AI 中文总结
该研究针对盖亚 DR3 XP 光谱的高斯过程分类器产生的白矮星主序双星候选体，探讨 BP 波段约 2%的残余量是否污染选择，通过分析得出不同过量情况下的虚假率等结果，并明确了干净可靠性信号及污染假设情况。
AI 中文摘要
在盖亚 DR3 XP 光谱上使用高斯过程分类器产生了约 30000 个白矮星主序（WD-MS）双星候选体，每个候选体都有一个概率，但没有似然性或拟合优度。校正后的 XP BP 波段保留了约 2%的局部蓝通量残余，其中热白矮星也会增加通量。我们询问该残余在其实际幅度下是否会污染选择。结果表明不会。答案取决于一个单位：沉积在窄蓝带中的总通量的 2%，在红色主序星上是中位数 55%的局部过量，是本地读取相同数量的 27 倍。以 2%的局部过量注入时，通过代表分类器的$\Delta\chi^2$阈值，在 0.05 的基线时会产生 0.08 的虚假率，摊销后的后验在 90%的覆盖率下保持 0.84，而干净样本为 0.88。只有在局部过量超过 10%-20%时选择才会失败，在接近 50%时达到 0.96，这是校正和$B\lt18$切割消除的原始未校正偏差。模型比较门仅在 WD 通量份额接近 0.05 以上时才认证双星性；在目录的中位数份额处，统计数据反转，虚假样本具有更大的$\Delta\chi^2$改善。干净的可靠性信号是冷却序列外的紫外线不足（GALEX FUV 检测率为 0.19 对 0.50），对距离控制具有鲁棒性。在校正留下的残余处，污染假设为空；开放区域是较暗的一半，那里该残余未被测量。审计给出了所需的幅度以及超过该幅度的失败形状。
英文摘要
A Gaussian-process classifier on Gaia DR3 XP spectra produced $\sim$30,000 white-dwarf main-sequence binary candidates (Li et al. 2025), with probabilities but no likelihood. The corrected spectra agree with external libraries to better than 2% below 400 nm and 1% redward (Huang et al. 2024), and that residual sits where a hot white dwarf also adds flux. We asked whether it contaminates the selection at its bound. It does, weakly. The spurious rate runs about 1.7 times the baseline there. Put into the blue band, 2% of total flux is a median 55% local excess across our MS templates, 27 times that number read locally. We inject a blue taper clipped to that bound, under the templates' published per-pixel errors, with a $Δχ^2$ threshold standing in for the classifier. The spurious rate is $0.09\pm0.02$ across 20 seeds on a 0.053 baseline, 19 of 20 above it. The unclipped 2% taper, above the redward bound, gives $0.11\pm0.03$. At that taper the amortized posterior's 90% coverage falls to 0.85 from a clean 0.88 on the flat-model leg, with PIT uniformity rejected at $p<10^{-5}$ in all four seeds. The selection fails in bulk only above a 10-15% local excess, and the rate reaches 0.96 near 50%. A model-comparison gate certifies binarity only above a white-dwarf flux share near 0.05. At share 0.01, near the catalog's median of 0.0096, the $Δχ^2$ preference inverts at a 20% local excess. Outside Huang's color-magnitude box, 40% of the catalog, the post-correction residual is not measured.
Comments18 pages, 3 figures, 1 table. v2: headline re-anchored on the Huang-consistent injection, title amended; see Data and code. Code, configurations, and the committed records behind every quoted number are at https://github.com/WizardEternal/wdms-xp-calibration
