The $\U$-Lagrangian of the Maximum Eigenvalue Function (Q4702298)

The author discusses the optimization problem \[ \inf_{x \in \mathbb{R}^m} \lambda_1(A(x)), \] where \(\lambda_1(\cdot)\) is the maximum eigenvalue function and \(A: \mathbb{R}^m\) \(\ni x \mapsto A_0 + \mathcal{A}\) \(\cdot x\) is an affine mapping with \(A_0\) and \(\mathcal{A}\) being a real \(n \times n\) symmetric matrix and a linear operator from \( \mathbb{R}^m\) to the space of \(n \times n\) symmetric matrices, respectively. Using the fact that the function \(f:= \lambda_1 \circ A\) is convex, the \(\mathcal{U}\)-Lagrangian theory with the geometrical interpretations of \(\mathcal{U}\)-objects is applied to this problem. It is proved that the Hessian of the \(\mathcal{U}\)-Lagrangian of \(\lambda_1\) can be explicitly computed, i.e. the second-order expansion of the \(\mathcal{U}\)-Lagrangian determines a second order expansion of \(\lambda_1\) along a characteristic smooth manifold given by the set of symmetric matrices with a fixed multiplicity of their maximal eigenvalues. The same result for the Hessian of \(\lambda_1\) is obtained if the transversality condition holds for the affine mapping \(A\). It is shown that the Hessian of the \(\mathcal{U}\)-Lagrangian of the function \(\lambda_1 \circ A\) corresponds to the reduced Hessian used in sequential quadratic programming. Second-order algorithms for minimizing the function \(\lambda_1 \circ A\) are constructed with the help of the \(\mathcal{U}\)-Lagrangian.