High resolution nonparametric inference of gravitational-wave populations in multiple dimensions
PixelPop:高维多维引力波群体非参数推断的高分辨率方法
The origins of merging compact binaries observed by gravitational-wave detectors remains highly uncertain. Several astrophysical channels may contribute to the overall merger rate, with distinct formation processes imprinted on the structure and correlations in the underlying distributions of binary source parameters. In the absence of confident theoretical models, the current understanding of this population mostly relies on simple parametric models that make strong assumptions and are prone to misspecification. Recent work has made progress using more flexible nonparametric models, but detailed measurement of the multidimensional population remains challenging. In pursuit of this, we present PIXELPOP---a high resolution Bayesian nonparametric model to infer joint distributions and parameter correlations with minimal assumptions. PIXELPOP densely bins the joint parameter space and directly infers the merger rate in each bin, assuming only that bins are coupled to their nearest neighbors. We demonstrate this method on mock populations with and without bivariate source correlations, employing several statistical metrics for information gain and correlation significance to quantify our nonparametric results. We show that PIXELPOP correctly recovers the true populations within posterior uncertainties and offers a conservative assessment of population-level features and parameter correlations. Its flexibility and tractability make it a useful data-driven tool to probe gravitational-wave populations in multiple dimensions.
展开 ▾创新点在于用最弱的平滑先验(相邻格点耦合)实现二维分布的高分辨率推断(100 bins/维),计算上避免了高斯过程协方差矩阵求逆的 O(B³) 难题;并用 Spearman 秩相关系数和拓宽系数量化非线性、非单调相关,在模拟 O4 数据集上成功识别质量比–有效自旋和红移–有效自旋的复杂相关。
引力波群体推断中,参数模型可能因设定错误而产生偏倚,近期非参数方法如 Golomb+ 2023、Callister+ 2024 等增加了灵活性,但推断多维联合分布仍具挑战。本研究采用仅在相邻格点间耦合的条件自回归(CAR)先验,以高分辨率刻画二维合并率密度,并通过 Spearman 秩相关系数与拓宽系数量化非线性相关,弥补了现有方法在高维受限的不足。模拟分析表明,PixelPop 能够可靠恢复 Callister+ 2021 与 Biscoveanu+ 2022 曾报告的质量比–有效自旋及红移–有效自旋相关,且在存在未建模三维相关时仍有包容性。随着 Heinzel+ 2025 及后续观测的增加,该方法可推广至更高维度,为光谱警笛宇宙学如 Farah+ 2025 提供无假设的群体建模工具;但仍需解决先验主导区域过估计及似然估计方差随样本增大的挑战。
预印本 2024-06-24 · 刊出 2025-03-17 · 收录 2026-10-12