Online learning over distributed low-rank networks via sequential power iteration

There has been a growing trend towards the modeling of distributed adaptive networks. In this paper, we propose a novel low-rank relation. By incorporating this prior knowledge, we develop a distributed constrained optimization problem at each node, subjecting to a low-rank constraint. These problems are solved through the projected gradient descent method, which includes the sub-problem of projecting onto a low-rank space. We suggest utilizing the power iteration sequentially in lieu of the expensive singular value decomposition to evaluate this projection. Further, a simple and efficient rank estimation strategy is presented. Moreover, single-task combination policies are employed to merge local counterparts within a neighborhood in order to enhance the precision of estimates. In conclusion, the effectiveness of the proposed algorithm is substantiated through the presentation of simulation results.