On optimal positioning of extrapolators in forecasting of random flows

The dependence of the quality of a linear forecast for a sum of random flows on the position of extrapolators in network elements is studied. The problem of optimal flow decomposition into groups is stated with the minimal r.m.s. forecast error as the criterion. The resultant problem is represented as that of nonlinear integer programming. A theorem on the structure of optimal solutions to this problem leads to an effective algorithm of optima extrapolator positioning.