Issue No. 03 - May/June (2002 vol. 17)
<p>The authors propose a method for multidimensional distribution analysis using a data compression technique. The method avoids the explosion in number of parameters (or coefficients) representing a multidimensional distribution even when the distribution has many dimensions (up to six dimensions or more). In the method, a multidimensional distribution is linearly expanded into a set of expansion coefficients. The expansion procedure neglects high-order cross-terms and reduces the total number of coefficients representing the distribution. This compression technique resemble DCT-based image data compression for computer vision.The authors applied the method to the knowledge-based mean-force potentials between residues for the analysis of protein sequence structure compatibility. They obtain the mean-force potentials by the multidimensional distribution of relative configurations (essentially 6D) between residues. The performance of the multidimensional mean-force potentials measured by native-structure-recognition tests was proved much higher than the performance of conventional 1D distance-based potentials derived from binned distributions.</p>
linear compression, mean-force potentials, multidimensional distribution, spherical Bessel, spherical harmonics
K. Onizuka, T. Noguchi, Y. Akiyama and H. Matsuda, "Using Data Compression for Multidimensional Distribution Analysis," in IEEE Intelligent Systems, vol. 17, no. , pp. 48-54, 2002.