Exact adjustment of the prediction network of an adpative DPCM-coder for transmission of picture signals is outlined. The usual criterion is minimization of power of coded picture signal. The theory itself is not presented, but there are discussed some special and simple solutions. They are valid if correlation of the picture content is exponentially decaying in space with respect to the horizontal and vertical directions as well as in time. The full history of the picture signal is considered for optimum prediction. If considering only the spatial part of correlation, the result is a modified planar prediction which uses only three samples in the present field. Including temporal correlation, only seven planar samples from the present and the previous field are needed for complete decorrelation of differential signal. The amplitudes of these samples, fed into the prediction network, and results for the optimum minimized differential signal are discussed with respect to well-known criteria.