A computational approach to non-smooth optimization by diffusion equations (Q2222070)

At first the authors show, that the optimization problem \[ \min_{x\in \mathbb{R}^n} P(x), \] where \(P: \mathbb{R}^n \rightarrow \mathbb R\) is only continuous, has a global minimal value \(p^*\) and a global minimizer \(x^*: P(x^*)=p^*\) , if \(P\) satisfies a growth condition: \[ \lim \inf_{x \rightarrow +\infty} \frac{P(x)}{||x||^2} > 0 . \] Because of that growth condition the variable \(x\) in the considered optimization problem may be considered to be bounded (\( ||x||\leq\alpha).\) The aim in this paper is to look for an approximation of \(p^*\). Therefore the author uses a suitable nonlinear diffusion equation for a function \(v(t,x)\) with a small parameter \(\epsilon > 0\) as coefficient for \( \Delta_x v(t,x)\). In order to include \(P\) one looks for solutions of the diffusion equation with boundary value \(v(1,x) = P\). The construction of the diffusion equation is so, that \(\epsilon \Delta_x v(t,x)\) is on the left side and on the right side there is a term which is especially interesting, when the left side is zero or near to zero.The term on the right side can get the total derivative of \(v(t,x_{\epsilon}(t))\), where \(x_{\epsilon}(t), t\in[0,1]\) is a welldefined state flow. And exactly such a situation is created: looking for solutions of the diffusion equation for \(\epsilon \rightarrow 0\). Then the right side teaches (Theorem 3.1) \(\lim_{\epsilon \rightarrow +0} P(x_{\epsilon}(1))=p^*,\) where \(p^* = \min_{||x||\leq \alpha} P(x).\) Now -- corresponding to the title of the paper -- three algorithms follow w.r.to \(v_x(t,x), x_{\epsilon}\) and approximation of \(p^*\). In the algorithms use is made of modifications of known numerical procedures and f. i. of the existence of second continuous partial derivatives of the solutions of the diffusion equation, of properties of the state flow and of the uniformly local convergence of the diffusion term \(\epsilon \Delta_x v(t,x)\) to zero as \(\epsilon \rightarrow 0.\) Remark 6.1 recalls, that by classical viscosity theory of PDE for the diffusion equation it holds: if the function P(x) is continuously differentiable, then the diffusion term converges to zero locally uniformly as \(\epsilon \rightarrow 0.\) Finally, several examples are given to show in detail how the algorithms work.