This paper studies expurgated error exponents for joint source-channel coding for discrete memoryless sources and channels. We consider a partition of the source messages into classes, where the codeword distributions depend on the class. We show that two carefully chosen classes suffice to achieve Csiszár’s expurgated exponent.

This paper provides a dual domain derivation of the error exponent of maximum mutual information (MMI) decoding with constant composition codes, showing it coincides with that of maximum likelihood decoding for discrete memoryless channels. The analysis is further extended to joint source-channel coding, demonstrating that the generalized MMI decoder achieves the same random coding error exponent as the maximum a posteriori decoder.