English

The Hannan-Quinn Proposition for Linear Regression

Statistics Theory 2010-12-21 v1 Statistics Theory

Abstract

We consider the variable selection problem in linear regression. Suppose that we have a set of random variables X1,...,Xm,Y,ϵX_1,...,X_m,Y,\epsilon such that Y=kπαkXk+ϵY=\sum_{k\in \pi}\alpha_kX_k+\epsilon with π{1,...,m}\pi\subseteq \{1,...,m\} and αkR\alpha_k\in {\mathbb R} unknown, and ϵ\epsilon is independent of any linear combination of X1,...,XmX_1,...,X_m. Given actually emitted nn examples {(xi,1...,xi,m,yi)}i=1n\{(x_{i,1}...,x_{i,m},y_i)\}_{i=1}^n emitted from (X1,...,Xm,Y)(X_1,...,X_m, Y), we wish to estimate the true π\pi using information criteria in the form of H+(k/2)dnH+(k/2)d_n, where HH is the likelihood with respect to π\pi multiplied by -1, and {dn}\{d_n\} is a positive real sequence. If dnd_n is too small, we cannot obtain consistency because of overestimation. For autoregression, Hannan-Quinn proved that, in their setting of HH and kk, the rate dn=2loglognd_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.

Keywords

Cite

@article{arxiv.1012.4276,
  title  = {The Hannan-Quinn Proposition for Linear Regression},
  author = {Joe Suzuki},
  journal= {arXiv preprint arXiv:1012.4276},
  year   = {2010}
}
R2 v1 2026-06-21T17:01:26.038Z