Slope intervals, generalized gradients, semigradients, slant derivatives, and csets (Q1826439): Difference between revisions

In classical optimization problems for smooth functions, automatic differentiation techniques are often used to provide gradients for descent algorithms. These can be combined with interval arithmetic ideas to obtain enclosures for gradients. For nonsmooth optimization problems, gradients are often replaced by slopes, semigradients, generalized gradients and slant derivatives. However it is not clear in general how to adapt the techniques of automatic differentiation and interval analysis to compute these quantities (with the exception of interval slopes, for which a calculus is available and automatic differentiation ideas can be used). This paper examines the relationship between these different approaches to non-smooth optimization, in particular when one-sided derivatives exist. Theorems can be proved and valid enclosures calculated for these cases.