On the convergence of iterative methods for the distribution balancing problem

Abstract

A general distribution balancing problem, specified by the given outflows and inflows and the factorial form of its solution, is formulated. Solution uniqueness and boundedness is discussed—primarily in graph theoretical terms. An entropy maximisation problem, subject to general linear constraints, is transformed into an unconstrained optimisation problem by application of standard duality theory, and a relevant general convergence theorem for iterative solution methods is given. The optimum solution in a special case is identified with the flow solution. When expressed in flow variables, the dual objective has a unique and bounded optimum solution and is an appropriate unifying concept for measuring the rate of convergence of different solution methods. By regarding the balancing procedure as an iterative optimisation method, a new and simple proof of its convergence is given, together with some asymptotic results, which are also compared with those of Newton's method. It is pointed out that there are two different forms of Newton's method, according to the choice of variables—untransformed or transformed— in the original distribution balancing problem. When Newton's method is applied to the whole system of equations simultaneously, the trajectory of iterates is observed to depend on this variable choice. For the transformed variables it is noticed that the convergence with the balancing procedure is quicker than with Newton's method applied to the outflow- and inflow- equations, alternately. To guarantee global convergence with Newton's method and to increase the rate a supplementary linear search routine is recommended.