In this paper, we develop a fast dynamic programming technique to construct the subband decomposition yielding the maximum coding gain. We first derive a unified coding gain applicable to arbitrary subband decompositions and arbitrary filter bases. Then, we prove that the unified coding gains are monotonically increasing function for the implementation complexities. Based on this phenomenon, we treat the implementation complexity and the coding gain in the same way as the rate and distortion function. Finally, we applied the developed algorithm to the wavelet packet based image coding, so as to verify coding performance improvement by the proposed algorithm.