The problem of linear minimum mean-square error (MMSE) multichannel de
convolution of sampled signals from noisy observations is approached v
ia matrix polynomial equations. The general solution is given in terms
of a left spectral factorization and a pair of bilateral Diophantine
equations. The first Diophantine equation is obtained by imposing opti
mality of the deconvolution filter whereas the second ensures stabilit
y of the filter, should the signal model be unstable. The proposed sol
ution encompasses classical Wiener as well as stationary Kalman filter
ing, prediction and fixed-lag smoothing. The duality with the polynomi
al equations for LQG regulation is also discussed.