BSM共振项目的搜索预算
The Search Budget of the BSM Resonance Program
浏览论文内容
中文总结 AI 辅助
该研究统计ATLAS BSM共振项目的试验因子，对比已发表扫描与完全组合扫描的成本及Z值，指出两阶段盲法可防范机器学习搜索的虚假信号。
中文摘要 AI 辅助
数据导向扫描、异常搜索和通用搜索可同时探测多个质谱，对于这类搜索，“别处效应”设定了发现阈值。我们仅基于公开输入，以有效独立分辨单元统计ATLAS BSM共振项目的试验因子，再以相同单位计算完全组合扫描的成本。已发表记录涵盖104个扫描质谱，总试验次数$N_{\text{trials}}=7.9 \times 10^{3}$次，因此当前5σ全局发现需对应局部$Z_{\text{local}}=6.55$。若改为扫描由最多4个重建对象构成的每个质量，每次进行一个事例级选择，结合Run 2与Run 3的成本，总试验次数达$3.6 \times 10^{5}$次，对应局部$Z_{\text{local}}=7.11$，即试验次数增加46倍仅使Z值提升0.56σ。若因估计量本身不完善（如机器学习搜索）无法枚举试验因子，两阶段盲法解除是防范虚假信号的保障措施：针对已发表的 bump-hunting 网络所测得的缺陷率，该方法是更灵敏的程序。
英文摘要
Data-directed scans, anomaly searches and general searches probe many mass spectra at once, and for them the look-elsewhere effect sets the discovery bar. We count the trials factor of the ATLAS BSM resonance program in effective independent resolution elements, from public inputs alone, and then work out what a fully combinatorial scan would cost in the same units. The published record, summed over the 104 spectra it scans, amounts to $N_{\mathrm{trials}} = 7.9 \times 10^{3}$ looks, so a $5σ$ global discovery today costs a local $Z_{\mathrm{local}} = 6.55$. Scanning instead every mass built from up to four reconstructed objects, one event-level selection at a time and priced on Run 2 and Run 3 together, amounts to $3.6 \times 10^{5}$ looks and $Z_{\mathrm{local}} = 7.11$. A factor 46 more looks therefore costs $0.56σ$. If the trials factor cannot be enumerated because the estimator is itself imperfect, as in machine-learning searches, two-stage unblinding is a safeguard against its spurious signals: at the defect rate measured for a published bump-hunting network it is the more sensitive procedure.
