When only two gravitons are involved, withπpropagators between them (diagram(a) in Section3.3.3), the associated modular graph functions read:
π·π(π) βΆ=
β«ξ±ππΊ1(π, π)πππ π2.
5See [33] and [94].
6Conjecture 1 is equivalent to Conjecture 2 in this case because Brown proved [21] that periods of mixed Tate motives overβ€are (up to invertingπ) MZVs.
The behaviour of this integral as π2 tends to infinity was already studied in [48].
Using the same ideas we generalize that result giving the following expansion of the functionsπ·πβs, which is of course a special case of Theorem1.4.1:
Theorem 4.2.1. For everyπβ₯2
π·π(π) = β
π,πβ₯0
π(π,π)
π (ππ2)ππππ, (4.7)
where for everyπ, π β₯0
ππ(π,π)(ππ2) =
2πβ1β
π=0
π(π,π)π (ππ2)πβπ
is a Laurent polynomial with coefficientsπ(π,π)π belonging to the algebraξ― of conical sums.
Proof. Using (3.20) and doing the change of variables π₯ = π2βπ2, we want to compute
π·π(π) = 1 2π
β
π=π+π
π!
π!π!
(ππ2)π
2π β«ξ±πB2(π₯)ππ(π, π)πππ1ππ₯,
where we recall thatξ±πdenotes the complex torus associated with the latticeπβ€+β€.
Since, by the definition (3.21), π(π§, π)π = β
πββ€π πββ€π0
1
|π|π2ππβπππ(πππ1+π1)πβ2ππ2βπ|ππ||ππβπ₯|, (4.8)
whereβ€0βΆ=β€β§΅{0}and|π|βΆ=|π1|β―|ππ|, we have
β«βββ€π(π, π)πππ1= β
πββ€π πββ€π0
πΏ0(π)
|π| π2ππ(πβ π)π1πβ2ππ2(β|ππ||ππβπ₯|),
where, for anyπ£ββ€π and for anyπββ€,πΏπ(π£) = 1if the sum of the coordinates isπ, and is zero otherwise. Moreover,π₯β π¦denotes the standard inner product ofβπ. In the interval[0,1],
B2(π₯)π= β
π+π+π=π
π!
π!π!π!
(β1)π 6π π₯2π+π. Therefore we have
π·π(π) = 1 4π
β
π+π+π+π=π
π!(2ππ2)πβπ π!π!π!π!
(β1)π 6π
β
πββ€π πββ€π0
πΏ0(π)
|π| π2ππ(πβ π)π1
β«
1 0
π₯2π+ππβ2ππ2(β|ππ||ππβπ₯|)ππ₯. (4.9)
To compute the last integral, let us fixπ1,β¦, ππ. Sinceπ₯β [0,1]we have
|ππβπ₯|=
{ ππβπ₯ ifππ>0 π₯βππ ifππβ€0, soβ|ππ||ππβπ₯|=β|ππ||ππ|+π₯β
sgn(βππ)|ππ|.
By repeated integration by parts one easily finds that for anyπββ>0,π½ ββ⧡{0}
andπ ββ
β«
π
0
π₯ππβπ½π₯ππ₯= π! π½π+1 β
βπ π=0
(π)π
π½π+1ππβππβπ½π, (4.10) where(π)π = π(π β 1)β―(π βπ + 1)is the descending Pochhammer symbol. In our caseπ½ = 2ππ2β
sgn(βππ)|ππ|,π= 1andπ = 2π+π. Note thatβ
sgn(βππ)|ππ|can be equal to zero. If it is not zero, since|ππ|+sgn(βππ) =|ππβ 1|, by (4.10) we get, as a result of the integral,
(2π+π)!πβ2ππ2β|ππ||ππ| (2ππ2β
sgn(βππ)|ππ|)2π+π+1 β
2π+πβ
π=0
(2π+π)!πβ2ππ2β|ππ||ππβ1| (2π+πβπ)!(2ππ2β
sgn(βππ)|ππ|)π+1, while ifβ
sgn(βππ)|ππ|= 0then we just get πβ2ππ2β|ππ||ππ|
2π+π+ 1 .
In both cases, once we fixπ1,β¦, ππ andπ1,β¦, ππ, putting everything together we are left with an expression of the kind
π(π,π)(ππ2)π2ππππ1πβ2πππ2, where π = β
ππππ β β€, π is a non-negative integer7 equal to either β|ππ||ππβ 1| or β
|ππ||ππ|, andπ(π,π)(ππ2) is a Laurent polynomial with rational (explicitly deter-mined) coefficients whose maximum power isπand minimum power is1 βπ.8
Note thatπ2ππππ1πβ2πππ2 =ππππ, withπ = (π+π)β2andπ = (πβπ)β2. Therefore we would like to show thatπ β₯ |π|and thatπ β‘2 π (this notation is a shorthand for the standardπβ‘πmod 2), in order to have thatπandπare non-negative integers. Ifπ =
β|ππ||ππ|the claim is trivial, so we have to take care only of the caseπ =β|ππ||ππβ1|. Note thatππππβ‘2|ππ||ππ|β‘2|ππ|(1 +|ππβ 1|)and thatβ
|ππ|β‘2
βππ= 0, so π=β
ππππβ‘2
β|ππ|+β
|ππ||ππβ 1|β‘2π.
Moreover|π|=|β
ππππ|+|β
ππ|β€|β
ππ(ππβ 1)|β€β|ππ||ππβ 1|=π.
To conclude our proof we have to analyse more carefully the rational coefficients ofπΜ(π,π)(ππ2), which are obtained by theπ(π,π)(ππ2)βs.
Any fixedπ = β|ππ||ππβπ|(π= 0,1) can be obtained with just finitely many π-tuples(π1,β¦, ππ), because|ππβπ|β€π for anyπ: otherwise, since for everyπ|ππ|β₯1, we would haveβ
|ππ||ππβπ|β₯ β
|ππβπ| β₯π. This means that for any(π, π), which is uniquely determined by a couple(π, π), one has to consider a finite rational linear combination of sums of the kind
π(π, π, πΌ) βΆ=β
π
πΏ0(π)
|π|(β
sgn(βππ)|ππ|)πΌ, whereπββ€π0 are such thatβ
sgn(βππ)|ππ|β 0andπΌis a positive integer. Note that the functionπ only depends onππβ|ππ|.
7This notation may generate some confusion, becauseπalso denotes e(π). What we mean will always be clear from the context.
8Whenπ = 0andπ = 1one can see that there is no contribution to the negative powers, and therefore one can assume thatπβ₯2to get this lower bound.
Since we can split the sum definingπ as a sum over cones such that the factors in the denominator are either bigger or smaller than zero, it follows that our coefficients are linear combinations of conical sums.
β‘
Specializing carefully this computation to the (π, π) = (0,0) case, one gets the proof, first obtained in [48], of Proposition3.3.2:
Proof of prop.3.3.2.Note that, sinceπ =β|ππ||ππβπ|withπ= 0or= 1, the only possibleπ-tuples(π1,β¦, ππ)which can giveπ= 0are(0,β¦,0)and(1,β¦ 1), and with them alsoπ= 0, becauseβ
ππ= 0. Using this and looking carefully9at the proof of the previous theorem one is then lead to the formula given in the statement, except for the hypergeometric coefficient of the leading term, which is more elegant than the term obtained with the integration process described in the proof, and can be deduced by making use of the identity
1
π₯(π₯+ 1)β―(π₯+π) =
βπ π=0
(β1)π π!(πβπ)!
1 π₯+π, easily obtained by partial fraction decomposition.
β‘
As remarked in the previous chapter, the functionπ(π, πΌ)was proven by Zagier in [83] to be equal to an explicit linear combination of MZVs, allowing to algorith-mically compute in terms of MZVs the non-exponentially small part ofπ·π:
Proposition 4.2.1(Zagier).
π(π, πΌ) =π!β
πβ₯0
β
π1,β¦,ππβ{1,2}
π1+β―+ππ=π+2
22(π+1)βπβπΌπ(π1,β¦, ππ, πΌ+ 2). (4.11)
We will not repeat the proof here, because it is essentially contained in the more complicated proof of Theorem4.3.2, that can be found in AppendixA.
Unfortunately, knowing the Laurent polynomialππ =ππ(0,0)is not enough to per-form the integration over the moduli space of complex tori, so one would like to un-derstand better the behaviour of the functionsπ·π. This can be achieved by looking at the more general Theorem4.2.1, because it allows us to predict other coefficients of the expansion (4.7).
To make an example, let us recall that the non-holomorphic Eisenstein series are defined, forπ ββwithβ(π )>1andπ ββ, by
πΈ(π , π) = (π2
π
)π β
(π,π)ββ€2⧡{(0,0)}
1
|ππ+π|2π . (4.12) Its expansion at the cusp is given, settingπ¦=ππ2, by
πΈ(π, π) = (β1)πβ1 B2π
(2π)!(4π¦)π+ 4(2πβ 3)!
(πβ 2)!(πβ 1)!π(2πβ 1)(4π¦)1βπ
+ 2
(πβ 1)!
β
πβ₯1
ππβ1π1β2π(π)(ππ+ππ)
βπβ1 π=0
(π+πβ 1)!
π!(πβπβ 1)!(4π π¦)βπ, (4.13)
9As explained in [48], one exploits the fact thatπ΅2(1 βπ₯) = π΅2(π₯) in order to get the nice look-ing formula in the corollary instead of the more complicated one that we would naively get just by performing the same steps as in the theoremβs proof.
whereBπ is the π-th Bernoulli number, andππ(π) = β
π|πππ is a finite power sum running over the positive divisors ofπ.
As we have mentioned in the last chapter, using the series representations (3.27) forπ·π(π) and (4.12) for the Eisenstein series one can immediately see that π·2(π) = πΈ(2, π)β16.
Let us briefly see how can we get the same result by comparing the expansion given by Theorem4.2.1and the expansion (4.13), which becomes in this case
πΈ(2, π) = π¦2
45+ π(3)
π¦ + 2β
πβ₯1
(2π+π¦β1)πβ3(π)(ππ +ππ).
Using the intermediate step (4.9) in the proof of the theorem, for π = 0 we get (1β16)(π¦2β45), which is the leading term of the non-exponential part, for π = 1we get zero, and forπ= 2(soπ=π=π= 0) we get
1 16
β
πββ€β§΅{0}
πββ€
π2ππππ1πβ2ππ2(|π|(|π1|+|π2|))
|π|2 β«
1 0
πβ2ππ2π₯|π|(sgn(βπ1)+sgn(βπ2))ππ₯,
where we denoteπ = π(π1 βπ2). When π = 0, i.e. when π andπ have the same power in the expansion given by the theorem, thenπ1 = π2, which implies that the argument of the exponential in the integral is never zero. Therefore, splitting the sum into theπ β₯ 1 part and theπ β€ 0 part, one gets two telescoping sums, both giving as a result
β
πββ€β§΅{0}
1 64|π|3π¦.
We conclude that the variableπ1does not appear only in the non-exponentially small part ofπ·2(π), which is the Laurent polynomialπ2(π¦) = (π¦2β45 +π(3)βπ¦)β16. This fits with the expansion ofπΈ(2, π).
Moreover, ifπ > 0one gets again telescoping sums inπ, but now we sum only over finitely manyπ, which are the divisors ofπ. We leave as an exercise to the reader to verify that one gets exactly the same expansion as we get for the non-holomorphic Eisenstein series.
In general it is too messy to repeat the same game as above and explicitly get the full expansion for otherπ·πβs, except maybe forπ·3(π), which is already known, as we have mentioned in eq. (3.31), to be equal to(πΈ(3, π) +π(3))β64. However, it is in principle possible to algorithmically get the coefficient ofππππfor any fixed(π, π), which would allow to check conjectures on the full expansion or to numerically ap-proximate the functions very precisely (the sum (4.7) converges much faster than the sum (3.27)).