Wavelet packet decomposition has been successfully applied to image analysis and classification. The most common approach for wavelet packet-based texture classification is to decompose texture images with wavelet packet transform and to extract energy values for all subbands as features for the subsequent classification. Due to the overcomplete representation provided by the wavelet packet transform, it is suitable to select a set of subbands for sparse representation of the texture for classification. For better classification results, it is desired that the energy features corresponding to the selected subbands are as independent from each other as possible. However, most of the current subband selection methods do not take the dependence between energy values from different subbands into account. In this paper, we investigate the dependence between energy values from different subbands, which may be from the same wavelet basis, or from different wavelet bases. Based on the theoretical analysis and simulation, we propose an information-theoretic measure, mutual information, for selecting subbands for sparse representation of textures for classification. Experimental results show that the proposed method yields a sparse representation of the textures and achieves lower classification error rates than the conventional methods, simultaneously.