Multidimensional scaling (MDS) is a collection of data analytic techniques for constructing configurations of points from dissimilarity information about interpoint distances. Classsical MDS assumes a fixed matrix of dissimilarities. However, in some applications, e.g., the problem of inferring 3-dimensional molecular structure from bounds on interatomic distances, the dissimilarities are free to vary, resulting in optimization problems with a spectral objective function. A perturbation analysis is used to compute first- and second-order directional derivatives of this function. The gradient and Hessian are then inferred as representers of the derivatives. This coordinate-free approach reveals the matrix structure of the objective and facilitates writing customized optimization software. Also analyzed is the spectrum of the Hessian of the objective.
R. M. Lewis, M. W. Trosset