MIPT/TH-17/23

IITP/TH-16/23
ITEP/TH-22/23

On factorization hierarchy of equations for banana Feynman amplitudes

V. Mishnyakova,b,c,1 , A. Morozova,c,d,2 , M.Revaa,c,3

a
MIPT, Dolgoprudny, 141701, Russia
b
Lebedev Physics Institute, Moscow 119991, Russia
c
NRC “Kurchatov Institute”, 123182, Moscow, Russia
d
Institute for Information Transmission Problems, Moscow 127994, Russia
arXiv:2311.13524v1 [hep-th] 22 Nov 2023

ABSTRACT

We present a review of the relations between various equations for maximal cut banana Feynman diagrams, i.e.
amplitudes with propagators substituted with δ-functions. We consider both equal and generic masses. There are three
types of equation to consider: those in coordinate space, their Fourier transform and Picard-Fuchs equations originating
from the parametric representation. First we review the properties of the corresponding differential operators themselves,
mainly their factorization properties at the equal mass locus and their form at special values of the dimension. Then
we study the relation between the Fourier transform of the coordinate space equations and the Picard-Fuchs equations
and show that they are related by factorization as well. The equations in question are the counterparts of the Virasoro
constraints in the much-better studied theory of eigenvalue matrix models and are the first step towards building a
full-fledged theory of Feynman integrals, which will reveal their hidden integrable structure.

1 Introduction
Lifting the theory of Feynman diagrams [1–7] to the level achieved in the study of the eigenvalue matrix
models [8–10], is one of the current tasks in theoretical physics. For a bird-fly view of the subject and its
connection to the celebrated Connes-Kreimer hidden symmetry [11] of Feynman graphs see [12]. Still, at
practical level the current stage of development is far more modest: majority of papers in the field [13–15]
are targeted at the study of Ward identities, i.e. differential equations on individual Feynman amplitudes,
considered as functions of external parameters. This is supposed to be in parallel with the Virasoro-constraint
stage [8, 10] of the matrix model studies.
Moreover, the study of these equations is focused mostly on the algebro-geometric interpretation, which
considers them as Picard-Fuchs (PF) equations for integral (periods) dependence of external parameters (moduli)
[13,14,16]. One natural family of equations is related to the nice families of certain toric Calabi-Yau spaces [3,17]
– which however does not cover Feynman integrals in all space-time dimensions.
In [18] it was suggested to approach the problem from a different side, and study the families of Feynman
graphs per se, starting from the simplest class of banana diagrams. The first simplification that this case brings
on is that, being two-point functions, these Feynman integrals depend on a single parameter x2 in the coordinate
or t = p2 in the momentum space representations. The second and important simplification from the point of
view of [18] is that in coordinate space the banana amplitude with n − 1 loops is just the product of n Green
functions. It therefore satisfies a conceptually trivial equation
(n)
Êx2 Bn (x) = 0 (1)

implied by the equation (□ + m2 )G = δ(⃗x) on a single Green function. For further simplification we omit the
δ-function at the r.h.s. and consider the propagators given by the solution of the ”homogeneous” equation. The
(n)
operator Êx2 is nothing but a tensor product of the basic Klein-Gordon operator. The important property of
1 mishnyakovvv@gmail.com
2 morozov@itep.ru
3 reva.ma@phystech.edu

1
this equation is the its solutions are only the n-fold products of propagators, therefore the banana amplitudes
Bn (x) are uniquely defined as the space of solutions to these equations. While banana amplitude is trivial in
position space, it is given by for more complicated n − 1 fold integrals in momentum space. On the other hand,
(n)
being a Fourier transform, they should satisfy equations given by Fourier transformed operators Êt :
(n) Fourier transform (n)
Êx2 −−−−−−−−−−−−−−−→ Êt (2)

The passing to ”homogeneous” equations implies the substitution of Feynman propagator by its cut (imaginary
part):
1
−→ δ(t − m2 ) (3)
t − m2
The corresponding amplitudes with all propagators replaced by δ-functions are called maximal cuts. On the
other hand these maximal cuts of banana integrals correspond to taking closed contour integration in the
(n)
parametric integrals and hence to periods and associated Picard-Fuchs equations P
d F t Bn (t) = 0. Here we
(n)
denote by P
d Ft the operator derived from the parametric integral. One should naturally wonder, whether the
(n) (n)
two differential equations given by Êt and P dF t are related.
(n)
To set up the conjecture we notice that it is quite easy to understand that the coordinate space operator Êx2
n
is a differential operator of exponential order – 2 in ∂x when all masses are different. This happens already in
the trivial case of D = 1 and is a manifestation of the simple fact that there are 2n independent n-fold products
of the original two solutions to the Klein-Gordon equation. Counting the order of the differential operator in
(n)
momentum space requires also keeping track of the dependence of Êx2 on x2 instead of just the derivatives.
(n)
Even this appears to be a complicated problem, and we can currently only conjecture, that the order of Êt
in ∂t also grows exponentially with n. We will discuss the known cases below. Similarly one could ask, what is
(n)
the order of P d F t ? This seems to be an even more sophisticated question and currently known examples do
not allow to differentiate between supposedly polynomial or exponential growth in n.
A general phenomenon that one encounters is that all of the discussed operators tend to factorize, often
dramatically dropping the order at coinciding masses or certain critical dimension, such as D = 2 in momentum
(n) (n) (n)
space. For example, for equal masses the position space operator Êx2 and both Êt and P d F t have order
n + 1 and n − 1 respectively. This phenomenon provides a plethora of differential operators even at given n and
hence a certain factorization hierarchy is present.
In view of the above discussion and based on the studied examples we conjecture that the relation between
(n) (n)
the Fourier transformed Êt and the P d F t is also that of factorization:
(n) (n)
Êt = B̂ · P
d F (4)

Where B̂ is some differential operator. Factorization of polynomials in the case of a single variable is fully
described by the fundamental theorem of algebra - any polynomial is factorizes over an algebraically closed field,
while factorization, say, over Q is not always possible, however, various methods and criteria are known. It is far
less simple for differential operators – even in a single variable [19]. Statements like Polya’s factorization [20,21]
are known for rather general operators:
n n 
∂i

X Y ∂
ai (t) i = + bi (t) (5)
i=0
∂t i=1
∂t

However, even if ai (t) are polynomial (rational) coefficients, the functions bi (t) can be complicated multivalued
functions. Here we will call an operator factorizable if it can be factorized into other operators with polynomial
(rational) coefficients. However criteria for existence of factorization are rather involved and algorithms are
known only in special cases .
Picard-Fuchs operators are reasonably studied in some cases, like Calabi-Yau operators [22–25], which has
only accidental intersections with banana Feynman amplitudes. However even in this case no factorization sim-
plifications are presented. Additional complication is that operator B̂ in (4) is not necessarily with polynomial
coefficients – in general they are rational functions of t, and a better formulation of factorization is

(n) (n)
P · Êt = B̂ · P
d F (6)

with a polynomial P (t) and an operator B̂(t, ∂t ) with polynomial coefficients.

2
 Our goal in this overview paper is to illuminate the current status of the relation between the operators
(n) (n)
Êt and Pd F t and, furthermore, their factorization properties at special points. Not much is known about
this factorization hierarchy even for banana diagrams, which is why we try to structure the known examples
and conjecture some general properties to set up the stage for further, more technical, works. It is a technically
involved problem, and an adequate language, which would make the answers as simple as the questions, is still
to be found. For technical details behind this overview see [26].

2 Equations Ê (n) in coordinate space

2.1 The basic case of D = 1
ˆ + m2 )Gm = 0, which is the simplest for D = 1: then
Cut propagators are solutions of the equation (□
±imx 2
G±m = e since □ = ∂x . A banana Feynman diagram in coordinate space is a product of n such functions:
n
Y
Bn (x) = e±mi x (7)
D=1
i=1

with no integration. These functions can clearly be uniquely defined as the set of solutions of the equation
Y n
Y
□ˆ + (m1 ± m2 ± . . . ± mn )2 G±mi = 0 (8)
± i=1

where the product is over all the possible combinations of + and − between the masses. This equation is of the
ˆ := ∂ µ ∂µ
order 2n−1 in □ = ∂x2 operator. For the equal mass case some of the alternate sums of the masses
D=1
become equal. The simplest example is m1 − m2 and m2 − m1 for n = 2 which both vanish at equal masses.
Hence the number of independent banana functions is now n + 1 and given by:
Bn (x) = e−nmx , e−(n−2)mx , . . . , ei(n−2)mx , einmx (9)
This set of functions is the space of solutions to the equation:
⌊n/2⌋  n
1+(−1)n Y Y
∂x 2 ˆ + (n − 2k)2 m2
□ G±m = 0 (10)
k=0 ±

The ∂x factor is responsible for the constant solutions which appear for even n.

As we can see the order drops greatly - from an exponential in n to linear. This happens through factorization
of the unequal mass operator. For example n = 2:

4
□4 □ + 4m2 = □3 (□ + 4m2 )3 ∂x ∂x □ + 4m2



(11)
n−k−1
When 2 P ≤ k ≤ n − 1 of n masses are the same,
Q the order is 2 (k + 1), when there P
are clusters of ks equal
masses, s ks = n, then the order in ∂x is s (ks + 1), modulo possible coincidences at i ±mi = 0 for several
even ks .

2.2 Beyond D = 1
Beyond D = 1 explicit dependence on ⃗x2 appears and equations become far more complicated. The most
practical way to write them is to use the dilatation operator:
Λ̂ = xµ ∂xµ (12)
together with the Laplacian
ˆ := ∂µ ∂ µ
□ (13)
Then the procedure to obtain the equation on a product of solutions of the original (□ + m2 )G = 0,
outlined in [18], gets a straightforward and easily computerizable realization. By definition, the order of the
resulting equation in Λ̂ remains the same: from n + 1 for equal masses and 2n for all masses different.
For example for n = 2 and equal masses one has:
ˆ(2)
Ex = Λ(□ + 4m2 ) + (D − 1) □ + 2m2


(14)

3
 √
As mentioned above, ar D ̸= 1 the equation is also graded by powers of x = xµ xµ . For further purposes
of studying the Fourier transform, it is important to keep track of this grading as well. For the equal mass case
it was studied in [18] - the order is n − 1 if we also count the power of x that enters the operator Λ. The same
question for the unequal mass case still remains unresolved.
√
Note that for the action on invariant functions of x = xµ xµ the operators Λ and □ can be rewritten as:

∂ ∂2 D−1 ∂
Λ=x , □= + (15)
∂x ∂x2 x ∂x
√
where the derivative is also over x = xµ xµ . We will use this representation to count the order of the equations
in x and derivatives. That way it is also easier to compare with D = 1. For other purposes the representation
in terms of Λ, □ and x2 seems to be more appropriate. This hierarchy of orders of the equations is summarized
in the following table (see Appendix 2 of [18]), where we list the leading term in both x and px :

Êx all masses equal all masses different

n
D=1 ∂xn+1 ∂x2

x2 ∂x4 at n = 2
x10 ∂x8 at n = 3
D>1 xn−1 ∂xn+1 x48 ∂x16 at n = 4
x226 ∂x32 at n = 5
n
x? ∂x2

The
Qnsimplest way to derive these equations is to act by powers of Λ̂ on the product of Green functions
Zn = i=1 Gmi (x2 ). Then, making use of the Leibnitz rule and the equation Λ̂2 Gm = −(D −2)Λ̂Gm −m2 x2 Gm
for individual Green functions, we obtain the relations of the form
Zν ,...,ν
z 1}| n {
X Yn
Λ̂s Zn = a{ν
s
1 ,...,νn }
Λ̂νi Gmi (16)
ν1 ,...,νn =0,1 i=1

Since there are 2n in dependent quantities Zν1 ,...,νn we conclude that the determinant of the matrix of the size
2n + 1 should vanish, e.g.

0 1 −Z1 0 1 −Z1
    
Λ̂G
2 {1} {0}
Λ̂ G = a2 Λ̂G + a2 G ⇐⇒  1 0 −Λ̂Z1   G =0 ⇐⇒ det  1 0 −Λ̂Z1  = 0
{1} {0} 2 {1} {0}
a2 a2 −Λ̂ Z1 1 a2 a2 −Λ̂2 Z1

for the zero-loop case n = 1 with Z1 = G. This provides the equation in coordinate space in the form of linear
relation between Λ̂s Zn with s = 0, . . . , 2n , which can be easily Fourier-transformed to the momentum space.
It is clearly of degree 2n in ∂x , but the degree in x2 is a little trickier. In the first approximation we can use
the fact that the coefficients as in (16) are of degree 2s in x2 , what implies that the degree of determinant
P2n  
cannot exceed than s=0 2s = 4n−1 . However, there are two corrections: on the one hand, the actual power
is slightly lower because of the overlaps between the powers of x2 in different lines, and on the other hand the
entire determinant appears to be proportional to some power of x2 , which can be factored out – actually it is
approximately one half of the net 4n . Moreover, the coefficient of Λ̂0 Zn also appears to be proportional to x2 ,
and, since the action of Λ̂ on functions of x2 is always proportional to x2 , Λ̂F (x2 ) = 2x2 F ′ (x2 ), this allows
to subtract one more power. This is a small correction, but it is essential for exact answers in above table.
The resulting power of x is slightly lower than 4n−1 , the general answer is still unavailable (and even the power

4
226 ≈ 44 = 256 is a result of a hard calculation).

n 4n−1 x2 − power of determinant common factor x2 − power of the equation

1 1 1 0 1−1=0
2 4 4 1 2−1=1
3 16 12 6 6−1=5
4 64 52 27 25 − 1 = 24
5 256 ? ? 114 − 1 = 113
···

For explicit examples of coordinate-space equations for banana diagrams see the appendices in [18].

2.3 Transformation of operators to momentum space

In order to go to momentum space we can do the Fourier transform directly in terms of the operators. The
rules follow from considering functions of invariants ⃗x2 and t = p⃗2 :
Z Z
∂ ipµ xµ µ
e f (t) = eipµ x (ipµ ) f (t)
∂xµ
Z Z   (17)
µ ipµ xµ ipµ xµ ∂
x e f (t) = e i f (t)
∂pµ

Hence, for invariant operators one has the following list of rules for transiting to momentum space:

Coordinate space Momentum space

□ = ∂µ ∂ µ −t = −pµ pµ

∂
Λ = x∂x = xµ ∂µ −(D − 2) − 2 t
∂t
∂ ∂
−(D + Λ) 2t = pµ µ
∂t ∂p
∂ ∂2
x2 = xµ xµ −2D − 4t 2
∂t ∂t

(n)
This allows us to compute the operators Êt . These operators now define the differential equations for
the momentum space banana integral, which is now indeed a loop integral and not a product of propagators.
One can explicitly show at low loop orders, that the result of integration indeed satisfies the equation. Some
examples were considered in [18], and some more will appear in sec. 4. Here we limit ourselves again to the
n = 2 equal mass case:
   
(2) d 2 2 2 d 2
Êt = D + 2t (t − 4m ) − 2(D − 1)(t − 2m ) = 2t(t − 4m ) + (t(4 − D) − 4m ) . (18)
dt dt

which should be compared to (14).

Here we do not intend to study the solutions of the equations; instead, as for the momentum space we want
to study their order for different cases. It is clear from the table above, that each multiple of x, whether it
enters through x2 or as part of the Λ operator, contributes a t derivative. Position space derivatives on the
other hand correspond just to powers of t. Having that in mind we can represent the order of the resulting
equation in the following table:

5
 Êt all masses equal all masses different

n
D=1 pn+1 p2

∂2
t2 at n = 2
∂t2
∂ 10
t4 10 at n = 3
∂t
∂ n−1 ∂ 48
D>1 tn t8 48 at n = 4
∂tn−1 ∂t
∂ 226
t16 226 at n = 5
∂t
n−1
t2 ∂t?

An interesting observation can be made about the order of the equations. For the equal mass case the order of
the differential operator is n − 1, whereas the position space operator had order n + 1. On the other hand, we
expect that the solutions to the momentum space equations are just the Fourier transforms of the momentum
space equations, so we expect the number of solutions to be the same. In fact, the counting appears to be more
intricate. This can already be seen for D = 1. At D = 1 the equations are not differential, just as for the single
cut propagator we have:
(p2 − m2 )G(t) = 0 (19)
On the other hand we have two solutions to the equations of motion which are clearly given by:
G(p) = δ(p ± m) (20)
We will not go further into this question here and leave it for a more technical paper. The takeaway, however,
is that comparing the orders of Eˆx and Êt is not naive.

The unresolved piece of the puzzle is the lower right corner of the table. The degree of the differential
operator Êt for unequal masses is unknown as it corresponds to the power of x in the respective position space
equation. The correspondence between the two unequal and equal mass case is again that of factorization.

3 PF equations in the momentum space

Our approach allows to derive differential equations for momentum space integrals starting only with the equa-
tions of motion. Another commonly discussed approach is to derive Picard-Fuchs equations for the parametric
representation of the integral, which will produce some differential operator in the t variable P
d F . In this section
we will discuss the origin and form of P F and discuss it’s properties. Most of the results obtained in this section
d
can also be found in [13, 27]

3.1 The origin and the form of PF equations

The parametric representation for momentum Feynman integrals and banana integrals specifically is well known
and relies on the 2 identities:
Z ∞
1 2 2

2 2
= dα e−α(p −m )
p −m 0
Z ∞ (21)
2 2
δ(p2 − m2 ) = dα e−iα(p −m )
−∞

For general derivations see [1]. The momentum space banana diagram is an n−1 loop integral and its parametric
form is given by:
n
! n n(2−D) Pn Qn
U (α) 2 δ ( i=1 αi − 1) i=1 dαi
Z X Y Z
In (t) := δ ⃗ki − p⃗ δ(ki2 − mi )2 dD⃗ki = (22)
 Pn 1+ (n−1)(2−D)
2
i=1 i=1 Rn 2
α1 . . . αn · t − U (α) · i=1 αi mi

6
where
n
Y X n
X
U (α) := αi αi−1 = α1 . . . α̌i . . . αn (23)
i=1 i i=1

and the denominator is usually denoted as:

 n
X 
F (α) = α1 . . . αn · t − U (α) · αi m2i (24)
i=1

The U and F are called the first and second Symanzik polynomials correspondingly. The integration goes over
the whole Rn (i.e. from −∞ to +∞) as opposed to only αi > 0 since we are considering the maximal cut.
After complexification the integral can be written as a contour integral:

I n−1 n(2−D)
Y U (α) 2
In (t) = dαi  1+ (n−1)(2−D) (25)
Pn 2
γ i=1 α1 . . . αn · t − U (α) · i=1 αi m2i Pn−1
αn =1− i=1 αi

where γ is some closed contour appropriately encircling the zero set of F . We could also work projectively to
get:
Z n(2−D)
U (α) 2 ω
In (t) =  1+ (n−1)(2−D) (26)
Pn 2
Γn 2
α1 . . . αn · t − U (α) · i=1 αi mi
Pn di ∧· · ·∧dαn is measure on CPn−1 and Γn = (α1 , . . . , αn ) ∈ CPn−1 |αi ∈ R .

where ω = i=1 (−1)i+1 αi dα1 ∧· · ·∧dα
It is important to note, that the integrand in (26) equals the one in (22) only on the integration domain. The
contour Γn is closed and to simplify calculations we can rewrite this integral as an integral over Cn :
I n(2−D) Qn
U (α) 2 dαii=1
In (t) =  1+ (n−1)(2−D) (27)
Pn 2
γ̄ α1 . . . αn · t − U (α) · i=1 αi m2i

where γ̄ is the fibration over γ with fibre S 1 . This equation instantly follows from the identity ( dα
α1 + · · · +
1

dαn Q n
αn ) ∧ ω = i=1 dαi and integration over the fibres. Thus in any case the integral is a “period” and satisfies
the Picard-Fuchs equation. Denoting the integrand n-form by Ω, we can say that the equation:
Z
P
d F Ω = 0 ⇐= P dF · Ω = dβ (28)
γ̄

follows from a differential equation where the l.h.s. contains derivatives in t, while the r.h.s. is made from
α-derivatives.

3.2 The case of equal masses

Notice that for D = 2 the parametric integral becomes simpler:
I n−1
Y 1
In (t) = dαi (29)
γ i=1 F (α)

In this case it is straightforwardly understood as a period of the manifold that is the vanishing locus of the
denominator:
F (α) = 0 (30)
It turns out to be a singular (for n > 3) n − 2 dimensional Calabi-Yau variety. As for our matter of deriving a
differential equation this case is also simpler. For example, at n = 2 (one-loop) we have:
  1  
∂  ∂ 1 − 2α1
t (t − 4m2 ) + t − 2m2 = m2 (31)
∂t F α2 =1−α1 ∂α1 F (α1 , α2 ) α2 =1−α1

7
 ∂4 ∂3 ∂2 ∂
n−1 ... ∂t4 ∂t3 ∂t2 ∂t
1

1 t(t − 4) t−2

t(t − 1)(t − 9) =
2 3t2 − 20t + 9 t−3
= t3 − 10t2 + 9t

t2 (t − 4)(t − 16) =
3 6t3 − 90t2 + 192t 7t2 − 68t + 64 t−4
= t4 − 20t3 + 64t2

t2 (t − 1)(t − 9)(t − 25)

2 · (5t4 − 140t3 + 25t3 − 518t2 + 1839t −
=

4 25t2 − 196t + 285 t−5

t5 −35t4 +259t3 −225t2 +777t2 − 450t) 450

.. .. .. .. .. .. ..
. . . . . . .

Table 1: Coefficient of the equal mass banana Picard-Fuchs operators. The underlined coefficient is always two
∂cn−1
times the derivative of the first coefficient cn−2 = n−1
2 ∂t , where n − 1 is the number of loops. We have set
m = 1 for brevity

For D = 2, the underlined coefficient t − 2m2 appears equal to one half of the t-derivative of the first coefficient
t(t − 4m2 ). Moreover, this property is true for bigger number of loops, where the coefficients of the PF equation
for D = 2 and all masses equal are given in table 1. The general structure of the operator is:
n−2
(2) ∂ n−1 X ∂i
D=2 : P
d F = cn + ci i (32)
∂tn−1 i=0
∂t

This agrees with the those from [13] and can be compared to momentum space operators in [18]. On the level
of PF equations the generalization to D ̸= 2 is straightforward and D appears polynomially in the coefficient
of the PF operator. Once again, for n = 2 and all masses equal:
U 2N
   
2 ∂

2 2 ∂ 1 − 2α1
t (t − 4m ) + (N + 1)t − 2m N +1
=m (33)
∂t F (α1 , α2 ) α2 =1−α1 ∂α1 F (α1 , α2 )N +1 α2 =1−α1
2−D
with N = 2 . Hence:  
∂ 4−D
P
d F 2 = t (t − 4m2 ) + t − 2m2 (34)
∂t 2
(n)
It is worth mentioning at this point already that for equal masses but generic D the operators Êt are
(n)
equal to P
d F . This statement was confirmed on all the calculated examples. Moreover at D = 2 [18] contains
a comparison of the leading and subleasing coefficient with the generic formulas of [13]. Hence the Fourier
transform is an easier method to compute these operators for many orders, when they are needed explicitly for
analysis. We have the first 8 loops of the position space operators listed in [18]

3.3 Different masses

Again, the simplest case is D = 2. One can explicitly obtain the PF operators and see that they are of the form
 ∂ ord(n)−1 ord(n)−1
(n) Y X ∂i
D=2: P
d F = c̃ord(n)−1 t − (m1 ± m2 ± . . . ± mn )2 ord(n)−1
+ ci i (35)
±
∂t i=1
∂t
| {z }
cord(n)−1

The structure of other coefficients is much more complicated now, however, what is important is that the order
also changes compared to the equal mass case. In particular it was conjectured in [13] that the order of the PF
equation at D = 2 for generic masses is:
 
n+1
ord(n) = 2n − (36)
⌊ n+1
2 ⌋

8
This leads to exponentially growing order, which is closer to agreement with how drastically the order grows
(n)
for the Êt operators.

Interestingly enough, the generalization to D ̸= 2 is now non-trivial. The only example, for which D only
deforms in the coefficients, is for n = 2:
 
(2) ∂
F = 2t(t − (m1 − m2 )2 )(t − (m1 + m2 )2 ) + (4 − D)t2 − 2(m21 + m22 )t + (D − 2)(m21 − m22 )2


P
d
∂t
 
∂
2t(t − (m1 − m2 )2 )(t − (m1 + m2 )2 ) + (4 − D)t2 − 2(m21 + m22 )t + (D − 2)(m21 − m22 )2 ·


∂t
(α1 + α2 )2−D
· 2− D2 =
2 2
(α1 + α2 )(m1 α1 + m2 α2 ) − tα1 α2
(α1 + α2 )3−D
 
∂ ∂
= 2 m22 (t + m21 − m22 ) + m21 (t + m22 − m21 ) 2− D2 =
∂α1 ∂α2 
2 2
(α1 + α2 )(m1 α1 + m2 α2 ) − tα1 α2
 
2 2 2 2 2 2 2−D
∂ 2 − m1 (t − m1 + m2 )α1 + m2 (t + m1 − m2 )α2 (α1 + α2 )
= 2− D2 (37)
∂α1 
tα1 α2 − (α1 + α2 )2−D (m21 α1 + m22 α2 )

where α2 = 1 − α1 . In fact, after this substitution the expression is simplified a little, but we keep it in this
form, which is generalizable to higher loops. The l.h.s. of this one-loop equation is in accordance with [14].

At higher n going to D ̸= 2 results in the growth of the order of the equation. In particular for n = 3 the
equation is of order 4:
4
(3) X ∂i
D ̸= 2 : P dF = ci (m1 , m2 , m3 , D) i (38)
i=0
∂t
This is in agreement with [27] and we can independently confirm the validity of their answer. To the best of
our knowledge there is no prediction on the order of the PF equation at D ̸= 2 and general n. Note that for
the cohomological reasoning about the closed contour integral to be valid we should require that D is even. On
the other hand after computation we clearly see that dependence of the operator coefficients on D is analytic.
Finally we summarize what is known about Picard-Fuchs equations for banana-graphs in the table:

all unequal mass, D = 2

equal mass, D = 2 factorization [13]
n+1

Order of P
d F : 2n − ⌊ n+1
Order of P
d F : n−1 2 ⌋
[13]

set D = 2
in coeffi- factorization
cients

all unequal mass, D ̸= 2

equal mass, D ̸= 2
Order of P
d F : unknown
factorization
Order of P
d F : n−1 n = 2 , ord = 1
+ n = 3 , ord = 4
D-dependent coefficients +
D-dependent coefficients

The most complicated equation is the bottom right square. Its reduction to the equal mass or D = 2 cases
is archived by factorization. However even looking at the orders we see that these are two separate phenomena.

9
3.4 Self-adjoint property
At D = 2 the Picard-Fuchs operators correspond to periods of Calabi-Yau manifolds and hence are supposed
to be self-adjoint [22–25]. Generically, for operator:
X ∂i
P = ai (39)
i
∂ti

Its adjoint is
X ∂i
P = (−1)i ai (40)
i
∂ti
Therefore the operator is self-adjoint when the following relation holds:

P g(t) = (−1)degP P g(t) = (−1)degP g(t)P (41)

for some g(t) ̸= 0. For g(t) = 1 this implies special explicit relations between all coefficients with even i, e.g.

(n − 1)(n − 2)(n − 3) ∂ 3 an−1 (n − 2)(n − 3) ∂ 2 an−2 n − 3 ∂an−3

an−4 = · 3
− · 2
+ · =
12 ∂t 4 ∂t 2 ∂t (42)
(n − 1)(n − 2)(n − 3) ∂ 3 an−1 n − 3 ∂an−3
=− · + ·
24 ∂t3 2 ∂t
The PF operators are self-adjoint for D = 2 and generic masses, which is in accordance with the expectation
that the Feynman integrals are periods of (n − 2)-dimensional Calabi-Yau variety. In particular for equal mass
D = 2 operators (see table 1):
n − 1 ∂cn−1
cn−2 = (43)
2 ∂t

For generic dimensions this property is broken. This implies that PF operators for generic banana diagrams
are not self-adjoint and are not of the Calabi-Yau type. This is clear already from the equal mass case for
D ̸= 2. For a nontrivial example, we can look at n = 4:

(4)
∂3
∂2
= t2 t − 16m2 t − 4m2 − 3t −10(D − 5)m2 t + (D − 4)t2 − 64m4


P
d F 3 2
+
 ∂t ∂t
1 ∂
+ −16(D − 4)Dm4 + ((88 − 7D)D − 216)m2 t + (D − 4)(11D − 36)t2 − (44)
4 ∂t
1
− (D − 3)(3D − 8) 2(D + 2)m2 + (D − 4)t


4
which can be checked to be not self-adjoint. Moreover, according to [18] (see also next section the deformation
D ̸= 2 of (43) is:  
(n − 1) n(D − 2) ∂cn−1
cn−2 = · cn−1 − (D − 3) (45)
2 2t ∂t
which true for all loops. The divergence of (45) from (43) on itself does not yet mean that the operator is not
self adjoint. We could potentially find a non-trivial g from (41). However, if a simple check shows that system
of equations on all the coefficients is inconsistent, hence the operator is not self-adjoint.
On the other hand the rather simple structure (45) suggest there could exist an appropriate deformation of
the self-duality condition. We leave this for future investigation.

4 Relation between Fourier transform of position space equations

and Picard-Fuchs equations
(n) (n)
As clear from the tables above the operators Êt and P d F are significantly different from the equal mass
case. One can clearly see, that for unequal masses the order of the operator obtained by Frourier transfrom of
the position space equations grows faster than those of the Picard-Fuchs operator. On the other hand one can

10
 (n)
check that the kernel of the PF operator is contained in the kernel of Êt . This means that the operators are
once again related by factorization. Why and how this happens is yet unclear and remains a problem for future
investigation.

The only truly tractable example is at n = 2. In position space one has [18]:
  
Êx(2) = x2 □ + (m1 + m2 )2 □ + (m1 − m2 )2 + 2(D − 1)Λ̂(□ + m21 + m22 ) + 2(D − 1)(D − 2)(□ + m21 + m22 ) (46)

and it’s momentum space version

 
(2) ∂ (2)
Êt = −2 P
d F =
∂t
  (47)
∂ ∂
2t(t − (m1 − m2 )2 )(t − (m1 + m2 )2 ) + (4 − D)t2 − 2(m21 + m22 )t + (D − 2)(m21 − m22 )2


−2
∂t ∂t
Here we can clearly trace the reduction of the order of operators from 2 to 1 according to the tables. Even at
(3) (3)
n = 3 we seem to lack the appropriate language to describe the relation between Êt and P d F . Mainly we
are supposed to have:
order 10 order 8 order 2
z }| { z}|{ z }| {
(3) (3)
D=2 : Êt = B × P
d F
(48)
order 10 order 6 order 4
z }| { z}|{ z }| {
(3) (3)
D ̸= 2 : Êt = B × P
d F

Of course knowing the explicit answer for the r.h.s factor we can manually check that it divides l.h.s operator.
(n)
However, we lack the understanding for the reason to expect Êt to be divisible and the techniques to find the
r.h.s factor independently without knowing the answer beforehand. Notice that if we want to stick to operators
with polynomial coefficients, then a modification of (48) is due. The operator B in both cases appears to have
rational coefficients, hence after multiplication of both sides we have exactly the relation announced in the
introduction:
(n) (n)
P · Êt = B̂ · P
d F (49)
A convenient way to study the relation between the operator and its factors is to apply it to a formal series
(3)
un tn and convert into recurrence relation [26]. For P
P d F at unequal mass and D ̸= 2 we obtain
n

3(n + 1)2 un + a(2) · un+1 + b(2) · un+2 + c(2) · un+3 + d(2) · un+4 + e(2) · un+5 + (n + 6)2 f (0) · un+6 = 0 (50)
(3)
where a, b, c, d, e, f are polynomials in n of degrees, shown in the superscripts. The Fourier transform Êt of
the coordinate space equation from A.3 of [18] converts into

(n + 1)2 (n + 2)2 (n + 3)2 75(n + 1)2 un + A(4) · un+1 + B (6) · un+2 + C (6) · un+3 + (n + 4)2 D(4) · un+4 +
 (51)
+ (n + 4)2 (n + 5)2 E (2) · un+5 + (n + 4)2 (n + 5)2 (n + 6)2 F (0) · un+6 = 0

where some factors are explicitly extracted from the coefficients – the superscripts refer to the power of n
which remains in the irreducible factors. The question is – in what sense is (51) divisible by (50)? Note that
in this representation both are difference equations of the same 6-th order, just coefficients are different. This
means that after division we are supposed to obtain an operator which is homogeneous in t. Looking at the
emergent structures we can actually obtain the polynomial multiplier in the r.h.s required for the polynomiality
of operators:
 10
P = 3t2 − 2 m21 + m22 + m23 t − m41 − m42 − m43 + 2m21 m22 + 2m21 m23 + 2m22 m23


(52)

5 Conclusion
The purpose of this note was to summarize the result of a technically complicated work, reported in [26], in
relatively simple words, emphasizing the various morals which can be extracted already at this stage. The

11
first takeaway is the somewhat well known fact that for generic values of parameters even the Picard-Fuchs
equations are much more sophisticated than those related to toric Calabi-Yau varieties. They form a different
family of operators and the respective family of varieties remains to be identified, The second takeaway is that
the position space approach which is conceptually simpler to formulate produces however even more complicated
equations in momentum space. They generically have a (much) greater order and are related to Picard-Fuchs
operators is via highly non-trivial factorization of the form (6). This relation remain to be made conceptually
clear, possibly requiring a new conceptual step. A positive example example is provided by the equal mass case
where even at generic D one can discover and prove general statements about the momentum space equation a
la (45).

Acknowledgements
We are indebted for valuable discussions and comments to P. Suprun.
Our work is partly supported by the BASIS foundation and RFBR grants 21-51-46010-ST-a and 21-52-52004.

References
[1] S. Weinzierl, Feynman Integrals. 1, 2022. arXiv:2201.03593 [hep-th].

[2] C. Duhr, “Mathematical aspects of scattering amplitudes,” in Theoretical Advanced Study Institute in
Elementary Particle Physics: Journeys Through the Precision Frontier: Amplitudes for Colliders,
pp. 419–476. 2015. arXiv:1411.7538 [hep-ph].
[3] P. Vanhove, “Feynman integrals, toric geometry and mirror symmetry,” in KMPB Conference: Elliptic
Integrals, Elliptic Functions and Modular Forms in Quantum Field Theory, pp. 415–458. 2019.
arXiv:1807.11466 [hep-th].
[4] C. Rella, “An Introduction to Motivic Feynman Integrals,” SIGMA 17 (2021) 032, arXiv:2009.00426
[math-ph].
[5] F. Brown, “Feynman amplitudes, coaction principle, and cosmic Galois group,” Commun. Num. Theor.
Phys. 11 (2017) 453–556, arXiv:1512.06409 [math-ph].

[6] F. Loebbert, “Integrability for Feynman Integrals,” 12, 2022. arXiv:2212.09636 [hep-th].
[7] R. P. Klausen, Hypergeometric feynman integrals. PhD thesis, Mainz U., 2023. arXiv:2302.13184
[hep-th].
[8] A. Morozov, “Integrability and matrix models,” Phys. Usp. 37 (1994) 1–55, arXiv:hep-th/9303139.

[9] A. Y. Morozov, “String theory: What is it?,” Soviet Physics Uspekhi 35 no. 8, (1992) 671.
[10] A. Mironov, “2-d gravity and matrix models. 1. 2-d gravity,” Int. J. Mod. Phys. A 9 (1994) 4355–4406,
arXiv:hep-th/9312212.
[11] A. Connes and D. Kreimer, “Renormalization in quantum field theory and the Riemann-Hilbert
problem,” JHEP 09 (1999) 024, arXiv:hep-th/9909126.
[12] A. Gerasimov, A. Morozov, and K. Selivanov, “Bogolyubov’s recursion and integrability of effective
actions,” Int. J. Mod. Phys. A 16 (2001) 1531–1558, arXiv:hep-th/0005053.
[13] P. Lairez and P. Vanhove, “Algorithms for minimal Picard–Fuchs operators of Feynman integrals,” Lett.
Math. Phys. 113 no. 2, (2023) 37, arXiv:2209.10962 [hep-th].
[14] S. Müller-Stach, S. Weinzierl, and R. Zayadeh, “Picard-Fuchs equations for Feynman integrals,”
Commun. Math. Phys. 326 (2014) 237–249, arXiv:1212.4389 [hep-ph].
[15] A. V. Kotikov, “Differential Equations and Feynman Integrals,” in Antidifferentiation and the Calculation
of Feynman Amplitudes. 2, 2021. arXiv:2102.07424 [hep-ph].

[16] V. Mishnyakov and P. Suprun, “Outlook on Differential Equations for Feynman Integrals (Brief
Review),” JETP Lett. 115 no. 8, (2022) 477–483.

12
[17] K. Bönisch, C. Duhr, F. Fischbach, A. Klemm, and C. Nega, “Feynman integrals in dimensional
regularization and extensions of Calabi-Yau motives,” JHEP 09 (2022) 156, arXiv:2108.05310
[hep-th].
[18] V. Mishnyakov, A. Morozov, and P. Suprun, “Position space equations for banana Feynman diagrams,”
Nucl. Phys. B 992 (2023) 116245, arXiv:2303.08851 [hep-th].
[19] J. van der Hoeven, “Around the numeric–symbolic computation of differential Galois groups,” Journal of
Symbolic Computation 42 no. 1, (2007) 236–264. Effective Methods in Algebraic Geometry (MEGA 2005).

[20] G. Pólya, “On the mean-value theorem corresponding to a given linear homogeneous differential
equation,” Transactions of the American Mathematical Society 24 no. 4, (1922) 312–324.
[21] P. Etingof, I. M. Gelfand, and V. Retakh, “Factorization of differential operators, quasideterminants, and
nonabelian toda field equations,” Mathematical Research Letters 4 (1997) 413–425,
arXiv:q-alg/97010082.
[22] M. Bogner, “Algebraic characterization of differential operators of Calabi-Yau type,” arXiv:1304.5434.
[23] G. Almkvist, “The art of finding Calabi-Yau differential equations. Dedicated to the 90-th birthday of
Lars Gärding,” arXiv:0902.4786.

[24] G. Almkvist, M. Bogner, and J. Guillera, “About a class of Calabi-Yau differential equations,”
arXiv:1310.6658.
[25] D. van Straten, “Calabi-Yau Operators,” arXiv:1704.00164.
[26] V. Mishnyakov, A. Morozov, M. Reva, and P. Suprun, To appear.

[27] L. Adams, C. Bogner, and S. Weinzierl, “The sunrise integral around two and four space-time dimensions
in terms of elliptic polylogarithms,” Acta Phys. Polon. B 46 no. 11, (2015) 2131, arXiv:1510.02048
[hep-ph].

13