This paper presents the original gradient method for the determination of the optimal signal constellation points disposition after quantization, when the error probability is minimized under the power constraint. This method consists of the nonlinear transformation and rescaling. The method differs from the method presented in [1], which can be used for uniform constellations with great numbers of points (where the continual approximation can be applied), since it can be applied for generating the nonequiprobable nonuniform constellations containing any number of points.