Certified spectral functions from lattice Monte Carlo data

Abstract: The Monte Carlo method, applied to lattice quantum field theory, gives access to Euclidean correlation functions with well-understood error bars. Recovering the observables one cares about, such as the spectral density, requires solving an ill-posed inverse problem, usually tackled with heuristics that lose rigorous control of the error. Instead of trying to find the ``best'' spectral density $ρ(ω)$, we ask how small or large linear functionals $$\int_{\mathbb{R}^+} G(ω) ρ(ω) \mathrm{d} ω$$ of it can be, given the Monte Carlo data and the reflection positivity of the lattice action. This is a convex but infinite-dimensional problem. We show how its dual can be rigorously relaxed into a hierarchy of finite semidefinite programs, solvable with standard solvers and enjoying strong convergence guarantees. The resulting bounds are rigorous even when the relaxation is not tight, and converge quickly to the regime where the error is entirely dominated by Monte Carlo statistics. The method also flags implausible Monte Carlo data, for instance underestimated error bars, through an infeasibility certificate. We demonstrate it on lattice $φ<sup>4$ theory in two dimensions.