Microphone array Wiener filtering has been studied for practical sound source enhancement in noisy environments. In our previous work, the power spectral density (PSD)-estimation-in-beamspace method was proposed for estimating PSDs of target sources and surrounding noises. Assuming that source signals are uncorrelated with each other, the relationships among the PSDs of multiple beamforming outputs (beamforming PSDs), those of sound sources (source PSDs), and the beam patterns can be modelled by a simple linear mixture in the power spectral domain. To estimate source PSDs, the inverse matrix of beam patterns (de-mixing matrix) is applied to beamforming PSDs. However, this process may suffer from estimation errors if the beamformers are poorly designed. To estimate source PSDs more accurately, we propose requirements for the de-mixing matrix using the property of the M-matrix. Through several simulation experiments, we confirmed that the proposed de-mixing matrix design performed well for speech enhancement.