The paper tackles a fundamental problem in mathematics and computer science: figuring out how "complex" certain mathematical objects called tensors really are. Tensors are multi-dimensional arrays of numbers that generalize matrices, and their complexity is measured by something called border rank, which roughly counts the minimum number of simple building blocks needed to approximate the tensor arbitrarily closely. Understanding border rank is deeply connected to figuring out how efficiently computers can multiply large matrices together, one of the most important open questions in theoretical computer science.
The key tool used is a technique called border apolarity, which provides a way to prove that a tensor's border rank cannot be smaller than some target number. The method works by examining algebraic objects called ideals associated with the tensor and checking whether they satisfy a specific list of mathematical properties. A previous algorithm by Conner, Harper, and Landsberg helped automate part of this process for tensors with a lot of symmetry, generating candidate ideals to check. However, all previous successful applications of the method only used a relatively weak version of the full test, meaning they never actually needed to check the harder conditions on those candidates.
The new contribution is being the first to go beyond that weak version. The authors find cases where the algorithm does produce candidate ideals, but they can then prove, using the full set of conditions, that none of those candidates actually work. This represents a genuine advance in the power of the technique. The payoff is concrete: they establish new border rank lower bounds for specific tensors relevant to matrix multiplication complexity, including one where the bound was previously unknown, and another connected to a longstanding open question about whether the theoretical optimal efficiency of matrix multiplication is achievable in practice.