Banana

A three-loop integral built on a K3 surface — run as a validation against the published answer, and solved at leading order by a two-term formula whose only constant is $7\zeta_3$.

The integral

The object is the family of three-loop two-point integrals

$$I_{\nu_1\nu_2\nu_3\nu_4} \;=\; \int \frac{d^d l_1\, d^d l_2\, d^d l_3}{\big(l_1^2-m^2\big)^{\nu_1}\,\big(l_2^2-m^2\big)^{\nu_2}\,\big(l_3^2-m^2\big)^{\nu_3}\,\big((l_1+l_2+l_3-p)^2-m^2\big)^{\nu_4}}\,, \qquad d = 2-2\varepsilon,$$

up to an overall normalization convention: four internal lines, all carrying the same mass $m$, strung between two vertices and bowing out like the segments of a banana. A single external momentum $p$ enters at one vertex and leaves at the other, so the integral depends on the one dimensionless ratio $t = p^2/m^2$ (here $m^2 = 1$), with $\varepsilon$ the dimensional regulator. Integration-by-parts reduction collapses the whole family onto four master integralsthe finite basis of independent integrals to which every integral of the family reduces via integration-by-parts identities. The object solved exactly here is the $\varepsilon^0$ term of the top master $m_1 = I_{1111}$ — one power of each of the four propagators — over the full disk $|t| < 4$ around the origin (the Euclidean window $-7/2 \le t < 0$ and the subthreshold window $0 < t \le 7/2$); the deeper $\varepsilon$ layers of all four masters are carried numerically by certified differential-equation transport. The same object is continued past the disk along the upper-half-plane path $t\to t+i0$: it is real between $t=4$ and $t=16$, matching the record at four points to at least 99 digits, and complex above 16, matching the record with the causal prescription at five points to at least 99 digits in the imaginary part as well; two fresh points, $t=10$ and $t=20$, predicted before they were computed, agree to 89 digits against the reference computed at a precision goal of 60 digits and to 58 against the one at goal 40 (the goal-40/goal-60 reference pairs themselves floor at 41 digits; --continue runs that tier).

At three loops the maximal cut — the integral with all four propagators put on shell — leaves elliptic territory: it is the holomorphic period of a one-parameter family of K3a K3 surface is a two-complex-dimensional Calabi-Yau manifold — the next step up from an elliptic curve surfaces, the first rung of the Calabi-Yau ladder. The four- and five-loop bananas climb higher on the same ladder.

The K3 in question is the Sym²the symmetric square: the geometry whose periods are the pairwise products of the periods of the underlying curve of the equal-mass sunrise's Γ₁(6)the congruence subgroup of SL(2,ℤ) that fixes the sunrise elliptic curve; level 6 sets which modular forms can appear elliptic curve, with Picard rank 19. Because it is a symmetric square, the Picard-Fuchs operator — the ordinary differential equation in $t$ satisfied by the periods of the K3 family — has order three rather than four, and the holomorphic period ϖ₀"varpi-zero", the holomorphic K3 period — the unique power-series solution at the MUM point, and the natural overall normalization of the integral is literally the sunrise period squared. Its Taylor coefficients are the Domb sequence1, 4, 28, 256, 2716, … (OEIS A002895): the integer sequence that generates the K3 banana period, originally found by Domb in lattice statistical mechanics. The Landau singularities sit at $t = 0, 4, 16, \infty$: the MUM cuspthe point of maximal unipotent monodromy, where all three Frobenius solutions degenerate into one power series plus logarithms — the natural expansion point for Calabi-Yau periods, the two-particle pseudo-threshold at $(2m)^2$, the four-particle normal threshold at $(4m)^2$, and the large-momentum point.

The integral itself is known (Bönisch-Fischbach-Klemm-Nega-Safari arXiv:2008.10574; Pögel-Wang-Weinzierl arXiv:2207.12893). It is the Calabi-Yau validation case of this catalog: the place to show the bootstrap handles CY geometry against their published answer before trying an unsolved one.

Why it matters

The equal-mass bananas are the canonical route into Calabi-Yau geometry for Feynman integrals: the two-loop sunrise sits on an elliptic curve, the three-loop banana's maximal cut is a K3 period, and each further loop climbs one complex dimension higher. Bönisch, Fischbach, Klemm, Nega and Safari mapped the analytic structure of the all-loop banana amplitudes, and Pögel, Wang and Weinzierl brought the three-loop case into ε-factorised form built from meromorphic modular forms, so this rung has a published answer to calibrate against. Most existing elliptic technique was built one rung lower, on the sunrise; the banana is where those techniques either generalize or break, and the rung from which the four- and five-loop Calabi-Yau climbs start.

Its particular K3 is a lucky one. Being the symmetric square of the sunrise curve means the modular machinery already certified for Γ₁(6) lifts almost verbatim: the same modular parameter, an alphabet that is the weight-doubled image of the sunrise's, periods that are products of objects already pinned down for the sunrise. A bootstrap that can do the sunrise has most of the ingredients for the banana already in hand.

What was hard

The maximal cut lives on a K3 surface, one complex dimension above the elliptic curves on which most multi-loop technology was built, so no ready-made class of special functions covers the answer — the modular toolkit had to be lifted from the sunrise one rung below. Second, a blind fit of sampled values against a graded basis of candidate functions floors at a few digits, because the holomorphic-period prefactor $\varpi_0$ effectively mixes the periods; the fit closed only after the finite basis of iterated-integral words was fixed by the differential equation — an ε-formthe connection matrix is strictly proportional to ε, so each order in ε becomes a finite iterated integral over the letters of the connection system over the six-letter Γ₁(6) modular alphabet $\{1, f_{2a}, f_{2b}, f_{4a}, f_{4b}, f_6\}$ of modular weights $(0, 2, 2, 4, 4, 6)$, solved by carrying the K3 periods through the tower of logarithmic solutions at the MUM point, a transport built for this integral. Third, one row of that differential equation initially treated a momentum-dependent numerator as constant; the error was caught by finite-difference checks against independent evaluations and the row re-derived before the system was accepted.

The result

Throughout, $t = p^2/m^2$ is the single kinematic ratio, $x = t/64$ is the toric variable in which the period series has integer coefficients, and $m_1^{(\varepsilon^0)}(t)$ is the top master $I_{1111}$ at order $\varepsilon^0$.

The answer at leading order is the two-term formula

$$m_1^{(\varepsilon^0)}(t) \;=\; 7\zeta_3\,\varpi_0(t)\;-\;\mathrm{Part}_{\rm reg}(t),$$

valid on the full disk $|t| < 4$. All four coefficients in the MUM basis $\{\varpi_0,\ \varpi_1^R,\ \varpi_2^R,\ \mathrm{Part}_{\rm reg}\}$ — the holomorphic period, the two regularized logarithmic periods, and the regularized particular solution — are returned by integer-relation search as exactly $(7\zeta_3,\ 0,\ 0,\ -1)$: the two log-period coefficients vanish, so only the holomorphic period and the particular solution survive. The constant 7ζ₃Apéry's constant, ζ(3) = 1.2020569…, the value of the Riemann zeta function at 3 and the most common weight-three transcendental in loop calculations is Broadhurst's $d = 2$ three-loop banana value at $p^2 = 0$.

Both pieces are explicit power series in the toric variable $x = t/64$. The first is the holomorphic K3 period, the Domb series

$$\varpi_0(x) = \sum_{n\ge 0} D_n\,x^n = 1 + 4x + 28x^2 + 256x^3 + 2716x^4 + \cdots$$

with integer coefficients fixed by the three-term recursion $(n+1)^3 D_{n+1} = 2(2n+1)(5n^2+5n+2)\,D_n - 64\,n^3 D_{n-1}$, $D_0 = 1$. The second, $\mathrm{Part}_{\rm reg}$, is the unique holomorphic particular solution of the same order-three Picard-Fuchs operator driven by the sub-banana source $24/(t^2(t-4)(t-16))$ — the inhomogeneity fed in by the sub-diagram obtained when one of the four lines shrinks away:

$$\mathrm{Part}_{\rm reg}(t) = \sum_{k\ge 1} b_k\,x^k, \qquad b_1 = 24, \qquad n^3 b_n = 2(2n-1)(5n^2-5n+2)\,b_{n-1} - 64\,(n-1)^3 b_{n-2},$$

so the series begins $0,\ 24,\ 216,\ \tfrac{18944}{9},\ \tfrac{204440}{9},\ \dots$ Both recursions, with reference values, live in the expression file below, and the standalone script evaluates the closed form at any point of the disk to any requested precision.

Downloads: banana-expression.md · banana-evaluate.py · MANIFEST.sha256 — the standalone evaluator of the closed form (Python with mpmath): the default run evaluates the three points $t=-3,\,-7/2,\,1/2$, none of them used in the fit, and agrees with the independent reference values to at least 110 digits at each point (119 at $t=-3$), in seconds, then repeats one point at twice the precision; --heldout18 checks all eighteen reference values of the record; --point and --dps take any point of the disk and any precision; --eps1 evaluates the $\varepsilon^1$ coefficient of the top master from the served graded system by the Frobenius recursion, checked against the eighteen reference strings of that order (at least 109 digits at every point), with the word list it reads in banana-eps1-words.json; --bessel-eps K recomputes the layers $\varepsilon^0,\ldots,\varepsilon^K$ of the top master at a Euclidean point directly from the one-dimensional Bessel representation of the four propagators, exact in $d$ and independent of the differential system, and at $t=-8$ reproduces the stored $\varepsilon^0,\ldots,\varepsilon^5$ reference values to 41 digits (reading banana-minkowski-gates.json).

First, the normalization: in the Pögel-Wang-Weinzierl canonical convention $I_2 = \varepsilon^3(\pi^2/\varpi_0)\,m_1$ the MUM constant is $\tfrac{4}{3}\zeta_3$; the two constants belong to different prefactor conventions and are not in conflict. Second, the deeper layers: at order $\varepsilon^1$ the letters $f_{2a}, f_{2b}$ switch on beside $1$ and $f_{4a}$ ($f_{4b}$ first enters at $\varepsilon^2$ and $f_6$ at $\varepsilon^3$). Because every master is analytic at $p^2=0$, the words collapse in $t$-space to $m_1^{(\varepsilon^1)}(t) = \alpha_1\,\varpi_0(t) + \mathrm{Part}^{(1)}(t)$, with $\mathrm{Part}^{(1)}$ the holomorphic particular solution of the same graded recursion and $\alpha_1$ a single boundary constant — the $\varepsilon^1$ coefficient of the $p^2=0$ vacuum banana, evaluated by quadrature, for which no closed form is established. With $\alpha_1$ so fixed and no fitted parameter, the layer reproduces the independent reference values at all eighteen points to at least 109 digits. The orders beyond $\varepsilon^1$ are carried numerically by the certified differential-equation transport of all four masters. The candidate transcendental for those layers is a classical complex-multiplication period: the equal-mass sunrise curve has CM by $\mathbb{Q}(\sqrt{-15})$ with weight-3 level-15 newform $f_3$ (Bönisch-Fischbach-Klemm-Nega-Safari), and its critical value $L(f_3,2)$ has a parameter-free Chowla-Selberg closed form as a product of $\Gamma$-values at fifteenths, $L(f_3,2) = \Gamma(\tfrac{1}{15})\Gamma(\tfrac{2}{15})\Gamma(\tfrac{4}{15})\Gamma(\tfrac{8}{15})\big/(120\sqrt{3}\,\pi)$. Its closed form matches an independent evaluation to 70 digits; the identification of the banana's own deeper-layer constant against it, in the integral's normalization, is pending (the arithmetic-boundary derivation route is discussed in the sunrise write-up).

The four-loop rung (Calabi–Yau)

The next member of the same family is the four-loop equal-mass banana, $I_{11111}$: its maximal cut is a period of a Calabi–Yau threefold, annihilated by an order-four Picard–Fuchs operator (AESZ-34) — the next complex dimension above the three-loop K3. The bootstrap delivers this rung in explicit symbolic form. With $z=-1/p^2$ and $L=\log z$, the five-master system closes as an exact first-order $\theta$-system with integer polynomial data and denominator $20(1+z)(1+9z)(1+25z)$, regular at the MUM pointthe family's most degenerate point, where the period expansion is anchored $z=0$ with a nilpotent residue $C_0$ of index 8, and its full graded solution is the explicit log-Frobenius series

$$F(z) \;=\; \sum_{n\ge0} z^n \sum_{j=0}^{7} F_{n,j}\, L^j, \qquad F_0(L) = e^{C_0 L}\, v,$$

with every coefficient $F_{n,j}$ generated by an exact integer recursion from the boundary vector $v$. That boundary vector is written out completely in the $\zeta$-ring — the tadpole block is $\Gamma(1+\varepsilon)^4 e^{4\gamma_E\varepsilon}$ exactly, and the three dotted top integrals carry vectors such as $\big(-\tfrac32,\,-1,\,-3\zeta_2,\,-2\zeta_2+2\zeta_3,\,-60+\tfrac{52}{3}\zeta_3-9\zeta_4\big)$ across $\varepsilon^{-4}\ldots\varepsilon^{0}$ — and the top master $I_{11111}$ carries no free boundary datum at all: its entire Laurent window is generated by the tadpole tower and the sub-master constants through the $C_0$ log-chains. The classical large-momentum boundary formula of Bönisch–Duhr–Fischbach–Klemm–Nega and Pögel–Wang–Weinzierl is imposed as one of the three conditions from which the boundary vector is derived, so this rung takes that published structure as input; the lowest log-tower then reproduces the formula layer by layer at two independent values of $L$ to full working precision, about 115 digits; the explicit form was additionally checked against two references that share no machinery with the integration-by-parts differential system — independent auxiliary-mass-flow evaluations (agreement to their own length, about 67 digits) and a Bessel-moment representation from the literature evaluated live by quadrature (116 digits) — and the check is stringent: perturbing one entry of the differential system by $10^{-6}$ drops the agreement from 67 digits to 3. An earlier purely numerical transport of the same system agreed three ways at 67 digits; the explicit form supersedes it.

Everything on this page was validated against independent evaluations at points never used in the construction, to at least 67 digits.

Downloads: banana-4loop-evaluate.py — the standalone evaluator of the explicit form (python3 + mpmath, no data files): the default run computes the full five-master window at $p^2 = -2$ to thirty digits in about a minute, certifying its own truncation and checking every layer against the independent references; --full recomputes the paper's table of values at points never used in the construction, at one hundred digits (--bessel adds the live Bessel-oracle rows) · reduced_statement.py — a stand-alone exact check of the $\varepsilon^0$ term of the top master: it reads the data block of the evaluator beside it, runs the log-Frobenius recursion in exact $\zeta$-ring arithmetic, and verifies that the AESZ-34 Picard–Fuchs operator applied to $(-p^2)\,I^{(0)}_{11111}$ gives $-5! = -120$ exactly, with the boundary log-tower $(0,\,80\zeta_3,\,0,\,0,\,-5)$ in powers of $L$ (python3 reduced_statement.py --reduced-statement, a few seconds). Change log: CHANGES.md.

References

← back to the gallery