Feynman diagrams (quantum field theory)
Thirty frontier Feynman integrals computed in closed form. Fifteen of these were already known, at least at the level of the $\varepsilon$ form, and are independently reproduced end-to-end with the BootLoops toolkit. Fifteen are new. Each integral has its own page and code to evaluate the integral to the BootLoops standard.
The 30 diagrams. Click each for details.
The BootLoops standard
An integral counts as computed here only when it meets three requirements, together called the BootLoops standard. First, an explicit closed form in a declared function class, in which every constant is a classical one or is defined by a convergent integral or series written out with the result. Second, agreement to thirty digits or more with independent numerical values at stated kinematic points that were withheld from the derivation, entering no fit and no boundary determination, with the comparison repeated at two working precisions. Third, a stand-alone evaluator, containing no reduction or auxiliary-mass-flow code, that anyone can run at any point and precision and that reproduces those checks. What counts as a closed form depends on the function class:
- Polylogarithmic: a combination, with rational coefficients, of multiple polylogarithms of stated weight whose arguments come from a declared alphabet of letters, the algebraic functions of the kinematics that mark every possible branch point.
- Elliptic: on a stated elliptic curve, iterated integrals of modular forms, elliptic polylogarithms, or iterated integrals of the curve’s two periods times powers of the first, with coefficients rational or in a stated ring of constants.
- K3: the order-three Picard–Fuchs operator of the stated K3 surface, its Frobenius basis of period solutions and their iterated primitives, the inhomogeneous terms from sub-diagrams, and a boundary vector of exactly identified constants.
- Calabi–Yau: the same data for the order-four operator of the stated threefold, expanded about its point of maximal unipotent monodromy, with the solution transported along a stated path to any kinematic point.
The problem
Feynman diagrams are the line drawings physicists use to compute what happens when particles collide. Each diagram stands for a definite integral over the momenta of the particles circulating inside it, whose value is the diagram’s contribution to the prediction. Diagrams with loops carry the quantum corrections: each added loop makes the prediction more precise and the integral much harder, and at the Large Hadron Collider, where new physics is sought as a discrepancy between a precise measurement and an equally precise prediction, loop integrals are the limiting step on the theory side.
In the conventions used here, a family of integrals with $L$ loops and $n$ internal lines is
$$ I_{\nu_1\cdots\nu_n}\big(d;\{s_{ij}\},\{m_i^2\}\big) \;=\; \int\prod_{l=1}^{L}\frac{d^dk_l}{i\pi^{d/2}}\;\prod_{j=1}^{n}\frac{1}{D_j^{\,\nu_j}}\,, \qquad D_j=q_j^2-m_j^2+i0\,. $$
Here the $k_l$ are the loop momenta in $d$ dimensions, $D_j$ is the propagator of the $j$-th internal line with momentum $q_j$ and mass $m_j$ (the $+i0$ fixes on which side of its pole each propagator is evaluated), the invariants $s_{ij}=(p_i+p_j)^2$ of the external momenta are the kinematic variables, and the integer powers $\nu_j$ select one member of the family. The integrals diverge in four dimensions, so they are computed in $d=4-2\varepsilon$ and expanded in $\varepsilon$; each result below is one coefficient of that expansion, usually the finite $\varepsilon^0$ term, as an explicit function of the masses and invariants.
How hard such an integral is depends on the class of functions its answer needs, and that class can be read off the diagram before anything is computed. Putting every internal line on its mass shellA particle is on its mass shell when its energy and momentum obey the relation $E^2=p^2+m^2$ of a real, observable particle; particles inside a loop need not. gives the maximal cut of the integral; the components of the loop momenta that those conditions do not fix sweep out a geometric object, and its shape determines the function class. When it is a sphere, the answer is built from polylogarithms, the functions obtained from the logarithm by repeated integration. Make enough of the internal lines massive and the sphere becomes a torus, that is, an elliptic curve, whose own functions the answer then needs: its periodsThe two numbers obtained by integrating the curve’s natural differential once around each of the two independent cycles of the torus; every function attached to the curve is built from them., elliptic generalizations of the polylogarithm, and iterated integrals of modular formsFunctions of the shape of the torus that transform in a prescribed way when the same torus is described by a different pair of cycles; their integrals supply the constants that elliptic Feynman integrals contain.. The standard example is the two-loop sunrise, one particle splitting into three massive particles that recombine, solved several times in different function languages. Adding loops to the sunrise gives the banana graphs: the torus gives way to a K3 surface at three loops and a Calabi–Yau threefold at four, for which still less of the calculus exists.
The one-loop problem was solved as a class: in 1979 Passarino and Veltman reduced every one-loop amplitude to a few standard integrals, all logarithms and dilogarithms. Beyond one loop, the integration-by-parts identities of Chetyrkin and Tkachov (1981) relate the integrals of a family, and Laporta’s algorithm (2000) solves them for a short list of master integralsA finite set of integrals of the family in terms of which every other member can be written with rational coefficients; computing the family means computing these.. These satisfy differential equations in the kinematic variables, so computing a family becomes solving that system from a point where the integrals are known (Gehrmann and Remiddi, 2000). The symbol of Goncharov, Spradlin, Vergu and Volovich (2010) encodes a polylogarithmic function as the sequence of logarithmic arguments it is built from, and Henn (2013) observed that a good choice of master integrals factors $\varepsilon$ out of the differential equations (the canonical form), which are then solved order by order by linear algebra plus a boundary value. Numerically, the auxiliary-mass-flow method of Liu and Ma (2022) gives any master integral to hundreds of digits at a chosen point, though not its analytic form.
Beyond the polylogarithms there is no general algorithm that brings the differential equations to canonical form, only constructions found family by family, and no finite set of logarithmic arguments from which to build a symbol. The reduction and numerical programs run the same way whatever the function class, and give no hint of which class it is. Sabry wrote down the two-loop sunrise in 1962; Laporta and Remiddi evaluated it analytically on its elliptic curve only in 2005. In 2023 Pögel, Wang and Weinzierl brought the differential equation of the four-loop banana, a Calabi–Yau integral, to $\varepsilon$-factorized form. Even the location of an integral’s singularities, governed by the Landau equations of 1959, became computable from the graph alone only with the Landau discriminants of Mizera and Telen (2021) and the programs PLD.jl (2023) and SOFIA (2025).

Figure 1. Landmarks in the computation of multi-loop Feynman integrals, 1947–2026, each drawn as a small diagram at its year. Black diagrams are polylogarithmic: the one-loop vertex behind the electron’s anomalous magnetic moment (1947) and its two-loop successor (1957, below the axis), the massless double box (1999), the hexagon that stands for the symbol (2010) and the two-loop five-point integrals (2018). Red doubled lines are massive: the sunrise and banana graphs, the same two-point diagram with one more massive line per loop, evaluated on an elliptic curve in 2005, a K3 surface in 2015 and a Calabi–Yau threefold in 2023. The gray arrow marks the four decades between the two-loop vertex and the massless double box. The 2026 mark stands for the thirty integrals summarized here, with the ice-cream cone drawn for them. (Drawn for this page.)
The word bootstrap here names a strategy: instead of integrating, write down every function the answer could be built from, then cut that set down with everything known about the answer’s analytic structure until one candidate is left. The strategy is mature for polylogarithmic amplitudes, and Morales, Spiering, Wilhelm, Yang and Zhang carried it to elliptic integrals at the level of the symbol in 2022. The two methods extended here both stopped at the elliptic curve, where there was no enumerable function space to fit: a Landau bootstrap, which constrains an integral from its behavior near its singularities, and analytic regression, which recognizes an integral’s exact coefficients from samples accurate to hundreds of digits. Meanwhile several integrals that collider predictions depend on had no closed form: the non-planar two-loop integrals for W-pair production ($q\bar q \to W^+W^-$) with the top-quark mass kept exact, the elliptic master integrals for $gg \to Z\gamma$, and the two-loop five-point integrals for a pair of vector bosons produced with a jet.
The approach
BootLoops computes each integral by one of two routes. The first, constrain-and-fit, starts from the function space that the geometry of the maximal cut allows. Exact constraints, namely where the answer may be singular, how it behaves there and which symmetries it has, remove almost every candidate, and the few remaining constants are fitted to high-precision numerical values of the integral and recognized as exact numbers by integer-relation (PSLQ) searches. The second, reduce-and-transport, reduces the family to master integrals, constructs their differential equation and solves it along a path from a point where the integrals are known exactly. Both routes rest on the differential equation in Henn’s canonical form,
$$ d\,\vec g \;=\; \varepsilon\, A\,\vec g \,, \qquad\qquad dA \;=\; A \wedge A \,.$$
Here $\vec g$ is the vector of master integrals, each normalized to uniform weight (the number of nested integrations), and $A$ is a matrix of one-forms in the kinematic variables. For polylogarithmic integrals $A=\sum_k A_k\, d\log \alpha_k$ with constant rational matrices $A_k$ and letters $\alpha_k$, algebraic functions of the kinematics that mark every possible branch point; on an elliptic curve or a Calabi–Yau manifold forms built from the periods replace the $d\log$ forms. Because $\varepsilon$ multiplies the whole right-hand side, the solution at each order in $\varepsilon$ is an iterated integral of these forms, so the function space can be enumerated. The second equation says that $A$ is a flat connection, so the result cannot depend on the path of integration; when there are several kinematic variables this is the strongest single constraint. The answer is then written as
$$ I \;=\; \mathrm{LS}(x) \sum_{w} c_w\, f_w(x)\,,$$
with $x$ the kinematic variables, $\mathrm{LS}(x)$ the leading singularity (the prefactor read off the maximal cut), $f_w$ the iterated integrals labeled by words $w$ in the letters, and $c_w$ the rational coefficients and boundary constants that the constraints and the fit determine.
The geometry of the maximal cut also determines which of the two routes applies. An elliptic curve has two independent periods, and the maximal cut supplies only the first. When a master integral’s normalization involves only that period, a fit of values can succeed. When the normalization also involves an integral of a modular form over the kinematic variable (an Eichler integralAn integral of a modular form, taken from the point where the torus degenerates; it plays the role for elliptic Feynman integrals that the dilogarithm plays at one loop., which no cut supplies), no fit of numerical values alone can succeed, and the integral is computed by the second route from a boundary value derived in closed form. The class of integrals these methods reach is described on a page of its own.
Every result below is held to the BootLoops standard, and each integral page carries its three items: the closed form, the checks and the evaluator. Fifteen of the thirty integrals are known results rederived by these routes with the published answer withheld until the final comparison; the method was first tested that way on the outer-mass double box of Caron-Huot and Henn (2014). Computer-algebra and numerical programs compute every number and every symbol. The language model (Claude) sets up each problem, chooses the route, runs those programs and assembles the result, and whether a result meets that standard never depends on the model’s judgment.
The results
In the table each name links to the integral’s own page and each function class to a short primer.
| integral | loops; legs; masses | function class | status |
|---|---|---|---|
| Collider amplitudes with exact masses | |||
| Non-planar two-loop W-pair production | 2; 4; two of mass $m_t$ | elliptic | new |
| $gg\to Z\gamma$ with exact top-quark mass | 2; 4; top-quark loop of mass $m_t$ | elliptic | new |
| Elliptic master integrals for $gg\to H$ | 2; 3; five of $m_t$, one of $M$ | elliptic | known result, rederived |
| Non-planar hexa-box with two massive legs | 2; 5; none (legs $p_1^2=p_3^2=m^2$) | polylog | new |
| Outer-mass double box | 2; 4; six of mass $m$ | polylog | known result, rederived |
| Massless planar double box | 2; 4; none | polylog | known result, rederived |
| Central-mass double box | 2; 4; one of mass $m$ | polylog | known result, rederived |
| One-loop massless pentagon | 1; 5; none | polylog | known result, rederived |
| Two-loop Sudakov form factor | 2; 3; none | polylog | known result, rederived |
| Three-loop ladder | 3; 3; none | polylog | known result, rederived |
| Three-loop QED light-by-light | |||
| Light-by-light box with a sunrise insertion | 3; 4; six of mass $m$ | elliptic | new |
| Light-by-light box with a kite insertion | 3; 4; one mass $m$ | elliptic | new |
| Ten-line QED light-by-light graphs, photon exchanged along one side | 3; 4; one mass $m$ | elliptic | new |
| Light-by-light box with a vacuum-polarization insertion | 3; 4; one mass $m$ | elliptic | new |
| Crossed light-by-light box | 3; 4; four of mass $m$ | polylog + K3 | new |
| Two-point families and their geometries | |||
| Equal-mass sunrise | 2; 2; three of mass $m$ | elliptic | known result, rederived |
| Sunrise, squared masses $(1,1,2)$ | 2; 2; squared masses $(1,1,2)$ | elliptic | known result, rederived |
| Sunrise, squared masses $(1,2,3)$ | 2; 2; squared masses $(1,2,3)$ | elliptic | known result, rederived |
| Sunrise, squared masses $(1,1,4)$ | 2; 2; squared masses $(1,1,4)$ | elliptic | known result, rederived |
| Sunrise, order $\varepsilon^1$ | 2; 2; the three mass sets above | elliptic | known result, rederived |
| Equal-mass kite | 2; 2; three of mass $m$ | elliptic | known result, rederived |
| Unequal-mass kite | 2; 2; squared masses $(1,1,2)$ | elliptic | new |
| Four-mass kite with a massless rung | 2; 2; squared masses $(1,4,9,2)$, rung $0$ | polylog | new |
| Equal-mass ice-cream cone | 2; 3; four of mass $m$ | elliptic | new |
| Ice-cream cone with generic masses | 2; 3; squared masses $(1,2,3,5)$ | elliptic | new |
| Ice-cream cone, order $\varepsilon^1$ | 2; 3; as the two cones above | elliptic | new |
| Three-loop equal-mass banana | 3; 2; four of mass $m$ | K3 | known result, rederived |
| Four-loop equal-mass banana | 4; 2; five of mass $m$ | Calabi–Yau | known result, rederived |
| Three-loop banana at the mass threshold | 3; 2; squared masses $(1,1,1,9)$ | K3 | new |
| Four-loop banana at the mass threshold | 4; 2; squared masses $(1,1,1,1,16)$ | Calabi–Yau | new |
The largest new result is the non-planar family of two-loop W-pair production, $q\bar q\to W^+W^-$ with the top-quark mass kept exact. In 2024 He and collaborators computed the master integrals of the three planar topologies, which they call branches, and left the non-planar one to future work. That branch has 76 master integrals, 23 of them not computed before, among them a group sharing a sub-diagram whose maximal cut is an elliptic curve absent from the planar branches. The non-planar branch is solved here in explicit form along a one-parameter slice of the kinematics (two of the three kinematic variables held fixed), with every boundary constant derived, and the closed form agrees with independent numerical evaluations to 37 digits at two points.
The two elliptic master integrals of gluon fusion into a Z boson and a photon with the full top-quark mass dependence are given in closed form, with every boundary constant derived analytically: the elliptic boundary value is an Eichler integral of a modular form for the group $\Gamma_0(4)$, and Catalan’s constant is the other transcendental that enters. The closed form agrees with independent numerical evaluations to 126 digits. For two vector bosons produced with a jet the integrals are non-planar two-loop five-point hexa-boxes (a pentagon loop joined to a crossed box). The simplest family, with the two massive legs at non-adjacent corners, is given in closed form, and 13 of its master integrals had not been computed before. Its three top-sector integrals, those with every line of the diagram present, are given as explicit layered iterated integrals along a path through the kinematic space rather than as polylogarithms.
Light-by-light scattering, photons scattering off photons through a loop of massive charged particles, requires at three loops a family of QED integrals for which no closed form has been published. Five members are computed here. Four involve a self-energy on one side of the one-loop massive box: the box with a sunrise inserted on one line, whose boundary constant is $2\,\mathrm{Cl}_2(\pi/3)$ with $\mathrm{Cl}_2$ the Clausen function; the box with a kite inserted, the kite being the two-loop self-energy drawn as a rhombus with one diagonal; the ten-line QED graphs with a photon exchanged along one side of the box, of which the kite insertion is a limiting case; and the box with a vacuum-polarization insertion. The last three follow from a dispersion relation: the inserted self-energy is written as an integral over single lines of squared mass $w$, weighted by a spectral density $\rho(w)$. In units where the electron mass is $m=1$, the kite’s spectral density below the three-particle threshold $w=9m^2$ is
$$\rho(w)=-\frac{\pi}{w}\Big[2\ln(w-1)\ln w+3\,\mathrm{Li}_2(1-w)\Big],\qquad m^2\le w\le 9m^2,$$
and the three-loop integral is $\frac1\pi\int_{m^2}^\infty\!\rho(w)\,\mathrm{Box}_1(s,t;w)\,\mathrm{d}w$, where $\mathrm{Box}_1(s,t;w)$ is the one-loop box whose inserted line carries squared mass $w$, a combination of dilogarithms of the two photon invariants $s$ and $t$. Above $w=9m^2$ the density is a one-fold integral over the period of the sunrise’s elliptic curve, and the elliptic functions of the result come only from that region. The fifth member is the crossed light-by-light box, in which the two photon lines cross inside the massive fermion loop. Its top-sector integral is an iterated integral over the period of a K3 surface, given here as an explicit list of words, iterated integrals over thirteen kernels built from that period.
Several of the new results are single numbers. A banana graph is at threshold, in the sense used here, when the mass of its heaviest line equals the sum of the other masses. With the masses fixed that way, the three-loop banana’s period, expanded in the external momentum squared $p^2$ about $p^2=0$, contains a term $c_{3/2}\,(p^2)^{3/2}$ with $c_{3/2}=-\sqrt{3}/(36\pi)$, which follows in closed form from the monodromyHow the solutions of the differential equation transform into one another when carried once around a singular point. there; in the four-loop banana the coefficients $c_{5/4}$ and $c_{7/4}$ of $(p^2)^{5/4}$ and $(p^2)^{7/4}$ lie in the ring of constants generated by $\Gamma(\tfrac14)$. The two-loop ice-cream cone, a massive triangle with a bubble on one edge, reduces to a single boundary constant that is a new period, written as an explicit convergent integral. For equal masses it lives on the sunrise’s own elliptic curve, yet integer-relation searches find no expression for it in the sunrise’s periods and classical constants. For generic masses the curve is not described by modular forms, and the searches again find nothing classical. The same closed form exhibits, as an explicit nonzero number, the cone’s violation of the hierarchical principle of Landshoff, Olive and Polkinghorne from the S-matrix program of the 1960s, the rule that a diagram should have no singularity that none of its sub-diagrams has, to which the ice-cream cone is the known counterexample.
Feynman geometry
In the threshold configuration of the sunrise and banana graphs the heaviest line’s mass equals the sum of the others (squared masses $(1,1,4)$ for the sunrise, $(1,1,1,9)$ and $(1,1,1,1,16)$ for the bananas), and at the point $p^2=0$ two of the graph’s singular points merge, so the modular-form expansions used at other special points do not exist. There the sunrise’s elliptic curve degenerates through a fiber of Type III in Kodaira’s classification of degenerating elliptic curves, with local monodromy of order four. In the ordinary (semistable) degeneration one cycle of the torus pinches to a point, and every degeneration of the equal-mass sunrise’s curve is of that kind; this one is not, and it is the first non-semistable degeneration identified so far for the elliptic curve of a Feynman integral. The order of the threshold monodromy alternates with the loop number: 4, 2, 4 at two, three and four loops, read from the exact Picard–Fuchs operators (the differential operators the periods satisfy), and 2 at five loops on a two-mass slice, whose reading as a threshold is conjectural. The threshold constants quoted above follow the same alternation: the $\Gamma(\tfrac14)$ ring of the four-loop coefficients already appears in the $(1,1,4)$ sunrise at that point, the other order-four case, while the three-loop constant $c_{3/2}$ contains no $\Gamma(\tfrac14)$.

Figure 2. The banana ladder. The same two-point graph gains one massive line per loop: the one-loop bubble, the two-loop sunrise, and the three- and four-loop bananas. The geometry of the maximal cut climbs with it, from a sphere (logarithms) to an elliptic curve (modular forms) to a K3 surface to a Calabi–Yau threefold.
Three of the results are proofs that an answer of a given kind does not exist. The first, used above to choose between the two routes, is that the $gg\to Z\gamma$ boundary value cannot be fitted from numerical values alone. The second is that the crossed light-by-light box’s top-sector integral has no polylogarithmic form in any basis, with the proof on its page. The third concerns periods of weight three, the kind a K3 surface carries. The K3 surface of the three-loop banana has periods built from two copies of the equal-mass sunrise’s elliptic curve, which has complex multiplication by $\mathbb{Q}(\sqrt{-15})$ (an extra arithmetic symmetry of special elliptic curves). A weight-three period without complex multiplication would be a new kind of number for Feynman integrals. The three-loop banana provably cannot produce one: anything assembled from pieces with complex multiplication inherits complex multiplication, and at equal mass Livné’s modularity theorem ties the surface’s weight-three part to a modular form with complex multiplication of its own. Details are in Bootstrapping elliptic and Calabi–Yau Feynman integrals.
The papers
- Automated computation of Feynman integrals with BootLoops (PDF) — Matthew D. Schwartz. A preliminary version of the paper.
- Bootstrapping elliptic and Calabi–Yau Feynman integrals (PDF) — Matthew D. Schwartz. A preliminary version of the paper.
- Polylogarithmic, elliptic, and Calabi–Yau Feynman integrals from a hybrid bootstrap (PDF) — Matthew D. Schwartz.
Supplementary material
- The integral pages — linked from the table above (results at order $\varepsilon^1$ share their graph’s page), each with the data files its evaluation script reads.
- Tools — the index of the programs used in these computations, one page each: the reduction, singularity-analysis, differential-equation, period and integer-relation programs, both the community packages adopted and the ones written for this work.
- The BootLoops integral class — which Feynman integrals these methods reach, what their answers look like on paper, and the diagnostics that tell which function class an integral belongs to.
- Polylogarithms — a primer: how the logarithm grows into the dilogarithm, how iterated integrals organize the whole family, and how the symbol turns identities among these functions into linear algebra.
- Elliptic functions and elliptic integrals — a primer on the elliptic curve, its periods and modular forms, and the Feynman integrals, beginning with the sunrise, whose answers are written in them.
- K3 integrals — a primer on K3 surfaces, one complex dimension above the elliptic curve, whose periods take over at the three-loop banana.
- Calabi–Yau integrals — a primer on Calabi–Yau threefolds and their periods, the geometry of the four-loop banana, where the answer is built from integrals over the geometry itself.
References
| G. Passarino and M. J. G. Veltman, One-loop corrections for $e^+e^-$ annihilation into $\mu^+\mu^-$ in the Weinberg model, Nucl. Phys. B 160 (1979) 151 | the reduction of every one-loop amplitude to a small set of standard integrals |
| K. G. Chetyrkin and F. V. Tkachov, Integration by parts: the algorithm to calculate $\beta$-functions in 4 loops, Nucl. Phys. B 192 (1981) 159 | the integration-by-parts identities that relate the integrals of a family |
| S. Laporta, High-precision calculation of multiloop Feynman integrals by difference equations, Int. J. Mod. Phys. A 15 (2000) 5087 [arXiv:hep-ph/0102033] | the algorithm that organizes those identities into solvable linear systems |
| A. B. Goncharov, M. Spradlin, C. Vergu and A. Volovich, Classical polylogarithms for amplitudes and Wilson loops, Phys. Rev. Lett. 105 (2010) 151605 [arXiv:1006.5703] | the symbol of a polylogarithmic function |
| J. M. Henn, Multiloop integrals in dimensional regularization made simple, Phys. Rev. Lett. 110 (2013) 251601 [arXiv:1304.1806] | the canonical, $\varepsilon$-factorized form of the differential equations |
| V. A. Smirnov, Analytical result for dimensionally regularized massless on-shell double box, Phys. Lett. B 460 (1999) 397 [arXiv:hep-ph/9905323] | the closed form of the massless double box, rederived here coefficient by coefficient |
| S. Caron-Huot and J. M. Henn, Iterative structure of finite loop integrals, JHEP 06 (2014) 114 [arXiv:1404.2922] | the outer-mass double box on which the method was first tested |
| S. Laporta and E. Remiddi, Analytic treatment of the two-loop equal-mass sunrise graph, Nucl. Phys. B 704 (2005) 349 [arXiv:hep-ph/0406160] | the analytic evaluation of the equal-mass sunrise on its elliptic curve |
| L. Adams, C. Bogner, A. Schweitzer and S. Weinzierl, The kite integral to all orders in terms of elliptic polylogarithms, J. Math. Phys. 57 (2016) 122302 [arXiv:1607.01571] | the equal-mass kite in elliptic polylogarithms, one of the elliptic results rederived here |
| S. Pögel, X. Wang and S. Weinzierl, Taming Calabi–Yau Feynman integrals: the four-loop equal-mass banana integral, Phys. Rev. Lett. 130 (2023) 101601 [arXiv:2211.04292] | the $\varepsilon$-factorized differential equation carried up to the Calabi–Yau banana |
| N. Schwanemann and S. Weinzierl, Electroweak double-box integrals for Møller scattering, SciPost Phys. 18 (2025) 172 [arXiv:2412.07522] | the central-mass double box whose published result the recomputation here matches |
| X. Liu and Y.-Q. Ma, AMFlow: a Mathematica package for Feynman integrals computation via auxiliary mass flow, Comput. Phys. Commun. 283 (2023) 108565 [arXiv:2201.11669] | the auxiliary-mass-flow method that supplies the high-precision values and the independent evaluations |
| S. Mizera and S. Telen, Landau discriminants, JHEP 08 (2022) 200 [arXiv:2109.08036] | the singularities of a Feynman integral computed from the graph by algebraic geometry |
| H. S. Hannesdóttir, A. J. McLeod, M. D. Schwartz and C. Vergu, Applications of the Landau bootstrap, Phys. Rev. D 111 (2025) 085003 [arXiv:2410.02424] | the Landau bootstrap, one of the two methods extended here, with the outer-mass double box as its worked example |
| O. Barrera, A. Dersy, R. Husain, M. D. Schwartz and X. Zhang, Analytic regression of Feynman integrals from high-precision numerical sampling, JHEP 01 (2026) 014 [arXiv:2507.17815] | analytic regression, the other parent method, with the elliptic case left for later work |
| W.-J. He, R.-Y. Zhang, L. Han, Y. Jiang, Z. Li, X.-F. Wang, S.-X. Li, P.-F. Li and Q.-h. Wang, Two-loop planar master integrals for NNLO QCD corrections to W-pair production in quark-antiquark annihilation, JHEP 12 (2024) 136 [arXiv:2409.08879] | the three planar branches of two-loop W-pair production with exact top-quark mass; the non-planar branch computed here was left open there |





























