In this paper we propose a sparse coefficient estimation procedure for
autoregressive (AR) models based on penalized conditional maximum likelihood.
The penalized conditional maximum likelihood estimator (PCMLE) thus developed
has the advantage of performing simultaneous coefficient estimation and model
selection. Mild conditions are given on the penalty function and the innovation
process, under which the PCMLE satisfies a strong consistency, local $N^{-1/2}$
consistency, and oracle property, respectively, where N is sample size.