PURPOSE: To reduce the number of times of calculations, and to prevent the amplification of an observation error by sampling an input signal by a frequency twice as high as an original signal band, and at the same time, adding a small random number matrix to the coefficient matrix of a linear equation when the original signal is modeled as the convolution integral of a certain input signal and a band limit filter.

CONSTITUTION: The original signal xn is modeled as the convolution integral of the impulse response hn of the band limit filter and the signal xDn whose sampling frequency is 2fC. Since the spectrum of the original signal xn is limited within a cut-off frequency fC, the spectrum of xDn, obtained by sampling the xn by the frequency of 2fC, at 0≤f≤fC is perfectly the same as the spectrum of the xn. Accordingly, the original xn can be perfectly restored by letting the xDn pass through the band limit filter. Next, the matrix H generally becomes to have small eigenvalues, which amplyfies the error superposed on xPTn. Against this, the random number matrix R is added to the H. By weakening the correlation between the elements of the matrix in this manners, the amplification of the error can be suppressed to be small.