Analytic Regression of Feynman Integrals from High-Precision Numerical Sampling

The Big Picture

Imagine trying to reverse-engineer a complex recipe by tasting it. Not just “salty and sweet,” but the precise gram measurements of every ingredient. Now imagine you have unlimited, infinitely precise bites, but the dish has thousands of components. That’s roughly the challenge physicists face when computing Feynman integrals, the mathematical machinery behind predictions in particle physics.

These integrals describe how particles interact at the quantum level. The decay of a Higgs boson, the anomalous magnetic moment of the electron, a particle scattering off another: all depend on evaluating Feynman integrals to high accuracy. Exact formulas are almost always what’s needed for comparison with experiment, but direct computation typically yields only messy decimal approximations.

Physicists have spent decades developing tricks to wrestle these objects into exact, compact symbolic expressions. Most approaches work “top-down,” exploiting the known mathematical structure of an integral to constrain its form before computing a single number. A team from Harvard and IAIFI has now shown a complementary strategy: work bottom-up. Compute the integral numerically, to extraordinary precision, at many carefully chosen input values, then use a classical algorithm from number theory to reconstruct the exact answer.

Key Insight: By combining high-precision numerical evaluation with prior knowledge of which functions can appear in the answer, it’s possible to pin down exact rational coefficients using lattice reduction, turning approximate numbers into exact analytic results.

How It Works

The idea is simpler than it sounds. Any Feynman integral f(x) can, in principle, be written as a linear combination of known basis functions Bᵢ(x): well-understood mathematical functions like logarithms and their higher-order generalizations. The catch is that the coefficients cᵢ must be exact rational numbers (like 2 or −7/3), not floating-point approximations. Matrix inversion alone can’t recover them reliably, because numerical error compounds quickly as the basis grows.

Three ingredients make the method work:

1. High-precision numerical integration using AMFlow, a modern tool that reduces multi-loop Feynman integrals to solving one-dimensional differential equations, yielding evaluations to 100 or more significant digits.

2. A basis of iterated integrals derived from the integral’s known singularity structure (the mathematical locations where the integral blows up or vanishes). Decades of progress in solving the Landau equations have made it possible to enumerate candidate basis functions before solving for anything.

3. Lattice reduction, specifically the Lenstra–Lenstra–Lovász (LLL) algorithm. This number-theory technique identifies exact rational numbers hiding inside long decimal approximations. Given enough precision, LLL distinguishes the true rational coefficient from numerical noise.

Here’s the workflow: evaluate the integral and all basis functions at n chosen points, assemble the data into a matrix, and feed it into LLL. If the precision is sufficient and the basis is correct, the algorithm returns exact integers (the coefficients cᵢ). A decimal like 1.9999999998… becomes, unambiguously, 2.

The approach scales well. The team mapped out how much precision is needed as a function of basis size, number of evaluation points, and coefficient magnitude. Precision requirements grow only linearly with the number of basis functions, so the method stays tractable even for complex multi-loop diagrams.

The paper works through four examples of increasing difficulty. The simplest is the one-loop triangle integral, whose exact answer involves dilogarithms with coefficients {2, −2, 1}, trivially recovered. The hardest is the two-loop outer-mass double box, a scattering diagram with massive particles and two kinematic scales. Its exact form involves dozens of basis functions drawn from weight-4 polylogarithms, and the LLL pipeline handles it cleanly, recovering all coefficients exactly. Several of those coefficients had eluded prior approaches.

Why It Matters

Feynman integrals are the engine of precision particle physics. Every theoretical prediction for the Large Hadron Collider, from Higgs measurements to searches for new physics, depends on computing these integrals to two, three, or four loops. Getting exact analytic results has long been the bottleneck.

The two main alternatives each have drawbacks. The Landau-bootstrap approach derives results purely from analytic constraints. It’s powerful, but requires considerable human ingenuity for each new integral. Numerical methods are fast but deliver approximations that can’t always be trusted at the precision experiments demand.

This paper opens a third path, one that is more automated. Once the function space is understood (and modern tools like SOFIA can now determine it semi-automatically for many integrals), the numerical-plus-LLL pipeline recovers exact answers with minimal hand-holding. The method also generalizes beyond Feynman integrals: any problem requiring exact functional decomposition over a sufficiently constrained function space could benefit, from algebraic geometry to number theory to symbolic regression in mathematical discovery.

Bottom Line: High-precision numerics plus lattice reduction can recover exact analytic answers for Feynman integrals that have resisted closed-form computation, opening a reproducible and automatable route through a long-standing bottleneck in theoretical physics.