Challenges of profile likelihood evaluation in multi-dimensional SUSY scans

Farhan Feroz, K. Cranmer, M. P. Hobson, Roberto Ruiz de Austri, Roberto TrottaJournal of High Energy Physics 2011 · DOI · arXiv

Introduction

In the hunt for new physics at the Large Hadron Collider (LHC), researchers navigate complex mathematical landscapes to test theories like Supersymmetry (SUSY). The goal is to identify model parameters that best match experimental data. However, a significant debate exists regarding the statistical tools used: Bayesian posteriors, which focus on probability mass, and Frequentist profile likelihoods, which focus on peak height. This paper addresses claims that MULTINEST, a popular Bayesian inference algorithm, is fundamentally incapable of accurately performing Frequentist profile likelihood scans in high-dimensional spaces.

Bayesian Volume vs. Frequentist Peaks

The fundamental difference between these two statistical frameworks lies in their focus. Bayesian inference identifies regions with high "probability mass" (volume), whereas Frequentist profile likelihoods seek the absolute "best-fit" point (peak height). In models like the Constrained MSSM (CMSSM), these best-fit points often reside in extremely narrow "spikes" of high likelihood.

Comparison of 1D and 2D marginal distributions for Bayesian and Profile Likelihood mappings.

Standard Bayesian algorithms are optimized to find volume. Consequently, a tiny spike with negligible volume has almost zero Bayesian probability and is typically ignored. However, for a Frequentist, such a spike represents the Global Maximum Likelihood. Previous studies suggested that MULTINEST was limited in finding these spikes, recommending Genetic Algorithms (GA) instead.

Tuning MULTINEST for Profile Likelihoods

The authors demonstrate that MULTINEST's perceived failure was not an algorithmic flaw but a configuration issue. By adjusting two key parameters, the algorithm can be repurposed for Frequentist scans:

  1. Increased Resolution: Increasing the number of "live points" ($n_{live}$) from 1,000 to 20,000.
  2. Stricter Termination: Lowering the tolerance parameter ($tol$) from 0.5 to $10^{-4}$.

This stricter tolerance forces the algorithm to continue searching deep into high-likelihood regions even after the Bayesian mass is accounted for.

1D profile likelihoods for a toy model comparing standard MULTINEST settings (blue) and tuned settings (black) against the analytical solution (red).

Results and Discussion

The tuned MULTINEST was tested on an 8D Gaussian mixture model and the real-world CMSSM parameter space. In the toy model, it perfectly reconstructed the profile likelihood, capturing narrow spikes missed by standard runs. In the CMSSM scan, MULTINEST found a global best-fit point with a $\chi^2$ of 9.26, outperforming the best result from Genetic Algorithms ($\chi^2 = 9.35$) reported in previous literature.

While this accuracy comes at a higher computational cost—requiring approximately 5.5 million likelihood evaluations compared to 163,000 for a standard Bayesian run—it allows a single run to provide both an accurate Bayesian posterior and a high-quality profile likelihood simultaneously.

Conclusion

The study concludes that MULTINEST is a robust tool for both Bayesian and Frequentist inference in SUSY scans, provided it is configured correctly. Future improvements may involve hybrid approaches, using MULTINEST to identify "basins of attraction" and local optimizers like MINUIT to refine the exact maximum. This work resolves a key methodological dispute in the particle physics community, ensuring more reliable interpretations of LHC data.

Summary updated

BibTeX
@article{feroz2011challenges,
  author = {Farhan Feroz and K. Cranmer and M. P. Hobson and Roberto Ruiz de Austri and Roberto Trotta},
  title = {{Challenges of profile likelihood evaluation in multi-dimensional SUSY scans}},
  journal = {Journal of High Energy Physics},
  year = {2011},
  doi = {10.1007/jhep06(2011)042},
}

Loading related content...

Figure