We focus on investigating the optimal MIMO transmission strategy and the optimal feedback strategy for forward channel capacity and forward error exponent of MIMO block fading channels when the feedback link is causal and has a capacity constraint. We assume a forward MIMO block fading channel where the channel state information is estimated at the receiver and partially fed back to the transmitter. The feedback link is assumed to be noiseless and causal with a feedback capacity constraint in terms of maximum number of feedback bits per fading block. We show that the design of the optimal feedback scheme is identical to the design of the vector quantizer - S.P. Loyld's algorithm (see IEEE Trans. Inf. Theory, 1982) - with a modified distortion measure. It is shown that in the general case, the optimal feedback strategy has a general form of power water-filling cascaded with a beamforming matrix as well. Furthermore, we show that the SNR gain with feedback is contributed by focusing transmission power on the active eigenchannel and temporal power water-filling. The former factor contributed at most log10(nT)nR dB SNR gain, when nT>nR in all SNR regions, while the latter contribution is significant only in the low SNR region. Finally, the MMSE receiver could be used to achieve the optimal capacity in the general case of partial feedback.