Scholarlink

<p>This paper studies the coherence of an high-dimensional observations matrix. Specifically, we describe the limiting distribution of the largest magnitude of correlations matrix associated to our data outside a central band which size depends of the sample size. Using the Chen–Stein method, we show the convergence of the normalized coherence towards a Gumbel distribution. We broaden previous results by considering a 3-regime band structure for the off diagonal covariance matrix, where the largest band is composed of asymptotically vanishing coefficients. We provide an hypothesis test on the covariance structure where the alternative shows a clear dichotomy on the vanishing band. Moreover, we provide numerical simulations illustrating the asymptotic behavior of the coherence with Monte-Carlo experiment. We use a splitting strategy computing correlation matrices by blocks in order to avoid the high-dimensional memory issue.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Largest magnitude for off-diagonal auto-correlation coefficients in high dimensional framework

  • 期刊论文

  • 发表日期：

    2025年5月10日

查看全文
  • Maxime Boucher,
  • Didier Chauveau,
  • Marguerite Zani

摘要

This paper studies the coherence of an high-dimensional observations matrix. Specifically, we describe the limiting distribution of the largest magnitude of correlations matrix associated to our data outside a central band which size depends of the sample size. Using the Chen–Stein method, we show the convergence of the normalized coherence towards a Gumbel distribution. We broaden previous results by considering a 3-regime band structure for the off diagonal covariance matrix, where the largest band is composed of asymptotically vanishing coefficients. We provide an hypothesis test on the covariance structure where the alternative shows a clear dichotomy on the vanishing band. Moreover, we provide numerical simulations illustrating the asymptotic behavior of the coherence with Monte-Carlo experiment. We use a splitting strategy computing correlation matrices by blocks in order to avoid the high-dimensional memory issue.