报告人：林尤武 副教授

时间：2026年7月15日下午16:00-17:00

地点：文昌校区第二教学楼320室

报告摘要：Partial correlation is a fundamental measure of conditional dependence, but valid high-dimensional inference remains challenging under dense dependence. The difficulty is that inference for this low-dimensional target depends on a high-dimensional nuisance structure whose recovery may be unreliable when sparsity is weak or absent. To address this problem, we develop a target-specific reduction framework for estimating and testing the partial correlation of interest. For estimation, we propose an optimal projection estimator, OPE, defined as the sample partial correlation after a target-preserving reduction. OPE therefore estimates the original partial correlation and is asymptotically normal without requiring sparsity. For testing, we propose an optimal projection test, OPT, based on a projected statistic with an exact finite-sample \(F\)-distribution under the null. A conditional representation under the alternative characterizes the signal retained by the projection, and maximizing this signal yields an optimal one-dimensional direction within the class of linear reductions considered. Under a rank-one alignment condition, this optimal direction also preserves the original partial correlation. To make this construction implementable, we further develop a feasible ridge-based direction and establish its asymptotic equivalence to the population optimal direction. Simulation studies demonstrate accurate estimation, reliable size control, and competitive power in non-sparse high-dimensional regimes.

报告人简介：林尤武，男，博士，副教授，硕士生导师。博士毕业于中南大学，获统计学博士学位。北京大学访问学者，访问导师王汉生教授。研究方向主要包括高维数据分析、共型预测与多重假设检验等。主持国家及省部级自然科学基金项目多项。在《Statistics and Computing》、《Journal of Nonparametric Statistics》及《Journal of Chemometrics》等重要期刊上发表高水平论文30余篇。此外，其开发的统计软件包累计下载量已超过70,000次，研究成果受到国内外学者的广泛关注，尤其所发展的变量选择方法OHPL被美国统计学者在《Statistics in Medicine》发表的综述性文章所正面评价。