Discrete-time gradient flows and law of large numbers in Alexandrov spaces (Q745561)

Let \(\left( M,\rho\right) \) be an Alexandrov space \(\left( M,\rho\right) \) of curvature \(\leq K\) (in which minimal geodesics depend continuously on their endpoints, i.e, \(\left( M,\rho\right) \) is an Alexandrov \(\Re_{K}\) domain also known as \(\text{CAT}\left( K\right) \) space) or an Alexandrov space of curvature \(\geq K^{\prime}\). The authors study discrete-time gradient flows for finding a point of minimum of a convex function \(f:M\rightarrow(-\infty,+\infty]\). The discrete-time gradient flow is constructed by means of the (metric) resolvent \(J_{\lambda }^{f}:M\rightarrow M\) by setting \(x_{k+1}=J_{\lambda}^{f}\left( x_{k}\right) \), \(k\in \mathbb{N}\) where \(\lambda>0\). In the case of curvature bounded above, the familiar Moreau-Yosida resolvent is used: \(J_{\lambda}^{f}\left( x\right) =\text{argmin}_{y\in M}\left\{ f\left( y\right) +\frac{1}{2\lambda }d^{2}\left( x,y\right) \right\} \); for \(K=0\), see [\textit{J. Jost}, Comment. Math. Helv. 70, No. 4, 659--673 (1995; Zbl 0852.58022)]. In the case of the lower curvature bound, \(J_{\lambda}^{f}\left( x\right) =\text{gexp}\left( \lambda\nabla\left( -f\right) \left( x\right) \right) \) where \(\text{gexp}\) is the gradient exponential map (Section 4). In Section 5, the authors generalize the result from [\textit{M. Bačák}, Isr. J. Math. 194, Part B, 689--701 (2013; Zbl 1278.49039)] by proving the following theorem for complete \(\Re_{K}\) domains: If \(f:M\rightarrow(-\infty,+\infty]\) is a convex, lower semi-continuous function, \(G\subseteq M\) is a closed, geodesically convex set containing a sublevel set of \(f\) satisfying \(\text{diam}\left( G\right) <\pi/2\sqrt{K}\) for positive \(K\), \(\left( \lambda_{k}\right) _{k=1}^{\infty}\) is a sequence of positive numbers such that \(\sum _{k=1}^{\infty}\lambda_{k}=+\infty\), \(x_{0}\in G\) and \(x_{k}=J_{\lambda_{k} }^{f}\left( x_{k-1}\right) \), \(k\in\mathbb{N}\), then \(\lim_{k\rightarrow\infty}f\left( x_{k}\right) =\inf_{y\in G}f\left( y\right) \). The authors also consider the case when \(f=\sum _{k=0}^{n}f_{k}\) where \(f_{k}\) are convex and Lipschitz in a locally compact Alexandrov space of curvature either \(\leq K\) or \(\geq K^{\prime}\) \ and prove that the proximal point algorithm produces the sequence converging to a point of minima of \(f\) when it exists (Theorem 5.5). In Section 6, the authors present results related to stochastic discrete-time gradient flow for a convex infinite (integral form) combination of convex functions and generalize Sturm's law of large numbers to Alexandrov spaces of curvature either \(\leq K\) or \(\geq K^{\prime}\) for arbitrary \(K\) and \(K^{\prime}\in\mathbb{R}\).