A new algorithm based on algebraic multigrid is presented for computing
the rank-
canonical decomposition of a tensor for small
. Standard
alternating least squares (ALS) is used as the relaxation method.
Transfer operators and coarse-level tensors are constructed in an
adaptive setup phase that combines multiplicative correction and
Bootstrap algebraic multigrid. An accurate solution is computed by an
additive solve phase based on the Full Approximation Scheme. Numerical
tests show that for certain test problems, our multilevel method
significantly outperforms standalone ALS when a high level of accuracy is
required.