Abstract:I mainly describe how we improve the deconvolution effect of nonminimum phase signal by using nancausal model and linear smoothing method in discrete-signal analysis. I also analyse different methods for determining model parameters by using respectively least square criterion, lp modulo criterion (Robust criterion is involved here), minimum entropy criterion, D modulo criterion and maximum entropy criterion in the case of two-sided predictive state model. Generally speaking, the regular equations obtained using these criteria are nonlinear equations. Besides, I present computation methods for the criteria, analyse the convergence of some iterative methods, and illustrate the deconvolution effect of different methods.