Our goal in this paper is to address channel estimations for spatially correlated multiple-input multiple-output with orthogonal frequency-division multiplexing (MIMO-OFDM) system upon the conditional mutual information (CMI). The design of superimposed training signal by maximizing the channel CMI is fully investigated. First, the interior-point-method is employed to obtain the optimal superimposed training signals for the general case of channel correlations. In addition, two efficient approximate designs are also proposed by invoking Karush-Kuhn-Tucker condition for optimality of convex programming. The first approximate design is obtained by proposing a iterativebisection-procedure (IBP), while the second one is in “waterfilling” closed-form. Numerical results demonstrate the excellent performance of the proposed designs and give the connection between channel estimation error and mutual information.