An algorithmic construction of optimal minimax designs for heteroscedastic linear models

We construct optimal designs to minimize the maximum variance of the fitted response over an arbitrary compact region. An algorithm is proposed for finding such optimal minimax designs for the simple linear regression model with heteroscedastic errors. This algorithm always finds the optimal design in a few simple steps. For more complex models where there is a symmetric error variance structure, we suggest a strategy to help find some hitherto elusive optimal minimax designs. (C) 2000 Elsevier Science B.V. All rights reserved.