A novel blind statistical algorithm for obtaining the optimal receive beam vector with maximum time-domain effective SNR in OFDMA systems under multipath environment is proposed. The basic procedure of the proposed approach is obtained via the coarse cyclic prefix (CP) detection first and the sample average of the correlation matrices is calculated from selected signal samples for later singular value decomposition (SVD) on the averaged correlation matrix. The proposed approach has the following advantages: First, it is accurate even in the presence of high noise power and when the process of the OFDMA sampled signals are not white (i.e. not all subcarriers are allocated for data transmission). Secondly, it can be operated in time-domain with no training signals and no pre-assumption of the antenna array structure. Hence, it could be applied directly to a variety of existing OFDMA systems and its design can be made relatively simple for implementation. Theoretical proofs of its optimality and the guarantee of convergence are also provided in this paper. Finally, simulation results verifies the theoretical analysis and shows that the proposed approach can accurately achieve a near optimal SNR gain even in low SNR environment and has a convergence speed within two OFDM symbols.