A dual approach for matrix-derivatives (Q1088347)

A dual approach for matrix by matrix derivatives is proposed, based on the work of \textit{P. Balestra} [La derivation matricielle. (1976; Zbl 0453.90001)] and the idea of \textit{P. S. Dwyer} and \textit{M. S. MacPhail} [Ann. Math. Stat. 19, 517-534 (1948; Zbl 0032.00103)]. The two derivative concepts are called B-type derivative for the form \(\partial B/\partial A=(\partial b_{kl}/\partial A)\), because it was studied fully in Balestra, and A-type derivative for the form \(\partial B//\partial A=(\partial B/\partial a_{ij})\). Both derivative concepts are linked by permutation matrices, which also reveil the duality aspect more clearly, and can be transformed to each other very easily. The derivatives are applied to least squares estimates and posterior means in the general regression model and the ``seemingly unrelated regression'' (SUR)-system, introduced by \textit{A. Zellner} [see J. Am. Stat. Assoc. 57, 348-368 (1962; Zbl 0113.349)]. The derivative with respect to the covariance matrix is related to the local sensitivity concept of \textit{E. E. Leamer} [Specification searches. (1978; Zbl 0384.62089)], while the derivative with respect to the data-matrix, also called local resistance, is linked with the robustness concept of \textit{J. W. Tukey} [Exploratory data analysis. (1977; Zbl 0409.62003)]. The newly defined B-derivative enables an easier interpretation of the results.