β as above, and IΓ (z) as in (88), we have IΓ (z) ∈ SolHA(β) (U).

Before proving Proposition 4.9, we encourage the reader to check that (75) is the GKZ
system for the beta integral. In that example, the toric ideal IA is 0. Notice that Proposition
4.9 is independent of the choice of cycle Γ (z).

Proof of Proposition

4.9. By the definition of the Cayley configuration (91), the constraint (89)
for u = ui,α i=1,...,ℓ , v = (vi,α )i=1,...,ℓ takes the following form :
α∈Ai α∈Ai

X X ℓ X
X ℓ X
X
ui = ui,α = vi,α (i = 1, . . . , ℓ) , and ui,α α = vi,α α . (92)
α∈Ai α∈Ai i=1 α∈Ai i=1 α∈Ai

34
 SciPost Phys. Lect. Notes 75 (2023)

By our construction of Γ (z) in (87), the integration contours ∆ p ⊂ X are independent of z,

and we may apply the operators ∂(ei ,α) to IΓ (z) by differentiating under the integration sign.
The sections τ p (z) : ∆ p → C differentiate to the corresponding branches of

xα .
−s1 −si −1 −sℓ
∂(ei ,α) • f −s = −si f1 · · · fi · · · fℓ (93)
Qℓ u Qℓ v
∂(ei,α,α) − ∂(ei,α,α) . We calculate
Q Q
The operators generating IA are i=1 α∈Ai i=1 α∈Ai
i i

ℓ Y ℓ uY
i −1
Z
ui,α α dx
Pℓ
ν+
P
u
Y Y
−s1 −u1 −sℓ −uℓ
∂(ei,α,α) • IΓ (z) = (−si − j) f1 · · · fℓ x i=1 α∈Ai
.
i=1 α∈Ai
i
i=1 j=0 Γ (z) x

Qℓ Q v
Doing the same for i=1 α∈Ai ∂(ei,α,α) and applying (92) we see that IA annihilates IΓ (z).
i
It remains to verify that the other operators in the GKZ system annihilate IΓ (z) as well:
X
zα ∂α • IΓ (z)α − IΓ (z)β = 0 . (94)
α∈A

The first ℓ-entries come from the homogeneity relation

!
X X
zi,α x α f i i =
−s −s −1 −s
−si f i i = −si zi,α ∂(ei ,α) • f i i .
α∈Ai α∈Ai

For any j = 1, . . . , n, the (ℓ + j)-th entry of (94) is derived by differentiating under the integral
sign and observing that the result is the pairing of Γ (z) with the exact n-form

ℓ
!
∧ · · · ∧ · · · ∧
 
dx 1 dx j−1 dx j+1 dx n
X X dx
∇ω (−1) j−1 =− si f i−1 zi,α x α α j + ν j .
x 1 · · · x j−1 x j+1 · · · x n i=1 α∈A
x
i

Here, α j is the j-th entry of α. This generalizes what happened for P2 in Example 4.2.

Proposition 4.9 implies the following homogeneity property for the function IΓ (z).

Lemma 4.10. Let A = {α1 , . . . , αN } ⊂ Zd , β ∈ Cd and let U ⊂ CN be an open subset. If

f : U → C is holomorphic, then (Aθ − β) • f (z) = 0 if and only if, for all z ∈ U and for all
u ∈ (C∗ )d such that (uα1 z1 , . . . , uαN zN ) ∈ U, we have

f (uα1 z1 , . . . , uαN zN ) = uβ f (z) . (95)

Proof. Suppose f (z) satisfies (95). We fix any 1 ≤ i ≤ d. Taking the derivative of (95)
with respect to ui and substituting u = (1, . . . , 1), we obtain the i-th entry of the identity
(Aθ − β) • f (z) = 0. For the other direction, suppose f (z) is annihilated by Aθ − β. To prove
(95), it is enough to prove it for u(i) = ui · ei with ui ∈ C∗ , and ei the standard basis vector.
For any z ∈ U, the functions φ1 : ui 7→ f (u(i)α1 z1 , . . . , u(i)αN zN ) and φ2 : ui 7→ u(i)β f (z) are
both annihilated by ui ∂∂u − βi . We also have φ1 (1) = φ2 (1). Unique solvability of the initial
i
value problem of an O.D.E. implies φ1 = φ2 , so (95) holds for u = u(i).

Often, in applications, one is interested in differential equations satisfied by IΓ (z) after

specializing the parameters z. This is the case for Feynman integrals, in which the coefficients
zi,α depend linearly on the Mandelstam invariants. For the case of Example 0.2, we now show
how the equations in these new variables can be derived from the GKZ system.

35
 SciPost Phys. Lect. Notes 75 (2023)

Example 4.11. We revisit the Euler integral (7) of Example 0.1. The differential equations
satisfied by this integral as a function of t 1 , t 2 , t 3 are the running example of [39]. In this case,
they can be derived in an easy way from the GKZ differential equations for
ν ν ν
x1 1 x2 2 x3 3
Z
dx 1 dx 2 dx 3
IΓ (z) = .
Γ (z1 x 1 + z2 x 2 + z3 x 3 + z4 x 2 x 3 + z5 x 1 x 3 + z6 x 1 x 2 )s x 1 x 2 x 3

Here ℓ = 1, n = 3, β = −(s, ν1 , ν2 , ν3 ) and A consists of the columns of

1 1 1 1 1 1
 
1 0 0 0 1 1
.
0 1 0 1 0 1

0 0 1 1 1 0

We denote these columns by α1 , . . . , α6 ∈ Z4 , and set ∂i = ∂αi for brevity. The toric ideal IA is
generated by P̃1 = ∂1 ∂4 − ∂3 ∂6 and P̃2 = ∂2 ∂5 − ∂3 ∂6 . These operators can be found using any
computer algebra software. For instance, in Macaulay2 [33], the commands are

needsPackage "Quasidegrees" 1

A = matrix{1,1,1,1,1,1},{1,0,0,0,1,1},{0,1,0,1,0,1},{0,0,1,1,1,0} 2

D = QQ[d_1..d_6] 3

T = toricIdeal(A,D) 4

The remaining operators constituting the GKZ system are

P6
k=1 θk + s, θ1 + θ5 + θ6 + ν1 , θ2 + θ4 + θ6 + ν2 , θ3 + θ4 + θ5 + ν3 , (96)

where θk = zk ∂k . Up to replacing R3+ with Γ , the integral IG in (7) is IΓ (1, 1, 1, −t 1 , −t 2 , −t 3 ).

Notice that here one could take Γ to be the regularization of R3+ , see Theorem* 3.24. To
turn our operators in z1 , . . . , z6 into differential operators in t 1 , t 2 , t 3 , we use the homogeneity
condition from Lemma 4.10:

IΓ (uα1 z1 , . . . , uα6 z6 ) = uβ IΓ (z) .

−ν −ν −ν
We adopted the usual notation uβ = u−s 1 2 3 −1 −1 −1
1 u2 u3 u4 . For u = (1, z1 , z2 , z3 ) this reads

z4 z5 z6
 ‹
ν ν ν
IΓ 1, 1, 1, , , = z1 1 z2 2 z3 3 IΓ (z) . (97)
z2 z3 z1 z3 z1 z2

We differentiate (97) with respect to z1 , and afterwards we substitute

z = (1, −t) = (1, 1, 1, −t 1 , −t 2 , −t 3 ). Using (∂5 IΓ )(1, −t) = −∂ t 2 (IΓ (1, −t)) and similarly
for ∂6 IΓ , we get

(∂1 IΓ )(1, −t) = − ν1 + t 2 ∂ t 2 + t 3 ∂ t 3 • IΓ (1, −t) ,

(∂2 IΓ )(1, −t) = − ν2 + t 1 ∂ t 1 + t 3 ∂ t 3 • IΓ (1, −t) ,

(∂3 IΓ )(1, −t) = − ν3 + t 1 ∂ t 1 + t 2 ∂ t 2 • IΓ (1, −t) .

We now eliminate ∂1 IΓ , ∂2 IΓ , ∂3 IΓ in terms of ∂ t 1 , ∂ t 2 , ∂ t 3 . E.g., the first equation of (90) is

P̃3 • IΓ (z) = (z1 ∂1 + · · · + z6 ∂6 + s) • IΓ (z) = 0 .

This is equivalent to the following relation for IΓ (1, −t) = IΓ (1, 1, 1, −t 1 , −t 2 , −t 3 ):

P3 • IΓ (1, −t) = 0, P3 = t 1 ∂ t 1 + t 2 ∂ t 2 + t 3 ∂ t 3 + ν1 + ν2 + ν3 − s . (98)

36
 SciPost Phys. Lect. Notes 75 (2023)

Likewise, we obtain a relation

(∂4 ∂1 IΓ )(1, −t) = −∂ t 1 • (∂1 IΓ )(1, −t) = ∂ t 1 • ν1 + t 2 ∂ t 2 + t 3 ∂ t 3 • IΓ (1, −t)

(99)
= −∂ t 1 • ν2 + ν3 − s + t 1 ∂ t 1 • IΓ (1, −t) .


(100)

Here, we used (98) when passing from (99) to (100). Repeating this for the monomials in P̃1 ,
P̃2 , we find that IΓ (1, 1, 1, −t 1 , −t 2 , −t 3 ) is annihilated by the following operators:

P1 = t 1 ∂ t2 − t 3 ∂ t2 + (1 − s + ν2 + ν3 )∂ t 1 − (1 − s + ν1 + ν2 )∂ t 3 ,
1 3

P2 = t 2 ∂ t2 − t 3 ∂ t2 + (1 − s + ν1 + ν3 )∂ t 2 − (1 − s + ν1 + ν2 )∂ t 3 ,
2 3

P3 = t 1 ∂ t 1 + t 2 ∂ t 2 + t 3 ∂ t 3 + ν1 + ν2 + ν3 − s .

These operators agree with the ones in [39, Equation (2.6)] after setting s = D/2. ⋄

The local solutions of HA(β) at a point z ∗ ∈ CA are given by the direct limit

SolHA(β),z ∗ = lim SolHA(β) (U) .

−→
z ∗ ∈U

Elements of SolHA(β),z ∗ are represented by holomorphic solutions of HA(β), defined on a suffi-

ciently small open neighborhood U of z ∗ . Theorem 4.12 describes it in terms of integrals. For
a facet Q of the polyhedral cone pos(A), we write rQ for the primitive ray generator of the dual
ray { y ∈ (Rd )∨ : y · q ≥ 0 for all q ∈ Q}. A complex vector β ∈ Cn+ℓ is said to be non-resonant
if rQ · β ∈
/ Z for any facet Q of pos(A). The following is [31, Theorem 2.10]:

Theorem 4.12. Let β = −(s, ν) ∈ Cℓ+n be non-resonant. For any z ∗ ∈ CA, the map
H n (X z ∗ , −ω(z ∗ )) → SolHA(β),z ∗ given by [Γ (z ∗ )] 7→ IΓ (z) is a vector space isomorphism.

Remark 4.13. Theorem 4.12 implies that, for generic s, ν (in the sense of the Vanishing The-
orem 3.13), the dimension of the local solution space SolHA(β),z ∗ equals the signed Euler char-
acteristic of X z ∗ . For z ∗ outside an algebraic hypersurface {EA = 0} ⊂ CA, this number equals
the normalized volume of the convex hull of A [1, Theorem 5.15]. The polynomial EA is the
principal A-determinant, as introduced by Gelfand, Kapranov and Zelevinsky [32, Chapter 6].

While the function IΓ is an integral over Γ (z) against a particular cohomology class [ dxx ],
the pairing 〈Γ (z), φ〉 is well-defined for any twisted cocycle [φ] ∈ H n (X z , ω). Let [φ1 ],. . . ,[φχ ]
be a basis for the twisted cohomology H n (X z , ω) for z ∈ U. Again, by Corollary 3.19, there
are χ = (−1)n · χ(X z ∗ ) basis elements. We assume the [φk ] have coefficients that are rational
functions in z. There exist χ × χ-matrices Pα (α ∈ A) such that

〈Γ (z), φ1 〉 〈Γ (z), φ1 〉
   
.. ..
∂α •  .  = Pα ·  . . (101)
〈Γ (z), φχ 〉 〈Γ (z), φχ 〉

Here ∂α acts entry-wise on vectors. These expressions form the so-called Pfaffian system. The
Pfaffian system can be derived from a system of differential operators, like a GKZ system. The
general procedure is explained in [20, Section 3].

Remark 4.14. Pfaffian systems lead to one of the most efficient ways of evaluating Feynman
integrals [37]. In practice, (101) can be solved by providing boundary conditions 〈Γ (z ∗ ), φi 〉
for i = 1, . . . , χ at some z = z ∗ and using path-ordered exponentiation of the matrices Cα to
evaluate the Pfaffian system at other values of z. See [38] for a pedagogical introduction.

37
 SciPost Phys. Lect. Notes 75 (2023)

5 Open problems
The previous sections provide an overview of the basics of Euler integrals. While this is a
classical topic, the theory is currently still very much in development. We conclude with a list
of open research problems, hoping that the reader will join this effort.

1. Evaluating integrals numerically. When the integral (13) converges, it can be evalu-
ated numerically using sector decomposition and Monte Carlo integration, see Remark 1.9.
When the real part of the exponents is large, it becomes increasingly important to con-
centrate the Monte Carlo samples in the close neighborhood of the critical point a from
Section 2.3. This could lead to effective numerical algorithms for evaluating convergent
Euler integrals, with applications in Bayesian statistics [13]. Likewise, in physics appli-
cations one often has to analytically continue in the parameters (s, ν) before numerical
evaluation, which can be achieved using the results of Sections 3.3 and 4.1.

2. Generic parameters. Theorems 2.1 and 3.13 make genericity assumptions on the expo-
nents s, ν. In Theorem 3.13 that means outside a closed algebraic subvariety, in Theorem
3.13 it means outside a countable union of hyperplanes. The former can be seen as a limit
of the latter, by driving the parameter δ from Section 2 to zero [47]. It is interesting to
describe these hyperplanes explicitly, and investigate this problem in more detail.

3. Non-generic parameters. In physics applications, one often encounters Euler integrals

with special coefficients and parameters. In particular, these are not generic in the sense of
Theorems 2.1 and 3.13. It would be interesting to develop the analogous theory applicable
to such cases, perhaps along the lines of [19, 49].

4. Regularized integration cycles. As mentioned in Section 3, a rigorous proof of Theorem*

3.24 for general Euler integrals is currently still missing.

5. Nice bases of cohomology. There are several reasons for which it is favorable to use basis
element for cohomology which are represented by dlog forms [57]. These are regular n-
forms obtained as dlog of a rational function. Another notion of a nice basis is related
to so-called canonical differentials equations for Feynman integrals [37]. In both cases, it
would be interesting to find criteria for such bases to exist.

6. Beyond Euler integrals. While our framework deals with Euler integral defined by (1),
there are other types of integrals that resemble it. The list includes exponential integrals
[27, 45], matrix hypergeometric integrals [35], and integrals over M g,n [26, 62]. Theorems
stated in this article mostly remain unsolved for these integrals.

7. Intersection pairing. The intersection pairing is a canonically defined operation on a

twisted cohomology, which can be used to reduce twisted cocycles to a basis [21, 48, 51].
Efficient evaluation of intersection pairing remains an important computational challenge.

8. χ-Stratification. In Section 4.2, the very affine variety X z depends on the coefficients z of
the Laurent polynomials. We propose to study the loci in coefficient space on which the
Euler characteristic χ(X z ) is constant. E.g., for which z ∈ (C∗ )A is |χ(X z )| minimal?

38
 SciPost Phys. Lect. Notes 75 (2023)

Acknowledgements
This article served as accompanying notes for a series of four lectures given by the third named
author at the Max Planck Institute for Mathematics in the Sciences in Leipzig. We thank Jörg
Lehnert for organizing these lectures, and Raluca Vlad for pointing out several typos in a
previous version.

Funding information This material is based upon work supported by the Sivian Fund and
the U.S. Department of Energy, Office of Science, Office of High Energy Physics under Award
Number DE-SC0009988. It is also supported by JSPS KAKENHI Grant Number 22K13930 and
partially supported by JST CREST Grant Number JP19209317.

References
[1] A. Adolphson, Hypergeometric functions and rings generated by monomials, Duke Math.
J. 73, 269 (1994), doi:10.1215/S0012-7094-94-07313-4.

[2] D. Agostini, T. Brysiewicz, C. Fevola, L. Kühne, B. Sturmfels, S. Telen and T. Lam, Likeli-
hood degenerations, Adv. Math. 414, 108863 (2023), doi:10.1016/j.aim.2023.108863.

[3] D. Agostini, C. Fevola, A.-L. Sattelberger and S. Telen, Vector spaces of generalized Euler
integrals, (arXiv preprint) doi:10.48550/arXiv.2208.08967.

[4] K. Aomoto, Les équations aux différences linéaires et les intégrales des fonctions multiformes,
J. Fac. Sci. Univ. Tokyo 22, 271 (1975).

[5] K. Aomoto, M. Kita, T. Kohno and K. Iohara, Theory of hypergeometric functions, Springer,
Tokyo, Japan, ISBN 9784431539124 (2011), doi:10.1007/978-4-431-53938-4.

[6] N. Arkani-Hamed, Y. Bai and T. Lam, Positive geometries and canonical forms, J. High
Energy Phys. 11, 039 (2017), doi:10.1007/JHEP11(2017)039.

[7] N. Arkani-Hamed, P. Benincasa and A. Postnikov, Cosmological polytopes and the wave-
function of the Universe, (arXiv preprint) doi:10.48550/arXiv.1709.02813.

[8] N. Arkani-Hamed, S. He and T. Lam, Stringy canonical forms, J. High Energy Phys. 02,
069 (2021), doi:10.1007/JHEP02(2021)069.

[9] C. Berkesch, J. Forsgård and M. Passare, Euler-Mellin integrals and A-hypergeometric func-
tions, Michigan Math. J. 63, 101 (2014), doi:10.1307/MMJ/1395234361.

[10] T. Bitoun, C. Bogner, R. P. Klausen and E. Panzer, Feynman integral relations from paramet-
ric annihilators, Lett. Math. Phys. 109, 497 (2018), doi:10.1007/s11005-018-1114-8.

[11] M. Borinsky, Tropical Monte Carlo quadrature for Feynman integrals, Ann. Inst. Henri
Poincaré Comb. Phys. Interact. (2023), doi:10.4171/AIHPD/158.

[12] M. Borinsky, H. J. Munch and F. Tellander, Tropical Feynman integration in the Minkowski
regime, Comput. Phys. Commun. 292, 108874 (2023), doi:10.1016/j.cpc.2023.108874.

[13] M. Borinsky, A.-L. Sattelberger, B. Sturmfels and S. Telen, Bayesian integrals on toric
varieties, SIAM J. Appl. Algebra Geometry 7, 77 (2023), doi:10.1137/22M1490569.

39
 SciPost Phys. Lect. Notes 75 (2023)

[14] P. Breiding and S. Timme, HomotopyContinuation.jl: A Package for Homotopy Continua-

tion in Julia, in Mathematical Software - ICMS 2018, Springer International Publishing,
Cham, ISBN 9783319964171 (2018), doi:10.1007/978-3-319-96418-8_54.

[15] R. Britto, S. Mizera, C. Rodriguez and O. Schlotterer, Coaction and double-copy proper-
ties of configuration-space integrals at genus zero, J. High Energy Phys. 05, 053 (2021),
doi:10.1007/JHEP05(2021)053.

[16] F. Brown and C. Dupont, Single-valued integration and superstring amplitudes in genus
zero, Commun. Math. Phys. 382, 815 (2021), doi:10.1007/s00220-021-03969-4.

[17] L. D. Brown, Fundamentals of statistical exponential families with applications in sta-

tistical decision theory, Institute of Mathematical Statistics, Beachwood, USA, ISBN
9780940600102 (1986).

[18] F. Cachazo, S. He and E. Y. Yuan, Scattering equations and Kawai-Lewellen-Tye orthogo-

nality, Phys. Rev. D 90, 065001 (2014), doi:10.1103/PhysRevD.90.065001.

[19] S. Caron-Huot and A. Pokraka, Duals of Feynman integrals. Part I. Differential equations,
J. High Energy Phys. 12, 045 (2021), doi:10.1007/JHEP12(2021)045.

[20] V. Chestnov, F. Gasparotto, M. K. Mandal, P. Mastrolia, S. J. Matsubara-Heo, H. J. Munch

and N. Takayama, Macaulay matrix for Feynman integrals: Linear relations and intersec-
tion numbers, J. High Energy Phys. 09, 187 (2022), doi:10.1007/JHEP09(2022)187.

[21] K. Cho and K. Matsumoto, Intersection theory for twisted cohomologies and
twisted Riemann’s period relations I, Nagoya Math. J. 139, 67 (1995),
doi:10.1017/S0027763000005304.

[22] S. C. Coutinho, A primer of algebraic D-modules, Cambridge University Press, Cambridge,

UK, ISBN 9780521551199 (1995), doi:10.1017/CBO9780511623653.

[23] P. Deligne, Équations différentielles à points singuliers réguliers, Springer, Berlin, Heidel-
berg, Germany, ISBN 9783540363736 (2006), doi:10.1007/BFb0061194.

[24] V. S. Dotsenko and V. A. Fateev, Conformal algebra and multipoint correlation functions in
2D statistical models, Nucl. Phys. B 240, 312 (1984), doi:10.1016/0550-3213(84)90269-
4.

[25] T. Duff, C. Hill, A. Jensen, K. Lee, A. Leykin and J. Sommars, Solving polynomial sys-
tems via homotopy continuation and monodromy, IMA J. Numer. Anal. 39, 1421 (2019),
doi:10.1093/IMANUM/DRY017.

[26] G. Felder and A. Varchenko, Integral representation of solutions of the ellip-

tic Knizhnik–Zamolodchikov–Bernard equations, Int. Math. Res. Not. 221 (1995),
doi:10.1155/S1073792895000171.

[27] J. Fresán, C. Sabbah and J.-D. Yu, Quadratic relations between periods of connections,
Tohoku Math. J. 75, 175 (2023), doi:10.2748/tmj.20211209.

[28] W. Fulton, Introduction to toric varieties, Princeton University Press, Princeton, USA, ISBN
9780691000497 (1993).

[29] E. Gawrilow and M. Joswig, polymake: A framework for analyzing convex poly-
topes, in Polytopes–combinaorics and computation, Birkhäuser, Basel, Switzerland, ISBN
9783764363512 (2000).

40
 SciPost Phys. Lect. Notes 75 (2023)

[30] I. M. Gelfand, General theory of hypergeometric functions, Sov. Math. Doklady 33, 537
(1986).

[31] I. M. Gelfand, M. M. Kapranov and A. V. Zelevinsky, Generalized Euler integrals

and A-hypergeometric functions, Adv. Math. 84, 225 (1990), doi:10.1016/0001-
8708(90)90048-R.

[32] I. M. Gelfand, M. M. Kapranov and A. V. Zelevinsky, Discriminants, resultants, and mul-

tidimensional determinants, Birkhäuser, Boston, USA, ISBN 9780817647704 (1994),
doi:10.1007/978-0-8176-4771-1.

[33] D. R. Grayson and M. E. Stillman, Macaulay2, a software system for research in algebraic
geometry, https://macaulay2.com/.

[34] P. Griffiths and J. Harris, Principles of algebraic geometry, Wiley & Sons, Hoboken, USA,
ISBN 9780471050599 (2014), doi:10.1002/9781118032527.

[35] H. Hashiguchi, Y. Numata, N. Takayama and A. Takemura, Holonomic gradient method

for the distribution function of the largest root of a Wishart matrix, J. Multivar. Anal. 117,
296 (2013) doi:10.1016/j.jmva.2013.03.011.

[36] G. Heinrich, Collider physics at the precision frontier, Phys. Rep. 922, 1 (2021),
doi:10.1016/j.physrep.2021.03.006.

[37] J. M. Henn, Multiloop integrals in dimensional regularization made simple, Phys. Rev. Lett.
110, 251601 (2013), doi:10.1103/PhysRevLett.110.251601.

[38] J. M. Henn, Lectures on differential equations for Feynman integrals, J. Phys. A: Math.
Theor. 48, 153001 (2015), doi:10.1088/1751-8113/48/15/153001.

[39] J. Henn, E. Pratt, A.-L. Sattelberger and S. Zoia, D-Module techniques for solv-
ing differential equations in the context of Feynman integrals, (arXiv preprint)
doi:10.48550/arXiv.2303.11105.

[40] J. Huh, The maximum likelihood degree of a very affine variety, Compositio Math. 149,
1245 (2013), doi:10.1112/S0010437X13007057.

[41] J. Huh and B. Sturmfels, Likelihood geometry, Springer, Cham, Switzerland, ISBN
9783319048697 (2014), doi:10.1007/978-3-319-04870-3_3.

[42] S. Lang, Undergraduate analysis, Springer, New York, USA, ISBN 9780387948416
(1996), doi:10.1007/978-1-4757-2698-5.

[43] R. N. Lee and A. A. Pomeransky, Critical points and number of master integrals, J. High
Energy Phys. 11, 165 (2013), doi:10.1007/JHEP11(2013)165.

[44] C. R. Mafra and O. Schlotterer, Tree-level amplitudes from the pure spinor superstring,
Phys. Rep. 1020, 1 (2023), doi:10.1016/j.physrep.2023.04.001.

[45] H. Majima, K. Matsumoto and N. Takayama, Quadratic relations for confluent hypergeo-
metric functions, Tohoku Math. J 52, 489 (2000).

[46] P. Mastrolia and S. Mizera, Feynman integrals and intersection theory, J. High Energy Phys.
02, 139 (2019), doi:10.1007/JHEP02(2019)139.

[47] S.-J. Matsubara-Heo and S. Telen, Twisted cohomology and likelihood ideals, (arXiv
preprint) doi:10.48550/arXiv.2301.13579.

41
 SciPost Phys. Lect. Notes 75 (2023)

[48] K. Matsumoto, Intersection numbers for logarithmic k-forms, Osaka J. Math. 35, 873
(1998).

[49] K. Matsumoto, Relative twisted homology and cohomology groups associated with Lauri-
cella’s F D , (arXiv preprint) doi:10.48550/arXiv.1804.00366.

[50] S. Mizera, Combinatorics and topology of Kawai-Lewellen-Tye relations, J. High Energy

Phys. 08, 097 (2017), doi:10.1007/JHEP08(2017)097.

[51] S. Mizera, Scattering amplitudes from intersection theory, Phys. Rev. Lett. 120, 141602
(2018), doi:10.1103/PhysRevLett.120.141602.

[52] S. Mizera, Aspects of scattering amplitudes and moduli space localization, Springer, Cham,
Switzerland, ISBN 9783030530099 (2020), doi:10.1007/978-3-030-53010-5.

[53] L. Nilsson and M. Passare, Mellin transforms of multivariate rational functions, J. Geom.
Anal. 23, 24 (2013), doi:10.1007/S12220-011-9235-7.

[54] Oscar – open source computer algebra research system, version 0.12.1-dev (2023), https:
//juliapackages.com/p/oscar.

[55] E. Panzer, Hepp’s bound for Feynman graphs and matroids, Ann. Inst. Henri Poincaré
Comb. Phys. Interact. 10, 31 (2022), doi:10.4171/AIHPD/126.

[56] T. Reichelt, M. Schulze, C. Sevenheck and U. Walther, Algebraic aspects of hypergeometric

differential equations, Contrib. Algebra Geom. 62, 137 (2021), doi:10.1007/s13366-020-
00560-1.

[57] K. Saito, Theory of logarithmic differential forms and logarithmic vector fields, J. Fac. Sci.
Univ. Tokyo Sect. IA Math. 27, 265 (1980).

[58] A. J. Sommese, C. W. Wampler, The numerical solution of systems of polynomials arising

in engineering and science, World Scientific, Singapore, ISBN 9789812561844 (2005),
doi:10.1142/5763.

[59] B. Sturmfels and S. Telen, Likelihood equations and scattering amplitudes, Alg. Stat. 12,
167 (2021), doi:10.2140/astat.2021.12.167.

[60] A. Varchenko, Critical points of the product of powers of linear functions and families of
bases of singular vectors, Compos. Math. 97, 385 (1995).

[61] S. Weinzierl, Feynman Integrals, Springer International Publishing, Cham, ISBN

9783030995577 (2022), doi:10.1007/978-3-030-99558-4.

[62] E. Witten, Superstring perturbation theory revisited, (arXiv preprint)

doi:10.48550/arXiv.1209.5461.

[63] G. M. Ziegler, Lectures on polytopes, Springer, New York, USA, ISBN 9780387943299
(1994), doi:10.1007/978-1-4613-8431-1.

42