Smooth L0 Penalty Approximations for Gaussian Graphical Models
Provides smooth approximations to the L0 norm penalty for
estimating sparse Gaussian graphical models (GGMs). Network estimation
is performed using the Local Linear Approximation (LLA) framework
(Fan & Li, 2001 ;
Zou & Li, 2008 ) with five penalty
functions: arctangent (Wang & Zhu, 2016 ),
EXP (Wang, Fan, & Zhu, 2018 ), Gumbel,
Log (Candes, Wakin, & Boyd, 2008 ),
and Weibull. Adaptive penalty parameters for EXP, Gumbel, and Weibull
are estimated via maximum likelihood, and model selection uses
information criteria including AIC, BIC, and EBIC (Extended BIC).
Simulation functions generate multivariate normal data from GGMs with
stochastic block model or small-world (Watts-Strogatz) network structures.