A blind decision feedback equalizer （DFE） algorithm was proposed. The subspace method which was based on the second-order statistics （SOS） was employed to give the parametric estimation of channels unsupervised. The DFE was initialized optimally in the sense of MMSE using the estimated channel coefficients. In the comparison with the conventional DFE with training sequence

the proposed algorithm has meaningful advantages in coping with the abrupt time-variance of channels by re-initialization during communication. The ill-convergence of DFE is avoided.