Greedy algorithms for high-dimensional eigenvalue problems (Q485351)

The authors work within the following context. They consider two Hilbert spaces \((V,\langle\cdot,\cdot\rangle_V)\) and \((H,\langle\cdot,\cdot\rangle)\) with the induced norms \(\|\cdot\|_V\) and \(\|\cdot\|,\) respectively, and suppose that (HV) the embedding \(V\hookrightarrow H\) is dense and compact. They consider also a symmetric continuous bilinear form \(a(\cdot,\cdot)\) on \(V\times V\) such that (HA) \(\exists \gamma,\nu >0\) such that \(a(v,v)\geq\gamma \|v\|^2_V-\nu \|v\|^2\) for all \(v\in V\) . It is known that, under the hypotheses (HV) and (HA), there exists a sequence \((\psi_k,\mu_k)_{k\geq 1}\) of solutions to the elliptic eigenvalue problem (*) find \((\psi,\mu)\in V\times \mathbb R,\, \|\psi\|=1\) such that \(a(\psi,v)=\mu \langle \psi,v\rangle\) for all \(v\in V\), and such that the sequence \((\mu_k)\) is nondecreasing with \(\mu_k\to \infty\) and \((\psi_k)\) forms an orthonormal basis of \(H\). The paper is concerned with the calculation of the smallest eigenvalue \(\mu_1\) and of an associated \(H\)-normalized eigenvector. The authors propose two greedy algorithms for the solution of this problem -- \textit{Pure Rayleigh Algorithm} and \textit{Pure Residual Algorithm} -- and prove some convergence results for these algorithms. Some numerical test cases (as, e.g., the computation of the buckling modes of a microstructured plate) are included.