Cosmological correlators (cosmology)

Quantum fluctuations in the early universe left correlations across the sky, and computing them beyond the simplest approximation requires loop integrals like those of particle physics. In the one-loop, three-site case the integrals are elliptic, yet the elliptic parts cancel and the answer is built from dilogarithms.

Primordial fluctuations and the wavefunction

In inflationary cosmology, quantum fluctuations of the primordial fields, stretched to astronomical size by the expansion of space, seeded the temperature variations of the microwave background and the clustering of galaxies. Inflation predicts their correlation functions, how strongly the fluctuation at one place or wavelength is tied to that at another, and surveys of the microwave background and of the distribution of galaxies measure them. The measurements improve with every survey, while the calculations lag behind. Beyond the simplest models and the lowest orders of perturbation theory the integrals have resisted computation, and for diagrams with a closed loopa closed ring of internal lines in a diagram; each loop adds an integral over an undetermined momentum and one order of quantum correction even the kind of function the answer should be is mostly unknown.

The object behind these calculations is the wavefunction of the universe, $\Psi[\phi]$: the quantum amplitude for the fields to end the primordial era with spatial profile $\phi$, where $\phi$ stands for the values of the fields throughout space. Expanding $\log\Psi$ in powers of the field defines wavefunction coefficients $\psi_n$, one for each number $n$ of field factors, computed graph by graph as scattering amplitudes are. Feynman integrals for amplitudes evaluate to restricted classes of functions: first the polylogarithmsiterated integrals that generalize the logarithm; the dilogarithm Li2, with two nested integrations, is the simplest new member and is said to have weight two, then integrals over elliptic curvesthe curve defined by the square root of a cubic or quartic polynomial; over the complex numbers it is a torus, and the integrals around its two independent loops, the periods, cannot be reduced to logarithms and beyond. Once the class is known, an amplitude can often be built from its singularities and symmetries without evaluating any integral. Beginning in 2018, Arkani-Hamed, Baumann, Lee and Pimentel carried that strategy into cosmology as the “cosmological bootstrap,” which needs the same input: the space of functions in which the answer lives.

Sites, edges and a triangle of momenta

The benchmark model here is a conformally coupled scalara field coupled to the curvature of spacetime with the special strength at which, once rescaled, it propagates as in flat, non-expanding space with polynomial self-interactions in a universe whose scale factor, the overall size of space, grows as a power of conformal time $\eta$. After the field is rescaled, the expansion survives only as a weight $(-\eta)^{-(1+\sigma)}$ on each vertex, where the exponent $\sigma$ depends on the background and the type of interaction; $\sigma=0$ for a cubic vertex in de Sitter space, the idealized inflationary spacetime. That weight can be traded for an integral over a shift of the energy entering the vertex, referred to below as an energy integral. The flat-space coefficient, the object computed here, thus determines its counterpart in de Sitter space and in every power-law background through these energy integrals.

Flat-space coefficients have a simple structure (Arkani-Hamed, Benincasa and Postnikov, 2017). A graph has sites (its vertices), each carrying the total energy $X_v$ of its external legs ($v$ labels the site), and internal edges, each carrying an energy $y_e$ ($e$ labels the edge). Two sites with energies $X_1$ and $X_2$ joined by an edge of energy $y$ give $\psi_2=1/[(X_1+X_2)(X_1+y)(X_2+y)]$: one pole wherever the energy flowing into a connected subgraph vanishes. At the pole where the total energy of a tree graph vanishes, the residuewhat remains of a function at a pole once the vanishing denominator is stripped off; the residues of an integrand constrain which functions the integral can produce is an ordinary flat-space scattering amplitude, and the main results below rest on that fact.

Two line drawings side by side on white. Left: three violet dots at the corners of a triangle joined by thick dark slate edges labeled y31 on the left edge, y23 on the right edge and y12 on the bottom edge; a thin gray stub runs outward from each dot to a violet label, X3 above the top dot, X1 below left and X2 below right. Right: three dark slate arrows form a closed triangle, P2 along the bottom pointing right, P3 from the lower right corner up to the upper left corner, P1 from the upper left corner down to the lower left corner; a violet dot labeled with a script letter l carrying a vector arrow sits above and to the right of the triangle, and three violet dashed lines run from it to the corners, labeled y12 to the lower left corner, y23 to the lower right corner and y31 to the upper left corner.

The object of the calculation. Left, the one-loop triangle graph: three sites (dots), each carrying the total energy $X_v$ of the external legs attached to it, joined in a ring by internal edges with energies $y_{12}$, $y_{23}$, $y_{31}$ (the subscripts name the two sites an edge joins). Right, the loop measure: the external momenta close into a triangle with sides $P_1$, $P_2$, $P_3$, the magnitudes of the total momenta entering the three sites, and the three edge energies are the distances (dashed) from the loop momentum $\vec\ell$ to its corners. In general $\vec\ell$ lies out of the plane of the triangle, so the four points span a tetrahedron. Trading $\vec\ell$ for the three distances puts the volume of that tetrahedron, proportional to the square root of a polynomial $B$ in the squared energies and momenta, into the denominator of the measure. That square root, together with the doubled-edge poles described below, is where the elliptic curve comes from. (Redrawn after panels (a) and (c) of Figure 1 of the paper.)

The triangle has three sites and three edges (left in the figure). Its loop integrand $\psi_3$, the flat-space coefficient before the integral over the loop momentum, is a positive quartic polynomial over ten linear factors. Seven are partial energies, the energies flowing into connected subgraphs: the total energy $q_{\mathcal G}=X_1+X_2+X_3$ of the graph, each single site, each pair of sites. The other three, $q_{\mathcal G}+2y_{12}$ and its two relabelings, count one edge twice. They come from the piece of the propagatorthe factor for each internal line of a diagram; here it carries an extra term that fixes the field at the end of time that enforces the wavefunction’s boundary condition at the end of time, and every elliptic curve in the triangle involves one of them. The loop momentum $\vec\ell$ enters only through the edge energies, the distances from $\vec\ell$ to the corners of the momentum triangle, whose sides $P_v$ are the magnitudes of the total momenta entering the sites (right in the figure). Used as integration variables, the edge energies bring a factor $B^{-1/2}$ into the measure, with $B(y;P)$ a Cayley–Menger determinant proportional to the squared volume of the tetrahedron with apex $\vec\ell$. The result of the loop integral is the one-loop triangle coefficient, $V(X;P)=\tfrac{q_{\mathcal G}}{4\sqrt2}\int\mathrm d^3\ell\;\psi_3$, a function of the three site energies $X$ and the three momentum magnitudes $P$; the factor $q_{\mathcal G}$ removes the pole at vanishing total energy, and $4\sqrt2$ is a convenient normalization. With one external leg per site $X_v=P_v$. With several legs $X_v$ exceeds $P_v$ (unless the legs are collinear), and the energy integrals of an expanding background probe that region.

Where elliptic functions appear

Tree-level coefficients of the benchmark model are polylogarithms (or close relatives when $\sigma$ is not an integer), and the simplest loop, the two-site bubble, is polylogarithmic too. In August 2024 Benincasa, Brunello, Mandal, Mastrolia and Vazão derived the differential equations of the triangle’s master integralsa finite set of basis integrals to which every integral of a family reduces by exact linear relations (integration by parts); the physical quantity is one particular combination of them, and their differential equations decide which functions can appear. A block of them with the denominator $q_{\mathcal G}+2y_{12}$ obeys a second-order equation that cannot be factored, with elliptic integrals as solutions. They concluded that the triangle coefficient involves elliptic functions, and that this “is likely to extend to all polygon graphs with a higher number of sites,” while noting that an explicit computation was needed. Cosmological loops would then differ from flat-space amplitudes, where elliptic integrals first appear at two loops, in the sunrise graph.

The curve can be read off the loop measure. Consider the sector with denominator $q_{\mathcal G}+2y_{12}$, the master integrals that share it. Stripping off that pole leaves an integral over $y_{23}$ and $y_{31}$ containing only $\sqrt B$. In the squares $y_e^2$ the polynomial $B$ is quadratic, which would integrate to logarithms. A master integral, however, is integrated against $\mathrm dy_{23}\,\mathrm dy_{31}$ rather than the squares, and when both grow large at fixed ratio $\tau=y_{31}/y_{23}$ its integrand becomes proportional to $\mathrm d\tau/w$, with $w$ the square root of a quartic polynomial in $\tau$:

$$w^2 \;=\; P_2^2\,\tau^4-\bigl(P_1^2+P_2^2-P_3^2\bigr)\,\tau^2+P_1^2\,.$$

The points $(\tau,w)$ obeying this equation, with $\tau$ allowed complex values, make up the curve, which is elliptic for any non-degenerate triangle of momenta. The curve depends on the momenta only, not on the site energies. It sits at infinity on the complexified surface $q_{\mathcal G}+2y_{12}=0$, where $y_{23}$ and $y_{31}$ grow large at fixed $y_{12}$, which is not a region of real loop momentum. Take the one-parameter family of triangles $X=P=(2\lambda,\lambda,1)$, with one leg per site and the three sides set by a single number $\lambda$ between $\tfrac13$ and $1$. On it the curve’s two periodsthe integrals of the curve’s natural differential around its two independent closed loops; the other integrals on the curve are built from these, their second- and third-kind companions, and elementary pieces are the complete elliptic integrals $\tfrac1\lambda K\bigl(\tfrac{1-\lambda^2}{8\lambda^2}\bigr)$ and $\tfrac1\lambda K\bigl(\tfrac{9\lambda^2-1}{8\lambda^2}\bigr)$, where $K$ is the complete elliptic integral of the first kind, and the sector’s master integrals are iterated integrals built on these periods. The pieces are elliptic. Whether their physical combination is elliptic is a separate question.

Why the periods cancel

The physical combination of the master integrals is not elliptic, and the reason lies on the surface $q_{\mathcal G}+2y_{12}=0$, where the curve sits. The factor $q_{\mathcal G}+2y_{12}$ is the total energy of the chain obtained by cutting the edge $y_{12}$ open, so the residue of the physical integrand there is a flat-space amplitude. It is a product of two Feynman propagators, $1/[2q_{\mathcal G}(y_{23}^2-E_2^2)(y_{31}^2-E_1^2)]$, where $E_1=\tfrac12(X_2+X_3-X_1)$ and $E_2=\tfrac12(X_1+X_3-X_2)$ are the energies flowing through the two remaining edges. The residue depends on $y_{23}$ and $y_{31}$ only through their squares, that is, through the loop momentum, and in the squares $B$ is merely quadratic: on the pole one meets a genus-zero curve (a sphere rather than a doughnut), whose integrals are logarithms. A single master integral is not a function of the squares alone, so the periods appear in the pieces and cancel between them. The statement is proved in general, for arbitrary site energies and generic momenta: of the 36 lines where two denominator planes meet, an elliptic curve appears on 21, each of them inside a plane $q_{\mathcal G}+2y_e=0$, and no residue of the physical integrand involves any of those curves.

Residues alone do not fix the function class, since the integral over $B\gt0$ (the region where the tetrahedron exists) is not a sum of residues, so the next step is the equation $V$ itself obeys on the line $X=P=(2\lambda,\lambda,1)$. Exact reduction of $V$ and its $\lambda$-derivatives to master integrals yields an operator of order 20 that annihilates $V$. It factors completely over the rational functions of $\lambda$ into twenty first-order factors $\bigl(\tfrac{\mathrm d}{\mathrm d\lambda}-\tfrac{r_k'}{r_k}\bigr)$, every $r_k$ a rational function of $\lambda$ and $r_k'$ its derivative. Solving such a product means twenty successive integrations against rational functions, so $V(\lambda)$ is an iterated integral of rational forms, a function of polylogarithmic type. An elliptic period would need an irreducible second-order factor, and there is none.

Three dilogarithms

The order-20 operator, solved exactly, has 20 independent solutions, all polylogarithms of weight at most two. $V$ is a combination of them with constant coefficients, fixed by analyticity at one flattened end of the line ($\lambda=1$) together with a high-precision evaluation of the integral, and recognized as exact numbers. Written in terms of the site energies, the result extends to any triangle with one leg per site:

$$\int\mathrm{d}^3\ell\;\psi_3 \;=\; \frac{\pi}{q_{\mathcal G}}\left[\frac{\sum_i X_i\,\mathrm{Li}_2(u_i)}{\xi_1\xi_2\xi_3}+\frac{8\,X_1X_2X_3}{q_{\mathcal G}\,\xi_1^2\xi_2^2\xi_3^2}\;\mathcal W\right],$$ $$\mathcal W \;=\; \sum_i X_i\xi_i\,\mathrm{Li}_2(u_i)+2\sum_{i\lt j}X_iX_j\,\ln(1-u_i)\ln(1-u_j)-\frac{\pi^2}{12}\sum_{i\lt j}\xi_i\xi_j\,,\qquad \xi_i=q_{\mathcal G}-2X_i\,,\quad u_i=\frac{\xi_i}{q_{\mathcal G}}\,.$$

Here $q_{\mathcal G}$ is the total energy, $X_i=P_i$ the momentum magnitudes ($i=1,2,3$ labels the sites) and $\mathrm{Li}_2$ the dilogarithm. Each $\xi_i=q_{\mathcal G}-2X_i$ measures how far the momentum triangle is from flattening at site $i$, and $u_i=\xi_i/q_{\mathcal G}$ is the same quantity as a fraction of the total energy; $\mathcal W$ collects the terms of the second fraction. The $\xi_i$ are positive by the triangle inequalities, and the expression stays finite for a flattened triangle ($\xi_i=0$) or a softof vanishing magnitude: a soft momentum is one taken to zero momentum. The coefficient $V$ is the right-hand side times $q_{\mathcal G}/(4\sqrt2)$. Not counting the overall factor of $\pi$, the coefficient thus has transcendental weight two (roughly, the number of nested integrations in its functions), the same weight as a one-loop amplitude in flat space. The only earlier closed form known for this coefficient, by Agui Salcedo, Lee, Melville and Pajer, covers the configuration in which one external momentum vanishes. On the line the only numerical input is the constants, and the formula reproduces every value of the integral that was not used to fix them, to the full accuracy of those values. For a general triangle the formula is an ansatza trial expression of fixed form with undetermined coefficients, which are then fixed by matching to known values whose 14 rational coefficients follow from the result on the line. It matches the integral to 30 digits or better at 25 further triangles, including generic, nearly flattened and soft ones, where an independent fit recovers the same coefficients.

Ten square roots and a tetrahedron

One leg per site is a special case, in which many of the square roots and curves in the master integrals degenerate. On a second line, $X=P+\tfrac13(1,1,1)$ with $P=(2\lambda,\lambda,1)$ as before, each site energy exceeds its momentum by $\tfrac13$, as it would with several legs per site. On this line there are 134 master integrals and fifteen sectors with irreducible elliptic blocks. None of these elliptic blocks enters the physical combination: $V$ is one component of the solution of a first-order system of rank 44 built entirely from logarithmic forms $\mathrm d\ln(\,\cdot\,)$ whose arguments are rational or carry a single square root. On that line $V$ comes out as ten terms, each a weight-two function over a square root, and that structure suggests the form of the coefficient everywhere. For each of the ten square roots there is exactly one admissible weight-two function whose branch points lie where a positive sum of energies and momenta vanishes, which leaves ten rational numbers. The exact system on the second line fixes them at $-\tfrac14$, $-\tfrac12$ and $\pm1$ with no fit, leaving one overall constant to be fixed numerically. The result is built from four independent functions, each a combination of functions of Bloch–Wignera combination of a dilogarithm, its complex conjugate and logarithms that is free of branch ambiguities; it is best known for giving the volume of an ideal tetrahedron in hyperbolic space type, $\mathrm{Li}_2(x)-\mathrm{Li}_2(\bar x)+\tfrac12\ln(x\bar x)\ln\frac{1-x}{1-\bar x}$, where $x$ contains one of the ten square roots and $\bar x$ is the same expression with the sign of that root reversed.

Part of the formula is solid geometry: four of the ten radicands, the expressions under the square roots, are $-2B$ at points where three energy factors vanish together, and there the three momenta and the three edge energies are the six edges of a flat tetrahedron. The first of the four functions is a sum of Bloch–Wigner-type terms, one for the perimeter of each of the tetrahedron’s four faces, minus one for each of its three closed paths through all four vertices. Each term is proportional, when the radicand is negative, to the volume of an ideal hyperbolic tetrahedron.

The formula agrees with the integral to 30 digits or better at every point with $X_v\ge P_v$ at which it was checked. It also continues analytically below one leg per site, to all $X_v\gt0$, with no numerical input. The continued formula agrees with the integral to the accuracy of the numerical values on that side as well, with no feature at $X_v=P_v$. The one-loop triangle coefficient is therefore a polylogarithmic function of weight two for all positive site energies, whose arguments carry square roots away from one leg per site. That statement is exact on the two lines; away from them it rests on this numerical agreement.

A single plot on white. Horizontal axis labeled delta, with site energies X v equals P v plus delta and momenta P equals (7/5, 7/10, 1) in parentheses, running from minus 0.3 to 1.5; vertical axis labeled q G squared times V, from 0.325 to just above 0.5. A dotted gray vertical line at delta equals zero carries the small labels X v less than P v on its left and X v greater than P v on its right. A smooth violet curve labeled closed form (ten square roots), continued to delta less than 0, rises from about 0.33 at the left edge, steeply at first and then more gently, to about 0.505 at delta 1.5. Open dark circles labeled direct numerical loop integral lie on the curve on both sides of zero, six of them at negative delta. An orange star labeled three-dilogarithm formula, one leg per site sits on the curve exactly at delta equals zero, near 0.397.

One smooth function through the one-leg-per-site point. The momenta are fixed at $P=(\tfrac75,\tfrac7{10},1)$ and every site energy is shifted by the same amount, $X_v=P_v+\delta$; the vertical axis is the coefficient scaled by the square of the total energy, $q_{\mathcal G}^2V$. The violet curve is the general ten-root formula, eq. (34) of the paper, continued past $\delta=0$ (dotted line) to negative $\delta$, where every $X_v\lt P_v$. Open circles are direct numerical evaluations of the loop integral on both sides of $\delta=0$. The orange star is the three-dilogarithm formula, which applies only at $\delta=0$, where all ten square roots become rational. There is no kink at $\delta=0$: the general formula, its continuation and the one-leg-per-site formula are one function, and it matches the integral on both sides to the accuracy of the numerical values. (Redrawn from Figure 4 of the paper.)

Correlators, and de Sitter space

A telescope measures correlators rather than the wavefunction, and correlators follow from $|\Psi|^2$, the probability of each field profile; at one loop the boundary value of the loop field is integrated over as well, which glues pairs of boundary legs together. The triangle’s correlator integrand is then the loop coefficient plus the three chains with one edge cut open, plus terms with more edges opened. In that sum the doubled-edge factors $q_{\mathcal G}+2y_e$ cancel, as was known from the cosmological tree theorem of Agui Salcedo and Melville (2023) and from Chowdhury, Lipstein, Mei, Sachs and Vanhove (2023). This cancellation also follows from the 2017 recursion relations and traces to the boundary term of the propagator: the elliptic curve is a feature of conditioning on the late-time field profile, which the correlator integrates over. For the triangle every elliptic line lies where the correlator integrand has no pole, so that integrand has no residue on any elliptic line. The integrated triangle correlator itself was already known to be built from dilogarithms, in flat space (Pimentel and Westerdijk) and in de Sitter space (Henn, Mei and Yang). The new result here is that the wavefunction coefficient is built from dilogarithms as well.

For a cubic vertex in de Sitter space the energy integrals make the loop integral diverge logarithmically at large loop momentum, so the coefficient needs a momentum cutoffan upper limit imposed on the loop momentum to make a divergent integral finite; here the answer depends on it only through a logarithm $\Lambda$, an upper limit on the comoving loop momentum. With one leg per site, symmetry fixes the late-time coefficient as a three-point function of dimension-two operators in a three-dimensional conformal theory (Bzowski, McFadden and Skenderis, 2013). If the cutoff is assumed to break the symmetry only through the logarithm, then the de Sitter triangle coefficient is $\psi_3^{\rm dS}=-\tfrac{\pi\zeta(3)}{2}\ln\tfrac{P_1+P_2+P_3}{\Lambda}+b_0$. Here $P_1+P_2+P_3$ is the sum of the three external momentum magnitudes, $\zeta$ is Riemann’s zeta function and $b_0$ is a number that depends on how the cutoff is defined; a numerical evaluation at two kinematic points agrees with this form. The correlator’s loop integral converges and symmetry allows only a constant, identified numerically as $-3\pi^2\zeta(3)$. With one leg per site, symmetry alone reduces the de Sitter triangle to a logarithm and two numbers, regardless of which functions appear in $V$. The functions in $V$ matter when several legs meet at a site; the de Sitter coefficient is then an energy integral of the general formula above.

One more site: the box

The four-site box, with sites 1 to 4 joined in a ring by edges with energies $y_{12}$, $y_{23}$, $y_{34}$ and $y_{41}$, is the natural next test, and unlike the triangle it is not polylogarithmic. Its loop integrand is $\psi_4$, and its one-loop coefficient is $V_4=\int\mathrm d^3\ell\;\psi_4$, with no normalization factor. The doubled-edge factors are harmless again, and the trouble comes instead from the pole where the energy of a single site vanishes. On that pole, for site 1, the loop momentum lies on a spheroid whose foci are the points where $y_{12}$ and $y_{41}$ vanish, and the residue involves the other two edge energies, $y_{23}$ and $y_{34}$, each the square root of a quadratic function on that surface. A single such root, as in the triangle, can be rationalized; two of them define a curve $Y^2=z_{23}\,z_{34}$ along a line on the spheroid, where $z_e=y_e^2$ are the two quadratics and $Y$ is the product of the two roots, and that curve has genus one: it is elliptic.

A single plot on white with a small inset drawing at upper left. Inset: a square drawn with dark slate edges labeled y12 along the top, y23 on the right, y34 along the bottom and y41 on the left; violet dots at three corners and an orange dot at the upper left corner labeled site 1 in orange italics; a short gray stub runs diagonally outward from each corner. Main plot: horizontal axis labeled site energy X equals 1 plus t (regular tetrahedron of unit momenta), logarithmic, from 1 to about 6000; vertical axis labeled X to the fourth power times C4, from 0 to about 23. A smooth violet curve labeled u plus Omega v (exact system, nothing fitted) rises from below 1 at X equals 1, climbs steeply between X of about 3 and 100, and flattens toward a dotted gray horizontal line labeled Omega equals 21.444... near the top. Thirty-four open dark circles labeled direct numerical loop integral lie on the curve along its whole length, from X near 1.1 to X equals 5000.

The four-site box and its correlator on a line. Inset: the one-loop box graph, four sites joined in a ring by edges with energies $y_{12}$, $y_{23}$, $y_{34}$, $y_{41}$. On the pole where the energy of site 1 (orange) vanishes, the residue keeps two square roots, the energies $y_{23}$ and $y_{34}$ of the two edges that do not touch that site, and the triangle’s mechanism fails. Main panel: the loop-integrated correlator $C_4$ on the line where the momenta form a regular tetrahedron of unit side and every site energy equals $X=1+t$, multiplied by $X^4$ and plotted against $X$ on a logarithmic axis. The violet curve is the closed form $u+\Omega\,v$ described below, two solutions of an exact differential system and one number $\Omega$, with nothing fitted. Open circles are direct numerical values of the loop integral, and the dotted line is the constant $\Omega\approx21.444$ that $X^4C_4$ approaches at large $X$. The curve passes through every circle over more than three decades in $X$. (Main panel redrawn from Figure 6 of the paper; inset after Figure 1(b).)

With one leg per site, $V_4$ has a branch cut, a line in the complex plane of the site energy $X_1$ across which it jumps. The cut starts at the unphysical value $X_1=-P_1$, with $P_1$ the momentum magnitude at site 1, but it controls the coefficient at physical energies, as the soft limit below shows. At that point the spheroid collapses onto the segment between its foci. The size of the jump there, the discontinuity $D_0$, is an integral along that segment on an elliptic curve set by the shape of the momentum quadrilateral (for momenta not in a plane). This discontinuity is part of the physical coefficient: as the momentum at site 1 goes soft, $V_4=-D_0\ln 2P_1$ plus terms analytic at $P_1=0$. The discontinuity $D_0$ can be evaluated in closed form, logarithms plus elliptic integrals of the first and third kind, and the coefficient of the first-kind integral does not vanish. The closed form agrees with direct numerical integration of the discontinuity. For momenta in a plane one pole of the integrand becomes a double pole and the formula gains a corresponding term; all poles are then real and no choice of branch enters. For momenta out of the plane the branches of the logarithms at a pair of complex poles come from an analytic continuation, which has been checked numerically but not proved.

An elliptic integral in a discontinuity does not by itself exclude a polylogarithmic total, as the triangle shows, so $D_0$ is tested directly. Carried around the points where its curve degenerates, $D_0$ comes back reshuffled by monodromythe mixing of a many-valued function’s branches when it is carried along a closed path around a singular point and back matrices whose eigenvalues are not roots of unity, which no function of polylogarithmic type allows. The triangle, run through the same computation as a control, shows no such obstruction, and an exact computation on the box’s residue confirms that the periods of its curve enter. The soft limit of the box with one leg per site therefore contains elliptic periods, which makes the box the first polygon of this theory that is not polylogarithmic.

Above the threshold at $X_1=-P_1$ the spheroid no longer collapses, and the discontinuity becomes an integral over its double cover defined by the two roots: a K3 surface, the two-dimensional analog of an elliptic curve. For regular-tetrahedron momenta its periods obey an equation of order five that is known exactly, whose solutions are not products of elliptic periods. For those momenta, with several legs per site, the periods of this K3 surface enter the box. At exactly one leg per site two pieces of evidence, one numerical and one exact, indicate that they persist for both the coefficient and the correlator, though generic momenta have not been examined in that case. The analysis can be pushed further on one line in kinematic space. There the momenta form a regular tetrahedron of unit side and every site energy equals $X=1+t$, so $t$ measures how far the site energies exceed the momenta. On this line exact reduction gives differential systems for the coefficient and the correlator, of dimension 34 and 28, whose solutions include the K3 periods and the periods of three more elliptic curves. An exact change of basis brings the correlator’s system to a form with simple poles only, located at the zeros of 19 polynomials in $X$ of degree at most four; the system has no other singular point at finite $X$. For the correlator this gives a complete answer:

$$C_4(t)\;=\;u(t)+\Omega\,v(t)\,,\qquad \Omega\;=\;\int\!\mathrm{d}^3\ell\;\frac{1}{y_{12}\,y_{23}\,y_{34}\,y_{41}}\;=\;21.44442951\ldots$$

Here $C_4(t)$ is the loop-integrated four-site correlator on that line, and $u$ and $v$ are two solutions of the exact system, fixed without reference to any value of $C_4$. The function $v$, the one physical solution free of logarithms at large $X$, turns out to be elementary, $v=4\sqrt{10}\,\big/\bigl[X\sqrt{(10X^2-1)(16X^4-10X^2+3)}\,\bigr]$, and $u$ is determined by expansion coefficients at large $X$ and boundary constants that are known exactly. $\Omega$ is a number, the loop integral of the inverse product of the four edge energies, the distances from $\vec\ell$ to the vertices of the unit tetrahedron. Feynman parameters reduce it to a one-dimensional integral over a product of two complete elliptic integrals of the first kind, which resisted further reduction, and a search for an integer relation between $\Omega$ and familiar constants found none. The differential system fixes $\Omega$ as well: $v$ changes sign when continued around the point $X^2=\tfrac1{10}$, where the loop integral itself is analytic, so $u+\Omega\,v$ is single-valued there for only one value of $\Omega$. That value agrees with the integral to 36 digits. With $\Omega$ fixed in this way, $u+\Omega\,v$ agrees with every direct numerical value of the correlator computed, to the accuracy of those values, over more than three decades in $X$ (the figure above).

The coefficient $V_4$ can be fixed the same way, with two exact conditions. The first is the one used for $\Omega$: the answer must come back to itself when carried around the point $X^2=\tfrac1{10}$. The second is that $V_4$ has no pole at $t=0$, where the loop integral is finite. Together they leave no undetermined constant, and the result agrees with direct numerical values of $V_4$ to 32 digits.

The paper

Supplementary material

References

N. Arkani-Hamed, P. Benincasa and A. Postnikov, Cosmological Polytopes and the Wavefunction of the Universe, arXiv:1709.02813 (2017)the anatomy of flat-space coefficients, one pole per subgraph energy, and their recursion relations
N. Arkani-Hamed, D. Baumann, H. Lee and G. L. Pimentel, The Cosmological Bootstrap: Inflationary Correlators from Symmetries and Singularities, JHEP 04 (2020) 105the cosmological bootstrap: correlators built from their singularities and symmetries
P. Benincasa, G. Brunello, M. K. Mandal, P. Mastrolia and F. Vazão, One-loop corrections to the Bunch-Davies wave function of the universe, Phys. Rev. D 111 (2025) 085016the triangle’s differential equations and elliptic sector; expected elliptic functions for every polygon
D. Baumann, H. Goodhew and H. Lee, Kinematic flow for cosmological loop integrands, JHEP 07 (2025) 131differential equations for one-loop integrands; the two-site bubble is polylogarithmic
S. Laporta and E. Remiddi, Analytic treatment of the two loop equal mass sunrise graph, Nucl. Phys. B 704 (2005) 349the sunrise graph: in flat space, elliptic integrals enter at two loops
T. Heckelbacher, I. Sachs, E. Skvortsov and P. Vanhove, Analytical evaluation of cosmological correlation functions, JHEP 08 (2022) 139a one-loop de Sitter four-point correlator in polylogarithms, its elliptic sector canceling
S. Agui Salcedo, M. H. G. Lee, S. Melville and E. Pajer, The Analytic Wavefunction, JHEP 06 (2023) 020the earlier closed form for the triangle coefficient when one external momentum vanishes
G. L. Pimentel and T. Westerdijk, On Cosmological Correlators at One Loop, arXiv:2601.00952 (2026)the flat-space in-in triangle in dilogarithms; the wavefunction’s own integrals expected to be harder
J. Henn, J. Mei and Q. Yang, A compact analytic formula for the one-loop triangle cosmological correlator, arXiv:2608.17877 (2026)the de Sitter triangle correlator of conformally coupled scalars in dilogarithms
S. Agui Salcedo and S. Melville, The cosmological tree theorem, JHEP 12 (2023) 076the tree theorem by which the doubled-edge factors cancel in correlators
C. Chowdhury, A. Lipstein, J. Mei, I. Sachs and P. Vanhove, The Subtle Simplicity of Cosmological Correlators, arXiv:2312.13803 (2023)an independent demonstration that the boundary-term poles drop out of equal-time correlators
A. Bzowski, P. McFadden and K. Skenderis, Implications of conformal invariance in momentum space, JHEP 03 (2014) 111conformal three-point functions in momentum space, which fix the de Sitter triangle’s form
K. Acres and D. Broadhurst, Empirical determinations of Feynman integrals using integer relation algorithms, arXiv:2103.06345 (2021)the practice followed here: fit high-precision values, then test on digits not used

← back to the web summaries