All loop scattering as a sampling problem

Abstract How to turn the flip of a coin into a random variable whose expected value equals a scattering amplitude? We answer this question by constructing a numerical algorithm to evaluate curve integrals — a novel formulation of scattering amplitudes — by a Monte Carlo strategy. To achieve a satisfactory accuracy we take advantage of tropical importance sampling. The crucial result is that the sampling procedure can be realized as a stochastic process on surfaces which can be simulated efficiently on a computer. The key insight is to let go of the Feynman-bias that amplitudes should be presented as a sum over diagrams, and instead re-arrange the sum as suggested by a dual decomposition of curve integrals. We attach an implementation of this algorithm as an ancillary file, which we have used to evaluate amplitudes for the massive Tr(ϕ)3 theory in D = 2 space-time dimensions, up to 10-loops. Interestingly, we observe experimentally that the number of sample points required to achieve a fixed accuracy remains significantly smaller than what the number of diagrams would suggest. Finally we propose an extension of our method which is inspired by ideas from artificial intelligence. We use the stochastic process to define a parametrization for a space of distributions, where we formulate importance sampling for an arbitrary curve integrand as a convex optimization problem.