Aspects of Correlation Matrices
Suppose are random variables, and are convex cones of Borel-measurable real-valued functions of with finite variance. The elements of are called transformations of the variable .
For instance, can be the cone of monotone transformations, or the subspace of splines with given knots, or the subspace of quantifications of a categorical variable
A transformation is standardized if and . Standardized transformations define a sphere .
Now suppose is a concave and differentiable function defined on the space of all correlation matrices between random variables. Suppose we want to minimize over all transformations .
Because is concave Collect the partials in the matrix A majorization algorithm can minimize over all standardized transformations, which we do with block relaxation using blocks. In each block we must maximize a linear function on a cone, under a quadratic constraint, which is usually not hard to do.
This algorithm generalizes ACE, CA, and many other forms of MVA with OS. It was proposed first by De Leeuw [1988a], with additional theretical results in De Leeuw [1988b]. The function can be based on multiple correlations, eigenvalues, determinants, and so on.