It is possible to reduce the overhead associated with downlink channel estimation in massive Multiple-Input-Multiple-Output (MIMO) networks by processing training sequences according to a transformation matrix. The transformation matrix maps a generic dictionary to a non-generic dictionary associated with an antenna geometry of a MIMO antenna array. The transformation matrix can be computed based on the two dictionaries. In one embodiment, the training reference signal is precoded to obtain a precoded training reference signal, which is then transmitted over a MIMO antenna array. The training precoder used to precode the training reference signal is designed according to the transformation matrix to mitigate a dependence that the training reference signal transmission has on the antenna geometry.