An SV reduction method for multi-class SVMs is provided with which a number of SVs contained in the multi-class SVMs can be reduced without becoming trapped in a local minimum optimization solution and the reduction of the SVs can be performed at high precision and high speed. The method includes a step of selecting, from a plurality of initially present support vectors, support vector pairs zi, zj (i, j=1, 2, . . . , NS); a step of preparing a single-variable objective function with a single global maximum and determining a maximum value k of the objective function; and a step of applying the maximum value k to the support vector pairs zi and zj to determine a temporary vector Ztemp[i] of small classification errors; and the support vector pairs zi, zj are represented by the temporary vector Ztemp[i].