Statistical Challenges of Global SUSY Fits

Roberto Trotta, K. CranmerarXiv 2011 · DOI · arXiv

Abstract: We present recent results aiming at assessing the coverage properties of Bayesian and frequentist inference methods, as applied to the reconstruction of supersymmetric parameters from simulated LHC data. We discuss the statistical challenges of the reconstruction procedure, and highlight the algorithmic difficulties of obtaining accurate profile likelihood estimates.

Introduction

Supersymmetry (SUSY) remains one of the most compelling extensions to the Standard Model of particle physics. However, its practical application is complicated by a vast parameter space, necessitating "global fits" to reconcile theoretical predictions with experimental observations from the Large Hadron Collider (LHC). These fits are historically hindered by two primary obstacles: the extreme computational cost of solving Renormalization Group Equations (RGEs) and the ongoing debate regarding the statistical reliability—specifically the "coverage"—of Bayesian versus frequentist inference methods.

Methodology: Neural Network Surrogates and Nested Sampling

To overcome the computational bottleneck, this study utilizes Artificial Neural Networks (ANNs) as surrogate models. A three-layer perceptron was trained on 4,000 samples to mimic the RGEs, mapping high-level theory parameters to observable particle masses. This approach achieved a correlation of 0.9999 with traditional physics codes while providing a staggering 30,000x speedup.

For exploring the complex, multi-modal parameter landscapes, the authors employed the MultiNest algorithm. This nested sampling technique is more efficient than traditional Markov Chain Monte Carlo (MCMC) methods for identifying disparate peaks in the likelihood surface. The accuracy of this surrogate-based reconstruction is shown in the comparison of posterior contours below.

Comparison of 2D posterior contours for CMSSM parameters using standard RGE solvers (black) versus the Neural Network surrogate (blue).

Statistical Reliability and Coverage

A central goal of this research was to evaluate "coverage": the probability that a calculated 95% confidence interval actually contains the true physical value. By performing 10,000 pseudo-experiments based on the ATLAS "SU3" benchmark point within the Constrained Minimal Supersymmetric Standard Model (CMSSM), the authors tested the robustness of both Bayesian posteriors and frequentist profile likelihoods.

Key Findings

  • Conservative Over-coverage: The study found that 1D confidence intervals were often "conservative," meaning they contained the true value more than 95% of the time. This phenomenon is largely attributed to the physical boundaries of the CMSSM; when a best-fit point attempts to migrate into an unphysical region, it is constrained to the boundary, which artificially expands the resulting confidence intervals.
  • Algorithmic Difficulty: Mapping the profile likelihood proved significantly more demanding than mapping the Bayesian posterior. Accurate profile likelihood estimates required approximately 10 times more computational resources—specifically, higher resolution in the sampling algorithm (20,000 live points in MultiNest) to avoid missing narrow, high-likelihood features.
Coverage properties of 95% profile-likelihood intervals for four CMSSM parameters, demonstrating conservative over-coverage where values meet or exceed the 0.95 threshold.

Conclusion

The integration of ANN surrogates allows for a scale of likelihood evaluation ($4 \times 10^{10}$) previously thought impossible, turning a task requiring over a millennium of CPU time into a manageable multi-week project. While the current findings show conservative coverage in the CMSSM, future work will explore more complex SUSY models where "under-coverage" (intervals that are too small) may be a risk. These advancements ensure that as more LHC data becomes available, the statistical foundations for potential new physics discoveries remain rigorous.

Summary updated

Abstract

We present recent results aiming at assessing the coverage properties of Bayesian and frequentist inference methods, as applied to the reconstruction of supersymmetric parameters from simulated LHC data. We discuss the statistical challenges of the reconstruction procedure, and highlight the algorithmic difficulties of obtaining accurate profile likelihood estimates.
BibTeX
@misc{trotta2011statistical,
  author = {Roberto Trotta and K. Cranmer},
  title = {{Statistical Challenges of Global SUSY Fits}},
  year = {2011},
  eprint = {1105.5244},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1105.5244},
}
Figure