Smoothing by mollifiers. II: Nonlinear optimization

Let \(M\) be the feasible set of the nonlinear programming problem of minimizing \(f(x)\) subject to \(g_{i}(x)\geq 0\), \(i=1,\dots ,p\), all functions being of class \(C^{2}\) on \(\mathbb{R}^{n}\). One has \(M=\{x\in \mathbb{R} ^{n}:G(x)\geq 0\}\), with \(G(x):=\min_{i=1,\dots ,p}g_{i}(x)\). Using the standard mollifier \(\eta \) defined by \(\eta \left( x\right) :=C\exp (\frac{1 }{\left\| x\right\| _{2}^{2}-1})\) for \(\left\| x\right\| _{2}<1\) and \(\eta \left( x\right) :=0\) elsewhere, with \(C>0\) chosen such that \(\int_{ \mathbb{R}^{n}}\eta \left( x\right) dx=1\) and \(\left\| \cdot \right\| _{2}\) denoting the Euclidean norm, for \(\varepsilon >0\) the authors consider the function \(\eta _{\varepsilon }(x):=\frac{1}{\varepsilon ^{n}}\eta (\frac{x}{ \varepsilon })\) and the mollification \(G^{\varepsilon }\) of \(G\) given by the convolution \(\eta _{\varepsilon }\ast G.\) The set \(M^{\varepsilon }:=\{x\in \mathbb{R}^{n}:G^{\varepsilon }(x)\geq 0\}\) can be regarded a smooth approximation of \(M\); assuming that \(M\) is nonempty and compact, that the Mangasarian-Fromovitz constraint qualification (MFCQ) holds everywhere in \(M\) and that all KKT points are nondegenerate (that is, strict complementary slackness and the second order sufficiency conditions hold), it is proved that \(M^{\varepsilon }\) converges to \(M\) in the Hausdorff distance as \(\varepsilon \longrightarrow 0\) and that, for all sufficiently small \(\varepsilon \), MFCQ also holds everywhere in \(M^{\varepsilon }\) and this set is homeomorphic with \(M \). The main results deal with the correspondence between KKT points of the original problem and those of the one obtained by replacing the feasible set by \(M^{\varepsilon }\). In particular, it is proved that for each KKT point \( \overline{x}\) there is a neighborhhood which, for sufficiently small \( \varepsilon \), contains exactly one KKT point of the smoothed problem; moreover, this KKT point is nondegenerate and its Morse index coincides with that of \(\overline{x}\).

Recommendations

Cites work

Cited in

(10)

This page was built for publication: Smoothing by mollifiers. II: Nonlinear optimization

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q934788)