K3 integrals

One complex dimension above the elliptic curve: a compact surface carrying a nowhere-vanishing holomorphic 2-form, whose periods take over from elliptic integrals when Feynman diagrams climb past the sunrise — and where most of the function-space comforts of the elliptic world fall away.

The content on this page was written by AI under human supervision.

The ladder: The BootLoops integral class · Polylogarithms · Elliptic · K3 · Calabi–Yau

One dimension up

The elliptic curve of the previous rung is a curve in the complex geometer's sense: one complex dimension, two real ones. Topologically it is the torus. A K3 surface is its two-dimensional sibling, and the definition asks for three things. A K3 is a compact complex surface — two complex dimensions, four real ones. It is simply connectedevery closed loop on the surface can be contracted to a point — unlike the torus, whose two independent loops cannot. And it carries a holomorphic 2-form $\Omega$ that vanishes nowhere.

Each clause mirrors the elliptic curve of the previous rung. An elliptic curve is compact, and it carries a nowhere-vanishing holomorphic 1-form — the $\omega = dx/y$ inside every elliptic integral. The one property it conspicuously lacks is simple connectivity: the two independent loops on the torus are precisely where its two periods come from. A K3 trades those loops away. Simple connectivity forces $H_1 = 0$ — no 1-cycles at all — so the transcendental life of the surface moves up a dimension, into closed 2-cycles. That single shift — transcendence carried by 2-cycles instead of loops — drives everything else about these surfaces.

The concrete example to keep in mind is the quartic surface in projective 3-space. ($\mathbb{P}^3$ is the space of complex lines through the origin of $\mathbb{C}^4$: a point is a ratio $[x_0:x_1:x_2:x_3]$, and a homogeneous polynomial equation cuts out a well-defined subset.) Take any homogeneous quartic $F$ whose zero set is smooth — the Fermat quartic

$$x_0^4 + x_1^4 + x_2^4 + x_3^4 \;=\; 0$$

is the classic choice. The zero set is a compact surface of two complex dimensions, and it comes with its 2-form for free. Work in the affine chart $x_0=1$ with coordinates $(x,y,z)$, so the surface is $F(x,y,z)=0$; then

$$\Omega \;=\; \frac{dx\wedge dy}{\partial F/\partial z}$$

is holomorphic and nowhere zero on the surface (where $\partial F/\partial z$ vanishes, solve for a different variable; the expression reappears with the roles permuted). Compare the elliptic case: write the curve as $F(x,y)=0$ and the 1-form is $\omega = dx/(\partial F/\partial y)$, which for $y^2 = x^3+ax+b$ is $dx/2y$ — the same recipe, one dimension down.

The degrees are no accident. A smooth cubic curve in $\mathbb{P}^2$ is an elliptic curve; a smooth quartic surface in $\mathbb{P}^3$ is a K3; a smooth quintic threefold in $\mathbb{P}^4$ is a Calabi–Yau threefold, the next rung of the sunrise → K3 → CY₃ ladder these pages climb. In each case the degree makes the pole of projective space's natural top-degree form cancel exactly against the defining polynomial, leaving a global holomorphic form with neither zeros nor poles — "trivial canonical class," in the algebraic geometer's vocabulary, the defining property of the whole Calabi–Yau family, with the K3 as its two-dimensional member.

The label "K3" is younger than the objects: André Weil coined it in a 1958 report on his research program, in honor of Kummer, Kähler, Kodaira, and the mountain K2 in Kashmir. Nothing in the name describes the geometry; it is a monument to three people and a mountain.

All K3 surfaces are, moreover, diffeomorphic to one another. As a smooth four-dimensional shape there is only one K3; what varies — and what a Feynman integral's kinematic variable moves through — is the complex structure carried by that one shape. And just as $\omega$ on the elliptic curve is unique up to a constant, $\Omega$ is unique up to scale: one holomorphic 2-form per surface, one basic transcendental object to integrate.

Periods on a surface

On the elliptic curve you integrated the 1-form over the two loops of the torus and got two numbers, the periods $\psi_1, \psi_2$. On a K3 the cycles are 2-cycles — closed two-dimensional membranes sitting inside the surface, the analogue of the torus loops one dimension up — and the periods are the integrals

$$\int_\gamma \Omega, \qquad \gamma \in H_2(\text{K3}, \mathbb{Z}).$$

An elliptic curve has 2 independent cycles; a K3 has 22, and this set of twenty-two cycles is the same for every K3 — another face of the fact that there is only one underlying smooth shape.

Twenty-two sounds like an explosion of new transcendental functions. It is not, for one structural reason: the 22 classes split into two kinds. Algebraic classes are represented by actual curves lying inside the surface — a line or a conic on the quartic, say — and over an algebraic class the integral of $\Omega$ vanishes identically (restrict $\Omega$ to a curve inside the surface and it vanishes: a 2-form needs two independent complex directions, and a curve has only one). The number of independent algebraic classes is the Picard numberthe count of independent algebraic curve classes on the surface — the part of $H^2$ that supports no periods $\rho$, and it can be as large as 20. Whatever is left — the transcendental lattice, of rank $22-\rho$ — is where the periods actually live.

The K3 surfaces these integrals produce are extraordinarily special in exactly this respect. A one-parameter family — one kinematic variable, like the external momentum-squared $t=p^2$ of the three-loop banana met below — generically has $\rho = 19$, leaving a transcendental piece of rank just 3. Three genuine period functions of $t$, out of a lattice of 22.

That small number makes the subject computable, through the same mechanism as on the elliptic rung. Fix a transcendental cycle $\gamma$ and differentiate the period $\varpi(t) = \int_\gamma \Omega(t)$ under the integral sign. Each derivative produces another 2-form living in a space of rank 3 — so $\varpi, \varpi', \varpi'', \varpi'''$ must satisfy a linear relation with coefficients rational in $t$. Every period of the family therefore obeys a third-order linear differential equation, the Picard–Fuchs equationthe linear differential equation in the kinematic variable that a family's periods obey — the analogue, one rung up, of the second-order equation behind the elliptic periods of the family. The elliptic curve, with its rank-2 (co)homology, gave a second-order equation; the generic one-parameter K3 family gives a third-order one. The order tracks the rank of the transcendental piece — hold that thought, because a lopsided enough family can leave more transcendental directions open and push the order higher, and two of the examples below do exactly that.

How you recognize the K3 rung

The diagnostic route carries over from the elliptic rung unchanged: derive the differential equation the integral satisfies in its parameter — for a Feynman graph, most cleanly through the maximal cutthe integral with every propagator put on shell — it strips away subtopologies and isolates the geometry underneath — and read the geometry off the operator. On the sunrise you found a second-order operator and an elliptic curve. The K3 signature is a third-order operator with one extra property: it is the symmetric square of a second-order one.

The symmetric square is best explained by example. Suppose $f$ and $g$ solve a second-order equation, normalized so the first-derivative term is absent: $y'' + q(t)\,y = 0$. Then the three products $f^2$, $fg$, $g^2$ all solve one third-order equation,

$$y''' + 4q\,y' + 2q'\,y = 0,$$

and that third-order operator is called the symmetric square of the original. Try it with $q=1$: $f=\cos t$ and $g=\sin t$ solve $y''+y=0$, and you can check by hand that $\cos^2 t$, $\sin t\cos t$, and $\sin^2 t$ all solve $y''' + 4y' = 0$. Products of solutions of an order-2 equation fill out exactly an order-3 solution space.

Now the physics. The maximal cut of the two-loop equal-mass sunrise is the period of an elliptic curve, with its second-order operator $L_2$. The maximal cut of the three-loop equal-mass banana — the same diagram with one more line — satisfies a third-order operator $L_3$, and $L_3 = \mathrm{Sym}^2(L_2)$ exactly: the check is exact polynomial algebra with zero residual. The geometry behind the operator is a one-parameter family of K3 surfaces, the holomorphic K3 period is the square of the sunrise's elliptic period, and its Taylor coefficients are the Domb numbers $1, 4, 28, 256, \ldots$ (OEIS A002895). Whenever a third-order operator appears above an elliptic one, this is the pattern to test for — and the test is exact algebra on the operator alone, no diagram required.

The squaring also explains the bookkeeping physicists attach to these functions: the weight jumps from 1 to 2. An elliptic period comes from $H^1$ of the curve and counts as weight 1; a K3 period comes from $H^2$ and carries weight 2, consistent with its realization as the square of a weight-1 period. Every rung of the Calabi–Yau ladder adds one to the weight of its basic period.

The genuinely harder part is the function space. For the elliptic curve there is a complete, worked-out theory of iterated integrals — the elliptic multiple polylogarithms, developed for Feynman integrals by Broedel, Duhr, Dulat, Penante, and Tancredi, with a parallel line by Adams, Bogner, and Weinzierl. Whatever elliptic answer a diagram produces, there is a named function space to expand it in. No comparably complete basis exists at K3 level — nobody has built the eMPL dictionary's analogue for iterated integrals on a K3 surface — so evaluations lean directly on the raw materials the geometry provides:

In practice, at K3 level the differential equation and the surface's arithmetic are the function theory. There is no all-purpose basis to fit against, which is why "value-fitting against known constants" — the workhorse closing move of the polylogarithmic world — can fail structurally rather than merely numerically. The crossed light-by-light box below turned that failure into a theorem.

The physics examples

The three-loop banana at threshold — the K3 rung. The equal-mass three-loop banana is the prototype: a three-loop integral built on a K3 surface, the diagram where the $\mathrm{Sym}^2$ pattern above was verified exactly. Its sharpest form is the threshold banana — a three-loop integral tuned, at masses-squared $(1,1,1,9)$, so its heaviest line sits exactly at the production threshold of the other three: the K3 period underneath develops a sign-flipping square-root branch, and the coefficient of that branch is known exactly, $c_{3/2} = -\sqrt{3}/(36\pi)$. At the threshold the local exponents of the governing operator — the leading powers of its solutions at the singular point — come out $\{1,1,1,\tfrac32\}$: four of them, because the lopsided masses raise this family's Picard–Fuchs order from the generic three to four. The lone half-integer exponent means one circuit of the threshold flips a period's sign — monodromythe linear reshuffling of a differential equation's solutions when the variable is carried once around a singular point and back of order two, the way one circuit of a Möbius band brings you back flipped.

The crossed light-by-light box. The crossed member of the three-loop light-by-light family is the one whose finite piece was proved beyond every closed-form attack — no basis of known constants can recognize it, and the obstruction is a genuine K3 period — and it now has an exact answer anyway: thirteen nested integrals whose integrands are built from the K3 period itself, with nothing fitted numerically. The integrands are exactly the raw materials listed above — one integrates $A(x)\,\varpi + B(x)\,\varpi'^2/\varpi$, period and quasi-period together — and the answer comes with a falsifiable 122-digit prediction at a point no independent method has ever evaluated. The K3 here is the same banana K3, at virtuality $u=-s-t$ underneath a polylogarithmic top sector.

Black-hole scattering at fifth post-Minkowskian order. In the two-body problem of general relativity, the perturbative series in Newton's constant climbs this same ladder: through third order everything is polylogarithmic, and at the fifth-order frontier, at second order in the mass ratio, the governing surface is a K3 of Apéry type — its periods generated by the very integer sequence Roger Apéry used in his celebrated 1978 proof that $\zeta(3)$ is irrational. (Beukers and Peters identified the K3 family behind Apéry's numbers in 1984; forty years later it turned up inside gravitational-wave theory.) The BootLoops paper behind that page built the order's complete function space and derived the disputed memory constant $c_M=1$, reading the physical threshold behind it — a velocity $v/c=\sqrt{8}/3$ — off the K3's own geometry.

The Watson integral — a 1939 condensed-matter question. The same class of surface settles a problem from entirely outside particle physics. The return probability of a random walker on a lopsided crystal lattice comes down to the Watson integral, whose exact evaluation at three genuinely different hop rates was hunted from 1939 on and judged hopeless in Zucker's 2011 survey of the problem. The formula hunted since 1939 never existed — the integral obeys a proved fifth-order equation whose symmetry rules out every formula built from elliptic integrals — and the true answer lives on a K3 surface, explicitly identified and proved not to split into elliptic pieces. Note the order: five. Breaking the equalities among the three hop rates drops the Picard number from 19 to 17, liberating two extra transcendental directions and pushing the Picard–Fuchs order from the generic 3 up to 5 — the promised effect at full strength: a family that sees more than the minimal transcendental piece. The K3 instruments then earn their keep: the hop rates are stored inside the surface's theta constants by a proved identity, and running that identity backwards to a complex-multiplication point produced the first closed form in the subject's history away from every symmetric configuration — gamma functions at sixteenths, $\Gamma(\tfrac1{16})\Gamma(\tfrac3{16})\Gamma(\tfrac5{16})\Gamma(\tfrac7{16})$, at hop rates about $13.2:7.5:1$. The walker's own number barely moves — a return probability of $0.3480\ldots$ on the paper's $\sqrt2{:}\sqrt3{:}\sqrt5$ benchmark lattice against Watson's $0.3405\ldots$ on the symmetric one — while the mathematics underneath transforms completely.

Where this shows up in BootLoops

Further reading