An efficient approach to nonlinear minimax problems (Q1203873)

The paper presents a new and efficient approach to the solution of nonlinear minimax problems: minimize \(\phi(x)=\max_{1\leq i\leq m}\{f_ i(x)\}\), where \(f_ i(x)\) are generally smooth functions of an \(n\)- dimensional vector \(x\). It is well-known that this is a class of non- smooth optimization problems since the ``maximum'' function \(\phi(x)\) has discontinuous first derivatives at points where two or more of the component functions \(f_ i(x)\) meet. One usually solves such problems by transforming them to some constrained nonlinear programming problems, but in doing this the unconstrained feature of the original problems is lost. By interpreting the Lagrange multipliers as probabilities of the component functions \(f_ i(x)\) that are probably equal to \(\phi(x)\) and introducing an additional term that represents Shannon's entropy function into the Lagrangean of a minimax problem, we solve an augmented dual problem and obtain a smooth function, which is referred to as an ``aggregate'' function and serves as a substitute of the maximum function. It is shown that the aggregate function approximates uniformly in the whole space and decreases monotonically to the maximum function as a controlling parameter tends to infinity. For the parameter being finite, an upper bound on the difference between the two functions and a relationship between the aggregate function and the \(L_ p\) norm are also proved. A significant advantage of the present approach is that one can solve a minimax problem by using standard unconstrained optimizations subroutines. Some examples are given to show that accurate solutions can be obtained by only one iteration for a moderately large parameter.