Three new multi-sensor optimal information fusion algorithms are presented, which are weighted in the minimum linear variance sense by scalars, vectors, and matrices, respectively. Based on these algorithms, the optimal information fusion distributed Kalman filters with two-layer fusion structures are given for discrete linear stochastic systems with multiple sensors. The algorithms can handle the optimal information fusion problems for systems with multiple sensors when the estimation errors of the local subsystems are correlated. There is no assumption of normal distributions, of identical size or measurement matrices, or of the initial estimation errors among the local subsystems. Instead of using the upper bound of the process noise variance matrix, the matrix itself is used. The netted parallel structure is presented to determine the cross-covariance matrix between any two sensors.