The theory of low-rank tensor-train approximation is well-understood when the approximation error is measured in the Frobenius norm. The entrywise maximum norm is equally important but is significantly weaker for large tensors, making the estimates obtained via the Frobenius norm and norm equivalence pessimistic or even meaningless. In this article, we derive a new, direct estimate of the entrywise approximation error that is applicable in such cases. Our argument is based on random embeddings, the tensor-structured Hanson--Wright inequality, and the core coherences of a tensor. The theoretical results are accompanied with numerical experiments: we compute low-rank tensor-train approximations in the maximum norm for two classes of tensors using the method of alternating projections.
翻译:暂无翻译