A variational approach to the generalized Fisher transformation of correlation matrices

Abstract

The generalized Fisher transformation maps a non-singular correlation matrix to an unconstrained real vector through the off-diagonal elements of its matrix logarithm. We show that the inverse transformation is the unique minimizer of a smooth, strictly convex and coercive function of the diagonal elements of the matrix logarithm, which yields a new and short proof that the transformation is one-to-one and onto. The Hessian of this function is bounded, at every point, between the extreme eigenvalues of the associated matrix exponential and is dominated by the diagonal of that exponential, a bound that is attained. These bounds identify the standard fixed-point iteration for the inverse as a quasi-Newton method whose worst-case local factor can approach one for near-singular correlation matrices, and guarantee that the Newton system is never worse conditioned than the matrix exponential. We propose GFT-FP+N, which augments the fixed-point iteration with matrix-free Newton steps computed by conjugate gradients, with no explicit Jacobian. In benchmark experiments GFT-FP+N converged in all tested instances, including instances in which Broyden's method failed, and reduced computation time by up to a factor of thirty relative to the fixed-point iteration.