Counting Feynman Master Integrals from the Faces of a Polytope

A multi-loop Feynman integral, the basic building block of precision particle-physics predictions, is never computed alone. Physicists rewrite every integral in a family as a combination of a small set of "master integrals" and then compute only those. How many masters a family has sets the size of the whole calculation. Today that number is usually discovered only after running a large linear-algebra reduction to completion. Twisted intersection theory, which treats integrals as pairings of differential forms, reframes the count as a topological number, the Euler characteristic of the space being integrated over (Status of Intersection Theory and Feynman Integrals, Vector Space of Feynman Integrals and Multivariate Intersection Numbers). For generic masses and momenta, that Euler characteristic is a classical polytope quantity: the normalized lattice volume of the Newton polytope of the integrand's polynomial. The Newton polytope is the convex hull of the exponent vectors of its monomials. The trouble is that real physics is never generic. Equal masses, massless lines and on-shell legs make the count drop, and the existing theory handles this only through case-by-case blow-ups or by giving up.

This proposal claims that the drop can be read off the face structure of the same polytope. Each face of the Newton polytope carries a smaller "face polynomial", made of the monomials lying on that face. The count falls below the volume exactly when some face polynomial becomes singular at the physical kinematics. We conjecture that the size of the drop decomposes into local contributions, one per degenerate face, each computable from that face alone. If this holds, three things follow. The exact master count becomes a cheap combinatorial computation. Enumerating the faces gives an a priori upper bound on how many independent singular loci (Landau singularities) a family can have. And enumeration bounds on polytope faces become direct bounds on the complexity of Feynman calculations.

Core Claim

Write the Lee–Pomeransky polynomial of a graph $G$ as $\mathcal{G} = \mathcal{U} + \mathcal{F}$, where $\mathcal{U}$ and $\mathcal{F}$ are its Symanzik polynomials. Let $P_G$ be its Newton polytope and $z$ the physical kinematic point. The background fact is that the integration space is the complement of $\{\mathcal{G}=0\}$ in the algebraic torus, and for generic coefficients its signed Euler characteristic equals the normalized volume $\mathrm{Vol}(P_G)$. Genericity fails precisely on the zero set of the principal $A$-determinant, which factors into discriminants of the face polynomials $\mathcal{G}_F$. This is the GKZ hypergeometric structure that Four lectures on Euler integrals places at the centre of Euler-integral theory. The file's summary of those lectures records the convergence polytope, the link between Euler characteristic and cohomology dimension, and the GKZ connection, but not this volume formula explicitly. We therefore treat it as standard background rather than as a cited result. Our conjecture is

$$ N_{\text{masters}}(G,z) \;=\; \mathrm{Vol}(P_G) \;-\; \sum_{F \preceq P_G,\; \Delta_F(z)=0} \delta_F(z), $$

where the sum runs over faces $F$ whose face discriminant $\Delta_F$ vanishes at $z$. Each local defect $\delta_F(z) \ge 0$ depends only on $\mathcal{G}_F$. For isolated degenerations we expect it to be a Milnor-number-type multiplicity of the singular point of $\mathcal{G}_F$, computed in the lower-dimensional torus attached to $F$.

A corollary does not depend on additivity. The number of distinct irreducible components of the Landau-type singular locus is at most the number of faces of $P_G$ whose discriminant is non-trivial. This turns face-enumeration bounds for the polytope family into singularity-count bounds for the Feynman family.

The closest existing work differs in three ways. Feynman polytopes and the tropical geometry of UV and IR divergences builds polytopes from $\mathcal{U}$ and $\mathcal{F}$ and reads leading UV/IR divergences from volumes of dual cones. It names the link between facet structure and Landau thresholds as open, and it never addresses master counts. The Euler-integral lectures supply the generic theory and list as open the classification of loci where the Euler characteristic jumps. That open problem is the one we attack. Combinatorics and topology of Kawai-Lewellen-Tye relations works at tree level with associahedra, a different polytope family playing a different role.

Research Program

The first component is a face enumerator for the polytopes $P_G$. It computes the full face lattice, the face polynomials and the normalized volume for graph families up to roughly fifteen Schwinger parameters. In that range exact lattice-volume computation is feasible, which matters because master counts are integers and must be computed exactly. The randomized volume estimation of Practical Polytope Volume Approximation reports errors below 2%, which does not determine an integer count of several hundred. It is therefore useful here only as a cross-check at sizes beyond exact reach, not as the main tool.

The second component is a degeneracy detector. For each face it tests whether $\mathcal{G}_F$ has a singular point in the torus at the specified kinematics, using elimination restricted to the face. This is far smaller than the full-graph elimination performed by SOFIA: Singularities of Feynman Integrals Automatized. That tool supplies the comparison data, and its reported bottleneck on maximally connected topologies is the motivation for working face by face.

The third component computes the local defects $\delta_F$ and tests whether they add up. We will look specifically for cases where degenerate faces share a lower-dimensional face and their contributions interact. Such cases would force a corrected formula, for instance a sum with inclusion–exclusion over chains in the face lattice.

The fourth component is enumerative. It establishes how the number of faces of $P_G$ grows along infinite families: ladders, fishnets, sunrises, and the traintracks of Traintracks through Calabi-Yau Manifolds. It then compares the resulting singularity bounds with what is actually observed.

The lattice-polytope results in the file play a constraining role, not a supplying one. The $2^d$ facet bound for maximal lattice-free sets (A proof of Lovász's theorem on maximal lattice-free sets) and the 4–14 vertex window for terminal toric Fano polytopes (Toric Fano three-folds with terminal singularities) apply to classes that $P_G$ generally does not belong to. We do not assume they transfer. What we take from them is the method: a parity or arithmetic rigidity argument that converts a lattice constraint into a face bound. The Symanzik exponents give us exactly such a constraint, because every exponent vector has bounded total degree with entries in $\{0,1,2\}$. That fits the setting of the $(0,k)$-polytope diameter bounds of On the Diameter of Lattice Polytopes, and we will test whether analogous coordinate-box arguments bound face counts of $P_G$.

Baselines

The naive reference point is the generic volume $\mathrm{Vol}(P_G)$ reported as the master count at every kinematic point. It is exact off the degenerate locus and wrong on it, so the measure of success is how much of its error on physical kinematics our correction removes.

The strongest existing method for the count itself is to run a full integration-by-parts reduction and read off the masters. Integration-by-parts reduction derives linear relations between integrals and solves them. We will use NeatIBP 1.1 with Kira, the syzygy-based generator that NeatIBP 1.0 reports produced 14,120 relations where FIRE6 needed 11,207,942 on its test diagram. Its output is the ground truth. It competes with us on wall-clock time, not accuracy, and we will run both on the same machine and report time to count. Because NeatIBP's spanning-cut checks are reported to fail with massive propagators, we will also run plain Kira reductions on the massive cases to keep the ground truth trustworthy.

The second strong competitor is direct critical-point counting at the physical kinematics. This uses the result of Vector Space of Feynman Integrals and Multivariate Intersection Numbers that, under smoothness assumptions, the number of masters equals the number of critical points of the logarithm of the Baikov polynomial. We will give it the same polynomial systems and the same solver budget. For singularity counts, the baseline is SOFIA's output on the PLD.jl topologies. SOFIA is reported at roughly ten minutes per complex diagram and up to 275 singularities on the $HZ+2$-jet and $H+$jet examples.

The ablation that isolates the key ingredient keeps the volume and the face lattice but restricts degeneracy detection to vertices and edges only, or to facets only. If defects from intermediate-dimensional faces do not matter, the face-lattice machinery is unnecessary.

On accuracy we found no reported result on Lacuna for any method predicting master counts at non-generic kinematics. Beating the baselines therefore means exact agreement with IBP ground truth on every tested family, at least an order of magnitude less time than full reduction on the largest families, and correctness on non-smooth points where critical-point counting miscounts.

Why not just…?

Why not just run Kira or NeatIBP and count the masters? This gives the right answer, and we use it as ground truth. It still has to build and solve the full relation system before revealing the count. It gives no reason for the count and no prediction across a kinematic family without rerunning. It also has documented failures of its fastest mode with massive lines. A formula that is cheap, explains which kinematic coincidence removed which masters, and extends to whole families is a different deliverable. If our timings are not clearly faster on the larger topologies, the speed argument fails and the contribution reduces to explanation and bounds. We will report that honestly.

Why not just count critical points numerically at the physical point? This is the most serious threat. Where the very affine variety stays smooth at physical kinematics, counting critical points numerically may be as accurate as our formula and simpler. The case for the polytope route rests on non-smooth points. There, critical points escape to infinity or collide, and the identification of the count with the Euler characteristic needs the blow-ups that Status of Intersection Theory and Feynman Integrals calls prohibitive. The experiments split test points into smooth and non-smooth classes. If critical-point counting matches IBP on the non-smooth class too, we will narrow the claim to the singularity-bound and family-scaling results.

Why not just use the generic volume as an upper bound and stop? An upper bound is already useful for budgeting a computation. But physical kinematics is the common case, not the exception, and the drop can be large, as on the equal-mass and massless strata. A bound that is loose exactly where calculations happen does not tell anyone which masters disappeared. The naive-baseline comparison measures how loose it actually is on the benchmark families.

Proposed Experiments

The validation set has three parts. The first is the box, sunrise and rectangular fishnet families analyzed in the Feynman-polytope paper, at generic, equal-mass and massless kinematics. The second is the topologies in the PLD.jl collection distributed with SOFIA, including the eikonal graviton and jet diagrams. The third is the low-loop members of the traintrack family, where Calabi–Yau geometry appears and degeneracies are expected to be rich. For each graph and kinematic stratum we record the generic volume, the predicted count, the IBP count, the critical-point count, and the time each method took.

The first decisive check is whether the defect sum reproduces the IBP count exactly on every stratum. A single robust mismatch that cannot be explained by interacting faces would falsify additivity. We would then fall back to an inclusion–exclusion version and test that.

The second decisive check compares the face-count bound on singular components with SOFIA's actual singularity counts on the same diagrams. The bound should never be violated. We will also measure how tight it is, since a bound off by orders of magnitude would be true but of little use.

The third check is enumerative. We fit the growth of face counts along each family as loop order increases and test whether a coordinate-box argument in the style of the $(0,k)$-polytope results proves a polynomial-in-edges face bound for $P_G$. If such a bound holds, it is the cleanest bridge between the two directions: a theorem bounding the number of possible Landau singularity components, and hence the worst-case complexity of singularity analysis, for every Feynman graph with a given number of edges.