Abstract
Statistical approaches for quantization of patterns are discussed together with its asymptotic properties.
To discuss the problem quantitatively, we introduce an information measure closely related to the recognition rate of a Bayesian pattern recognizer.
Using this measure, we evaluate the information loss resulting from quantizing patterns as the amount of decrease of the measure.
We adopt the valuation as the performance index of quantization and consider optimum methods in the sense of the valuation for various kinds of quantization problems.
Experimental computer simulations are done in several cases and the relation to the Bayesian recognition rate is shown.
Further, to examine properties of these methods, we notice asymptotical properties in regard to the number of quantizing levels.
Consequently, their validity and the relation to the usual method based on the mean square error criterion are made clear.
