When does subtracting a rank-one approximation decrease tensor rank? (Q7020879)

For a given cone \(X\), a chain of varieties of tensors called data loci is defined inductively as follows:\N\N\(DL_1 = X\),\N\N\(DL_r\) is constructed from \(DL_{r-1}\) by subtracting a critical rank-one approximation.\N\NEach data locus contains the previous one, so the chain is ascending.\N\NAfterwards the geometry of the data loci is studied, namely the decomposition of \(DL_r\) into irreducible components. Also the intersection of the components of \(DL_r\) is proven to contain all weakly orthogonally decomposable tensors of rank at most \(r\).\N\NFinally the paper focuses on the situation when \(X\) is the cone over a Veronese variety. In this case the data locus \(DL_r\) turns out to be exactly the variety of weakly orthogonally decomposable tensors of appropriate rank. Moreover the ascending chain of data loci stabilizes and also the rank at which it becomes stable is given.