Series: 1996-20, Preprints
Abstract:
The numerically optimal designs for cubic multiple regression on a
ball (centered at the origin) are supported only by two sheres one of
which is the surface of the ball. However, their support sizes increase
rapidly when the number of regressors increases, So a practically im-
portant problem is to find equivalent designs (i.e., designs sharing the
same information matrix) can be found which have a smaller support.
The present paper solves this problem within the class of designs which
are invariant under the coordinate permutation and sign change trans-
formation groups. We develop a procedure for obtaining a minimum
support invariant design associated with the optimal moment matrix.
Only a small number of competing designs have to be inspected, thus
the procedure is numerically highly efficient.
Keywords:
This paper was published in:
Journal of statistical pl