An easily calculated bound on condition for orthogonal algorithms

Orthogonal search techniques are often used in training generalized single-layer networks (GSLNs) such as the radial basis function (RBF) network. Care must be taken with these techniques in order to avoid ill-conditioning of the required data matrix. The usual approach is to impose an arbitrary lower limit, say d/sub min/, on the norms of the orthogonal expansion terms, or equivalently on the diagonal values in the Cholesky decomposition matrices, which are calculated by the algorithms in question. In the paper, a bound on the condition number of the data matrix in terms of these qualities is given, and is used to derive a model-dependent guideline for d/sub min/.