Necessary conditions for constrained optimization problems with semicontinuous and continuous data (Q4382364)

The authors present necessary conditions for optimality in the form of ``fuzzy multiplier rules'', with one version in reflexive Banach spaces and another in smooth Banach spaces. Instead of using nonsmooth calculus (which requires constraint qualifications) to extend elementary necessary conditions using penalization, they apply the smooth variational principle of Borwein and Preiss directly to the given optimization problem to produce multiplier rules in terms of smooth (Fréchet or bornological) subderivatives. When limiting procedures are applied to these necessary conditions, one obtains familiar necessary conditions under very mild constraint qualifications (see, e.g., Corollary 2.6); many known results are special cases. Consider the case of a reflexive space \(X\). The problem, \((P)\), is to minimize \(g_0(x)\) over all points of some closed set \(C\subseteq X\) satisfying the constraints \(g_i(x)\leq 0\), \(i=1,\ldots,M\) and \(g_i(x) = 0\), \(i=M+1,\ldots,N\). Here \(g_0,\ldots,g_M\) are lower semicontinuous and extended-real-valued, while the remaining \(g_i\) are continuous. If \(\overline x\) is a minimizer in problem \((P)\), then (Theorem 2.1) for any \(\varepsilon>0\) and any weak \(0\)-neighbourhood \(V\) in \(X^*\), there are points \(x_0,\ldots,x_{N+1}\) within \(\varepsilon\) of \(\overline x\), multipliers \(\mu_i\geq 0\) for \(i=0,\ldots,N\) obeying \(\sum \mu_i=1\), and sign variables \(\tau_{M+1},\ldots,\tau_N\in\left\{ -1,+1\right\}\), such that \(\left| g_i(x_i)-g_i(\overline x)\right|<\varepsilon\) for \(i=0,\ldots,N\), and \[ 0\in \sum_{i=0}^M \mu_i D_F g_i(x_i) + \sum_{i=M+1}^N \mu_i D_F(\tau_i g_i)(x_i) + N_F(C,x_{N+1}) + V. \] Here \(D_F\) denotes the Fréchet subgradient, and \(N_F\) the Fréchet normal cone; the conclusion is called ``fuzzy'' because it involves evaluation points \(x_i\) that may differ slightly from the minimizer \(\overline x\), and because the addition of \(V\) expands the right side beyond a set completely specified by the data of problem \((P)\).