betaFeedback

WorksPhysics and Astronomy

Multiloop Unitarity from Airy Zeros

Abstract

Hidden zeros are non-trivial cancellations that make amplitudes or integrands vanish on specific kinematic loci away from propagator poles. They have been shown to uniquely determine tree amplitudes and one-loop surface integrands in \Tr(ϕ^3) theory, without assuming factorization. Here we extend both the zeros and the uniqueness results to two, three and four loops at arbitrary multiplicity. We show that two types of zeros naturally enter the all-loop problem. First, tree-level zeros extend to all loop orders and fix relative weights within families of subgraphs attached to vacuum ``loop cores.'' We then show that Kontsevich's matrix Airy function can be used to generate a rich structure of loop level zeros that fix the relative weights of the loop cores themselves. The relative factorization and cut data required by unitarity therefore follow from the combination of tree and Airy loop zeros. These results bring within reach the emergence of unitarity to all orders in perturbation theory.

Introduction

Unitarity ties scattering processes together: the singularities of an amplitude must describe consistent propagation and scattering of intermediate particles. On-shell recursion and loop unitarity methods use factorization poles and cuts as inputs [1, 2, 3]. Hidden zeros offer another kind of input. They are non-trivial cancellations for which an amplitude or integrand vanishes on a specific kinematic locus, while the individual diagrams remain generically nonzero and no propagator is singular [4].

Zeros also connect to exact splitting formulas, which express amplitudes as products on regular kinematic loci [4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19]. Related zeros and splitting relations have been studied in a broad range of amplitudes, form factors and cosmological wavefunctions [20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37].

In Tr⁡(ϕ3)\operatorname{Tr}(\phi^3) theory, hidden zeros were shown to determine tree amplitudes and one-loop surface integrands up to normalization without assuming unitarity, making factorization a consequence of the zeros [38, 39]. At tree level, they also determine locality within a prescribed planar pole alphabet [40]. These results are part of a broader program to derive locality and unitarity from properties such as gauge invariance, soft behavior and large-momentum scaling [41, 42, 43, 44, 45, 46, 47]. Most of the results have focused on tree amplitudes, with the above mentioned one-loop uniqueness from hidden zero opening the possibility for beyond tree results. Establishing this emergence at all loop orders is a significant open question, which is now within reach.

The passage from tree amplitudes to one loop already required a new kind of zero with no obvious origin. Remarkably, the one-loop zero unifies the different tree hidden zeros, which follow from it on appropriate residues [39]. At higher loops the possible graph topologies and propagators become much more varied, and the ansatz grow rapidly. Finding such cancellation patterns, valid at every multiplicity, is therefore a substantial challenge by itself, even before attempting to prove uniqueness. To make

progress at arbitrary loop order, we need a principle that guides the search for zeros strong enough to determine the integrand.

In this Letter we propose an approach that is demonstrated through four loops and can in principle be extended to arbitrary loop order. We organize the general problem around two results, illustrated in Figure 1. First, we show that tree zeros can be extended to arbitrary loop order as boundary zeros. These zeros then fix the relative weights of external trees attached to the same loop core. The core is the vacuum diagram obtained by removing all external legs and tree propagators. The problem then reduces to finding loop zeros that relate the different cores and their sign weights.

Figure 1. The two steps of reconstruction. (a) Boundary zeros compare positions of the orange external leg, fixing attachment weights for a fixed blue loop core and sign sector. (b) Airy zeros and their deformations relate the remaining core and sign weights. The arrows indicate coefficient relations. External insertions transfer the core constraints to higher multiplicity.

Second, we show that Kontsevich’s matrix Airy function provides a systematic source of zeros comparing these higher loop cores [48]. We assign a weight to each face of the graph, multiplied by its puncture sign s=±s=\pm for internal faces, and set every inverse propagator to the sum of the weights on its two sides. The resulting object is a matrix-model correlator, and summing over the ss gives a non-trivial zero at every multiplicity and every planar loop order L≥2L \ge2.1

However, this "Airy zero" alone is not sufficient to relate all core weights, but remarkably, it can be used to generate further families of loop zeros, by simple deformations of the propagator assignments. With locality and cubic power counting as inputs, this gives a common reconstruction argument for the two-, three- and four-loop integrands at arbitrary multiplicity.

Together, these zeros therefore determine the factorization and cut relations required by unitarity, up to normalization. Our approach also opens a clear path to proving uniqueness at any loop order. At each fixed loop order, the boundary zero argument reduces the all-multiplicity question to a finite problem. Finding sufficiently constraining deformations of the Airy zero would establish the emergence of unitarity to all orders in perturbation theory.

The local ansatz

We define the integrands through the duality between cubic graphs and triangulations underlying the surface construction [49, 50, 51, 52, 53, 54]. An nn-particle, LL-loop planar cubic graph corresponds to a triangulated disk with nn boundary edges and LL punctures. Each internal curve crosses one Feynman diagram propagator, and each triangle contains a cubic vertex.

We name the inverse propagators according to the curves they cross.

  • Di,dD_{i,d} joins marks ii and i+di+d, with boundary indices read cyclically. Together with the intervening dd boundary edges it bounds a region containing no punctures, and represents an external tree. Here 2≤d≤n2 \le d \le n.

  • Xij,SX_{ij,S} joins boundary marks ii, jj with punctures on both sides. For i<ji<j, the nonempty proper subset SS lies between the curve and the directed boundary interval from ii to jj. For i=ji=j, we choose the zero-length interval, so SS is the set inside the loop based at ii.

  • Yi,psY^{s}_{i,p} joins boundary mark ii to puncture pp.

  • ZpqstZ^{st}_{pq} joins distinct punctures pp, qq, with signs ss, tt at its endpoints.

  • Tp,SsT^{s}_{p,S} starts and ends at pp and encloses a nonempty subset SS of the other punctures.

Figure A1 in the appendix illustrates these various curve types. Each puncture carries a sign s=±1s=\pm1, shared by every curve ending there; in ZpqstZ^{st}_{pq}, ss and tt are the signs at pp and qq, respectively [39, 53, 55, 56]. The two-loop one-point examples are shown in Figure 2. The full integrand sums all 2L2^L choices of these signs. The relation of these objects to usual momentum-space integrands is described in Refs. [49, 57].

Figure 2. Two-loop one-point triangulations and their dual cubic graphs, up to puncture exchange 1↔21 \leftrightarrow2. The curve labels distinguish boundary–puncture rays YY, puncture connections ZZ, returning curves TT, and boundary curves XX. External-tree channels DD first occur at two points. Restoring the exchanged partners gives eight graphs and seven projected products, with multiplicity two for (d). These one-point objects are auxiliary surface functions.

Beyond one loop, curves can wind around punctures in infinitely many ways, through the action of the mapping class group (MCG). Following the surface-function construction [57], we identify routes related by the label-preserving MCG. Figure A1 in the appendix illustrates different routes assigned the same variable. In each fixed sign sector this is the planar scalar surface function GS=∑Γ∏C∈ΓxCG_S=\sum_{\Gamma}\prod_{C\in\Gamma}x_C of Ref. [57], evaluated at xC=1/QCx_C=1/Q_C.

We write In,L\mathcal{I}_{n,L} for the full integrand, as a graph sum over all puncture-sign sectors. To obtain our local cubic ansatz, we now allow each distinct denominator product to carry an unknown coefficient cαc_\alpha

Mn,L=∑αcα∏γ∈αQγ,∣α∣=n+3L−3.(1)M_{n,L}=\sum_\alpha\frac{c_\alpha}{\prod_{\gamma\in\alpha}Q_\gamma}, \qquad|\alpha|=n+3L-3. \tag*{(1)}

Locality requires the curves QγQ_\gamma in each product to belong to one triangulation α\alpha, and cubic scalar power counting makes its coefficient a constant. The coefficients are independent in every sign sector. For simplicity, we assume reflection symmetry, assigning equal coefficients to a diagram and its mirror image (see Appendix A for details).

Boundary zeros fix the external attachments

Boundary zeros are the first zero class we introduce. These extend the familiar tree-level zeros to arbitrary loop order and arbitrary multiplicity.2 These zeros also follow from the all-order tree–loop split of Ref. [8]: the integrand splits into a four-point tree amplitude times a lower-point loop integrand. Setting the tree factor to zero gives the boundary zero, with the loop factor kept regular. We will show they involve the positions of an external leg while leaving the internal loop propagators arbitrary, and thereby fix the relative weights of attachments to each loop core.

At tree level, the only variables are Xij=(pi+pi+1+⋯+pj−1)2X_{ij}=(p_i+p_{i+1}+\cdots+p_{j-1})^2, with Xi,i=Xi,i+1=0X_{i,i}=X_{i,i+1}=0, and particle 1 corresponds to the boundary edge between marks 1 and 2. Its skinny zero compares channels beginning at these two marks,

c1j=X1j+X2,j+1−X1,j+1−X2j=0,(2)c_{1j}=X_{1j}+X_{2,j+1}-X_{1,j+1}-X_{2j}=0, \tag*{(2)}

for j=3,…,n−1j=3,\ldots,n-1. On the above locus, the nn-point Tr⁡(ϕn)\operatorname{Tr}(\phi^n) amplitude vanishes [4]. The above conditions can equivalently be written as

X2j−X1j=η,j=3,…,n,η=−X13=X2n.(3)X_{2j}-X_{1j}=\eta,\quad j=3,\ldots,n,\qquad\eta=-X_{13}=X_{2n}. \tag*{(3)}

At loop level, the zero selecting the same external leg 1 imposes the analogous channel differences. Suppressing puncture signs,

Y2,p−Y1,p=η,X2j,S−X1j,S=η(j=3,…,n),(4)Y_{2,p}-Y_{1,p}=\eta,\qquad X_{2j,S}-X_{1j,S}=\eta\quad(j=3,\ldots,n), \tag*{(4)}

for each puncture pp and directed subset SS. Curves with both endpoints at marks 1, 2 satisfy X12,S−X11,S=X22,S−X12,S=ηX_{12,S}-X_{11,S}=X_{22,S}-X_{12,S}=\eta. The neighboring tree channels satisfy

D1,2=−η,Dn,2=η.(5)D_{1,2}=-\eta,\qquad D_{n,2}=\eta. \tag*{(5)}

All Z,TZ,T remain free. Together with the remaining DD equations in Appendix A, Eq. (19), these relations define the boundary zero B1B_1, and cyclic relabelling gives BiB_i for each external particle. The zero holds separately in each sign sector, and leaves all propagators non-zero.

Now we want to determine the effect of imposing zero BiB_i on a local ansatz Mn,LM_{n,L}, more precisely what relations the cαc_\alpha need to satisfy for BiB_i to be a zero of Mn,LM_{n,L}. This is easily understood graphically, and the argument is similar to the purely tree level one [38].

We delete leg ii from every diagram, and group resulting identical diagrams together. Boundary zero BiB_i then requires all diagrams that map into the same child diagram to have equal weight. Appendix A gives the residue proof (and explains the low-point exceptions that will be important later).

For example, the four parents in Figure 3 contribute

Figure 3. The four two-loop attachments in Eq. (6). The orange leg is added to the same one-point graph. The first and last terms attach an external tree; the middle terms split the upper or lower outer-loop edge. The common internal factor [Tp,{q}Zpq]−1[T_{p,\{q\}}Z_{pq}]^{-1} is suppressed below the graphs. The boundary zero equates c0,c1,c2,c3c_0,c_1,c_2,c_3.

Matt=1Tp,{q}Zpq[c0D1,2Y1,p2+c1Y1,p2Y2,p+c2Y1,pY2,p2+c3D2,2Y2,p2].(6)M_{\mathrm{att}}=\frac{1}{T_{p,\{q\}}Z_{pq}}\left[\frac{c_0}{D_{1,2}Y_{1,p}^2}+\frac{c_1}{Y_{1,p}^2Y_{2,p}}+\frac{c_2}{Y_{1,p}Y_{2,p}^2}+\frac{c_3}{D_{2,2}Y_{2,p}^2}\right]. \tag*{(6)}

Their common T,ZT,Z poles isolate this group. Substituting Y2,p=Y1,p+ηY_{2,p}=Y_{1,p}+\eta, (D1,2,D2,2)=(−η,η)(D_{1,2},D_{2,2})=(-\eta,\eta) and demanding the sum to vanish requires c0=c1=c2=c3c_0=c_1=c_2=c_3.

Imposing all boundary zeros for each external leg relates the attachments to a fixed loop core, obtained by removing all external trees and legs (Figure 4).

Figure 4. From attachments to the Airy vacuum sum. (a) Removing external attachments and smoothing bivalent vertices exposes a vacuum core. (b) The four two-loop cores carry one exterior-face weight bb and two internal-face weights. Their additive propagators give Eq. (9). Boundary zeros fix the attachments; the Airy conditions compare the cores.

Starting at a loop-dependent multiplicity n≥nLn\geq n_L, the boundary zeros require

Mn,L=∑rwr,n,LCr,n,L,n≥nL,(7)M_{n,L}=\sum_r w_{r,n,L}C_{r,n,L},\qquad n\geq n_L, \tag*{(7)}

Here rr labels a core and its puncture-sign sector, and Cr,n,LC_{r,n,L} sums all ways of attaching the nn external particles, including external trees, to that core. We choose an independent set of these complete sums as the basis.

The thresholds are n1=2n_1=2, n2=n3=3n_2=n_3=3 and $n_4=4, as proved in Appendix A. At general L≥2L\geq2, a simple argument bounds the finite range to check by NL=10L−12N_L = 10L - 12, but the starting multiplicity nLn_L can be smaller when the remaining cases are excluded sooner. The reason Eq. (7) only holds for n≥nLn \ge n_L is that at lower multiplicity there are not enough boundary zeros. Each complete sum Cr,n,LC_{r,n,L} obeys every boundary zero separately, so its overall weight wr,nw_{r,n} remains free. The remaining task is to find loop zeros that relate these core and sign weights.

Loop zeros from the matrix Airy function

The Airy zero at every multiplicity

We now use Kontsevich’s matrix Airy function to construct zeros that relate different loop cores [48]. Originally introduced to compute intersection numbers on moduli spaces of marked curves, this cubic matrix integral generates ribbon Feynman graphs in perturbation theory. Our vacuum loop cores appear in its planar expansion, so its exact formulas evaluate the graph sums needed here. Each face carries a weight, and the quadratic term gives an edge γ\gamma separating faces of weights λL\lambda_L and λR\lambda_R the propagator

1Qγ→Airy assignment1λL+λR.(8)\frac{1}{Q_{\gamma}} \xrightarrow{\text{Airy assignment}} \frac{1}{\lambda_L+\lambda_R}. \tag*{(8)}

This makes the matrix integral useful here: its identities describe cancellations among the cores on this common propagator assignment. For the four two-loop cores in Figure 4(b), let bb be the exterior-face weight and ap,aqa_p,a_q the internal-face weights. Denote their vacuum sum by A3A_3, where the subscript counts the three labelled faces. The three propagators of the first core have inverse values b+apb+a_p, b+aqb+a_q and ap+aqa_p+a_q. In the other cores, a bridge borders the same face on both sides and therefore has inverse value 2b2b, 2ap2a_p or 2aq2a_q. The four contributions give

A3(b;ap,aq)=1(b+ap)(b+aq)(ap+aq)+12b(b+ap)(b+aq)+12ap(b+ap)(ap+aq)+12aq(b+aq)(ap+aq)=12bapaq.(9)\begin{aligned} A_3(b;a_p,a_q)={}&\frac{1}{(b+a_p)(b+a_q)(a_p+a_q)}\\ &+\frac{1}{2b(b+a_p)(b+a_q)}\\ &+\frac{1}{2a_p(b+a_p)(a_p+a_q)}\\ &+\frac{1}{2a_q(b+a_q)(a_p+a_q)} =\frac{1}{2ba_pa_q}. \tag*{(9)} \end{aligned}

The four different propagator structures have collapsed to a single product of face weights. The result changes sign when either internal-face weight changes sign. Introducing the puncture signs {s,t}=±1\{s,t\}=\pm1, and replacing the face weights by sapsa_p, taqta_q gives st/(2bapaq)st/(2ba_pa_q), so the four sign sectors cancel. This is a zero at generic face weights, where every individual propagator remains finite.

Kontsevich’s formula has the same odd dependence on each internal-face weight at every planar loop order L≥2L \ge2 [48]. Adding external legs along the outer face splits propagators while preserving the face-sum assignment and the cancellation between puncture signs. Restoring external trees multiplies these vanishing sums by their D−1D^{-1} factors, so all DD can remain independent. Appendix B1 gives the insertion argument.

We can therefore write the Airy zero directly for the nn-point integrand. Give each boundary mark its own weight bib_i and each puncture its signed weight sapsa_p. Applying the face-sum rule to every non-tree propagator gives

Yi,ps=bi+sap,Zpqst=sap+taq,Y^s_{i,p}=b_i+sa_p,\qquad Z^{st}_{pq}=sa_p+ta_q,
Xij,s=bi+bj,Tp,ss=2sap,(10)X_{ij,s}=b_i+b_j,\qquad T^s_{p,s}=2sa_p, \tag*{(10)}

with the external-tree inverse propagators DD left arbitrary. We denote this assignment by A∘\mathcal{A}^{\circ} and obtain

In,L∣A∘=0,L≥2.(11)\mathcal{I}_{n,L}\big|_{\mathcal{A}^{\circ}}=0,\qquad L\ge2. \tag*{(11)}

This remarkably simple Airy assignment therefore produces a non-trivial zero valid at any multiplicity and any loop order! However, if our goal is uniqueness, this identity is not sufficient, and so we must seek larger zero families. This can be done in a semi-systematic fashion by varying the propagators while preserving the reference cancellation.

Two-loop mountain deformations

A surprisingly simple two-loop choice consists of just two one-parameter deformations. Keep X,Z,TX,Z,T at the Airy values in Eq. (10), leave every DD independent, and impose

Yi,ps=bi+sap+sξ,Yi,qt=bi+taq+tσξ,(12)\begin{aligned} Y^s_{i,p}&=b_i+sa_p+s\xi,\\ Y^t_{i,q}&=b_i+ta_q+t\sigma\xi, \tag*{(12)} \end{aligned}

where we can choose σ=+1\sigma=+1 or −1-1, and both assignments reduce to Airy at ξ=0\xi=0. Appendix B2 gives a more general zero family.

Three-loop deformations

At three loops, two simple deformation rules suffice. For the first, keep Y,XY,X at their Airy values and every DD independent. Select one sign γp\gamma_p at each puncture and shift

Zpqst=sap+taq+ϵδs,γpδt,γq,Tp,ss=2sap+ϵδs,γp.(13)\begin{aligned} Z^{st}_{pq}&=sa_p+ta_q+\epsilon\delta_{s,\gamma_p}\delta_{t,\gamma_q},\\ T^s_{p,s}&=2sa_p+\epsilon\delta_{s,\gamma_p}. \tag*{(13)} \end{aligned}

Here δs,γp\delta_{s,\gamma_p} is one for matching signs and zero otherwise.

The second rule starts again from Airy and changes only one chosen return,

Tp,Ss=2sap+τ,(14)T^{s}_{p,S}=2sa_p+\tau, \tag*{(14)}

Both rules recover the Airy locus when their displacement vanishes. Choosing the signs γp\gamma_p in the first rule, or the return (p,S,s)(p,S,s) in the second, specifies a particular zero family. Three choices of each type suffice to fix the three-loop core weights. Appendix B3 gives a more general zero family.

Four-loop zeros

At four loops, the explicit zeros are given in Appendix B4.

Lifting uniqueness by boundary zeros

We first explicitly check uniqueness at nL=2=nL=3=3n_{L=2}=n_{L=3}=3 and nL=4=4n_{L=4}=4, then extend it by a residue argument similar to Ref. [39].

The finite starting cases

At n=nLn=n_L, we proved the boundary zeros leave one weight for each independent core attachment sum and sign sector. We impose the loop-zero assignments introduced in the previous section on these complete nLn_L-point sums, which can be verified to leave one normalization in each case:

MnL,L=anL,LInL,L,(15)M_{n_L,L}=a_{n_L,L}\mathcal{I}_{n_L,L}, \tag*{(15)}

for n2=n3=3n_2=n_3=3, and n4=4n_4=4. At two loops the two zero families reduce the sixteen weights left by the three-point boundary equations to one. At three loops, the five vacuum topologies in Figure 5, with their embeddings, exterior-face choices and puncture labels, give 59 independent core sums per sign sector. The six three-loop families fix their 59×8=47259 \times8=472 weights. At four loops, the zero families in Appendix B4, with their puncture and sign images and boundary insertions, fix the 1624×16=259841624 \times16=25984 four-point weights.

Figure 5. The five cubic vacuum topologies at three loops. The core weights also distinguish planar embeddings, the exterior face and puncture labels.

Adding one external particle

Assume uniqueness at nn points, with n≥nLn\ge n_L. Taking a Di,2D_{i,2} residue of an (n+1)(n+1)-point ansatz removes a two-particle tree and leaves an nn-point ansatz at the same loop order, as in Figure 6. Both the boundary and loop zeros descend to this residue, in the sense that they act as lower-point zeros. Induction then gives

Figure 6. Removing the two-particle tree at the orange propagator preserves the core and its parent weight wr,n+1w_{r,n+1}.

Rn,L:=Res⁡Di,2=0Mn+1,L=an+1,LIn,L,(16)R_{n,L}:=\operatorname{Res}_{D_{i,2}=0}M_{n+1,L}=a_{n+1,L}\mathcal{I}_{n,L}, \tag*{(16)}

after relabelling the remaining particles. Repeating this for every ii fixes all two-particle residues, and the boundary equations relate their normalizations.

These residues also determine diagrams without a DD propagator, which are missed by the above argument. The boundary equations have already tied every attachment to the weight of its core. Removing a two-particle tree preserves those weights, and the resulting lower-point core sums are independent for n≥nLn\ge n_L. Equation (16) therefore fixes each core weight to its reference value times an+1,La_{n+1,L}, which determines all its attachments in the parent. Appendix A gives a more detailed argument.

Together with the finite checks for n<nLn<n_L, this proves

Mn,L=an,LIn,L,for L=2,3,4,(17)M_{n,L}=a_{n,L}\mathcal{I}_{n,L},\qquad\text{for }L=2,3,4, \tag*{(17)}

so that zeros have uniquely fixed the local ansatz, up to overall normalizations. Lower multiplicities are checked directly using the zeros in Appendices B2, B3 and B4. The accompanying Wolfram Language package checks L=2,3L=2,3 at n=1,2,3n=1,2,3, using the main-text families where sufficient and the additional low-point families in Appendix B [59].

Factorization and loop cuts from zeros

Equation (17) fixes every relative residue and cut to that of the reference graph sum. At a simple separating pole Q=DQ=D or XX,

Res⁡Q=0Mn,L=an,LInA,LAInB,LB,(18)\operatorname{Res}_{Q=0}M_{n,L}=a_{n,L}\mathcal{I}_{n_A,L_A}\mathcal{I}_{n_B,L_B}, \tag*{(18)}

with nA+nB=n+2n_A+n_B=n+2 and LA+LB=LL_A+L_B=L. A DD cut gives a tree factor times a loop factor; in particular, Di,2D_{i,2} leaves In−1,L\mathcal{I}_{n-1,L} because the three-point tree is one. An XX cut separating loop regions gives two factors with positive loop order. Other separating cuts likewise give products on the two resulting surfaces.

Cutting a ray Yi,pY_{i,p} opens puncture pp, giving an (n+2)(n+2)-point graph sum at loop order L−1L-1, with inherited channel values and signs. Since Eq. (17) fixes the coefficients of the graph sum, it also fixes these cut sums and their relative multiplicities. Factorization and loop cuts therefore follow from the zeros, without imposing either as a condition on the ansatz.

Outlook

We have shown that up to four loops two kinds of regular cancellation can determine the relative weights of perturbative diagrams. Boundary zeros first fix the external attachments to each loop core, and deformations of the Airy zero then relate the remaining core and sign weights. The boundary construction and the Airy zero are available at arbitrary loop order, while the general bound NL=10L−12N_L=10L-12 ensures finite multiplicity checks can start the inductive step at any finite loop order. The same induction argument can therefore establish uniqueness at every higher multiplicity. This already makes higher loop proofs accessible, while further optimizations of the construction can realistically lead to an all loop-order proof.

An all-orders uniqueness theorem would show that the full hierarchy of relative factorization and cut relations follows from hidden zeros at regular kinematics. This would establish hidden zeros as a new organizing principle for perturbative scattering. Extending the same mechanism to other theories is a next step of this program.

Understanding deformations of the Airy function directly in the matrix model could make their construction systematic. The factorization relations of surface integrands may in turn suggest new ways to organize ribbongraph correlators and their relation to intersection theory and two-dimensional gravity [48, 60, 61, 62].

Acknowledgments LR is supported by the National Natural Science Foundation of China General Program No. 12475070 and the Beijing Natural Science Foundation International Scientist Project No. IS24014. AI tools were used as research and coding assistants for exploratory calculations, checks, and manuscript editing.

References

Boundary zeros, core reduction and lifting

We give the boundary argument at arbitrary multiplicity. Summing all insertions of one external leg proves the reference zero; comparing pole coefficients of the weighted sum then relates its parent diagrams. We use these relations to reduce the ansatz to core weights and to lift the finite uniqueness results.

The boundary identity at arbitrary multiplicity

We prove that the reference integrand satisfies the boundary zero at any multiplicity and loop order. The proof groups diagrams according to the graph obtained by removing one external leg. Within each group, the diagrams differ only in where that leg is attached, and their propagator products cancel by telescoping. We work throughout in a fixed puncture-sign sector.

The zero locus. For the boundary zero associated with adjacent marks n−1,nn-1,n, choose η≠0\eta\ne0. At n≥3n \ge3, impose the following relations among the nn-point inverse propagators:

Yn,ps−Yn−1,ps=η,Y^{s}_{n,p}-Y^{s}_{n-1,p}=\eta,
Xi,n,S−Xi,n−1,S=η(1≤i≤n−2),X_{i,n,S}-X_{i,n-1,S}=\eta\qquad(1\le i\le n-2),
Xn−1,n,S−Xn−1,n−1,S=η,X_{n-1,n,S}-X_{n-1,n-1,S}=\eta,
Xn,n,S−Xn−1,n−1,S=2η,X_{n,n,S}-X_{n-1,n-1,S}=2\eta,
Dn−1,2=−η,D_{n-1,2}=-\eta,
Dn−2,2=η,D_{n-2,2}=\eta,
Di,n−i−Di,n−1−i=η(1≤i≤n−3),D_{i,n-i}-D_{i,n-1-i}=\eta\qquad(1\le i\le n-3),
Dn,d−Dn−1,d+1=η(2≤d≤n−1),D_{n,d}-D_{n-1,d+1}=\eta\qquad(2\le d\le n-1),
Dn,n−Dn−1,n=2η.(19)D_{n,n}-D_{n-1,n}=2\eta. \tag*{(19)}

The Y,XY,X relations apply to every puncture, puncture sign and allowed subset SS. All remaining channels, including every Z,TZ,T, are free. At two points, use the same Y,XY,X relations and set D1,2=−ηD_{1,2}=-\eta, D2,2=ηD_{2,2}=\eta. Cyclic relabelling gives the other boundary zeros. All propagators are kept nonzero.

Grouping the diagrams. Remove the selected external leg from an nn-point diagram. The result is an (n−1)(n-1)-point graph at the same loop order, which we call the child. All diagrams that give this same child form one group of parents. Hatted variables used below denote channels of the child.

To recover the parents, insert the removed leg in every position compatible with its original place in the cyclic external ordering. The new leg must remain between the same two neighboring external particles. In the child, these particles bound a segment of the exterior face, shown schematically in Figure A3.

Figure A3. All allowed positions of one new external leg between two fixed neighboring external particles. Each orange segment represents a separate parent diagram. The interior positions split an existing propagator. The two end positions attach to a neighboring external leg and create a two-particle tree. Both end insertions point inward to preserve the cyclic ordering. The rest of the child graph is arbitrary.

There are two kinds of insertion. On an internal edge along this segment, add a cubic vertex and attach the new external leg to it. The original propagator is then replaced by two propagators. Alternatively, attach the new leg to either of the two neighboring external legs. This creates a cubic vertex carrying two external legs, connected to the rest of the graph by a new propagator. These are the two external-tree attachments, with channels Dn−1,2=−ηD_{n-1,2}=-\eta and Dn−2,2=ηD_{n-2,2}=\eta.

Number the internal-edge insertions 1,…,m1,\ldots,m, starting from the end with tree channel −η-\eta. Label the two external-tree insertions 00 and m+1m+1. If an edge occurs twice along this boundary, its two insertion positions are counted separately. For a one-point child, the two end positions are on opposite sides of its sole external leg.

The telescoping cancellation. Let P0P_0 be the product of propagators in the first parent after removing its new tree factor 1/Dn−1,21/D_{n-1,2}. Now pass through the ordered internal insertion positions. At each position, the relations in Eq. (A1) change one inverse propagator by η\eta. Define PkP_k from Pk−1P_{k-1} by making this change at position kk.

More explicitly, let QkQ_k be the inverse propagator before this change, and let BkB_k be the product of all the other propagators. Then

Pk−1=BkQk,Pk=BkQk+η.P_{k-1}=\frac{B_k}{Q_k},\qquad P_k=\frac{B_k}{Q_k+\eta}.

The parent obtained by inserting the leg at position kk contains both propagators 1/Qk1/Q_k and 1/(Qk+η)1/(Q_k+\eta). Its contribution is therefore

Ak=BkQk(Qk+η)=Pk−1−Pkη.(20)A_k=\frac{B_k}{Q_k(Q_k+\eta)}=\frac{P_{k-1}-P_k}{\eta}. \tag*{(20)}

Thus each internal insertion gives the difference of two successive products. The negative PkP_k from position kk cancels the positive PkP_k from position k+1k+1.

The two external-tree parents supply the terms needed at the ends of this sequence:

A0=−P0η,Am+1=Pmη.A_0=-\frac{P_0}{\eta},\qquad A_{m+1}=\frac{P_m}{\eta}.

Consequently the whole group vanishes,

∑k=0m+1Ak=−P0η+∑k=1mPk−1−Pkη+Pmη=0.(21)\sum_{k=0}^{m+1}A_k=-\frac{P_0}{\eta}+\sum_{k=1}^{m}\frac{P_{k-1}-P_k}{\eta}+\frac{P_m}{\eta}=0. \tag*{(21)}

The child graph was arbitrary, so the cancellation holds independently of its multiplicity and loop structure. Every parent belongs to the group selected by deleting the chosen leg. Summing these vanishing groups therefore proves the boundary zero for the full reference integrand. When different diagrams have the same projected denominator product, their contributions combine with their graph multiplicities.

For the lifting argument, we also need the same sum with Dn−1,2=s0D_{n-1,2}=s_0 left independent. Keep every other relation in Eq. (19) unchanged. Only the first external-tree contribution changes, giving

Figure A1. The channel variables. Open circles are boundary marks; filled dots are punctures. Shading selects the boundary-interval side of DD, XX and the enclosed subset of TT. Thus (c) shades the outside of the full-circuit D1,5D_{1,5}, while (d) shades the inside of X11,{1}X_{11,\{1\}}. The winding routes in (g,h) share Y1,1sY^{s}_{1,1}.

Figure A2. The two three-loop mirror pairs. Each face retains its label and sign. Reflection reverses the boundary order while transporting the curves and their chosen sides. The two pairs reduce 61 core weights to 59 per sign sector.

P0s0+∑k=1mPk−1−Pkη+Pmη=(1s0+1η)P0.(22)\frac{P_0}{s_0}+\sum_{k=1}^{m}\frac{P_{k-1}-P_k}{\eta}+\frac{P_m}{\eta}=\left(\frac{1}{s_0}+\frac{1}{\eta}\right)P_0 . \tag*{(22)}

Summing over children and puncture signs gives

In,L∣split=(1s0+1η)In−1,L(Q^),(23)\mathcal{I}_{n,L\mid\mathrm{split}}=\left(\frac{1}{s_0}+\frac{1}{\eta}\right)\mathcal{I}_{n-1,L}(\widehat{Q}) , \tag*{(23)}

where Q^\widehat{Q} are the child channel values. This is the four-point-tree specialization of the tree–loop split in Ref. [8]. The prefactor is the cubic tree amplitude A4(s0,η)=1/s0+1/ηA_4(s_0,\eta)=1/s_0+1/\eta. Setting s0=−ηs_0=-\eta therefore gives every boundary zero used here, at arbitrary multiplicity and loop order, with the lower-point factor regular. Keeping s0s_0 independent gives the insertion identity used in the lifting argument below.

Coefficient comparisons for arbitrary parents

We now apply the identity to unknown diagram coefficients. Take residues in unchanged propagators to select the child structure being compared. Weight its interior insertions by c1,…,cmc_1,\ldots,c_m and its two external-tree insertions by c0,cm+1c_0,c_{m+1}. Equation (A4) at s0=−ηs_0=-\eta becomes

1η[(c1−c0)P0+∑k=1m−1(ck+1−ck)Pk+(cm+1−cm)Pm]=0.(24)\frac{1}{\eta}\left[(c_1-c_0)P_0+\sum_{k=1}^{m-1}(c_{k+1}-c_k)P_k+(c_{m+1}-c_m)P_m\right]=0 . \tag*{(24)}

Distinct shifted poles compare neighboring coefficients. When the products are distinguished by these poles, the comparisons give c0=c1=⋯=cm+1c_0=c_1=\cdots=c_{m+1}. This is the tree-level diagram-subset argument [38] applied to an arbitrary loop-containing child.

The winding projection can identify some of the products in Eq. (24). We then use another deleted leg or another pole to separate the coefficients. At low multiplicity there may be too few such comparisons. The next step determines when the boundary equations nevertheless fix every attachment to a given core.

Reducing the remaining freedom to core weights

We now show that the boundary equations leave only the weights of complete core attachment sums, once the multiplicity is sufficiently large. We first determine the terms containing external trees, then exclude any remaining terms.

Terms containing external trees. A residue in Di,2D_{i,2} removes a two-particle tree and replaces it by one external leg. On a boundary zero, two neighboring triangle channels take the values η\eta and −η-\eta. They cannot occur in the same triangulation, so the coefficient of 1/η1/\eta equates their residues, with the remaining particles labelled in the same order. Repeating this comparison relates all triangular residues.

These residues satisfy the lower-point boundary equations. For example, for n≥4n\geq4 the boundary assignment gives

Dn,2=Q+η,Dn−1,3=Q.(25)D_{n,2}=Q+\eta,\qquad D_{n-1,3}=Q. \tag*{(25)}

Here QQ is an independent channel value. Taking the pole at Q=−ηQ=-\eta selects the residue in Dn,2D_{n,2}. After removing that two-particle tree, the remaining propagators satisfy the boundary assignment at n−1n-1 points. This also applies at four points.

Suppose, by induction, that the common residue is a linear combination of complete core sums. Attach the removed particle in all positions on those same cores, with the same weights, and subtract the resulting nn-point sum from the ansatz. The difference has zero residue in every Di,2D_{i,2}. Restoring a specified two-particle tree is unique, so these residues eliminate every term containing a triangle channel. Every external tree contains two external legs meeting at a cubic vertex, and hence such a channel. The difference therefore contains no DD propagators.

Terms without external trees. Consider a term with three consecutive legs attached to the same side of a core edge. Delete the middle leg and apply the insertion comparisons of Eq. (24). The retained neighbors fix the insertion order. Since the external-tree terms are absent, the two end coefficients are c0=cm+1=0c_0=c_{m+1}=0, and the pole equations force the intervening coefficients to vanish as well. When propagators repeat, the same comparison uses the highest-order poles after partial fractions in the lower-point channels.

A surviving term can thus have at most two legs on each side of a core edge facing the exterior. A cubic vacuum core has 3L−33L-3 edges, with 6L−66L-6 sides in total. Its LL internal faces use at least one side each, leaving at most 5L−65L-6 exterior sides. Consequently, for L≥2L\geq2, a nonzero remainder is possible only when

n≤NL:=2(5L−6)=10L−12.(26)n\leq N_L:=2(5L-6)=10L-12. \tag*{(26)}

It remains to solve a finite set of boundary equations. For each core, distribute at most two legs on each exterior side and substitute every cyclic placement of Eq. (19). Partial fractions in the independent channel variables give a linear equation for each distinct pole product. Full column rank of this system means that no nonzero combination of the remaining diagrams survives. The actual starting multiplicity nLn_L is the point above which these equations exclude every remainder; it can be much smaller than NLN_L.

The starting multiplicities. At two loops, the finite equations exclude all remainders for 3≤n≤83\leq n\leq8.

At three loops, the five topologies in Figure 5 permit a direct case analysis. First consider cores without a bridge bordering the exterior face. If only one puncture borders that face, the boundary assignment leaves its rays arbitrary, so the remainder vanishes. If two punctures do, choose qq to occur on only one exterior side and call the other pp. After clearing denominators, the numerator is at most linear in each Yi,qY_{i,q}. On each cyclic boundary zero,

di=di+1,di=Yi,q−Yi,p.d_i=d_{i+1},\qquad d_i=Y_{i,q}-Y_{i,p}.

The numerator must therefore contain ∏i(di−di+1)\prod_i(d_i-d_{i+1}), which is quadratic in each Yi,qY_{i,q} for n≥3n\geq3. It must be zero.

For the exterior triangle, at most two legs per side leaves only (2,2,0)(2,2,0), (2,1,1)(2,1,1), (2,2,1)(2,2,1) and (2,2,2)(2,2,2) above three points. Deleting an end leg and comparing pole coefficients eliminates these placements.

The four sides of the remaining core border internal faces 1,2,3,21,2,3,2, in order around the exterior. The same comparisons remove all placements except one with a leg on each side. For this last case, apply the boundary zero at marks 3,43,4. After suppressing the common internal factors and the rays at marks 1,21,2, the two products contributing to the relevant pole are

cAY3,2Y3,3Y4,2Y4,3+cB(Y3,2)2Y3,3Y4,2.(27)\frac{c_A}{Y_{3,2}Y_{3,3}Y_{4,2}Y_{4,3}} + \frac{c_B}{(Y_{3,2})^2Y_{3,3}Y_{4,2}}. \tag*{(27)}

Set Y4,2=Y3,2+ηY_{4,2}=Y_{3,2}+\eta and Y4,3=Y3,3+ηY_{4,3}=Y_{3,3}+\eta. The coefficient of 1/(Y3,2Y3,3)1/(Y_{3,2}Y_{3,3}) in partial fractions is (cA−cB)/η2(c_A-c_B)/\eta^2, so cA=cBc_A=c_B. The second product has two legs on one side bordering puncture 2 and was already eliminated by the preceding pole comparisons. Hence the last coefficient vanishes as well.

For cores with bridges, the same insertion argument limits each attached one-puncture component to two legs. In a chain, the pole comparisons can fail only when the deleted leg was the sole attachment to an end component. There are at most two such legs. For n≥4n \ge4, another choice of deleted leg therefore sets the coefficient to zero.

At four loops, the bound of two legs applies to an entire tadpole together with the edge connecting it to the rest of the core. This branch occupies three exterior sides: the tadpole and the two sides of its connecting edge. If a core has mm exterior sides and tt such disjoint branches, the bound is therefore

n≤2(m−3t)+2t=2(m−2t).(28)n \le2(m-3t)+2t=2(m-2t). \tag*{(28)}

The four-loop core list gives m−2t≤8m-2t \le8, reducing the range from N4=28N_4=28 to n≤16n \le16. Applying the finite boundary equations above to these allowed distributions gives full column rank modulo 65521 for 4≤n≤164 \le n \le16, and hence also over the rationals. These remainder calculations are separate from the low-point uniqueness checks in the accompanying WL package.

Together with the finite starting checks, these arguments give n2=n3=3n_2=n_3=3 and n4=4n_4=4. At one loop the remainder is simply c/∏iYi,1sc/\prod_i Y_{i,1}^{s}, which a boundary zero forces to vanish, giving n1=2n_1=2.

Below these multiplicities, boundary zeros can leave additional coefficients. At two loops the boundary equations leave 7, 5, 4 weights per sign at n=1,2,3n=1,2,3, while the core sums have dimension four. At four loops, the three-point boundary space has 1644=1624+201644=1624+20 weights per sign, whereas four points leave precisely the 1624 core weights.

Passing a loop zero to the residue

We now explain how to construct the higher-point loop conditions used in the main-text induction. The boundary equations have already restricted the parent ansatz. Apply the insertion assignment to Mn+1,LM_{n+1,L}, keeping the new triangle channel s0s_0 independent. Only terms containing this triangle depend on s0s_0, and it occurs once in each such term. Hence the result is

Mn+1,L∣split=Rn,L(Q^)s0+C,(29)\left.M_{n+1,L}\right|_{\mathrm{split}}=\frac{R_{n,L}(\widehat{Q})}{s_0}+C, \tag*{(29)}

where Rn,LR_{n,L} is the triangle residue and CC does not depend on s0s_0. The notation means substitution of Eq. (19), with n+1n+1 parent particles. Its boundary-zero value s0=−ηs_0=-\eta fixes C=Rn,L/ηC=R_{n,L}/\eta. Therefore

Mn+1,L∣split=(1s0+1η)Rn,L(Q^).(30)\left.M_{n+1,L}\right|_{\mathrm{split}}=\left(\frac{1}{s_0}+\frac{1}{\eta}\right)R_{n,L}(\widehat{Q}). \tag*{(30)}

This uses the boundary equations to express the other insertions in terms of the residue.

Choose any required nn-point loop zero and put the hatted channels on it. For the reference sum, Rn,L=In,LR_{n,L}=I_{n,L}, so the right-hand side vanishes. This constructs a valid (n+1)(n+1)-point zero. Impose it on the unknown parent with s0≠−ηs_0\ne-\eta. The prefactor is nonzero, and the residue must obey the chosen lower-point loop zero. Repeating for all required families puts the residue under the full lower-point uniqueness conditions.

Finally, boundary reduction has written the parent as ∑rwr,n+1Cr,n+1,L\sum_r w_{r,n+1}C_{r,n+1,L}. Removing a two-particle tree from a complete attachment sum gives the lower-point sum on the same core, with the same parent weight. Thus lower-point uniqueness gives

∑rwr,n+1Cr,n,L=an+1,LIn,L.(31)\sum_r w_{r,n+1}C_{r,n,L}=a_{n+1,L}I_{n,L}. \tag*{(31)}

For n≥nLn\ge n_L, these core sums are independent, so each parent weight is fixed to its reference value times an+1,La_{n+1,L}. This fixes all parent attachments, including those without a DD propagator. The argument concerns the weights of this parent ansatz; its overall normalization is still free at each multiplicity.

Explicit loop-zero families

We first explain how external insertions extend the Airy zero to arbitrary multiplicity, then list the larger families containing the main-text zeros and the additional assignments used at low multiplicity. The construction starts from the Airy face sums. Connections between punctures and the returns around them are varied together; cutting a return identifies groups whose cancellation allows that propagator to remain free. The prescriptions below specify the resulting zeros. All free parameters are chosen away from propagator poles, and a sign exchange acts on every channel ending at the chosen puncture.

Adding external legs to the Airy zero

We show how the vacuum Airy cancellation gives the all-nn assignment in Eq. (10). We first attach legs directly to the loop graph, without external trees. Let Hn,LH_{n,L} be the sum of these diagrams on the Airy assignment in a fixed puncture-sign sector, and let AL+1A_{L+1} be the corresponding vacuum sum. Their internal-face weights are left implicit.

Adding the first leg splits one outer edge into two. The factor 1/(b+sap)1/(b+s_{a p}) becomes 1/(b+sap)2=−∂b[1/(b+sap)]1/(b+s_{a p})^2=-\partial_b[1/(b+s_{a p})]. Differentiating the vacuum sum therefore adds a leg at every position along the outer face. An edge bordering that face on both sides contributes twice, once for each side. Thus H1,L=−∂bAL+1H_{1,L}=-\partial_b A_{L+1}.

Adding another leg divides one boundary region into two, to which we assign independent weights b,b′b,b'. split edge now contributes

1(b+sap)(b′+sap)=1b′−b(1b+sap−1b′+sap).(32)\frac{1}{(b+sa_p)(b^{\prime}+sa_p)}=\frac{1}{b^{\prime}-b}\left(\frac{1}{b+sa_p}-\frac{1}{b^{\prime}+sa_p}\right). \tag*{(32)}

This is why differences appear: they express the two propagators created by inserting a leg. Figure A4 shows the two possible insertions on a pair of adjacent edges. Applying Eq. (32) to the split edge in each diagram cancels the mixed product 1/[(b+sap)(b′+taq)]1/[(b+sa_p)(b^{\prime}+ta_q)]. Only the original product evaluated at bb and at b′b^{\prime} remains. The same cancellation continues along a longer sequence of edges, giving

Figure A4. Adding one external leg to two adjacent edges on the outer face. The orange leg can attach to either edge, dividing the boundary region of weight bb into regions of weights b,b′b,b^{\prime}. The internal faces retain weights sapsa_p, taqta_q, and the rest of the graph is unchanged. Summing these two insertions gives the difference of the original two-edge product at bb and b′b^{\prime}, divided by b′−bb^{\prime}-b. The same cancellation applies to any number of adjacent edges.

Hn+1,L(…,b,b′)=Hn,L(…,b)−Hn,L(…,b′)b′−b.(33)H_{n+1,L}(\ldots,b,b^{\prime})=\frac{H_{n,L}(\ldots,b)-H_{n,L}(\ldots,b^{\prime})}{b^{\prime}-b}. \tag*{(33)}

Deleting the new leg reverses the construction, so the insertions generate every higher-point diagram with the chosen cyclic ordering. We repeat this operation whenever we add a particle, introducing one new boundary weight each time. Every new edge still carries the sum of its adjacent face weights, so the resulting diagrams are precisely on the higher-point Airy assignment. The differences act only on boundary weights and preserve oddness in each internal-face weight. The puncture-sign sum therefore vanishes after every insertion. Coincident boundary weights follow by taking b′→bb^{\prime}\to b, when the difference becomes a derivative.

At two loops, Eq. (9) gives the explicit example

H1,2st(b)=st2b2apaq,(34)H_{1,2}^{st}(b)=\frac{st}{2b^{2}a_pa_q}, \tag*{(34)}
H2,2st(b,b′)=st(b+b′)2b2(b′)2apaq.H_{2,2}^{st}(b,b^{\prime})=\frac{st(b+b^{\prime})}{2b^{2}(b^{\prime})^{2}a_pa_q}.

Both retain the same factor stst, so summing the four sign sectors gives zero. Further insertions preserve this sign dependence in the same way.

Finally, fix an external-tree structure. Removing its trees leaves one of these vanishing sums at the surviving attachment marks. Restoring the trees multiplies that sum by their product of D−1D^{-1} factors. Summing over tree structures proves the full zero, with every DD independent.

Two loops

The full mountain contains the two simple deformations used in the main text. For ordered punctures p,qp,q, choose independent uiu_i, viv_i, hh, zz, ee and assign

\begin{aligned} Y_{i,p}&=u_i v_i+h,\\ Y_{i,q}&=u_i+h v_i, \end{aligned} \qquad \left(Z_{pq}^{st}\right)= \begin{pmatrix} z-h & \\ h & e \end{pmatrix}. \tag*{(35)}\end{aligned}
Xij,{p}=ui+vj+h,Xij,{q}=uj+vi+h,Tp,{q}+=z−h,Tp,{q}−=e+h,Tq,{p}+=z+h,Tq,{p}−=e−h.(36)\begin{aligned} X_{ij,\{p\}}&=u_i+v_j+h,\\ X_{ij,\{q\}}&=u_j+v_i+h,\\ T_{p,\{q\}}^{+}&=z-h,\qquad T_{p,\{q\}}^{-}=e+h,\\ T_{q,\{p\}}^{+}&=z+h,\qquad T_{q,\{p\}}^{-}=e-h. \tag*{(36)} \end{aligned}

The rows and columns of ZZ carry s,t=+,−s,t=+,-; all DD remain free. Set ap=(z−h)/2a_p=(z-h)/2, aq=(z+h)/2a_q=(z+h)/2, bi=(ui+vi+h)/2b_i=(u_i+v_i+h)/2, ξi=(ui−vi−z)/2\xi_i=(u_i-v_i-z)/2 and ϵ=e+z\epsilon=e+z. Airy is ξi=ϵ=0\xi_i=\epsilon=0. The common-shift family has ξi=ϵ=0\xi_i=\epsilon=0; exchanging the qq sign gives the opposite-shift family.

The internal rule is particularly simple: each return is a row or column sum of the ZZ matrix. Allowing independent uiu_i, viv_i then separates the two boundary orientations while preserving the cancellation. The four core contributions cancel for arbitrary ordered external insertions, and restoring external trees leaves every DD free.

The common- and opposite-shift families suffice at every multiplicity used in the finite checks. At n=1n=1, their 28 signed coefficients reduce to one; at n=2n=2, both boundary placements leave 20 coefficients, which the two families again reduce to one. At n=3n=3, they fix the sixteen remaining core weights. No additional low-point zero is needed.

Three loops

We give the choices used in the main text and the larger families needed for the one- and two-point checks. Write EγE_\gamma for the uniform internal shift in Eq. (13), with one selected sign γp=±1\gamma_p=\pm1 per puncture. Write RpsR^s_{p} for a single-return displacement in Eq. (14). The six sufficient three-point choices are

E+++, E++−, E+−+,R1,{3}−, R2,{1}+, R3,{2}+.(37)E_{+++},\ E_{++-},\ E_{+-+}, \qquad R^{-}_{1,\{3\}},\ R^{+}_{2,\{1\}},\ R^{+}_{3,\{2\}}. \tag*{(37)}

No further images are required for this choice. Each selected return receives the same displacement, independently of the size of its enclosed subset.

There is a larger version of the collective family: all six returns enclosing both other punctures can remain independently free. To see why, group the vacuum diagrams by these propagators. Each occurs at most once, and the coefficient of each is a lower-loop cut sum that vanishes on the deformed assignment; the part containing none vanishes as well. For a single-return displacement starting from Airy, the same cut isolates a sign-summed lower-loop cancellation. This explains why varying that return preserves the zero. Ordered external insertions and independent external-tree factors extend these identities to arbitrary multiplicity.

Additional finite families. For the low-point checks, choose h1=0h_1 = 0, h2=hh_2 = h and independent AA, BB, hh, h3h_3, UiU_i, ViV_i. Begin with

Yi,p+=Ui+hp,Y−i,p−=Vi−hp,Zpq++=A+hp+hq,Z−pq−−=B−hp−hq,Zpq+−=hp−hq,Zpq−+=hq−hp,Tp,{q}+=A+2hp,Tp,{q}−=B−2hp.(38)\begin{aligned} Y^{+}_{i,p} &= U_i + h_p, & Y^{-}_{-i,p} &= V_i - h_p,\\ Z^{++}_{pq} &= A + h_p + h_q, & Z^{--}_{-pq} &= B - h_p - h_q,\\ Z^{+-}_{pq} &= h_p - h_q, & Z^{-+}_{pq} &= h_q - h_p,\\ T^{+}_{p,\{q\}} &= A + 2h_p, & T^{-}_{p,\{q\}} &= B - 2h_p. \tag*{(38)} \end{aligned}
Xij,S={Ui+Vj,i=j or 1∈S,Uj+Vi,i<j and 1∉S.(39)X_{ij,S} = \begin{cases} U_i + V_j, & i = j\ \text{or}\ 1 \in S,\\ U_j + V_i, & i < j\ \text{and}\ 1 \notin S. \end{cases} \tag*{(39)}

The connection formulas use p<qp < q. The following completion also replaces Y1,1+Y^{+}_{1,1} by an independent ww:

Z12++=A+h+ϵ,T1,S+=A+ϵ(2∈S),T2,S+=A+2h+ϵ(1∈S).(40)\begin{aligned} Z^{++}_{12} &= A + h + \epsilon,\\ T^{+}_{1,S} &= A + \epsilon&& (2 \in S),\\ T^{+}_{2,S} &= A + 2h + \epsilon&& (1 \in S). \tag*{(40)} \end{aligned}

The four other larger returns, T1,{2,3}−T^{-}_{1,\{2,3\}}, T2,{1,3}−T^{-}_{2,\{1,3\}} and T3,{1,2}−T^{-}_{3,\{1,2\}}, remain free.

At one point use two families. The first has no completion, (U1,V1)=(U,V)(U_1,V_1) = (U,V), and leaves all six larger returns and all three X11,SX_{11,S} with ∣S∣=2|S| = 2 independent; singleton XX values are U+VU + V. The second uses the completion with (U1,V1)=(0,−A)(U_1,V_1) = (0,-A), w=A+hw = A + h and ϵ=h\epsilon= h, so every X=−AX = -A. Include all puncture and sign images. Rational substitution proves both identities; together they reduce the 1008 coefficients to one.

At two points use the completion with (U1,V1)=(0,−A)(U_1,V_1) = (0,-A), (U2,V2)=(u,v)(U_2,V_2) = (u,v) and D1,2=−uD_{1,2} = -u. The parameters ww, ϵ\epsilon, D2,2D_{2,2} and the four larger returns are independent. Include both boundary placements and all puncture and sign images. With boundary zeros, this family fixes the two-point ansatz.

The following larger family is an alternative for n≥3n \ge3; it remains available in the package, while the default three-point check uses the six simpler families above:

\begin{array}{c|cc} \text{mark} & U_i & V_i\\ R = 1 & 0 & -A\\ I = 2,\ldots,n-1 & A + v_i & v_i\\ J = n & A - h + v_n & v_n \end{array} \qquad \epsilon= h,\quad w = A + h. \tag*{(41)}\end{array}

Use Eqs. (38)–(40); every viv_i, DD and the four larger returns are independent. Include cyclic placements and all puncture and sign images. The RR, II, JJ profiles coordinate the modified ray with the boundary curves. Substitution gives the three-point zero, including the sums obtained by deleting external trees. Splitting an II mark repeats its profile and gives a divided difference in vv, extending the identity to every RIn−2JRI^{n-2}J.

Four loops

We list three families at three points and the separate one- and two-point prescriptions. Their boundary insertions give the four-point starting zeros and the higher-point families. Let P={1,2,3,4}P = \{1,2,3,4\}, h1=0h_1 = 0, h2=hh_2 = h, with h3h_3, h4h_4 independent.

Three-point families. Choose a distinguished puncture r∈Pr \in P and set

Yi,p+=Ui+hp,Yi,p−=Vi−hp,Zpq++=A+hp+hq,Zpq−−=B−hp−hq,Zpq+−=hp−hq,Zpq−+=hq−hp,Tp,{q}+=A+2hp,Tp,{q}−=B−2hp.(42)\begin{aligned} Y^{+}_{i,p} &= U_i + h_p, & Y^{-}_{i,p} &= V_i - h_p,\\ Z^{++}_{pq} &= A + h_p + h_q, & Z^{--}_{pq} &= B - h_p - h_q,\\ Z^{+-}_{pq} &= h_p - h_q, & Z^{-+}_{pq} &= h_q - h_p,\\ T^{+}_{p,\{q\}} &= A + 2h_p, & T^{-}_{p,\{q\}} &= B - 2h_p. \tag*{(42)} \end{aligned}
Xij,S={Ui+Vj,i=j or r∈S,Uj+Vi,i<j and r∉S.(43)X_{ij,S} = \begin{cases} U_i + V_j, & i = j\ \text{or}\ r \in S,\\ U_j + V_i, & i < j\ \text{and}\ r \notin S. \end{cases} \tag*{(43)}

Connections use p<qp < q and boundary endpoints i≤ji \le j. All six DD and all 32 signed returns with ∣S∣≥2|S| \ge2 are independent. Its relation to Airy is explicit under

ap=A/2+hp,bi=(Ui+Vi)/2,a_p = A/2 + h_p,\qquad b_i = (U_i + V_i)/2,
δi=(Ui−Vi−A)/2,Δ=A+B.(44)\delta_i = (U_i - V_i - A)/2,\qquad\Delta= A + B. \tag*{(44)}

Then Yi,p±=bi±(ap+δi)Y^{\pm}_{i,p} = b_i \pm(a_p + \delta_i), Zpq−−=−ap−aq+ΔZ^{--}_{pq} = -a_p - a_q + \Delta and Tp,{q}−=−2ap+ΔT^{-}_{p,\{q\}} = -2a_p + \Delta. The other ZZ channels and positive singleton returns keep their Airy values; the two XX values are bi+bj±(δi−δj)b_i + b_j \pm(\delta_i - \delta_j). Setting δi=Δ=0\delta_i = \Delta= 0 and the larger returns to 2sap2sa_p recovers Airy.

A second family deforms a pair a<ba < b and its associated returns:

Zab++=A+ha+hb+ϵ,Z^{++}_{ab}=A+h_a+h_b+\epsilon,
Ta,S+=A+2ha+ϵ(b∈S),T^{+}_{a,S}=A+2h_a+\epsilon\qquad(b\in S),
Tb,S+=A+2hb+ϵ(a∈S).(45)T^{+}_{b,S}=A+2h_b+\epsilon\qquad(a\in S). \tag*{(45)}

The replacements include larger returns. A third family uses r=1r=1 in the first assignment and frees one ray:

(U1,U2,U3)=(u,A+v,A+t−h),(U_1,U_2,U_3)=(u,A+v,A+t-h),
(V1,V2,V3)=(−A,v,t),Y1,1+=w,(46)(V_1,V_2,V_3)=(-A,v,t),\qquad Y^{+}_{1,1}=w, \tag*{(46)}
D1,3=−(A+v),D1,2=t−v.(47)D_{1,3}=-(A+v),\qquad D_{1,2}=t-v. \tag*{(47)}

Keep uu in every other channel; all other DD and larger returns stay free. Take the first family with r=1r=1, the pair 12 deformation with r=1r=1 or 3, and the free-ray family. Include their puncture permutations and coherent sign exchanges.

One and two points. The following conditions supply the separate low-point checks. Use the internal assignment above and the pair deformation, with

(A,h2,h3,h4)=λ(7,2,5,11)orλ(11,3,7,17),(48)(A,h_2,h_3,h_4)=\lambda(7,2,5,11)\quad\text{or}\quad\lambda(11,3,7,17), \tag*{(48)}

where λ≠0\lambda\ne0, while BB, ϵ\epsilon and the 26 unaffected larger signed returns are free. Include all puncture and sign images.

At one point deform pair 34, take independent U,VU,V, and use

Y1,p+=U+hp,Y1,p−=V−hp,Y^{+}_{1,p}=U+h_p,\qquad Y^{-}_{1,p}=V-h_p,
X11,S=U+Vif∣S∩{3,4}∣=1.(49)X_{11,S}=U+V\quad\text{if}\quad|S\cap\{3,4\}|=1. \tag*{(49)}

The six other XX channels remain free. The constrained XX curves separate the deformed pair. Cutting these curves divides the diagram into lower-loop pieces; their cancellations permit the other six XX channels to remain free.

At two points deform pair 12, set C={3,4}C=\{3,4\}, and use

Y1,1+=w,Y1,p−=−A−hp,Y^{+}_{1,1}=w,\qquad Y^{-}_{1,p}=-A-h_p,
Y1,p+=hp+τ1p∈C(p≠1),Y^{+}_{1,p}=h_p+\tau1_{p\in C}\qquad(p\ne1),
Y2,p+=A+v+hp,Y2,p−=v−hp.(50)Y^{+}_{2,p}=A+v+h_p,\qquad Y^{-}_{2,p}=v-h_p. \tag*{(50)}

Here 11 is the membership indicator. Let r(S)=1r(S)=1 if 1∉S1\notin S and S∩C≠∅S\cap C\ne\varnothing, and zero otherwise. With Sc=P∖SS^c=P\setminus S, set

X11,S=−A+τ[r(S)+r(Sc)],X_{11,S}=-A+\tau\left[r(S)+r(S^c)\right],
X12,S=v+τr(S),X22,{p}=A+2v,X_{12,S}=v+\tau r(S),\qquad X_{22,\{p\}}=A+2v,
D1,2=−(A+v).(51)D_{1,2}=-(A+v). \tag*{(51)}

The parameters vv, τ\tau, ww, D2,2D_{2,2} and the ten nonsingleton X22,SX_{22,S} are free. Include both boundary placements, complementing SS on X12,SX_{12,S} when exchanging marks, and impose ordinary boundary zeros. Substitution and grouping by the free channels proves the zero. These one- and two-point families each leave one normalization in the local ansatz.

Starting checks and package coverage. The three-point four-loop families fix the full local ansatz there. At four points, the boundary equations leave the core attachment sums, whose triangular residues are the corresponding three-point core sums. Insert one leg into the three-point zeros using Appendix A. They fix these same weights and establish the four-point starting case. The subsequent induction is the one already described above.

The program-only WL package checks n=1,2,3n=1,2,3 at two and three loops. It uses the two mountains throughout at two loops, the additional finite families at three-loop n=1,2n=1,2, and precisely Eq. (37) at three-loop n=3n=3. The four-loop calculations are separate. Every finite starting check leaves one overall normalization.

References

  1. [1]R. Britto, F. Cachazo, B. Feng, and E. Witten, Phys. Rev. Lett. 94, 181602 (2005), arXiv:hep-th/0501052.
  2. [2]Z. Bern, L. Dixon, D. C. Dunbar, and D. A. Kosower, N=1 super-Yang-Mills amplitudes, Nucl. Phys. B 425, 217 (1994), arXiv:hep-ph/9403226.
  3. [3]R. Britto, F. Cachazo, and B. Feng, Efficient pseudo, Nucl. Phys. B 725, 275 (2005), arXiv:hep-th/0412103.
  4. [4]N. Arkani-Hamed, Q. Cao, J. Dong, C. Figueiredo, and S. He, Positive geometry, J. High Energy Phys. 2024 (10), 231, arXiv:2312.16282.
  5. [5]F. Cachazo, N. Early, and B. Giménez Umbert, J. High Energy Phys. 2022 (08), 252, arXiv:2112.14191.
  6. [6]Q. Cao, J. Dong, S. He, and C. Shi, Phys. Lett. B 856, 138934 (2024), arXiv:2403.08855.
  7. [7]Q. Cao, J. Dong, S. He, C. Shi, and F. Zhu, J. High Energy Phys. 2024 (09), 049, erratum: J. High Energy Phys. 01 (2025) 126, arXiv:2406.03383.
  8. [8]N. Arkani-Hamed and C. Figueiredo, J. High Energy Phys. 2025 (10), 077, arXiv:2405.09608.
  9. [9]A. Guevara and Y. Zhang, New factorizations of Yang-Mills amplitudes (2024), arXiv:2406.08969.arxiv.org/abs/2406.08969
  10. [10]Y. Zhang, J. High Energy Phys. 2025 (02), 074, arXiv:2412.15198.DOI
  11. [11]C. R. T. Jones and S. Paranjape, J. High Energy Phys. 2025 (07), 251, arXiv:2505.02520.
  12. [12]T. Azevedo, H. Gomez, and R. Lipinski Jusinskas, J. High Energy Phys. 2026 (06), 001, arXiv:2512.20790.DOI
  13. [13]L. Zhang and K. Zhou, J. High Energy Phys. 2026 (06), 105, arXiv:2601.00297.arxiv.org/abs/2601.00297
  14. [14]Y. Li and L. Rodina, From Hidden Zeros to Exact Splittings and New Zeros (2026), arXiv:2610.01775.arxiv.org/abs/2610.01775
  15. [15]Y. Li, L. Rodina, and Z.-Z. Wang, Exact Splitting of Amplitudes and One-Loop Integrands: A Classical Origin of Zeros (2026), arXiv:2611.06810.arxiv.org/abs/2611.06810
  16. [16]K. Zhou, J. High Energy Phys. 2025 (03), 154, arXiv:2411.07944.arxiv.org/abs/2411.07944
  17. [17]K. Zhou, J. High Energy Phys. 2026 (09), 112, arXiv:2604.23680.arxiv.org/abs/2604.23680
  18. [18]B. Feng, L. Zhang, and K. Zhou, J. High Energy Phys. 2026 (02), 204, arXiv:2508.21345.arxiv.org/abs/2508.21345
  19. [19]B. Feng, L. Zhang, and K. Zhou, J. High Energy Phys. 2025 (08), 205, arXiv:2504.14215.arxiv.org/abs/2504.14215
  20. [20]C. Bartsch, T. V. Brown, K. Kampf, U. Oktem, S. Paranjape, and J. Trnka, Phys. Rev. D 111, 045019 (2025), arXiv:2403.10594.arxiv.org/abs/2403.10594
  21. [21]Y. Li, D. Roest, and T. ter Veldhuis, J. High Energy Phys. 2025 (04), 121, arXiv:2403.12939.arxiv.org/abs/2403.12939
  22. [22]Y. Li, T. Wang, T. Brauner, and D. Roest, Phys. Rev. Lett. 135, 181601 (2025), arXiv:2412.14858.arxiv.org/abs/2412.14858
  23. [23]H. Huang, Y. Yang, and K. Zhou, Eur. Phys. J. C 85, 685 (2025), arXiv:2502.07173.arxiv.org/abs/2502.07173
  24. [24]K. Zhou, J. High Energy Phys. 2026 (02), 039, arXiv:2510.11070.arxiv.org/abs/2510.11070
  25. [25]M. Carrillo González and F. Ward, Phys. Rev. D 113, 125018 (2026), arXiv:2601.16860.arxiv.org/abs/2601.16860
  26. [26]Y. Zhang and P. Zhu, Fermionic hidden zeros (2026), arXiv:2609.06552.arxiv.org/abs/2609.06552
  27. [27]Y.-C. Chang, H.-Y. Chen, and Y. t. Huang, All zeros of (super) String Theory (2025), arXiv:2506.15299.arxiv.org/abs/2506.15299
  28. [28]S. He, Y. Zhang, F. Zhu, and P. Zhu, String Amplitudes on Integer Rectangles (2026), arXiv:2610.31197.arxiv.org/abs/2610.31197
  29. [29]C. Bartsch, K. Kampf, J. Novotný, and J. Trnka, Phys. Rev. D 112, 085002 (2025), arXiv:2505.01501.arxiv.org/abs/2505.01501
  30. [30]S. De, S. Paranjape, A. Pokraka, M. Spradlin, and A. Volovich, J. High Energy Phys. 2025 (07), 174, arXiv:2503.23579.arxiv.org/abs/2503.23579
  31. [31]Y. Li and L. Rodina, Cosmological Wavefunctions as Amplitudes: Dual Shuffle Factorization and Uniqueness from New Hidden Zeros (2026), arXiv:2604.01133.
  32. [32]Q. Cao, Phys. Rev. Lett. 135, 021603 (2025), arXiv:2504.15702.DOI
  33. [33]L. Zhang, J. High Energy Phys. 2026 (01), 103, arXiv:2509.12564.arxiv.org/abs/2509.12564
  34. [34]B. Giménez Umbert and B. Sturmefols, J. High Energy Phys. 2025 (04), 076, arXiv:2501.10805.arxiv.org/abs/2501.10805
  35. [35]N. Early, J. High Energy Phys. 2025 (04), 030, arXiv:2502.08016.arxiv.org/abs/2502.08016
  36. [36]X. Li and K. Zhou, J. High Energy Phys. 2026 (01), 010, arXiv:2508.12894.arxiv.org/abs/2508.12894
  37. [37]N. Arkani-Hamed, C. Figueiredo, and G. N. Remmen, J. High Energy Phys. 2025 (04), 039, arXiv:2412.20639.
  38. [38]L. Rodina, Phys. Rev. Lett. 134, 031601 (2025), arXiv:2406.04234.DOI
  39. [39]J. V. Backus and L. Rodina, Phys. Rev. Lett. 135, 131601 (2025), arXiv:2503.03805.DOI
  40. [40]L. Rodina, G. Sellaroli, and Y. Ni, A Combinatorial Origin of Locality (2026), arXiv:2608.23008.arxiv.org/abs/2608.23008
  41. [41]N. Arkani-Hamed, L. Rodina, and J. Trnka, Phys. Rev. Lett. 120, 231602 (2018), arXiv:1612.02797.
  42. [42]L. Rodina, J. High Energy Phys. 2019 (09), 084, arXiv:1612.06342.DOI
  43. [43]C. Cheung, K. Kampf, J. Novotny, and J. Trnka, Phys. Rev. Lett. 114, 221602 (2015), arXiv:1412.4095.DOI
  44. [44]L. Rodina, Phys. Rev. Lett. 122, 071601 (2019), arXiv:1807.09738.DOI
  45. [45]L. Rodina, J. High Energy Phys. 2019 (09), 078, arXiv:1612.013885.
  46. [46]J. J. M. Carrasco and L. Rodina, Phys. Rev. D 100, 125007 (2019), arXiv:1908.08033.DOI
  47. [47]L. Rodina, Phys. Rev. D 102, 045012 (2020), arXiv:2005.06446.DOI
  48. [48]M. Kontsevich, Commun. Math. Phys. 147, 1 (1992).
  49. [49]N. Arkani-Hamed, H. Frost, G. Salvatori, P.-G. Plamondon, and H. Thomas, J. High Energy Phys. 2025 (08), 194, arXiv:2309.15913.
  50. [50]N. Arkani-Hamed, H. Frost, G. Salvatori, P.-G. Plamondon, and H. Thomas, J. High Energy Phys. 2025 (09), 033, arXiv:2311.09284.
  51. [51]N. Arkani-Hamed, Q. Cao, J. Dong, C. Figueiredo, and S. He, J. High Energy Phys. 2025 (04), 078, arXiv:2401.00041.
  52. [52]N. Arkani-Hamed, C. Figueiredo, H. Frost, and G. Salvatori, J. High Energy Phys. 2025 (05), 051, arXiv:2402.06719.
  53. [53]N. Arkani-Hamed, Q. Cao, J. Dong, C. Figueiredo, and S. He, Phys. Rev. Lett. 134, 171601 (2025), arXiv:2408.11891.
  54. [54]C. Bartsch, K. Kampf, J. Novotný, and J. Trnka, J. High Energy Phys. 2026 (10), 099, arXiv:2602.15102.arxiv.org/abs/2602.15102
  55. [55]N. Arkani-Hamed, Q. Cao, J. Dong, C. Figueiredo, and S. He, Phys. Rev. D 110, 065018 (2024), arXiv:2401.05483.
  56. [56]N. Arkani-Hamed and C. Figueiredo, J. High Energy Phys. 2025 (09), 189, arXiv:2403.04826.
  57. [57]N. Arkani-Hamed, H. Frost, and G. Salvatori, J. High Energy Phys. 2026 (06), 102, arXiv:2412.21027.
  58. [58]K. Zhou, J. High Energy Phys. 2026 (10), 030, arXiv:2604.13810.arxiv.org/abs/2604.13810
  59. [59]See Appendices A–B4 for the zero assignments and reconstruction arguments. The accompanying Wolfram Language package AiryUniqueness.wl checks L = 2, 3 at n = 1, 2, 3. The four-loop assignments and finite checks are described in Appendix B4.
  60. [60]E. Witten, Surveys Diff. Geom. 1, 243 (1991).
  61. [61]D. J. Gross and M. J. Newman, Nucl. Phys. B 380, 168 (1992), arXiv:hep-th/9112069.DOI
  62. [62]P. Di Francesco, C. Itzykson, and J.-B. Zuber, Commun. Math. Phys. 151, 193 (1993), arXiv:hep-th/9206090.DOI

Paper details

vlnc:W_42xujrkp
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 7

downloads · 1

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents