Tensors are multi-dimensional arrays of numbers that generalize matrices, and decomposing them into sums of simpler pieces is a fundamental problem in mathematics, data science, and engineering. One key question is whether the "rank" of a tensor, meaning the minimum number of simple pieces needed to build it exactly, equals its "border rank," which is the minimum number of pieces needed to approximate it arbitrarily closely. These two notions can differ in general, but knowing when they agree is very useful because border rank is often easier to work with theoretically.
A classical result by Kruskal gives a simple numerical condition on a tensor decomposition that guarantees it is the unique shortest possible decomposition. Tensors satisfying this condition are called Kruskal tensors. The paper proves that for Kruskal tensors, border rank and rank are always equal, meaning you cannot do better with approximations than with exact decompositions. This is a clean and practically useful result because it rules out any shortcut through approximation for this large and well-studied family of tensors.
The paper also extends these ideas to a related setting involving "alternating" or skew-symmetric tensors, which are tensors with a special sign-flipping symmetry that appears naturally in areas like geometry and physics. The authors develop an analogue of Kruskal's uniqueness condition for this skew setting, prove that border skew rank equals skew rank for the corresponding family of alternating Kruskal tensors, and provide a concrete algorithm for finding the minimal skew decomposition. Together, these results give a more complete and unified picture of when exact and approximate tensor decompositions coincide, with practical algorithmic consequences.