Iterative methods based on soft thresholding of hierarchical tensors (Q2407677)

The article presents a quasi-optimal alternative procedure to the very successful and largely used low-rank decomposition in order to provide a numerical approximation of high-dimensional problems, such as partial differential equations on high-dimensional domains. The technique is a variation to the truncation of the ranks of a hierarchical singular value decomposition, and it relies on soft thresholding of the singular values. It achieves similar good features in terms of complexity and quasi-optimality properties but it results to be non-expansive. The results apply to a broad class of fixed-point iterations on general contractive mappings even though the article focuses in particular on high-dimensional elliptic problems. The article contains a detailed discussion on the influence of the thresholding parameters on the approximated solution, and it proposes an algorithm to choose them while a test section illustrates, through the classic cases of a high-dimensional Poisson problem and a parametric diffusion problem, the effectiveness of the technique.