Adaptive exact recovery in sparse nonparametric models

Area of Science:

• Statistics
• Machine Learning
• High-Dimensional Data Analysis

Background:

• Observing an unknown function f(t) of d variables in a Gaussian white noise model.
• Assuming f(t) is a sum of k-variate functions (1 <= k <= s), with only a few being nonzero.
• Addressing the challenge in high-dimensional settings where d -> infinity and s can also grow.

Purpose of the Study:

• To identify the nonzero components of the unknown function f(t).
• To develop a variable selection procedure for high-dimensional models with growing complexity.
• To determine conditions for successful and impossible exact variable selection.

Main Methods:

• Utilizing a Gaussian white noise model with intensity epsilon > 0.
• Developing a variable selection procedure adaptive to model sparsity (parameter beta).
• Deriving theoretical conditions for exact variable selection.

Main Results:

• Established conditions under which exact variable selection is possible.
• Proposed an adaptive selection procedure that achieves exact variable selection.
• Identified conditions that preclude exact variable selection.

Conclusions:

• The developed procedure enables exact variable selection in high-dimensional, sparse settings.
• The findings provide a theoretical framework for understanding variable selection limitations.
• This work advances the field of statistical inference for complex models.