scientific article; zbMATH DE number 1167545

From MaRDI portal
Publication:4396804





Let \(A(x)\) be an operator acting on a Banach space, where \(x\) is a parameter ranging over another Banach space \(X\). Suppose that, for each \(x\), there exists an ordered sequence of eigenvalues \(\lambda_1(x) \leq \lambda_2(x) \leq \ldots\) for the operator \(A(x)\). The authors consider the question of differentiability of the function \(\lambda_n(x)\). More precisely, the setting is as follows. Let \(V\) and \(H\) be separable Hilbert spaces, where \(V\) is compactly and densely embedded in \(H\). Let \(\Omega\) be an open subset of the Banach space \(X\). Given (in \(V \times V\)), a family \((a(x, . , .))_{x \in \Omega}\) of continuous, symmetric, \(V\)-elliptic bilinear forms and a family \((b(x, . , .))_{x \in \Omega}\) of equivalent scalar products on \(H\), the authors consider the following spectral problem (SP): find \(u \in V, u \not=0\) and \(\lambda\) such that \(a(x,u,v) = \lambda b(x,u,v)\) for every \(v \in V\). To each \(x \in \Omega\) there corresponds a sequence \(0<\lambda_1(x) \leq \lambda_2(x) \leq \ldots\) of eigenvalues (repeated according to their multiplicity) for this spectral problem. The differentiability properties of \(\lambda_n(x)\) are given in the following theorem, where \(\lambda_{n}^o(x,e)\) is the Clarke generalized directional derivative of \(\lambda_n\) at \(x \in \Omega\) (in the direction \(e \in X\)), \(\partial \lambda_n(x)\) is the generalized gradient and \(D^{+} \lambda_n(x,e)\) is the upper Dini derivative of \(\lambda_n\) at \(x \in X\) (in the direction \(e \in X\)). Theorem. Assume that \(x \mapsto a(x,u,v)\) (respectively, \(b(x,u,v)\)) is Gateaux differentiable for each \(u,v \in V\), and its derivative \(a'\), (respectively \(b'\)) is continuous from \(\Omega \times V \times V\) into the dual \(X^*\) endowed with the weak-star topology. Then (i) \(x \mapsto \lambda_n(x)\) is locally Lipschitzian. (ii)\(D^{+}\lambda_n(x, .) = \lambda_{n}^o(x, .)\), for each \(x \in \Omega_n = \{ x \in \Omega, \lambda_n(x) < \lambda_{n+1}(x)\}\). (iii) For \(x \in \Omega\) such that \( \lambda_{k-1}(x) < \lambda_{k}(x)= \ldots= \lambda_n(x) < \lambda_{n+1}(x)\) one has \(\partial \lambda_j(x) \subset \partial \lambda_k(x) = \lambda_n(x)\), \(j=k, \ldots, n\). (iv) For \( x \in \Omega_n\), \[ \partial \lambda_n(x) = co\{e^* = a'(x,u,u) - \lambda_n(x)b'(x,u,u), u \in V_n(x), b(x,u,u) =1 \}, \] where \(V_n(x)\) denotes the eigenspace associated with \(\lambda_n\). As the authors mention, the motivation for this work comes from the desire to find the generalized gradient of eigenvalues of the Laplacian with respect to domain perturbation.











This page was built for publication:

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4396804)