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 2511.09431math.STstat.TH
收藏

脑连接图中差异的新型检验方法

A Riemannian Generalization of Wilks' Lambda for Testing Differences Among Brain Connectomes

Perelman School of Medicine, University of Pennsylvania(宾夕法尼亚大学佩雷尔曼医学院);Indiana University(印第安纳大学)

Nicolas Escobar-Velasquez, Jaroslaw Harezlak

AI总结:

本文提出基于流形结构的新型ANOVA检验方法,利用对称正定矩阵流形上的Mahalanobis分布进行参数估计,提升统计检验效能。

摘要:

非欧几里得空间上的统计分析通常依赖距离作为构建似然函数的主要工具。然而,流形值数据除了Riemannian距离外还具有更丰富的结构。我们证明可以在对称正定(SPD)矩阵流形上构造简单且可处理的模型,这自然出现在脑连接分析中。具体而言,我们强调了流形值Mahalanobis分布,这是一个扩展经典多元概念到SPD流形的参数族。我们开发了该分布的估计器并建立了其渐近性质。在此框架基础上,我们提出了一种新的ANOVA检验,利用流形结构获得更能捕捉数据维度的检验统计量。我们理论证明,我们的检验在统计功效上优于基于距离的Fréchet ANOVA方法。

英文摘要:

Frćhet analysis of variance is a well-established tool for comparing samples of non-Euclidean data such as brain connectomes, symmetric positive definite (SPD) matrices that encode coactivation of different brain regions. But by relying solely on various squared distances, it can fail to detect real group differences. We introduce a Riemannian generalization of Wilks' Lambda for SPD-valued data. Our central result is a power comparison. Against a location shift of size $θ$, the non-centrality of the proposed test behaves as $O(θ^2)$ while that of Fréchet ANOVA behaves as $O(θ^4)$, so the new test is more powerful for small-to-moderate effects. We confirm this on Human Connectome Project data, where the test detects sex and acquisition-site differences in functional connectivity that Frćhet ANOVA misses while the two tests agree at negligible and large effects. The method is implemented in an open-source R package.

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。

↑