Gradient projection method for optimization problems with a constraint in the form of the intersection of a smooth surface and a convex closed set (Q2314191)

The well-known gradient projection method used to the problem: \(\phi(x) \to \min\), where \(x \in X \subset E^n\), \(\phi(x)\) is a smooth function and \(X\) is a convex closed set is generalized to the case of \(X\) being the intersection of a smooth surface \(S\) and the convex closed set \(F\). The idea of the algorithm is to construct a tangent hyperplane \(\Lambda(x_{k})\) to the surface \(S\) at the point \(x_{k} \in F \cap S\) at every iteration step and to determine the point \(z(x_{k})\) by the projection of the point \(x_{k} - \beta \phi^{\prime}(x_{k})\) onto \(F \cap \Lambda(x_{k})\), where \(\beta\) is a positive constant. Then a number \(\alpha_{k} \in (0,1]\) is specified and an arbitrary projection of the point \(x_{k} + \alpha_{k}(z(x_{k})- x_{k})\) onto \(F \cap S\) is determined. The obtained projection is used as the next approximation \(x_{k+1}\). The author shows that, under certain additional assumptions, the sequence \(\{x_{k}\}\) converges to points satisfying the necessary condition for a local minimum of \(\phi\) in \(F \cap S\). It is worth to stress that the problem of projecting \(x_{k} - \beta \phi^{\prime}(x_{k})\) onto \(F \cap \Lambda (x_{k})\) always has a unique solution, which is not always easy to find.