Gaussian Process Eigenmodes for Statistical and Systematic Uncertainties in Template Fits

TL;DR

This paper introduces a Gaussian process eigenmode approach to efficiently estimate statistical and systematic uncertainties in template fits, improving over traditional bin-by-bin methods at the LHC.

Contribution

It proposes a unified eigenmode decomposition using Gaussian processes to replace multiple uncertainty parameters with a small set of constrained amplitudes.

Findings

Eigenmode basis encodes both statistical and systematic uncertainties.

Truncating to leading eigenmodes simplifies uncertainty estimation.

Method contains Barlow-Beeston as a special case and bounds variance.

Abstract

Template histograms are the foundation of statistical inference at the Large Hadron Collider. The HistFactory likelihood encodes template uncertainty through per-bin Barlow-Beeston gamma factors for Monte Carlo statistical error and through interpolation-based modifiers for systematic shape variations. These two mechanisms scale with the number of bins, which becomes problematic for multi-dimensional analyses and for templates constructed from limited Monte Carlo samples. We propose the use of eigenmode decomposition for efficiently estimating statistical and systematic uncertainties when replacing histogram templates with smooth functional representations derived from log-Gaussian Cox process posteriors fitted to the Monte Carlo data. The posterior covariance, augmented by rank-1 updates for each systematic shape variation, provides a unified eigenmode basis that encodes both statistical and systematic template uncertainty. Truncating to the leading eigenmodes replaces the full set of per-bin gamma factors and interpolation parameters with a small number of Gaussian-constrained amplitudes. We prove that this construction contains Barlow-Beeston as a limiting case and that the Gaussian Process posterior variance is bounded above by the Barlow-Beeston variance at every bin.