Kurdyka-Łojasiewicz property of zero-norm composite functions (Q2026719)

Let \(\|x\|_{0}\) denote the zero-norm (cardinality of non-zero coordinates) of \(x\in\mathbb{R}^{p}\) and let \(h:\mathbb{R}^{p}\rightarrow [0,+\infty]\) be either the function \(x\mapsto\nu \|x\|_{0}\) (\(\nu>0\)) or the indicator function of the set \(\{x\in\mathbb{R}^{p}:\,\|x\|_{0}\leq k\}\) where \(k\in\{1,\ldots,p\}.\) Let further \(f:\mathbb{R}^{p}\longrightarrow\mathbb{R}\) be a smooth function and \(\theta:\mathbb{R}^{p}\longrightarrow(-\infty,+\infty]\) be a lower semicontinuous function. The authors are interested on a class of zero-norm composite optimization problems of the form \(\underset{x\in\mathbb{R}^{p}}{\min}\Theta(x)\) (where \(\Theta(x):=f(x)+\theta(x)+h(x),\) capturing sparsity, due to the presence of the function \(h\)). For this class of (nonconvex nonsmooth) composite problems, the authors are interested in additional assumptions on the functions \(f\) and \(\theta\) guaranteeing that the function~\(\Theta\) satisfies the KŁ property of exponent \(1/2\) and provide some examples to illustrate situations in which the proposed assumptions are satisfied.