COVARIANCE ASSISTED SCREENING AND ESTIMATION.

Area of Science:

• Statistics
• Machine Learning
• Time Series Analysis

Background:

• Linear models (Y = Xβ + z) with Gaussian noise are common.
• Variable selection aims to identify non-zero coefficients in β.
• Challenges arise with non-sparse Gram matrices (X'X) and rare, weak signals.

Purpose of the Study:

• To develop a novel procedure for variable selection in challenging linear models.
• To address situations where the Gram matrix is non-sparse but sparsifiable.
• To achieve optimal convergence rates for variable selection under rare and weak signal conditions.

Main Methods:

• Introduced Covariance Assisted Screening and Estimation (CASE) procedure.
• CASE employs linear filtering to transform the regression setting, yielding a sparse Gram matrix.
• Utilizes a graph-based approach for multivariate screening and a 'patching' technique to handle information leakage.

Main Results:

• CASE successfully decomposes the problem into smaller subproblems by leveraging signal sparsity and graph structure.
• The 'patching' technique effectively overcomes information leakage issues.
• CASE achieves the optimal rate of convergence for variable selection in a broad class of non-sparse but sparsifiable Gram matrices.

Conclusions:

• CASE provides an effective solution for variable selection in complex statistical models.
• The method demonstrates strong performance in challenging scenarios with rare and weak signals.
• Successfully applied to long-memory time series and change-point models, showcasing practical utility.