We consider a multiuser MIMO downlink system with NT transmit antennas and N R users each with one antenna. We assume that the channel is quasi-static flat fading and the perfect channel state information is available at the transmit side. So the system model can be written as where x., E ceNT X1 is the transmitted signal, y c E ceNR x 1 is the received signal. He E ceNRXNT is the channel matrix with the entry hc,ml representing the complex transfer function from transmit antenna l to the mth user. Furthermore, hc,ml is circularly symmetric complex Gaussian distributed with unit variance. n., E ceNR x 1 denotes i.i.d. additive white Gaussian noise with zero mean and variance a~. Without loss of generality, we assume that NT = N R = K.It is noted that the real model has advantage over the complex model [11]. So we transform the complex system model to the equivalent real-valued system model by stacking real which is much simpler than the SD algorithm. Although the LRA method has the same diversity with SD, there has much performance gap between them. Our work attempts to fill the performance gap between the two methods above, thus sacrifices some of complexity advantage over LRA method. The original LRA method reduces the lattice only once to obtain searching space for perturbation vector, which may cause searching errors. In order to reduce searching errors, we alternately fix each of the elements of the original perturbation vector to the possible values and reduce the formulated lattice to get subspaces of searching space. For each subspace, there is a perturbation vector to be solved. Therefore, a set of candidates of perturbation vectors is created and the optimum one is chosen from this set. The searching procedure shows that our algorithm requires to reduce lattice several times which decided by the numbers of antennas, so that we name our method as multi-Iatticereduction aided (MLRA) algorithm. Although the proposed method has slightly more complexity than LRA algorithm, simulation shows that it leads to immense improvement of performance closing to SD only with polynomial complexity.Abstract-A multi-lattice-reduction aided(MLRA) vector precoding algorithm is proposed in this paper. Compared with traditionallattice-reduction-aided(LRA) method, the new method tries to find a set of candidates of perturbation vectors through reducing the formulated lattice to get subspaces of searching space. Within each subspace, a candidate vector is solved. Then the optimum vector is found in this candidate set. Although the searching procedure has slightly more complexity than LRA method, simulation results show that it outperforms the LRA method and fills the performance gap to the sphere decode algorithm with only polynomial complexity.