On using Kullback Leibler distance to estimate vector quantization performance
Septimiu Mischie, Liviu Toma · 2009
This paper studies the high-rate performance of a VQ (vector quantization) system designed from data having a given pdf and evaluated on data having a different pdf. The expected number of extra bits required by this VQ to achieve the same performance as a VQ designed and evaluated on the same data is computed. It is shown that the expected number of extra bits can be approximated by the Kullback Leibler distance if the two pdf's are multivariate Gaussian distributed. Experimental results which confirm the theoretical results are also presented.