Optimality conditions for portfolio optimization problems with convex deviation measures as objective functions (Q1026966)

The authors consider the following optimization problem \[ \inf_{g(x)\leq 0,\;x\in S}f(Ax), \tag{P} \] where \({\mathcal Z}\) is a Hausdorff locally convex space, \(S\subseteq {\mathbb R}^n\) a nonempty convex set, \(f:{\mathcal Z}\to\overline{\mathbb R}\) a convex function, \(A:{\mathbb R}^n\to{\mathcal Z}\) is a linear mapping and \(g:{\mathbb R}^n\to {\mathbb R}^m\) is a vector-valued function with the components being convex functions. In Section 3, by means of the Fenchel-duality theory, they construct the conjugate dual problem \[ \sup_{\lambda\in {\mathbb R}^m_+,\;x^*\in{\mathcal Z}^*}\{-f^*(x^*)-(\lambda^Tg)^*_S(-A^*x^*)\}. \tag{D} \] They prove, without any assumptions of convexity, a weak duality result and, assuming a regularity condition, the strong duality (Theorem 3.2). In particular, they derive necessary and sufficient optimality conditions for the primal optimization problem \((P)\) (Theorem 3.3) In Section 4, these results are applied to special portfolio optimization problems. In the first application, the authors deal with the minimization of the generalized variance (the minimization of the variance will be a special case). The primal problem \((P_v)\) is defined as: \[ \inf_{x\in G}\left\|\sum_{i=1}^n x_iR_i-E\left(\sum_{i=1}^nx_iR_i\right)\right\|_p^a, \tag{\(\text{P}_v\)} \] where \[ G=\left\{x\in{\mathbb R}^n:x\geq 0,\sum_{i=1}^nx_i-1=0,B-E\left(\sum_{i=1}^nx_iR_i\right)\leq 0\right\}, \] with \(a>0\) and \(R_i:\Omega\to{\mathbb R},\) \(R_i\in L_p\) for all \(i=1,2,\dots, n\). The \(n\)-tuple \(R(\omega)=(R_1(\omega),\dots,R_n(\omega)),\) \(\omega\in \Omega,\) containing the random returns for the considered assets, is supposed to satisfy \[ B\leq \max_{i\in \{1,\dots, n\}}E(R_i). \] The problem \((P_v)\) can be written as \(\inf_{x\in G}d_1(Ax),\) where \[ d_1(X)=\|X-E(X)\|_p^a \quad (a>1) \] is the so-called generalized variance and \(A:{\mathbb R}^n\to L_p,\) \(Ax=\sum_{i=1}^nx_iR_i\) is a linear mapping. Using Theorems 3.2 and 3.3, they state a strong duality result and derive necessary and sufficient duality conditions for \((P_v)\) and its dual \((D_v)\) that, after some calculations, can be written as \[ \sup_{X^*\in L_q, E(X^*)=0,c\in {\mathbb R}, \gamma\in{\mathbb R}_+} \left\{ (1-a)\left\|\frac1a(X^*-c)\right\|^{a\over a-1 }_q+\min_{i\in\{1,\dots, n\}}(\text{cov}(X^*,R_i)-\gamma E(R_i))+\gamma B\right\}.\tag{\(\text D_v\)} \] Another application concerns the generalized lower semivariance that, unlike generalized variance, measures only negative deviations.