On the optimization properties of the correntropic loss function in data analysis (Q2448161)

The authors present generalized convexity properties for correntropic loss functions used in data analysis methods. After some preliminary remarks about correntropic measures, the loss functions (with fixed parameters \(\sigma_i> 0\), \(i= 1,\dots, n\)) \[ \begin{aligned} {\mathcal F}^\sigma_C(x) &= \sum^n_{i=1} \left[1-e^{\left(-{1\over 2\sigma^2_i}\right)}\right]^{-1} \left[1- e^{\left(-{x^2_i\over 2\sigma^2_i}\right)}\right],\qquad x_i\in \mathbb{R},\quad i= 1,\dots, n,\\ {\mathcal G}^\sigma_C(x) &= \sum^n_{i=1} \left[1-e^{\left(-{1\over 2\sigma^2_i}\right)}\right]^{-1} \left[1- e^{\left(-{\| x_i\|^2\over 2\sigma^2_i}\right)}\right],\qquad x_i\in\mathbb{R}^m,\quad i=1,\dots,n,\end{aligned} \] are discussed. It is shown that in the one-dimensional case \((n=1)\) both loss functions are pseudoconvex. Unfortunately, the sum of pseudoconvex functions is not pseudoconvex in general. Nevertheless it is proved that for the \(n\)-dimensional case the loss functions are invex and locally (under additional assumptions even global) pseudoconvex. The paper is concluded with remarks to the application of these results and to future research directions.