arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2403.00753nucl-th
收藏

似然函数在弹性散射不确定性量化中的作用

The role of the likelihood for elastic scattering uncertainty quantification

C. D. Pruitt, A. E. Lovell, C. Hebborn, F. M. Nunes

AI总结:

本文比较光学模型弹性散射分析中的频率学派与贝叶斯不确定性量化方法,揭示似然函数及后验非高斯性对参数不确定性和预测可靠性的关键影响。

摘要:

背景:利用光学模型(OMP)进行弹性散射分析在核反应研究中被广泛使用。目的:以往工作比较了传统频率学派方法与贝叶斯方法在量化OMP不确定性方面的表现。本研究重新审视这一比较,并考察分析中所采用似然函数的作用。方法:我们将用于χ²最小化的Levenberg-Marquardt算法与用于获得参数后验分布的Markov Chain Monte Carlo采样进行比较。沿用以往工作,我们考察在似然函数中使用χ²/N时结果如何受到影响,其中N为数据点数量,以此说明模型中可能存在的相关性以及数据误差被低估的情况。结果:我们首先分析一个简单的线性模型,随后转向弹性散射角分布的OMP分析,分别采用a)一个5参数模型和b)一个6参数模型。在线性模型中,频率学派与贝叶斯方法给出了一致的最优值和不确定性估计。对于5参数OMP分析,二者在定性上也是如此。对于6参数OMP分析,参数后验不再能由高斯分布良好近似,基于协方差的频率学派预测变得不可靠。在所有情况下,当贝叶斯方法在似然中使用χ²/N时,不确定性会增大√N倍。结论:当参数后验近似为高斯分布且使用相同似然函数时,频率学派与贝叶斯方法能够恢复一致的参数不确定性估计。如果参数后验具有显著的高阶矩,仅依赖协方差的频率学派方法将变得不可靠,此时应采用贝叶斯方法。经验覆盖率可作为不确定性估计的重要内部检验,为不确定性分析提供警示信号。

英文摘要:

Background: Analyses of elastic scattering with the optical model (OMP) are widely used in nuclear reactions. Purpose: Previous work compared a traditional frequentist approach and a Bayesian approach to quantify uncertainties in the OMP. In this study, we revisit this comparison and consider the role of the likelihood used in the analysis. Method: We compare the Levenberg-Marquardt algorithm for $χ^{2}$ minimization with Markov Chain Monte Carlo sampling to obtain parameter posteriors. Following previous work, we consider how results are affected when $χ^{2}$/N is used for the likelihood function, N being the number of data points, to account for possible correlations in the model and underestimation of the error in the data. Results: We analyze a simple linear model and then move to OMP analysis of elastic angular distributions using a) a 5-parameter model and b) a 6-parameter model. In the linear model, the frequentist and Bayesian approaches yield consistent optima and uncertainty estimates. The same is qualitatively true for the 5-parameter OMP analysis. For the 6-parameter OMP analysis, the parameter posterior is no longer well-approximated by a Gaussian and a covariance-based frequentist prediction becomes unreliable. In all cases, when the Bayesian approach uses $χ^{2}$/N in the likelihood, uncertainties increase by $\sqrt{N}$. Conclusions: When the parameter posterior is near-Gaussian and the same likelihood is used, the frequentist and Bayesian approaches recover consistent parameter uncertainty estimates. If the parameter posterior has significant higher moments, the covariance-only frequentist approach becomes unreliable and the Bayesian approach should be used. Empirical coverage can serve as an important internal check for uncertainty estimation, providing red flags for uncertainty analyses.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑