This paper discusses an information theoretic approach for Bayesian dichotomization of patterns.
In the first section, we introduce a new measure which is closely related to the recognition rate of Bayesian pattern recognizer and is easy for practical use. This measure is defined as the correlation between the teacher's response and Bayesian discriminant function.
Further, it is shown that the measure is composed of three non-negative parts well known in information theory and statistics. These parts are analyzed in connection with Bayesian pattern recognition.
In the second section, the relation between the recognition rate and the measure is discussed by the aid of central limit theorem. This relation assures the usefulness of the measure in the synthesis of pattern recognizer and the quantization of patterns.
Experimental computer simulation ascertains the above discussions and results.