Low rank Tucker-type tensor approximation to classical potentials (Q2464684)

In section 2, the authors estimate the complexity of tensor-product implementations of the inner, outer and contracted products as well as the Hadamard and convolution products of \(d\)-th order tensors. They present three versions of the orthogonal alternating least-squares algorithm (OALSA) to compute the best rank-(\(r_1,\dots,r_d\)) Tucker-type decomposition: (1) decomposition of the full-format tensor into the Tucker format (OALSA(\({\mathcal F}\rightarrow {\mathcal T}_r\))); (2) input data presented as components in the CP model (OALSA(\({\mathcal C}_R\rightarrow {\mathcal T}_r\))); and (3) input data given in the Tucker format (OALSA(\({\mathcal T}_R\rightarrow {\mathcal T}_r\))) which can be interpreted as rank reduction methods in the CP and Tucker models, respectively. The mixed CP-Tucker and Tucker-Tucker models are discussed. Lemma 2.4 proves the super-convergence of the best rank-(\(r_1,\dots,r_d\)) Tucker-type approximation to the Frobenius norm. Lemma 2.5 shows that the CP decomposition of a tensor represented in the Tucker format can be reduced to the CP approximation of the corresponding ``small size'' core tensor. In Section 3, the Tucker model is applied numerically to classical potentials. Numerical results show that mixed models stabilize the alternating least-squares iteration in the case of ``ill-conditioned'' tensors and illustrate exponential convergence for tri-linear decompositions.