Outlier Resistant Predictive Source Encoding for a Gaussian Stationary Nominal Source

摘要

A sequence of qualitatively robust predictive source encoders, for a Gaussian stationary source with outlier contaminated observation data, is proposed and analyzed. Performance measures include mean difference-sequence distortion and output entropy at the nominal Gaussian source, as well as breakdown point and influence function. The proposed sequence of predictive encoders attains strictly positive breakdown point and uniformly bounded influence function, at the expense of increased mean difference-squared distortion and differential entropy, at the Gaussian nominal source.