Universal divergence estimation for finite alphabet sources

This paper studies universal estimation of divergence from the realizations of two unknown finite-alphabet sources. Two algorithms that borrow techniques from data compression are presented. The first divergence estimator applies the Burrows-Wheeler block sorting transform to the concatenation of the two realizations; consistency of this estimator is shown for all finite-memory sources. The second divergence estimator is based on the Context Tree Weighting method; consistency is shown for all sources whose memory length does not exceed a known bound. Experimental results show that both algorithms perform similarly and outperform string-matching and plug-in methods.