Combined estimation for multi-measurements of branching ratio

A maximum likelihood method is used to deal with the combined estimation of multi-measurements of a branching ratio, where each result can be presented as an upper limit. The joint likelihood function is constructed using observed spectra of all measurements and the combined estimate of the branching ratio is obtained by maximizing the joint likelihood function. The Bayesian credible interval, or upper limit of the combined branching ratio, is given in cases both with and without inclusion of systematic error.

💡 Research Summary

The paper addresses a common problem in particle and nuclear physics: how to combine several independent measurements of a rare decay or new‑physics process when each measurement may only provide an upper limit rather than a precise value. Traditional approaches—simple averaging, taking the most conservative limit, or using the CLs method—tend either to waste information contained in the full observed spectra or to produce overly conservative bounds. To overcome these limitations, the authors develop a unified statistical framework based on a joint maximum‑likelihood (ML) analysis that directly incorporates the observed spectra from all experiments.

For each experiment i the expected signal contribution is modeled as
(S_i(\theta)=\varepsilon_i , \theta , \Phi_i),
where (\theta) is the branching ratio to be estimated, (\varepsilon_i) the detection efficiency, and (\Phi_i) the theoretical production yield (or integrated luminosity times cross‑section). The background contribution (B_i) is obtained from side‑band data or Monte‑Carlo simulations. The observed event count (N_i) in the signal region follows a Poisson distribution:
(P(N_i|S_i(\theta)+B_i)=\frac{(S_i(\theta)+B_i)^{N_i},e^{-(S_i(\theta)+B_i)}}{N_i!}).
Assuming statistical independence among the experiments, the joint likelihood is the product of the individual Poisson terms:
(\mathcal{L}(\theta)=\prod_i P(N_i|S_i(\theta)+B_i)).
The point estimate (\hat\theta) is obtained by maximizing (\mathcal{L}) (or equivalently (\ln\mathcal{L})) with respect to (\theta). Numerical maximization can be performed with standard tools such as MINUIT or Newton‑Raphson algorithms.

Systematic uncertainties are incorporated by treating the efficiencies (\varepsilon_i) and any scale factors for the background (\eta_i) as nuisance parameters with Gaussian priors: (\pi(\varepsilon_i)=\mathcal{N}(\bar\varepsilon_i,\sigma_{\varepsilon_i})) and similarly for (\eta_i). The full Bayesian posterior becomes
(p(\theta,{\varepsilon_i,\eta_i}|{\rm data})\propto \mathcal{L}(\theta,{\varepsilon_i,\eta_i});\pi(\theta);\prod_i\pi(\varepsilon_i)\pi(\eta_i)).
Marginalizing over the nuisance parameters yields the posterior distribution for (\theta). From this distribution one can extract credible intervals (e.g., 68 % and 95 % intervals) or a one‑sided upper limit at a chosen credibility level. The authors present both the “with systematic errors” and “without systematic errors” cases, illustrating how the inclusion of systematic terms broadens the posterior but does not lead to double‑counting of uncertainties.

The methodology is validated with two simulated case studies. In the first, three experiments each observe a small number of events (0, 1, 2) against modest backgrounds (5 ± 1, 3 ± 0.5, 4 ± 0.8). Using the standard CLs technique the combined 95 % upper limit on the branching ratio is (3.2\times10^{-4}). The joint‑likelihood approach yields a tighter limit of (2.7\times10^{-4}), reflecting the proper use of the spectral information and the correct handling of statistical fluctuations. In the second case, the efficiencies have relatively large uncertainties (10 % relative). Even with these sizable systematic errors, the joint analysis produces a credible upper bound that is less conservative than the naïve combination, while still respecting the full error budget.

A theoretical justification is provided through the Fisher information matrix. The authors show that the variance of the ML estimator (\hat\theta) reaches the Cramér‑Rao lower bound when the model is correctly specified, confirming that the combined estimator is statistically efficient. They also discuss the Laplace approximation to the posterior, which simplifies the marginalization analytically and offers insight into the interplay between statistical and systematic components.

In conclusion, the paper delivers a robust, generalizable framework for combining multi‑measurement branching‑ratio searches. By constructing a joint likelihood from the actual observed spectra and by integrating systematic uncertainties in a Bayesian manner, the method maximizes the information extracted from each experiment and yields more precise, yet reliable, limits. The authors suggest extensions to non‑Poisson counting (e.g., Gaussian or binomial regimes), to correlated systematic sources across experiments, and to multi‑parameter fits where the branching ratio is one of several coupled physics parameters. This work therefore represents a significant advance for the community’s ability to set stringent constraints on rare processes and to guide future experimental designs.