The Hannan-Quinn Proposition for Linear Regression

TL;DR

This paper extends the Hannan-Quinn criterion for variable selection to linear regression, establishing the optimal penalty rate for consistent model selection in this setting.

Contribution

It proves that the Hannan-Quinn penalty rate of 2 log log n ensures strong consistency in linear regression variable selection.

Findings

Hannan-Quinn criterion applies to linear regression.

The penalty rate 2 log log n is optimal for consistency.

The result differs from autoregression setting.

Abstract

We consider the variable selection problem in linear regression. Suppose that we have a set of random variables $X_1,...,X_m,Y,\epsilon$ such that $Y=\sum_{k\in \pi}\alpha_kX_k+\epsilon$ with $\pi\subseteq \{1,...,m\}$ and $\alpha_k\in {\mathbb R}$ unknown, and $\epsilon$ is independent of any linear combination of $X_1,...,X_m$. Given actually emitted $n$ examples $\{(x_{i,1}...,x_{i,m},y_i)\}_{i=1}^n$ emitted from $(X_1,...,X_m, Y)$, we wish to estimate the true $\pi$ using information criteria in the form of $H+(k/2)d_n$, where $H$ is the likelihood with respect to $\pi$ multiplied by -1, and $\{d_n\}$ is a positive real sequence. If $d_n$ is too small, we cannot obtain consistency because of overestimation. For autoregression, Hannan-Quinn proved that, in their setting of $H$ and $k$, the rate $d_n=2\log\log n$ is the minimum satisfying strong consistency. This paper solves the statement affirmative for linear regression as well which has a completely different setting.